xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam 2200409Ka 4500
controlfield tag 001 001912470
007 cr bnu uuuuu
008 071009s2005 flu sbm 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0001305
Verrill, Stephen W.
Social structure and social learning in delinquency :
b a test of Akers' social structure-social learning model
h [electronic resource] /
by Stephen W. Verrill.
[Tampa, Fla] :
University of South Florida,
ABSTRACT: Social learning theory (Akers, 1973, 1977, 1985, 1998; Burgess & Akers, 1966) is an established general theory of criminal, deviant, and conforming behavior that finds substantial empirical support (e.g., Akers, Krohn, Lanza-Kaduce & Radosevich, 1979; Akers, La Greca, Cochran & Sellers, 1989; Alarid, Burton & Cullen, 2000; Krohn, Skinner, Massey & Akers, 1985). Although the theory provides insight into the processes that influence criminal behavior, the theory does not speak to the environments that produce such behavior--the domain of structural theories. Akers (1998) has suggested that social learning theory accounts for differences in crime rates through its mediation of structural effects on individual criminal behavior. He postulated that social structure acts as the distal cause of crime, affecting an individuals exposure to norm and norm-violating contingencies through the social learning process. Although the integrated cross level social structure-social learning^theory (Akers, 1998) has received empirical attention, criminologists have not adequately tested the model (Akers, 1998; Bellair, Roscigno, & Vlez, 2003; Lanza-Kaduce & Capece, 2003; Lee, 1998; Lee, Akers & Borg, 2004). Akers (1999) and colleagues (Lee et al., 2004) have suggested that future research should test models that incorporate broader social structural measures, especially those derived theoretically. The present research contributes to the theoretical body of literature through its more complete measurement of the macrosocial correlates and theoretically defined structural causes dimensions posited by Akers (1998). Secondly, the study introduces possible linkages between social structure and the social learning process in an attempt to address the concerns of Krohn (1999), who suggested that the theory does not adequately do so, and Sampson (1999), who suggested that the theory is incapable of producing a priori, refutable macrosocial propositions. Although finding a relationship between social structure and social learning, the study finds no support for Akers (1998) use of the mediation descriptor. Instead, the present research finds support for several moderator hypotheses, concluding that the social structure-social learning statement requires modification.
Dissertation (Ph.D.)--University of South Florida, 2005.
Includes bibliographical references.
Text (Electronic dissertation) in PDF format.
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
Title from PDF of title page.
Document formatted into pages; contains 307 pages.
Adviser: Christine S. Sellers, Ph.D.
t USF Electronic Theses and Dissertations.
Social Structure and Social Learning in Delinquency: A Test of AkersÂ’ Social Structure-Social Learning Model by Stephen W. Verrill A dissertation submitted in partial fulfillment of the requirement s for the degree of Doctor of Philosophy Department of Criminology College of Arts and Sciences University of South Florida Major Professor: Christine S. Sellers, Ph.D. John K. Cochran, Ph.D. Wilson R. Palacios, Ph.D. L. Thomas Winfree, Jr., Ph.D. Date of Approval: November 14, 2005 Keywords: differential association, re inforcement, disorganization, community, cross-level, mediation Copyright 2005, Stephen W. Verrill
Dedication To Diane. Thank you for your encouragement and support.
Acknowledgements I thank my wife Diane for helping guide my career interests and experiences into the scholarly pursui ts subsequently undertaken. Her support and commitment has directly influenced th e completion of my goals. I acknowledge the hard work of my co mmittee: Dr. Chris Sellers, Dr. John Cochran, Dr. Wilson Palacios, and Dr. Tom Winfree. I thank Chris, my doctoral and dissertation committee chair, for in itially engaging me in the pursuit of theoretical understanding, helping me learn and adhere to a high standard of scholarship, and her support thro ughout my graduate education. I thank John for also engaging me in theory and research, and for his guidance in exploring ideas br oader than that reflected in this dissertation. I thank Wilson for agreeing to work with me on a topic that probably at first seemed outside his area of primary interest. I thank Tom for agreeing to work with someone he did not initially know. Each member contributed to my development as a scholar, and I would not have complet ed this project without their individual effort and support.
i Table of Contents List of Tables iv List of Figures vii Abstract x Chapter One 1 Introduction 1 Aims of the Research 4 Dissertation Overview 6 Chapter Two 7 Social Learning and Social Stru cture Theoretical Framework 7 Differential Associati on Theoretical Statement 7 Social Learning Theoretical Statement 30 Social Structure-Social Learni ng (SSSL) Theoretical Statement 49 Theoretical Critiques 55 Empirical Validity 69 Chapter Three 76 Crime Rate Determinants 76 Criminal Behavior and Environment 76 Social Structural Crime Correlates and Explanations 78 Background 78
ii Empirical Research 79 Applicability to Social St ructure-Social Learning 99 Chapter Four 102 Rationale for the Present Study 102 Overview 102 Study Objectives 104 Mediation and Substantial Medi ation versus Moderation 108 Functional Relationships 120 Hypotheses 126 Chapter Five 142 Research Design and Analytic Strategy 142 Sample 142 Measures 149 Dependent Variable 149 Microsocial Independent Variables 159 Macrosocial Independent Variables 166 Procedure 175 General Issues and Moderation 175 Mediation 178 Structural Equation Modeling 182 A Priori Measures 189
iii Chapter Six 197 Results 197 Preliminary Evidenc e on Relationships 197 Bivariate Correlations 197 OLS Regression Models 199 Direct and Indirect Effects 213 Initial and Revised measurement Models 213 Structural Models 227 Chapter Seven 233 Discussion 233 Summary of the Problem 233 Implications of the Present Research 238 Nuances of the Research Question 238 Overview of the Findings 242 Reconciliation of the Resu lts with Previous Research 247 Nuances of the Findings 253 Modification of the Theoretical Statement 257 Limitations of the Present Research 269 Conclusion 272 References 275 About the Author End Page
iv List of Tables Table 1 Missing Values Analysis 149 Table 2 Frequency Distribution and Cumu lative Percentages for SelfReported Delinquency ( N = 1062) 151 Table 3 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Dif ferential Associations Index (Range 2-20) 163 Table 4 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Costs Index (Range 4-32) 164 Table 5 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Rewards Index (Range 4-32) 165 Table 6 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Costs Index (Range 4-32) 166 Table 7 Descriptive Statistics for Variables Under Analysis ( N = 1062) 174 Table 8 Inter-correlations Among Explanatory Variables ( N = 1062) 175 Table 9 Zero-Order Correlations fo r the Explanatory Variables and Log10 Delinquency ( N = 1062) 198 Table 10 OLS Regression Dimension I (Population Density) Moderator Models ( N = 1062) 200 Table 11 OLS Regression Dimension I (Log10 Race Composition) Moderator Models ( N = 1062) 201 Table 12 OLS Regression Dimension I (Sex Composition) Moderator Models ( N = 1062) 202
v Table 13 OLS Regression Dimension I (Age Composition) Moderator Models ( N = 1062) 203 Table 14 OLS Regression Dimension I (Near Poverty) Moderator Models ( N = 1062) 204 Table 15 OLS Regression Dimension II (Individual Sex) Moderator Models ( N = 1062) 205 Table 16 OLS Regression Dimension II (Individual Race) Moderator Models ( N = 1062) 206 Table 17 OLS Regression Dimension II (Individual Age) Moderator Models ( N = 1062) 207 Table 18 OLS Regression Dimension III (SES) Moderator Models ( N = 1062) 208 Table 19 OLS Regression Dimension III (Log10 Ethnic Heterogeneity) Moderator Models ( N = 1062) 209 Table 20 OLS Regression Dimensi on III (Residential Mobility) Moderator Models ( N = 1062) 210 Table 21 OLS Regression Dimension III (Family Disruption) Moderator Models ( N = 1062) 211 Table 22 Goodness of Fit Indices fo r the Social Structure-Social Learning Measurement Model ( N = 1062) 218 Table 23 Goodness of Fit Indices for the Differential Social Organization Measurement Model ( N = 1062) 219 Table 24 Goodness of Fit Indices fo r the Revised Diffe rential Social Organization Measurement Model ( N = 1062) 220 Table 25 Goodness of Fit Indices for the Differential Location in the Social Structure Measurement Model ( N = 1062) 222 Table 26 Goodness of Fit Indices for the Theoretically Derived Structural Causes Measurement Model ( N = 1062) 223 Table 27 Goodness of Fit Indice s for the Revised Social
vi Structure-Social Lear ning Measurement Model ( N = 1062) 226 Table 28 Properties of the Final Differential Social Organization, Differential Location in the Soci al Structure, and Theoretically Defined Structural Causes Measurement Models ( N = 1062) 227 Table 29 Goodness of Fit Indices fo r the Social Structure-Social Learning Structural Models ( N = 1062) 231
vii List of Figures Figure 1 Social StructureSocial Learning Model 52 Figure 2 Social Structure and Social Learning Theoretical Models 106 Figure 3 Social Structure-Social Learning Theoretical Model 106 Figure 4 Theoretical Model of the Rela tionship Between Social Structure and Delinquency with Social Learning as a Mediator 109 Figure 5 Path Diagram of Hypotheses Depicting Social Learning as a Moderator and a M ediator of the Social Structural Effects on Delinquency 117 Figure 6 Path Diagram for SSSL Dimension I (Population Density) Hypotheses 127 Figure 7 Path Diagram for SSSL Dimension I (Race Composition) Hypotheses 128 Figure 8 Path Diagram for SSSL Dimension I (Sex Composition) Hypotheses 129 Figure 9 Path Diagram for SSSL Dimension I (Age Composition) Hypotheses 130 Figure 10 Path Diagram for SSSL Dim ension I (Poverty) Hypotheses 131 Figure 11 Path Diagram for the So cial Structure-Social Learning Dimension I Hypothesis that So cial Learning Mediates the Effect of Differential Social Organization on Delinquency 132 Figure 12 Path Diagram for SSSL Dimension II (Individual Sex) Hypotheses 133 Figure 13 Path Diagram for SSSL Dimension II (Individual Race) Hypotheses 134
viii Figure 14 Path Diagram for SSSL Dimension II (Individual Age) Hypotheses 135 Figure 15 Path Diagram for the So cial Structure-Social Learning Dimension II Hypothesis that Social Learning Mediates the Effect of Differential Location in the Social Structure on Delinquency 136 Figure 16 Path Diagram for SSSL Dim ension III (SES) Hypotheses 137 Figure 17 Path Diagram for SSSL Dim ension III (Ethnic Heterogeneity) Hypotheses 138 Figure 18 Path Diagram for SSSL Dim ension III (Residential Mobility) Hypotheses 139 Figure 19 Path Diagram for SSSL Dimension III (Family Disruption) Hypotheses 140 Figure 20 Path Diagram for the So cial Structure-Social Learning Dimension III Hypothesis that Social Learning Mediates the Effect of Theoretically De fined Structural Causes on Delinquency 141 Figure 21 Path Diagram for the So cial Structure-Social Learning Hypothesis that Social Learni ng Mediates the Effect of Social Structure on Delinquency 141 Figure 22 Path Analysis Illustrations with Manifest and Latent Variables 185 Figure 23 Path Diagram for the So cial Structure-Social Learning Dimension I (Population Dens ity), Social learning (Differential Associations), and Delinquency 214 Figure 24 Social Structure-Soci al Learning Measurement Model 217 Figure 25 Differential Social Or ganization Measurement Model 219 Figure 26 Differential Location in t he Social Structur e Measurement Model 221 Figure 27 Theoretically De rived Structural Causes Measurement Model 223
ix Figure 28 Revised Social Structure-Social Learni ng Measurement Model 225 Figure 29 Differential Social Organizati on Multifactor Structural Model ( N = 1062) 229 Figure 30 Differential Location in t he Social Structure Multifactor Structural Model ( N = 1062) 230 Figure 31 Theoretically De fined Structural Causes Multifactor Structural Model ( N = 1062) 231
x Social Structure and Social Learning in Delinquency: A Test of AkersÂ’ Social Structure-Social Learning Model Stephen W. Verrill ABSTRACT Social learning theory (Akers, 1973, 1977, 1985, 1998; Burgess & Akers, 1966) is an established general theory of criminal, deviant, and conforming behavior that finds substantial empirica l support (e.g., Akers, Krohn, LanzaKaduce & Radosevich, 1979; Akers, La Gr eca, Cochran & Sellers, 1989; Alarid, Burton & Cullen, 2000; Krohn, Skinner Massey & Akers, 1985). Although the theory provides insight into the processe s that influence criminal behavior, the theory does not speak to the environments that produc e such behaviorÂ—the domain of structural theories. Akers (1998) has suggested that so cial learning theory accounts for differences in crime rates through its mediat ion of structural effects on individual criminal behavior. He postulated that social structure acts as the distal cause of crime, affecting an individualÂ’s exposure to norm and norm-violating contingencies through the social learning process. Although the integrated crosslevel social structure-soci al learning theory (Akers, 1998) has received empirical attention, criminologists have not adequately tested the model (Akers, 1998;
xi Bellair, Roscigno, & Vlez, 2003; Lanz a-Kaduce & Capece, 2003; Lee, 1998; Lee, Akers & Borg, 2004). Akers (1999) and colleagues (Lee et al., 2004) have suggested that future research should test models that incorporate broader social structural measures, especially those derived theoretically. The present research contributes to the theoretical body of literature through its more complete measurement of the macrosocial correlates and theoretically defined structural causes dimensions posited by Akers (1998). Secondly, the study introduces possible linkages between social structure and the social learning process in an atte mpt to address the concerns of Krohn (1999), who suggested that the theory does not adequately do so, and Sampson (1999), who suggested that t he theory is incapable of producing a priori, refutable macrosocial propositions. Although finding a relationship between social structure and social learning, the study finds no support for AkersÂ’ (1998) use of the mediation descriptor. Instead, the present resear ch finds support for several moderator hypotheses, concluding that the social structure-social learning statement requires modification.
1 Chapter One Introduction Social learning theory (Akers, 1973, 1977, 1985, 1998; Burgess & Akers, 1966) integrates operant conditioning and co gnitively oriented psychological and sociological theories to ex plain criminal, deviant, and conforming behavior. It is a general theory that describes the learning process involved in an individualÂ’s history and opportunity for crime (Akers, 1998). Social learning theory has received mu ch empirical attention, and its concepts and variables find moderate to strong support with survey, official, cross-sectional, and longitudinal data (e .g., Akers & Lee, 1996; Akers, Krohn, Lanza-Kaduce & Radosevich, 1979; Conway & McCord, 2002; Haynie, 2002; V. Johnson, 1988; Winfree, Mays & Backstro m, 1994). When researchers employ theory competition, social learning theory concepts and pr opositions generally find more support than those derived from other simultaneously tested theories (e.g., Akers & Cochran, 1985; Alar id, Burton & Cullen, 2000; Benda, 1994; Kandel & Davies, 1991; Burton, Cullen, Evans & Dunaway, 1994; Matsueda & Heimer, 1987; Rebellon, 2002; White, Johnson & Horowitz, 1986). When scholars apply social learning concepts and propositions to integrated theory, social learning variables generally have the strongest ef fect (e.g., Conger, 1976;
2 Elliott, Huizinga & Ageton, 1985; R. J ohnson, Marcos & Bahr, 1987; Marcos, Bahr & Johnson, 1986; Thornberry, Lizo tte, Krohn, Farnworth & Jang, 1994; White, Pandina & LaGrange, 1987). Despite the large body of research, there is still much unknown about the social learning process, and scholars conti nually seek to test social learning theoryÂ’s scope. Much of the social learni ng body of science involves explaining minor forms of juvenile offending and subs tance use (Akers et al., 1979; Krohn, Skinner, Massey & Akers, 1985; Winfree & Bernat, 1998). One direction research has taken has been to examine broader of fenses and populations of offenders. For example, social learning variables partially accounted for illegal computer behavior (W.F. Skinner & Fream, 1997) and intimate partner violence (Sellers, Cochran & Winfree, 2003) in samples of college students, deviance in police officers (Chappell & Piquero, 2004), drin king behavior in people 60 years old or older (Akers, La Greca, Cochran & Sellers, 1989), marijuana use in rural middle school students (Winfree & Griffith s, 1983), and alcohol and drug use in American Indian youths (Winfree, Griffiths & Sellers, 1989). The vast body of research on social learning theory has demonstrated that individual deviant behavior varies dependi ng on the individualÂ’s associations, definitions, reinforcements, and to some ex tent, imitation of deviant models. The theory appears to identify with a fair degree of accuracy the basic mechanism by which individuals learn deviant behavior. As satisfactory as the theory might be,
3 though, it still has limitations. In its strictly social psychological (p rocessual) form, social learning cannot answer why some individuals and not ot hers encounter confi gurations of the social learning elements conducive to dev iant behavior. Such a solution requires the integration of macro-sociological (s tructural) concepts into social learning theory. Akers (1998) has proposed such an integration, terming the social learning model elaboration Â“social structure-social learning.Â” In this latest explicatio n of the theory, Akers (1998) suggests that social learning theory mediates social structur al influences on individual criminal behavior and ultimately on crime rates. Akers postulates that social structure acts as the distal cause of crime, affecti ng an individualÂ’s exposure to norm and normviolating contingencies. The social lear ning variables differential association, definitions, imitation, and differential reinforcement, and other discriminative stimuli, mediate social structureÂ’s effe ct on individual behavior, providing the proximate causes of crime. Although a comprehensive explanation of crime and criminal behavior addresses both individual differences in crime formation and the structure that shapes the process (Akers, 1968; Shaw, Zorbaugh, McKay & Cottrell, 1929), there are barriers to testing such a model. Notably, data allowing for the simultaneous examination of macrosoc ial and microsocial variables are uncommon (Lanza-Kaduce & Capece, 2003).
4 Despite these hindrances, there are three tests of the social structuresocial learning elaboration in the litera ture (Bellair et al., 2003; Lanza-Kaduce & Capece, 2003; Lee et al., 2004; see also Hoffmann, 2002). In one study with limited structural measures, researchers concluded that family well being and social learning partially mediated t he impact of occupational structure on adolescent violence (Bellair et al., 2003). In the second study, researchers concluded that social learning partia lly mediated the relationship between structural variables and binge drinki ng (Lanza-Kaduce & Capece, 2003). In the third study, researchers concluded that social learning partially mediated the relationship between structural variables and adolescent substance use (Lee et al., 2004). Although measured imperfectly and utilizing varying and limited statistical techniques, each of the re searchers reported findings that are suggestive that social learning variables mediate structural influences on individual behavior. Aims of the Research As the tests in the literature have not incorporated strong social structural measures, Akers (1998) and colleagues (Lee et al., 2004) suggest that research on the social structure-social learning model should test models that include broader indicators of social structure, especially theoretically derived measures. It is this suggestion on which the present study focuses. The present research contributes to the theoretical body of literature
5 through its more complete measurement of the macrosocial correlates and theoretically defined structural causes di mensions. Notably, the study measures race, poverty, and family disruption, thr ee variables that Pr att and Cullen (2005) identified in a macro-leve l predictors meta-analysis as Â“among the strongest and most stable predictors Â“ (p. 373) of crime, and which so me researchers think of as indicators of a Â“concentrated dis advantageÂ” construct (e.g., Sampson & Raudenbush, 1999; Sampson, Raudenbush & Ea rls, 1997). Further, the present study measures social disorganization theory variables in a manner similar to that used by Sampson (Sampson & Groves 1989), one of the social structuresocial learning modelÂ’s more vocal skeptics (Sampson, 1999). Secondly, the study introduces possible linkages bet ween social structure and the social learning process in an attempt to address the concerns of Krohn (1999), who suggested that the theor y does not adequately do so, and Sampson (1999), who suggested that the theory is incapabl e of producing a priori, refutable macrosocial propositions. The present research also critically examines AkersÂ’ (1998) notion that social learning mediates the relations hip between social st ructure and crime, introducing the possibility that social learning may inst ead moderate social structureÂ’s effect on crime and criminal behavior. The study argues that clarifying this distinction may contribute to under standing how exactly social structure might influence the social le arning process. Combined, t he two aims of the study,
6 utilizing more complete social structur al measures and explaining how social structure might impinge on t he social learning process, respond to AkersÂ’ (1999) plea to help specify the most under developed portion of the model. Dissertation Overview The dissertation comprises se ven chapters. Chapter Two introduces the background and theoretic al framework for the research question. Chapter Three examines macrosocial crime correlates and theoretical explanations, serving as the foundation for the studyÂ’s later measurement of social structural variables. Chapter Four presents the rationale for the present research, expl aining how the study differs from that in the extant liter ature, and including a spec ification of the studyÂ’s hypotheses. Chapter Five presents the studyÂ’s research design and analytic strategy. Chapter Six descri bes the analytic results, and Chapter Seven presents a discussion of the fi ndings, limitations of the study, and recommendations for fu ture research.
7 Chapter Two Social Learning and Social Structure Theoretical Framework Differential Association Theoretical Statement In order to understand the complexity of the social learning model, as well as its social structural el aboration, it is first necessa ry to trace its historical development, beginning with the inceptio n of SutherlandÂ’s (1939, 1947) differential association theory. Suther land (1939) sought a general theory of crime that would resolve failings in t he literature, advance criminology as a science, and provide for the meaningful control of crime (Sutherland, 1924). Sutherland (1939) believed that prevailing theories of criminal behavior were inadequate to provide meaningful understanding and control, resulting instead in a scattered body of knowledge that provided little practical application. One approach, for example, viewed crime as a product of a variety of individual factors. As individual crim inal behavior derived from t hese situationally different factors, the approach did not allow for general explanations that would hold without exception (see historical discu ssions in Matsueda, 1988; Sutherland & Cressey, 1970). Sutherland (1939, 1973a) was concerned that such a multiplefactor approach was not scientific, resulting in unsound theorizing. Sutherland (1939) instead fa vored general statements of criminal behavior
8 that would aid in both the understanding and control of crime. Rather than view crime as the particularistic product of numerous factors (Sut herland & Cressey, 1974), Sutherland (1939) sought a set of uni versal statements. He believed that an organized, scientific theor y of criminal behavior, however tentative, was necessary to bring discussion and underst anding to bear on issues that would otherwise go unsolved if not advanced unt il theoretically complete. Sutherland considered his theory t entative and hypothetical, needing future examination against data, but necessary to start a discussion based on science. Building off his sociologic al training and notion that a theory of criminal behavior should center on learning, intera ction, and communication, Sutherland (1973a) sought an account of all crime c ausation facts. He wished to express general statements that accounted for all k nown correlates of criminal behavior, without exception, from a sociological viewpoint. In formulating his theory, Suther land (1939) followed three guidelines. First, comprehensive criminological theor y must acknowledge and consider all reasonable explanations for cr iminal behavior. Sutherland classified existing explanations for crime into two groups: individual and situational or cultural. Sutherland (1939) suggest ed that individual ex planations emphasized inherited or acquired traits, such as feeblemindedness and anatomical or emotional deviations. Individual explanations were concerned with the differences of people, viewing criminal behavior as derived from individual
9 defects (see Sutherland, 1973b) and consid ering such personal abnormalities as the primary cause of crime (see Sutherland, 1973c). The situational or cultural differ ence perspective emphasized social processes. Sutherland (1939) characteriz ed these processes as occurring either at the small group level, such as fam ilies and neighborhoods, at the institutional level, reflected in economic and political systems, or more generally in the form of differential associations, cultural conf licts, and societal social disorganization. Situational and cultural difference viewpo ints considered crime as part of a process (see Sutherland, 1924). SutherlandÂ’s second theory-construc tion guideline hinged on the notion of desire. Sutherland (1939) suggested that crime involved more mechanisms than offender needs and restraints, and that ma ny theories focused too narrowly on desire and inhibition. He belie ved that a general theory of criminal behavior must additionally account for more elements, such as result s, external restraints, public opinion, possibility of detection and punishment, technical ability, and other related factors (see Sutherland, 1939). Third, Sutherland (1939) acknowledged the multiple-factor viewpoint that criminal behavior is sometimes advent itious, but he reasoned that criminal behavior is only beyond analytic possibility at the complex, individual circumstances level. He equated that not ion with the chance inherent in a coin flip coming up heads or tails. Sutherl and reasoned that the coinÂ’s outcome,
10 similar to behavior involving individual ci rcumstances, is not without cause but that the cause is too complicated to dist inguish at the level of occurrence. He carried the analogy further, suggesting that unlike the two limit ed outcomes of a coin toss, and instead like the roll of loaded dice, individually circumstanced behavior involves numerous outcomes, some of which although not certain, are more probable than other behaviors. Suther land concluded that a general theory of crime must focus on systematic crim inal behavior, rather than adventitious, individually circumstanced behavior, in order to discover general and uniform processes (see Suther land, 1939). Methodologically, Sutherland (1939) embraced LindesmithÂ’s (1938) application of analytic induction to test for necessary and sufficient causes. The approach specified a case-by-case search for exceptions to a hypothesis and upon finding one, necessitated either a modification of the hypothesis or a redefinition of the universe of cases. The idea was t hat after investigating a number of segments of crimi nality and finding no exception, the series of general propositions about those s egments would lead, with practical certainty to a general body of criminological theory (Sutherland, 1939). Sutherland (1939) dealt with the problematic issue of multiple causal factors that differ individually by abstr acting individual criminal behavior to systematic criminal behavior Sutherland was vague on the termÂ’s meaning, but as he used adventitious and systematic to distinguish opposing viewpoints, it is
11 likely that Sutherland defined adventitious criminal behavior as sporadic and multi-sourced, contrast ed with systematic criminal behavior as planned and regular (see Sutherland, 1973a). Sutherland (1939) intended syst ematic criminal behavior to serve as the framework for the formulati on of scientific statements about individual behavior. He acknowledged criminal behavior as adventitious when considered from the point of view of individual circumstanc es, but as he sought universal statements, he abstracted the behavior under study in or der to avoid the consideration of trivial crimes with imm easurable causes. Sutherl and evaded the question of multiple crime causes, adventitious crim e, by defining crim e in a way that emphasized behavioral commonalities and i gnored individually specific factors that he viewed as rare (see Sutherland, 1973a). Believing it impossible to account fo r all situations that might lead a specific individual to commit a specific crime, Sutherland (1939) reasoned that a theory that explained system atic criminal behavior w ould accordingly explain specific acts generally. He used organized cr iminal behavior and criminal careers as examples of systematic crim inal behavior, and he believed that practically all criminals would fall into the category (Suther land, 1973a). Sutherland created the concept of systematic cr iminal behavior as a matter of convenience (see Sutherland, 1973a), perhaps redef ining the universe up front so that he would not have to modify the hypotheses based on trivial, incidental exceptions.
12 In the first statement of his t heory, Sutherland organized scientific characteristics of crime into a general explanation that addressed both the epidemiology and etiology of crime and criminal behavior. Sutherland (1939) stated, First, the processes which result in systematic criminal behavior are fundamentally the same in form as the processes which result in systematic lawful behavior If criminality were specifically determined by inheritance, the laws and principles of inheritance would be the same for criminal behavior and for lawful behavior. The same is true of imitation or any other genetic process in the development of behavior. Criminal behavior differs from lawful behavior in the standards by which it is judged but not in the principles of the genetic process. (p. 4) Second, systematic criminal behavior is determined in a process of association with those who commit cr imes, just as systematic lawful behavior is determined in a process of association with those who are law-abiding Any person can learn any pattern of behavior which he is able to exercise. He inevitably assimilates such behavior from the surrounding culture. The pattern of behavior may cause him to suffer death, physical injury, loss of friendship, or loss of money, but it may neverthele ss be followed with joy provided he has learned that it is the thing to do. Since criminal behavior is thus developed in association with crimi nals it means that crime is the cause of crime. In the same manner war is the cause of war, and the Southern practice of dropping the Â“rÂ” is the cause of the Southern practice of dropping the Â“r.Â” This proposition, stated negatively, is that a person does not participate in systematic criminal behavior by inheritance. No individual inherits tendencies which inevitably make him criminal or inevitably make him lawabiding. Also, the person who is not already trained in crime does not invent systematic criminal beh avior. While personality certainly includes an element of inventiv eness, a person does not invent a system of criminal behavior unles s he has had training in that kind of behavior, just as a person does not make systematic mechanical inventions unless he has had training in mechanics. (pp. 4-5) Third, differential association is the s pecific causal process in the development of systema tic criminal behavior The principles of the
13 process of association by which cr iminal behavior develops are the same as the principles of the pr ocess by which lawful behavior develops, but the contents of the patterns presented in association differ. For that reason it is ca lled differential association. The association which is of primary im portance in criminal behavior is association with persons who engage in systematic criminal behavior. A person who has never hear d of professional shoplifting may meet a professional shopli fter in his hotel, may become acquainted with and like him, learn from his techniques, values, and codes of shoplifting, and under this tutelage may become a professional shoplifter. He could not become a professional shoplifter by reading newspaper s, magazines, or books. The impersonal agencies of communicati on exert some influence but are important principally in determini ng receptivity to the patterns of criminal behavior when they are presented in personal association, and in producing incidental offens es. These patterns are presented through the impersonal agencies of communication to everyone in our culture. Every child capable of learning inevitably assimilates knowledge regarding property rights and thefts in the simpler situations. It is probably for this reason that everyone is somewhat criminal. College students, with a few exceptions doubtless due to poor memories, report an average of ei ght thefts or se ries of thefts during their lifetimes; a series of thefts in this case may include scores of incidents, such as steali ng fruit from neighborsÂ’ trees from the age of seven to twelve. These thefts were reported equally for males and females, and continued in most cases to the age at which the reports were made. In t he later years they generally took the form of theft of books from the library, of equi pment from the gymnasium or laboratory, or of souvenirs from hotels and restaurants. Students do not regard such thefts as especially reprehensible; they regard them as amusing. Similarly, boys in the delinquent areas of cities do not regar d thefts of automobiles or the burglary of stores as reprehensib le, and business or professional men do not regard their frauds and tricky manipulations as reprehensible. A person engages in t hose criminal acts which are prevalent in his own groups, and he a ssimilates them in association with the members of th e groups. (pp. 5-6) Fourth, the chance that a person will participate in systematic criminal behavior is determ ined roughly by the frequency and consistency of his contacts with the patterns of criminal behavior If a person could come into contact only with lawful behavior he would inevitably be completely law-abiding. If he could come into
14 contact only with criminal behavior (which is impossible, since no group could exist if all of its behavior were criminal) he would inevitably be completely criminal The actual condition is between these extremes. The ratio of criminal acts to lawful acts by a person is roughly the same as the ratio of the contacts with the criminal and with the lawful behavior of others. It is true, of course, that a single critical experience may be t he turning point in a career. But these critical experiences are gener ally based on a long series of former experiences and they produce their effects generally because they change the personÂ’s associations. One of these critical experiences that is most important in determining criminal careers is the first public appearance as a criminal. A boy who is arrested and convicted is thereby publicly defined as a criminal. Thereafter his associations with lawf ul people are restricted as he is thrown into associations with ot her delinquents. On the other hand a person who is consistently crimi nal is not defined as law-abiding by a single lawful act. Every person is expected to be law-abiding, and lawful behavior is taken for granted because the lawful culture is dominant, more extensive, and mo re pervasive than the criminal culture. (p. 6) Fifth, individual differences among peopl e in respect to personal characteristics or social situations cause crime only as they affect differential association or frequen cy and consistency of contacts with criminal patterns Poverty in the home ma y force a family to reside in a low-rent area w here delinquency rates are high and thereby facilitate association with delinquents. Parents who insist that their boy return home imme diately after school and who are able to enforce this regulation may prevent the boy from coming into frequent contact with delin quents even though the family resides in a high delinquency area A child who is not wanted at home may be emotionally upset, but the significant thing is that this condition may drive him away from the home and he may therefore come into contact with delinquents. A boy who is timid may be kept from association with rough delinquent s. It is not necessary to assume a generic difference between persons by reason of which some are generally receptive to cr iminality and others not receptive. Such an assumption would be far-fetched and unjustified. There may be receptivity at a particular mo ment to a particu lar stimulation, but the elements are so complex that no generalization regarding such receptivity is possible. The closest approach to a generalization is to say that this specific receptivity is determined principally by the frequency and cons istency of previous contacts
15 with patterns of delinquency and that beyond this the delinquent behavior is adventitious. (pp. 6-7) Sixth, cultural conflict is the underlying cause of differential association and therefore of systematic criminal behavior Differential association is possible because society is composed of various groups with varied cultures. These differences in culture are found in respect to many values and are generally regarded as desirable. They exist, also, with reference to the values which the laws are designed to protect, and in that form are generally regarded as undesirable. This criminal culture is as real as lawful culture and is much more prevalent than [is] usually believed. It is not confined to the hoodlums in slum s or to professional criminals. Prisoners frequently state and undoubt edly believe they are no worse than the majority of people on the outside. The more intricate manipulations of business and prof essional men may be kept within the letter of the law as interpreted but be identical in logic and effects with the criminal behavior which results in imprisonment. These practices, even if they do not result in public condemnation as crimes, are a part of the criminal culture. The more the cultural patterns conflict, the more unpr edictable is the behavior of a particular person. It was possible to predict with almost complete certainty how a person reared in a Chinese village fifty years ago would behave because there was only one way for him to behave. The attempts to explain the behavio r of a particular person in a modern city have been unproductive because the influences are in conflict and any particular influen ce may be very evanescent. (pp. 7-8) Seventh, social disorganization is the basic cause of systematic criminal behavior The origin and the persist ence of culture conflicts relating to the values express ed in the law and of differential association which is based on the cultural conflicts are due to social disorganization. Cultural conflict is a specific aspect of social disorganization and in that sens e the two concepts are names for smaller and larger aspects of the same thing. But social disorganization is important in anot her sense. Since the law-abiding culture is dominant and more ex tensive, it could overcome systematic crime if organized fo r that purpose. But society is organized around individual and sma ll group interests on most points. A law-abiding person is more interested in his own immediate personal projects than in abstract social welfare or justice. In this sense society permi ts crime to persist in systematic
16 form. Consequently systematic cr ime persists not only because of differential association but also because of the reaction of general society toward such crime. W hen a society or a smaller group develops a unified interest in crimes which touch its fundamental and common values, it generally succeeds in eliminating or at least greatly reducing crime. This o ccurred for instance, when baseball players in the world series took bribes for throwing away a game they could have won. This affected so many people in a manner which they regarded as vital, and they reacted in such evident opposition, that crime, so far as is known, has never been repeated. Also, when many wealthy people were kidnapped and held for ransom at the end of t he prohibition period, our society reorganized the legal and administrat ive system in violation of the slogans and myth of state sovereignty and such kidnappings practically ceased. However, in previous times when poor and helpless people were victims of kidnappings, as in the slave trade, imprisonment of sailors, shanghaiin g of sailors by crimps, and unjustifiable arrests, it took generations and in some cases centuries for society to become su fficiently aware and interested to stop kidnappings in those fo rms. When a gang starts in a disorganized district of a city it keeps growing and other gangs develop. But when a delinquent gang started on a business street adjacent to Hyde Park, a good resident ial district in Chicago, the residents became concerned, form ed an organization, and decided that the best way to protect them selves was by providing a club house and recreational facilities for the delinquents. This practically eliminated the gangs. T herefore, whether syst ematic delinquency does or does not develop is dete rmined not only by associations that people make with the criminals, but also by the reactions of the rest of society toward systematic cr iminal behavior. If the society is organized with reference to the val ues expressed in the law, the crime is eliminated; if it is not organized, crime persists and develops. The opposition of the society may take the form of punishment, of reformation, or of prevention. (pp. 8-9) SutherlandÂ’s (1939) seven general statem ents refer to systematic criminal behavior, a concept he created to allo w for the formulation of universal statements about crimi nal behavior (propositions one, two, three, four, and five) and crime rates (propositions six and sev en). Sutherland was interested in the
17 causes of criminal behavior generally, t he gross facts regarding crime (Cressey, 1960), as he believed that incidental cr ime, although causally similar to systematic criminal behav ior, would contain exceptional cases due to its adventitious character (Sutherland, 1939, 1973b). Regardless of the concept ual unit of analysis, SutherlandÂ’s (1939) ideas represented a formal organization of hi s earlier approaches to the subject, inherent in the hypotheses, First, any person can be trained to adopt and follow any pattern of behavior which he is able to execute. Second, failure to follow a prescribed pattern of behavior is due to the inconsistencies and lack of harmony in the influences which direct the individual. Third, the conflict of culture is therefore the fundamental principle in th e explanation of crime. (Sutherland, 1934, pp. 51-52) Sutherland (1939) sugges ted that both lawful and unlawful behavior developed from differing me ssages gained during the proc ess of associating with others. Etiologically, Suther land identified differential association, association with people who engage in system atic criminal behavior, as the proximate cause of systematic criminal behavior. Sutherland (1924) reasoned t hat at birth, individuals are born with both innate physiological tendencies and general tendencies that vary by social conditions. Sutherland posited that human nature comprised both individual and group phenomena. Focusing on g eneral tendencies, he argued that intellectual expressions, anger, sympathy, imitation, and the like derive from contacts with others. Although physiologic al tendencies such as sneezing and frowning are
18 innate, and may occur in complete isol ation from others, general tendencies are general expressions of social events that only derive from social interaction (see Sutherland, 1932). Sutherland (1924) main tained that these general expressions would not occur in complete isolat ion from others, and because social interactions vary, both lawful and unlaw ful behavior represent expressions of human natureÂ—expressions of varied soci al interactions that are developed through the same social process (Sutherland, 1932). Influenced by the epidemiology of the Chicago School, Sutherland (1939) viewed social disorganization as the dist al cause of systematic criminal behavior. He argued that historically, society pr ovided uniform and consistent societal influences. As society moved away from small communities, mobility, competition, and conflict resulted in a stat e of social disorganization. Sutherland marks the colonization of America as a starting point to soci al disorganization, particularly noting the industrial revo lution, capitalism, competition, and democracy as strong fact ors. He commented, This sequence of events necessarily resulted in an immense increase in crime. In the first place t he large family and the homogeneous neighborhood, which had been the principal agencies of social control, disintegrated, primarily as the result of mobility. They were replaced by the small family, consisting of parents and children, detached from other relatives, and by a neighborhood in which the mores were not homogeneous, and the behavior of one person was a matter of relative indifference to other persons. Thus the agencies by which control had been secured in almost all earlier so cieties were greatly weakened. (Sutherland, 1939, p. 71) Sutherland (1939) viewed crime as a social phenomenon comprising three
19 elements: appreciated value by a politically im portant group; cult ural conflict by part of the group, resulti ng in unappreciated or le ss appreciated value; and coercion by those who appreciate the va lue against those who do not appreciate the value. Simply, to Sutherland, crime represented the descripti on of events that occurred when one important group sanc tioned mores that were otherwise acceptable behavior to others. Sutherland suggested that all crimes contained this set of relationships when viewed at the group, rather than the individual, level, and he adopted the view that cr ime was an antagonistic action of an individual against oneÂ’s group. Influenced by his work with Sellin (1938), Sutherland (1939) expressed culture conflict as an underlying cause of differential associat ion and therefore a special case of social disorganizat ion. Culture conflict reflects the characterization of the groups creating and punishing the violation of mores, versus the groups not in agreement with t he mores. Culture conflict provides the link between individual criminal behavior t hat stems from differe ntial associations, and crime rates that stem from social disorganization. Sutherland (1939) consider ed culture conflict a sma ller representation of social disorganization. If not for a societal organization of conflicting cultures, a small part of the larger gr oup disagreeing over mores, individuals would have no opportunity to associate with others holdi ng differing values. Culture conflict enables social disorganization to resu lt in systematic criminal behavior.
20 Sutherland emphasized that crime exists only when the violation of such mores does not result in public condemnation, a consensus from the whole group, suggesting that if society organized itse lf against systematic crime, criminal behavior could not exist. Sutherland (1939) intended hi s theory as a tentative statement on criminal behavior and crime, and he invited critic ism. Sutherland (1973a) focused his evaluation of critiques in nine areas: (1) the relationship between differential association, social organization, and cult ure conflict, (2) the distinction between systematic and adventitious crime; (3) the significance of the term differential; (4) the relationship between differential asso ciation theory and TardeÂ’s (1912) theory of imitation; (5) what specifically is l earned in association wit h others; (6) whether non-criminals can invent crim e; (7) the origin of crim e; (8) the modalities of association with criminal versus non-crim inal patterns; and (9) the relationship between personal traits and culture in the genesis of criminal behavior. Further, Sutherland (1973d) vigorously ar gued his notion of the best case against differential association theory in an originally unpublished paper, honing in on opportunity, intensity of need, crime and alternate behaviors, and methodologies (e.g., sufficient causal ity). Sutherland (1947) subsequently revised the theory, incorporating his responses to what he believed to be important criticisms, whether acceptance or refutation, in the groundwork section leading up to his formal propositions the propositions themselves, the
21 commentary immediately fo llowing the propositions, and the remainder of his book. First, Sutherland (1947) focused attention on methods of scientific explanation. He specifi ed that he was searching for necessary and sufficient causes, organized in the form of universal statements that, still consistent with analytic induction, contained no exceptions. To achieve these universal proposi tions, Sutherland (1947) noted the desirability of abstracting the multiple fa ctors that operate at the instant of occurrence to their common elements. Such abstract propositions treated criminal behavior as a class of event s, emphasizing the interrelations among various patterns of behavior (see Su therland, 1973d). Sutherland sought the intervening mechanisms (see Matsueda, 1988) that occurred in the genesis of criminal behavior, the hi story of behavior that wa s present just before the instance of expressed needs, values, goals, and the like (Sutherland, 1947; Sutherland, 1973d). Sutherland (1947) sought to distingui sh criminal from noncriminal behavior (Suther land & Cressey, 1969), argui ng that general needs and values require explanation because both criminal and non-criminal behavior represent an expression of general needs and values. Sutherland (1947) suggested t hat it was essential to a universal statement of criminal behavior to reinte rpret concrete factors know n to correlate with crime, such as race, urbanicity, and offender age, so that their abstract mechanisms
22 became apparent. Sutherland noted that otherwise, a general statement about these correlations would be incorrect because the correlations contain exceptions. For instance, not all Afric an Americans commit crime, not all city dwellers commit crime, nor do all juven iles. Sutherland insisted that knowing about these correlations was important, but that a useful theory, one offering universal statements, mu st identify the commonalit ies between the correlates and crime. A useful, universal theory mu st identify the comm onalities present in criminal behavior yet absent in non-cr iminal behavior (Sutherland & Cressey, 1969). Sutherland (1947) offered abstrac tion as a tool for this purpose. Next, Sutherland (1947) di fferentiated levels of ex planation. He delimited the problem under analysis to a small par t of the larger problem, removing macrosocial statements from his criminological theor y and thus restricting his propositions to the individua l level. He was interested in the chronology of the criminological problem, and viewed it des irable to hold constant earlier causal processes in the expression of indivi dual criminal behavior (Sutherland & Cressey, 1969). Sutherland (1947) dispens ed with formally seeking distal universal statements as to why an individual has differential associations, the proximate cause of criminal behavior, instead readdr essing that issue elsewhere in the book. Sutherland argued that such restrict ed causal analysis was necessary in order to find valid generaliz ations. He sought a simple, temporal statement that
23 distinguished criminal behav ior from non-criminal beha vior, suggesting that it made no difference in the quest for valid generalizationsÂ—the derivation of universal statementsÂ—how the behaviors themselves came to be. After specifying the methodology, Sutherland (1947) described two potential research avenues for explaining criminal behavior: explain the instant causes of criminal behavior, the proc esses operating at the moment of crime (Sutherland & Cressey, 1969), or explai n the processes working in the earlier history of criminal behavio r. Sutherland referred to the instant causes approach as mechanistic, situational, or dynamic (Sutherland and Cressey, 1969), and he dismissed the approach as falsely separatin g the individual from the situation, falsely separating the individual from life experiences that define certain situations as opportunities for law br eaking (Sutherland, 1947; Sutherland & Cressey, 1969). Conceding that a situati onal explanation w ould be superior to other explanations if achievable in a useful manner, Sutherland (1947) considered instant causes the particulari stic product of multiple factors. He believed it impossible to isolate and derive universal stat ements from such personal and social pathologies. Sutherland (1947) instead fa vored the earlier history approach, labeling it genetic or historical. The genetic approac h examined the processes working in the earlier history of crimi nal behavior, identifying crim inological antecedents in the genesis of criminal behavior (S utherland, 1973a). Drawing on symbolic
24 interactionism (Mead, 1934; see Dewey, 1931) and his work on criminal life histories (Sutherland, 1937), Sutherland (1 947) held that t he individualÂ’s life experience is important to engagement or not in crime. SutherlandÂ’s revised statement of differential a ssociation theory is concer ned with explaining criminal behavior from the perspective of the individuals engaging in the behavior, maintaining that criminal acts occur w hen individuals define presented situations as appropriate for the criminal act. In his earlier statement of the theory, Sutherl and (1939) created the term systematic criminal behavior in order to ignore instant processes that he believed to be rare and incidental. He argued t hat had he looked at behavior generally, rather than systematic behavio r, trivial exceptions would have prevented the derivation of universal st atements (see Sutherland, 1973a). In the revision, Sutherland (1947) tackled the issue of multiple factor s in individual criminal behavior in a way that allowed him to e liminate systematic criminal behavior as a proxy for that behavior. Sutherland (1973a) realized that he was unclear in his original statement and that critics misunderstood the term syst ematic criminal behavior. Moreover, he found that researchers had difficulty distinguishing systematic criminal behavior from adventitious criminal behav ior. Sutherland (1947) still viewed abstraction as the solution to making universal statements about behavior with multiple causes at the inst ant of occurrence, but in t he revision, he abstracted
25 these multiple factors to their co mmonalities without l abeling such phenomena systematic. Sutherland used the same argument, elaborating a bit on the rationale, but he abandoned the term systematic. As he had originally used the term out of convenience, and realizing that that it no longer held utility (see Sutherland, 1973a), for few understood w hat he meant, Sutherland (1947) advanced his theory revision as pertaining to all crime. His final statement of differential association, with his inclusive commentary, postulated, Genetic Explanation of Criminal Behavior. The following statement refers to the process by which a particular person comes to engage in criminal behavior. 1. Criminal Behavior Is Learned. Negatively, this means that criminal behavior is not inherited, as such; also, the person who is not already trained in crime does not invent criminal behavior, just as a person does not make mechanical inventions unless he has had a training in mechanics. 2. Criminal behavior is learned in interaction with other persons in a process of communication. This communication is verbal in many respects but includes also Â“the communication of gestures.Â” 3. The principal part of the lear ning of criminal behavior occurs within intimate personal groups. Negatively, this means that the impersonal agencies of communication, such as picture shows and newspapers, play a relative ly unimportant part in the genesis of criminal behavior. 4. When criminal behavior is lear ned, the learning includes (a) techniques of committing the crim e, which are sometimes very complicated, sometimes very simple ; (b) the specific direction of motives, drives, rationalizations, and attitudes. 5. The specific direction of motives and drives is learned from definitions of the legal codes as favorable or unfavorable. In some societies an individual is surrounded by persons who invariably define the legal codes as rules to be observed, while in others he is surrounded by persons whose definitions are favorable to the violation of the legal codes. In our American society these definitions ar e almost always mixed and
26 consequently we have culture conflict in relation to the legal codes. 6. A person becomes delinquent because of an excess of definitions favorable to violat ion of law over definitions unfavorable to violation of law This is the principle of differential association. It refers to both criminal and anti-criminal associations and has to do with counteracting forces. When persons become criminal, they do so because of contacts with criminal patterns and also because of isolation from anti-criminal patterns. Any person inevitably assimilates the surrounding culture unless other patterns are in conflict; a Southerner does not pronounce Â“rÂ” because other Southerners do not pronounce Â“r.Â” Negatively, this proposition of differential association means that associations which are neutra l so far as crime is concerned have little or no effect on the genesis of criminal behavior. Much of the experience of a person is neutral in this sense, e.g., learning to brush oneÂ’s teeth. This behavior has no negative or positive effect on criminal behavio r except as it may be related to associations which are conc erned with the legal codes. This neutral behavior is important espec ially as an occupier of the time of a child so that he is not in contact with criminal behavior during the time he is so engaged in the neutral behavior. 7. Differential associations may vary in frequency, duration, priority, and intensity. This means that associations with criminal behavior and also associations with anti-criminal behavior vary in those respects. Â“FrequencyÂ” a nd Â“durationÂ” as modalities of associations are obvious and need no explanation. Â“PriorityÂ” is assumed to be important in the sense that lawful behavior developed in early childhood ma y persist throughout life, and also that delinquent behavior de veloped in early childhood may persist throughout life. This tendency, however, has not been adequately demonstrated, and priori ty seems to be important principally through its selective influence. Â“IntensityÂ” is not precisely defined but it has to do with such things as the prestige of the source of a criminal pattern and with emotional reactions related to the associations. In a precise description of the criminal behavior of a per son these modalities would be stated in quantitative form and a ma thematical ratio be reached. A formula in this sense has not been developed and the development of such a formula would be extremely difficult. 8. The process of learning crimin al behavior by association with criminal and anticriminal patterns involves all of the mechanisms that are involved in any other learning. Negatively, this means
27 that learning of criminal behavior is not restricted to the process of imitation. A person who is seduced, for instance, learns criminal behavior by associati on but this process would not ordinarily be described as imitation. 9. While criminal behavior is an expression of general needs and values, it is not explained by those general needs and values, since noncriminal behavior is an expression of the same needs and values Thieves generally steal in order to secure money, but likewise honest laborers work in order to secure money. The attempts by many scholars to explain criminal behavior by general drives and values, such as the happiness principle, striving for social status, the m oney motive, or frustration, have been and must continue to be futile since they explain lawful behavior as completely as they explain criminal behavior. They are similar to respiration, whic h is necessary for any behavior but which does not differentiate criminal from non-criminal behavior. (Sutherland, 1947, pp. 6-8) SutherlandÂ’s (1947) ni ne statements combine to form a general explanation of the indi vidual formation of criminal behavior. Differential association theory offers a broad explanation of crim inal behavior by advancing universal crime causes that exist regardle ss of earlier social or instant individual conditions (Sutherland & Cre ssey, 1970; Matsueda, 1988). Sutherland (1947) discounted typologic al (proposition one) and micro strain implications of anomie theory (p roposition nine), instead drawing on the symbols and gestures (language, action, appearance) implied by symbolic interaction (proposition two), and the broad sociological suppos ition of learned behavior. Sutherland considered proposit ion six, an excess of criminal definitions, the central st atement of the microsoc ial theory. Differential association theoryÂ’s primary assertions ar e that heredity plays no role in crime,
28 and that criminal behavior is learned in differential association with influential groups holding contradictory definit ions of law violation. Although Sutherland (1947) st ated that the revision was restricted to the individual level of analysis, he did revisi t his earlier exposition of crime rates (Sutherland, 1939) in his commentary imme diately following the revised general propositions. Moreover, Sutherland (1947) retained the concept of culture conflict, using it to expound on the proposit ion five notions of favorable and unfavorable definitions of the legal code as a manifestation of groups holding contradictory definitions of law. Conseque ntly, despite the qualifications on levels of analysis, and in a different form, Suther land (1947) did implicit ly maintain that criminal behavior derives from a set of complex interrelationships between differential associations, culture c onflict, and social disorganization (see Sutherland, 1939). Although Sutherland (1947 ) specified a distinct microsocial explanation for criminal behavior, t he theory remained consistent with the macrosocial explanation fo r crime rates afforded by the idea of social disorganization (see Cressey, 1960; Matsueda, 1988). Sutherland (1947; 1973a) plac ed differential associati ons into the context of what he called Â“differ ential social organizationÂ” or Â“differential group organization,Â” his preferred terms for Shaw and McKayÂ’s (1942) description of social disorganization. Ag reeing with the notion of social disorganization, Sutherland (1973a) thought the te rm itself reflected a parti cularistic point of view.
29 He thought the term differential social organization better captured both types of group organizationÂ—groups organized for criminal behavior and groups organized against criminal behavior. Sutherland (1947) suggest ed that in a uniform organization of people, there is only one behavioral pattern. In groups (communities) with no uniform organization, such as those developed thr ough mobility or culture conflict, crime may occur. Sutherland viewed culture c onflict as Â“the basic principle in the explanation of crimeÂ” (Sutherland, 1973a, p. 20). He view ed crime, enabled by culture conflict, as an expression of social disorganization. He viewed differential social organization as an explanation for crime rate s (the collective sum of individual crimes) and differential associat ions as the explanation of individual criminal behavior. Sutherland (1947) suggested that differential social organization provides the opportunity for differential associations to occur. By removi ng social structural statements from the explicit propositions of the final version of the theory, however, Sutherland did not formally ex press the links between social structure and criminal behavior. He continued to s uggest that social disorganization and normative conflict (Cressey, 1960; Matsueda, 1988) play a role in the formation of individual criminal behavior, bu t he abstracted the concepts to the term differential social organization, and he expressed no specific postulates. Differential association theory is conceptual. Sutherland (1939, 1947)
30 proposed theoretical relationships between sociological concepts, but he did not operationalize or test his propositionsÂ—he offered no data, but rather advanced a theory he believed would fi nd support when tested. Although research supported the majo r differential association theory theme (Glaser, 1954; Glueck & Glueck, 1950; Short, 1957, 1958; Reiss, 1951; Reiss & Rhodes, 1964; Voss, 1964; see Glaser, 1960), some researchers expressed concerns that the theory oversimplified the process of learned behavior because it did not fully specify the learning mechanisms that affect behavior (Ball, 1957; see Short, 1960; for a thorough discussion of literary and theoretical critiques, see Cressey, 1960; Sutherland & Cressey, 1970, 1974). The theoryÂ’s propositions combine for a genetic (historical) explanation of the processes that affect engagement in criminal behavior (Sutherland, 1947). Although stressing an individualÂ’s definition of situations, the process that allows an individual to view various situations as opportunities for law violation, the theory proposes that criminal behavior involves all of the mechanisms involved in learning other kinds of behavior. Howeve r, differential association theory does not identify those mechanisms. Social Learning Theoretical Statement Burgess and Akers (1966) addressed the task of specifying the learning process left implicit by Sutherland ( 1947). They were influenced by Cressey (1960), who commented, [Differential association theory cr iticism] ranges from simple
31 assertions that the learning proc ess is more complex than the theory states or implies, to the idea that the theory does not adequately take into account some specific type of learning process, such as differential identification. Between these two extremes are assertions that the theory is inadequate because it does not allow for a process in which criminality seems to be Â“independently inventedÂ” by the ac tor. I am one of the dozen authors who have advanced this kind of criticism, and in this day of role theory, reference group theory, and complex learning theory, it would be foolhardy to assert that this type of general criticism is incorrect. But it is one thing to [c riticize] the theor y for failure to specify the learning process a ccurately and another to specify which aspects of the learning pr ocess should be included and in what way. (pp. 53-54) Cressey (1960) dismissed research-free criticisms as proposals for research, rather than valid critiques of di fferential association theory. Initially called differential associat ion-reinforcement theory (Burgess & Akers, 1966), social learning theory (A kers, 1973, 1977, 1985, 1998) draws from psychological behavioral and social cognitive theories to specify the differential association learning process. Unlike Jeffery (1965), who also tried to operationalize the learni ng process, Burgess and Ak ers kept the core of SutherlandÂ’s (1947) theory intact. They restated differential association theory statement by statement in behavioral te rms in a numbered format that coincided with the nine differential associat ion theory statement s (statement one concurrently addressed differential asso ciation theory stat ements one and eight). Burgess and Akers proposed, 1. Criminal behavior is learned a ccording to the principles of operant conditioning. (Bur gess & Akers, 1966, p.146) 2. Criminal behavior is learned both in nonsocial situations that are reinforcing or discriminative and through that social interaction
32 in which the behavior of other persons is reinforcing or discriminative for criminal behav ior. (Burgess & Akers, 1966, p.146) 3. The principal part of the learni ng of criminal behavior occurs in those groups which comprise or control the individualÂ’s major source of reinforcements. (Burgess & Akers, 1966, p.146) 4. The learning of criminal behavio r, including specific techniques, attitudes and avoidance procedu res, is a function of the effective and available reinforcers, and the existing reinforcement contingencies. (Burgess & Akers, 1966, p.146) 5. The specific class of behav iors which are learned and their frequency of occurrence are a func tion of the reinforcers which are effective and available, and the rules or norms by which these reinforcers are applied. (Burgess & Akers, 1966, p.146) 6. Criminal behavior is a function of norms which are discriminative for criminal behavior, the learning of which takes place when such behavior is more highly reinforced than noncriminal behavior. (Burgess & Akers, 1966, p.146) 7. The strength of criminal behavio r is a direct function of the amount, frequency, and probability of its reinforcement. (Burgess & Akers, 1966, p.146) 9. (Omit from theory.) (B urgess & Akers, 1966, p.146) Burgess and Akers (1966) argued that SutherlandÂ’s (1947) supposition that learning occurs through interacti on with others in social environments was compatible with the operant theory notion that environm ent shapes individual behavior. Burgess and Akers subscribed that if one accepted the notion that differential association theory was essentia lly a learning theory, and that criminal behavior and non-criminal behavior are l earned through the same process, then it was reasonable to incorporate moder n learning knowledge into the theory. They further believed that by incorporat ing previous changes to differential association theory (Cressey, 1953; Hart ung, 1965; Jeffrey, 1965; Sykes & Matza,
33 1957), with their blending of the symbolic interactionist and behaviorist traditions, their reformulation offered a testable general theory of human behavior (Akers, 1998). Burgess and Akers (1966) suggested that modern lear ning theory had sufficiently advanced to the point that Su therlandÂ’s (1947) implicit mechanisms were specifiable. They emphasized that whereas SutherlandÂ’s differential social organization had sufficiently made sens e of crime rates through the idea of normative conflict, the explanation offered for the individual level process was less satisfying because, making use of Vold (1958), psychology and social psychology had not previously advanced enough to distinguish such qualitative differences in human behavior. Sociology did not sufficiently understand determining variables at the individual le vel of analysis (Burgess & Akers, 1966). Burgess and Akers (1966) offered diffe rential association-reinforcement theory as an explanation for why some persons exposed to normative conflict engage in criminal behavior. They, like Su therland (1947), viewed their theory revision as consistent with sociologic epidemi ological explanations for variation in crime rates. However, differential a ssociation-reinforcement theory, like differential association theory, sought an etiological explanation for criminal behavior. Akers (1973, 1977, 1985) cl arified and revised the seminal differential association-reinforcement model and renamed it social learning theory, tweaking
34 the serial propositions along the way. Soci al learning theory expands differential association theory. It is not a compet ing explanation. It offers a broader explanation, specifying the learning process and behavioral mechanisms for all types of deviant behavior, but it does not invalidate the co re supposition of differential association theory. Empirica l support for differential association theory, therefore, supports social learning theory (Akers, 1998). Social learning theory no longer relies on the serial statements that tied it to classic differential association theor y. Instead, the most recent statement describes the social learning process na rratively. Akers (1998) postulated, The probability that persons will engage in criminal and deviant behavior is increased and the probability of their conforming to the norm is decreased when they differ entially associate with others who commit criminal behavior and es pouse definitions favorable to it, are relatively more exposed in -person or symbolically to salient criminal/deviant models, define it as desirable or justified in a situation discriminativ e for the behavior, and have received in the past and anticipate in the current or future situation relatively greater reward than punishment for the behavior. (p. 50) Social learning theory st resses four concepts. Differential association is an elaboration of that presented in differential association theory (Sutherland, 1947), and it provides the social context for t he other three concepts (Akers et al., 1979), the context for the mechanisms inher ent in the social learning of behavior (Akers & Sellers, 2004). Differential a ssociation refers to exposure to the attitudes and behaviors of others. Such ex posure may be direct or indirect and verbal or nonverbal (Akers, 1998).
35 Differential association is mainly a lat ent construct of interactional (direct associations with the behavior of others) and normative (exposure to patterns of norms and values) dimensions (Akers, 1998) Associations occur in primary and secondary reference groups such as fam ily, peers, school, work, church, and the like. Each reference group contribute s to the learning process through association modalities (Akers, 1998; Sut herland, 1947), providing the context for behavior. Akers (1998) relies on the four modalities of association initially identified by Sutherland: frequency, duration, priority, and intensity (Akers, 1998; Sutherland, 1947). Frequency refers to how often one associ ates with another, whereas duration identifies the amount of time spent in those associations. Priority time-orders the influence of a ssociations, and intensity estimates their importance (e.g., how clos e one feels to another). There is much research on peers and delinquency, with peer association usually measured as the summation of t he number or a proporti on of friends who engage in delinquent behavior. However, a comprehensive measure of differential association captures more than the single-item measure of the number of deviant friends. The concept invo lves influential associations broadly to include more groups than friends al one, as well as varied modalities of association (e.g., Akers et al., 1979; Lee et al., 2004). Akers and colleagues (1979) comment, [P]rincipal behavioral effects come from interaction in or under the
36 influence of those groups which control indivi dualsÂ’ major sources of reinforcement and punishment and expose them to behavioral models and normative definitions The most important of these groups with which one is in differential association are the peer-friendship groups and the family but they also include schools, ch urches, and other groups. (p. 638) The literature reports a consistent correlation between delinquent behavior and delinquent friends (Akers et al., 1979; Brownfield & Thompson, 2002; Elliott et al., 1985; Glueck & Glueck, 1950; Hirschi, 1969; Jaquith, 1981; R. Johnson et al., 1987; Matsueda & Anderson, 1998; Short, 1958; Voss, 1964; Zhang & Messner, 2000). The number of delinquent fr iends one has is the best external predictor of an individualÂ’s criminal behav ior (Akers et al., 1979; Elliott et al., 1985; R. Johnson et al., 1987; Warr, 2002). The best external predictor of an adolescentÂ’s incidence and amount of drug use is the extent of association with others who use drugs (Elliott et al ., 1985; Jaquith, 1981; see also Flom, Friedman, Kottiri, Neaigus & Curtis 2001; Urberg, 1997). Scholars differ, however, on their interpretation of peer associations. Some scholars view differential asso ciation (Akers, 1998; Sutherland, 1947) as associating with bad companions. T he supposition is that Â“birds of a feather flock togetherÂ” (Glueck & Glueck, 1950, p. 164). Scholars suggest that delinquents may seek out other delin quents because of common interests (Glueck & Glueck, 1950; M. Gottfredson & Hirschi, 1987; Hirschi, 1969). Besides the social selection effect (Robbins, 1974) they also note that delinquent acts often occur in groups (Erickson & Jensen, 1997; Gold, 1970; see also Warr,
37 1996, 2002). In such interpretations, the onset of delinquency precedes the onset of exposure to deviant others. Furt her, some scholars suggest that the relationship between delinquent behavio r and delinquent friends may be spurious. Indirect measures of peer delinquency may represent the same construct as self-reported delinquen cy (M. Gottfredson & Hirschi, 1987; M. Gottfredson & Hirschi, 1990; Kandel, 1996; see also Regnerus, 2002; Urberg, 1992; Zhang & Messner, 2000). Other scholars view the onset of expo sure to deviant friends as occurring before the onset of delinquency (Akers 1998; Bandura, 1977; Burgess & Akers, 1966; Elliott & Menard, 1996; Sutherland, 1947). Further, so me scholars do not view peer delinquency as an ar tifact of self-reporting m easures, but rather view self-reported delinquency and r eporting of peer deviancy as distinct measures of delinquency (Flom et al., 2001). Moreover perceived peer behavior may be as important as actual peer behav ior (Iannotti & Busch, 1992). Social learning theory suggests that the onset of exposure to deviant friends typically occurs before the onset of delinquency (Akers, 1998). However, the theoryÂ’s reciprocal model does not preclude delinquents from forming associations with other delinquents (Ake rs & Lee, 1996; Elliott & Menard, 1996; Warr, 2002). Rather, social learning t heory predicts (Akers, 1998) and research supports (Farrell & Danish, 1993; Jesso r, Jessor & Finney, 1973; Kandel & Davies, 1991; Krohn, Lizotte, Thornberry Smith & McDowall, 1996; Oetting &
38 Beauvais, 1987; Sellers & Winfree, 1990; Warr, 1993) peers influencing each other mutually (but see discussion in Sampson, 1999). Social learning theory addresses the c ausal ordering of peer associations and deviancy through the differential a ssociations concept, and its various modalities of association. The notion of priority (A kers, 1998; Sutherland, 1947) suggests that associations formed earlier in life may have greater influence than later-formed associations. Families prov ide early contingencies for reinforcement and punishment (Patterson & Dishion, 198 5), typically providing normative orientations (Bauman, Foshee, Linzer & Ko ch, 1990; Elliott et al., 1985; Kandel & Andrews, 1987; Patterson & Dishion, 1985). Family associations precede peer associations, except in rare circ umstances, and may s pan a greater period (Akers, 1998). However, frequency, durat ion, and intensity also influence behavior, and parents are typically more in fluential in early adolescence than in later years (Allen, Donohue, Griffin, Ry an & Mitchell-Turner, 2003), a time when peers have more influence (Jang, 1999, 2002). Although association m easures are the most common social learning variables used to test the theory, and often the only measure included in research (Akers, 1998), the other three concepts offer import ant understanding of the social learning process. The second social learning concept, definitions is also an elaboration of that presented in differential associati on theory (Sutherland, 1947). Definitions
39 refer to an individualÂ’s (Akers, 1998; Sut herland, 1947) attitudes toward deviant or conforming behavior (Akers, 1998), yet t hey allow that the attitudes of others may also be important (Akers, 1998). De finitions occur through contingencies of reinforcement, and they may generally or specifically favor deviancy (positive definitions), oppose deviancy ( negative definitions), or ju stify or excuse deviancy under certain conditions despite generally opposing certain behavior (neutralizing definitions). Once formed, definitions serve as cues (discriminative stimuli) to anticipated reinforcement or punishment for certain behavior (Akers, 1998). Social learning researchers have thus fa r identified, or in corporated, four definition dimensions (see Ak ers, 1998): beliefs (Hirschi, 1969; see Akers, 1998), attitudes (Burgess & Akers, 1966; Cr essey, 1953; Sutherland, 1947), justifications/rationalizations (Cressey, 1953; Sutherland, 1947; Sykes & Matza, 1957), and orientations (Sutherland, 1947). Measurements of general lawabiding or law-violating attitudes (e.g ., Akers et al., 1979), approval or disapproval of specific acts (e.g., Akers et al., 1979), and justifications or excuses for specific behavior (e.g., Akers et al ., 1979; Sykes & Matza, 1957) index the definitions concept. Imitation the third social learning conc ept, stems from social cognitive theory (Bandura, 1977). Imitati on represents an incorporati on of modern learning theory ideas that alte r SutherlandÂ’s (1947) view that imitation plays little role in
40 criminal behavior. Imitation involves the idea that indivi duals note and model the behavior of admired others. By watching others and not ing the outcomes, individuals are able to deduce probable outcomes from adopting the behavior. Imitation may be more important to the onset of deviant behavior as opposed to its effect on the continuance or desistance of behavior (A kers, 1998). Measurem ents of admired models who engage in certain behaviors index im itation (e.g., Akers et al., 1979). The fourth social learning concept, differential reinforcement stems from behavioral theory (B.F. Skinner, 1953) and refe rs to the instrumental conditioning of behavior. Individuals anticipate the out come of present or future behavior based on the reward or punishment of past or present behavior (Akers, 1998). Measurements of social and nonsocial expec tations of the rewards or costs of a certain behavior index differ ential reinforcement (e.g., Akers et al., 1979). Social learning theory identifies f our concepts involved in learned behavior, but they are not equally import ant. Further, behavior is complex and the theory anticipates t hat the concepts feedback into one another through the individual thought process, affecting future behavior (Akers, 1998). Social learning theory postulates that behavio r is determined by the frequency, amount, and probability of past and present environm ental consequences. Akers (1998) comments, The typical process of initiation, continuation, progression, and desistance is hypothesized to be as follows:
41 1. The balance of past and current associations, definitions, and imitation of deviant models, and the anticipated balance of reinforcement in particular situat ions, produces or inhibits the initial delinquent or deviant acts. 2. The effects of these variables continue in the repetition of acts, although imitation becomes less import ant than it was in the first commission of the act. 3. After initiation, the actual social and nonsocial reinforcers and punishers affect the probability that the acts will be or will not be repeated and at what level of frequency. 4. Not only the overt behavior, but also the definitions favorable or unfavorable to it, are affected by the positive and negative consequences of the initial acts. To the extent that they are more rewarded than alternative behavior, the favorable definitions will be strengthened and the unfavorabl e definitions will be weakened, and it becomes more li kely that the deviant behavior will be repeated under si milar circumstances. 5. Progression into more frequent or sustained patterns, rather than cessation or reduction, of criminal and deviant behavior is promoted to the extent that rein forcement, exposure to deviant models, and norm-violating definitions are not offset by negative formal and informal sanctions and norm-abiding definitions. (pp. 53-54) Akers (1998) advances four separate testable hypotheses, explaining, The individual is more likel y to commit violations when: 1. He or she differentially associates with others who commit, model, and support violations of social and legal norms. 2. The violative behavior is differentially reinforced over behavior in conformity to the norm. 3. He or she is more exposed to and observes more deviant than conforming models. 4. His or her own learned defin itions are favorable toward committing the deviant acts. (p. 51) A comprehensive examination of social learning theory indexes each of the theoretical concepts (Akers, 1998). Diffe rential associations are so important to the statement of the theory and the resulting research, however, that some
42 scholars (Stafford & Ekland-Olson, 1982; Strickland, 1982) question the analytic path implied by the Akers and colleagues (1 979) model. Still others question the need to measure differential associati ons simultaneously with definitions, imitation, and differential re inforcement (Krohn, 1999). Strickland (1982) suggested that the direct effect of differential associations is the most important predictor of delinquent behavior. LanzaKaduce, Akers, Krohn, and Radosevich ( 1982) pointed out that Akers and colleagues (1979) did not order the internal components of the social learning process. Beyond identifying theoretically derived causal linkages, they noted that the hypotheses did not order these linkages. Akers and colleagues instead suggested that there should be a high degree of interc orrelation between the social learning concepts and that sorting out the interrelationships would require longitudinal research. Krohn (1999) added to the complexity of the social learning variable ordering debate. He noted t hat there is a problem wit h thinking of differential associations as a summary concept and including combined measures of it with its definitions, imitation, and differ ential reinforcement components. When viewing differential associations as a summary concept, and typically the most powerful predictor of de linquency in models measuri ng it, Krohn suggested that measuring its component parts is unnec essary. Krohn suggested measuring the component mechanisms absent associatio n measures as an alternative,
43 preferred approach. The first approach k eeps differential association theory as originally advanced, whereas the alter native recognizes social learning theoryÂ’s contribution. Akers (1999) responded to this suggesti on by stressing that each of the four concepts mutually comprise the ma jor components of social learning. He remarked that social learning theory is not as concerned with how precisely the concepts interrelate than it is with explai ning criminal and deviant behavior. Akers suggests that removing measures of associ ations from empirical tests will result in less understanding of such behavio r. Akers (1999) comments, To say that an empirical meas ure can both index differential association and have the added benefit of functioning as a summary index of unmeasured proc esses does not mean that it can perform as a complete proxy m easure for all of the other major concepts. It does not mean that there is no need to measure anything else in social learning or that its presence in empirical models renders all other measures of social learning variables redundant. (p. 488) Akers instead suggested that a more prudent approach is to continue developing measures of the four majo r concepts, as well as identifying and exploring other learning mechanisms. Recently, Akers (see Lee et al., 2004) has tested social learning as a latent construct comprising the indicators differential association, definitions, and differential reinforcement. Although he did so without much explanation, and the approach may have been utilized for conven ience in order to use structural equation modeling to test social lear ning as a mediator of macrosocial
44 dimensions, what may seem at first to be an apparent departure in positions may not be inconsistent with hi s previous arguments. Akers (1999) posits that each of the four social learning concepts, as well as other unidentified measures together produces social l earning, and that it is inappropriate in cross-sectional research to employ structural equation modeling to parse out causality. He instead prefers to view soci al learning as a combined process, more important in its sum t han in its component parts. This is not necessarily inconsistent with his earlie r comments (see Lanza-Kaduce et al., 1982) explaining that the social learning measures have notable overlap with one another and cannot be easily parsed into a causal model as attempted by Strickland (1982). Akers (Lanza-Kaduce et al, 1982) has previously stated that causal modeling implies a closed system that does not allow for inadequate measures and excluded variables, but he stresses t hat the causal approach is desirable when acceptable data exist. Moreover, Aker sÂ’ (Lee et al., 2004) use of social learning as a latent constr uct comprised of differentia l associations, differential reinforcement, and definitions, rather than trying to parse out causality, instead takes the notion of a social learni ng mechanism whose component parts are unnecessary one step further. Akers, in using social learning as a latent construct, whatever his intent, effectiv ely advances rather than retracts his argument that how precisely the social learning concepts interrelate is less
45 important than how well they expl ain criminal and deviant behavior. Beyond the social learning model, anot her important debate relevant to the present study is that of rival te sts and integrated theor y. No single theory accounts for all the variation in crim e; thus, more than one explanation is possible. Although behavior is comple x and one theory may have difficulty identifying the causes underlying all dev iance (A. Cohen, 1962; Glueck, 1956; Glueck & Glueck, 1950; Hirschi & Selvin, 1967; Sutherland, 1924; Tittle, 1985, 1989), multiple theories undermine the role of theory as a means of organizing ideas to advance research (Bernard, 1990, 2001; Bernard & Ritti, 1990; Bernard & Snipes, 1996; Gibbs, 1972). Theory competition (Liska, Krohn & Messner, 1989) is a common approach to reducing multiple theoretica l explanations that promotes testing competitive theories against each other to aid in falsification (Bernard & Snipes, 1996; Liska et al., 1989). The assumption is that some theories (e.g., strain, control, differential association) are fundamentally incompat ible (Hirschi, 1969, 1979; Kornhauser, 1978). Incompatible theories produce contradictory hypotheses, and tests of t hese hypotheses using the same data result in a crucial test (Hirschi, 1989; Liska et al., 1989). Incompatible hypotheses cannot be correct simultaneously, thus the theory garnering more suppor t must be more believable (Elliott, 1985; Liska et al., 1989). For example, HirschiÂ’s (1969) control theory (referred to by Akers as social
46 bonding theory; for a thorough discussion of its empirical status see Kempf, 1993) is arguably the most important social learning theory rival. Researchers commonly pit the two theories against eac h other in the liter ature. Further, Hirschi and Akers have debated the theoretical adequacy of their oppositional theories, measurement conc epts, derived propositions, empirical findings, the notion of peer associations, culture conf lict, and theory competition versus theory integration. There is much research in the literatur e that examines social learning and social bonding variables, among others, simultaneously on the same data. When researchers employ theory competit ion, social learning concepts and propositions typically find more su pport than those derived from other simultaneously tested theories (Akers & Cochran, 1985; Alarid et al., 2000; Benda, 1994; Benda & Corwyn, 2002; Brow nfield & Thompson, 2002; Burton et al., 1994; Dembo, Grandon, La Voie, Schmeidler & Burgos, 1986; Kandel & Davies, 1991; Krohn, Lanza-Kaduce & Akers, 1984; Matsueda & Heimer, 1987; Rebellon, 2002; White et al., 1986; Winfree & Bernat, 1998). Some scholars argue that empirical t heory competition is an unsatisfactory approach to theory reduction (Bernard, 2001; Bernard & Snipes, 1996; Elliott, 1985; Elliott, Ageton & Cantor, 1979; Elliott et al., 1985). They suggest that pitting theories against each other may not be useful because testable hypotheses are not often rival. Predictions are often vague, and accepting one theoryÂ’s
47 hypothesis does not necessarily require rejecting another theoryÂ’s hypothesis (Elliott, 1985). Further, crime and delinquency causal processes may be more complex than the explanations offered by criminological theory (E lliott, 1985; Tittle, 1995). Many tests of theories find small statistical significance with questionable substantive meaning (Elliott, 1985). Thus, th ere are many believable theories that account for little variation in crime (Elliott, 1985; Tittle, 1995). Theory competition has not significant ly reduced the number of competing criminological explanations (Bernard, 2001; Bernard & Snip es, 1996). Theory integration is an alternative approach that promotes wide-ranging explanations by linking more than one t heory together (Bernard, 2001; Bernard & Snipes, 1996; Liska et al., 1989). The goal of theor y integration is to unify theory into comprehensive explanations having greater explanatory power than constituent theories (Farnworth, 1989). The assumption is that although competing theories offer different predictions, the predicti ons are not necessarily contradictory (Bernard & Snipes, 1996; Elliott, 1985). Although theory integration offers an alternative to theory competition, theory elaboration (Thornberry, 1989) o ffers a compromise between theory competition and theory integration. In su ch an approach, the scholar seeks broad implications of a theory through modifi cation and refinement (Thornberry, 1989; Tittle, 1995). The goal of theory elaboration is to extend a theory to its limit by
48 incorporating compatible concepts and propositions as needed, increasing the preexisting theoryÂ’s explanatory power (T hornberry, 1989). At its outer reaches, especially in its outcome (Thornberry, 1989), theory elaboration is similar to theory integration (Ber nard, 2001; Bernard & Snipes, 1996) and may be necessary to progress to su ch a level (Tittle, 1995). Several elaborated and int egrated theories exist in the literature, varying by their incorporation of added conc epts, propositions, and variables. For example, scholars have integrated element s from such theories as control and social learning (Akers & Lee, 1999; Kr ohn, 1986; Thornberry, 1987); strain, control, and social learning (Akers & Cochran, 1985; Elliott et al., 1985; Hoffmann, 2002); labeling, c ontrol, and social learning (Braithwaite, 1989); and rational choice, control, and soci al learning (Tittle, 1995). When researchers apply social lear ning concepts and propositions to integrated theory, social learning variabl es typically have the strongest effect (Conger, 1976; Elliott et al., 1985; R. J ohnson et al., 1987; Lanza-Kaduce & Klug, 1986; Lewis, Sims & Shannon, 1989; Marcos et al., 1986; Thornberry et al., 1994; White & LaGrange, 1987; see also Michaels & Miethe, 1989; H. Kaplan, Martin & Robbins, 1984). Further, schol ars have noted overl ap between social learning theory and several alternative t heories, suggesting that their concepts and propositions are special cases of social learning concepts. Examples of such theories include control (Akers, 1973, 1977, 1989; Pearson & Weiner, 1985),
49 self-control (Akers, 1998), anomie/strai n (Akers, 1973, 1977, 1989; Pearson & Weiner, 1985), labeling (Akers, 1973, 1977; Pearson & Weiner, 1985), normative conflict (Akers, 1973, 1977; Pearson & We iner, 1985), deterrence (Akers, 1977, 1985, 1990; Pearson & Weiner, 1985), rational choice (Akers, 1990), economic (Pearson & Weiner, 1985), routine activities (Pearson & Weiner, 1985), neutralization (Pearson & Weiner, 1985), and relative deprivation (Pearson & Weiner, 1985). Most attempts to integrate social learning theory with other theories has maintained a single-level explanation: Individuals with weak social bonds, for example, are more likely to associate with delinquent peers, from whom they learn delinquent behavior (Elliott et al ., 1979; Elliott et al., 1985). However, recalling that Sutherland ( 1939, 1947) initially intended to address both structural and processual elements of the learning of crime and criminal behavior, it seems a natural fit to attempt a cross-level in tegration of social learning theory, a processual explanation that expanded SutherlandÂ’s microsocial theory, with macro-sociological or structural theories. Social Structure-Social Learni ng (SSSL) Theoretical Statement In 1998, Akers revisited SutherlandÂ’s early line of inquiry by specifying a learning approach to deviancy and conformity that crosses levels of explanation. He offered Â“an integrated t heory of social organization and associationÂ” (Akers, 1998, p. 325) that formalized the fr agmented ideas about the relationship
50 between the epidemiology of crime and etiology of crim inal behavior that he and others had advanced over the years (e.g., Akers, 1968, 1973, 1977, 1985, 1989, 1992; Akers & La Greca, 1991; Akers et al., 1979; Burgess & Akers, 1966; Cloward, 1959; Cressey, 1960; Krohn et al., 1985; McKay, 1960). Although accepting the research approach that separat es structure from behavior in order to develop theory, Akers (1998) saw value in a cross-level integrated theory that addressed the social structural situat ions that shape individual behavior. Akers (1998) suggested that social learning theory mediates social structural influences on individual behavio r and thus by extension crime rates. The social learning variables differential association, definit ions, imitation, and differential reinforcement, with other di scriminative stimuli, mediate social structureÂ’s effect on individ ual behavior, providing the pr oximate causes of crime. Akers proposed that social structure provides the environment that shapes behavior through the learning process. Refe rring to the social learning theory elaboration as social structuresocial learning, he commented, Its basic assumption is that social learning is the primary process linking social structure to individual behavior. Its main proposition is that variations in the social st ructure, culture, and locations of individuals and groups in the social system explain variations in crime rates, principally through their influence on differences among individuals on the social learning variablesÂ—mainly, differential association, different ial reinforcement, imitation, and definitions favorable and unfav orable and other discriminative stimuli for crime. The social structur al variables are indicators of the primary distal macro-level and meso -level causes of crime, while the social learning variables reflec t the primary proximate causes of criminal behavior by individuals that mediate the relationship between social structure and crime ra tes. Some structural variables
51 are not related to crime and do not explain the crime rate because they do not have a crime-relevant effect on the social learning variables. Deviance-producing environment s have an impact on individual conduct through the operation of learning mechanisms. The general culture and structure of society and the particular communities, groups, and other cont exts of social interaction provide learning environments in wh ich the norms define what is approved and disapproved, behavioral models are present, and the reactions of other people (for exampl e, in applying social sanctions) and the existence of other stimuli attach different reinforcing or punishing consequences to individual sÂ’ behavior. Social structure can be conceptualized as an arra ngement of sets and schedules of reinforcement contingencies and other social behavioral variables. The family, peers, schools, churc hes, and other groups provide the more immediate contexts that promot e or discourage the criminal or conforming behavior of the individual. Differences in the societal or group rates of criminal behavior ar e a function of the extent to which cultural traditions, norms, social organization, and social control systems provide socializat ion, learning environments, reinforcement schedules, opportuniti es, and immediate situations conducive to conformity or devi ance. (Akers, 1998, pp. 322-323) Social structure-social learning theory specifies four structural dimensions that indirectly influence individual behavio r through social learning variables. Figure 1 depicts AkersÂ’ (1998) model.
52 Akers (1998) calls the first social structural dimension Â“ social structural correlates: differential social organization Â” (p. 332). This dimension captures aggregate-level characteristics that empi rically influence whether a community has low or high rates of crime. The c oncept includes empirical correlates that researchers have used as statistical controls in previous social structural studies, as well as correlates that represent social structural indicato rs of a theoretical construct (Lee at al., 2004). The differential social organization di mension further re fers to social structural characteristics (A kers, 1998) that contribute to what Sutherland (1947) viewed as a societal organization for or against crimeÂ—SutherlandÂ’s notion that crime has its origin in social organ ization and is an expression of that organization. The dimension refers to known and unknown social structural correlates that empirically influence cr ime rates. Societal social organization creates environments and opport unities that differentia lly influence micro-level social learning variables. Examples of such aggregate social structural characteristics that influence micr osocial learning environments include Social StructureSocial LearningIndividual Behavior Differential Social OrganizationDifferential AssociationsCriminal Behavior Differential Location in the Social StructureDefinitions Theoretically Defined Structural CausesImitationDifferential Social Location in Primary, Secondary & Reference GroupsDifferential Reinforcement Source. Derived from Akers (1998, p. 331) Social Structure-Social Learning Model Figure 1 Group Rates Crime Rates
53 community size or population density (Akers, 1998); age, sex, or racial composition of a population (Akers, 1998; Akers & Sellers, 2004; Lee at al., 2004); and other regional, geographic, or economic social systems (Akers & Sellers, 2004; Lee at al., 2004). Akers (1998) labels the second social structure social learning concept Â“ sociodemographic/socioeconomic correlates: differential location in the social structure Â” (p. 333). This dimension refers to social differentiation. Akers (1998) notes that social groupings and descriptive characteristi cs of individuals, such as sociodemographic and socioeconomic corr elates, differentially locate people within a larger social structure. Al though recognizing age, gender, race, class, religion, marital status, occupation, and ot her individual-level characteristics as important descriptive characteristics, Akers views the collectivities of these properties as important social structures. The differential location in the so cial structure dimension taps the aggregate of individual characteristics in order to capture social categories that correspond with differing crime rates (Akers, 1998; Lee et al., 2004). Akers (1998) models the aggregate groupings of indi vidual attributes such as family (Akers, 1998, Sutherland, 1947), age (Ake rs, 1998, Cressey, 1960; Sutherland, 1947), sex (Akers, 1998, Sutherland, 1947) class (Akers, 1998, Sutherland, 1947), race (Akers, 1998, Cressey, 1960; Sutherland, 1947), poverty (Akers, 1998, Cressey, 1960), educational stat us (Akers, 1998, Cressey, 1960),
54 urbanization (Akers, 1998, Cressey, 1960), and the like as direct indicators of various categories of individuals in the social structure. The third social structural dimension is Â“ theoretically defined structural causes: social disorganization and conflict Â” (Akers, 1998, p. 333). This concept refers to structural causes of crime that researchers have theoretically advanced in the literature. Unlike t he structural correlate dim ension, which oftentimes utilizes the same variables, this dimension refers specifically to conceptually defined conditions that explain t he correlation between crime rates and sociodemographic or socioeconomic conditions (Akers, 1998). The theoretically defined structural causes dimension lumps together explanations that link observed, elev ated crime rates to observed, elevated abstract social conditions (Akers, 1998). T he dimension taps theor etically distinct social explanations for the correlation between crime rates and social conditions such as race, class, gender, region, city, neighborhood, and population size, density, and composition. This theoretical dimension generally views social order as implying agreement with societal no rms and values, and it suggests that low levels of disruptive conflict produce conf ormity, or rather non-conformity comes from high levels of disruptive conflict i nherent in social disorder (Akers, 1998). Although Akers (1998) views anomie, social disorganization, and conflict theories as well known examples of theories belongin g in this dimension, other theoretical examples include class oppression and pat riarchy (Akers, 1998; Akers & Sellers,
55 2004). The fourth social structural dimension, Â“ differential social location in primary, secondary, and reference groups Â” (Akers, 1998, p. 334), refers to small groups with whom individuals associate. Examples of this dimension include family, peers, school, work, and church. Such personal networks provide the immediate environment that shapes behavior through the informal control of social environments, situations, and opportuni ties for criminal behavior (Akers, 1998). The four structural dimensions co mbine to affect individual behavior through social learning variables. Social structure acts as the distal cause of crime, affecting an individualÂ’s exposure to norm and norm-violating contingencies, and ultimately crime rates. Theoretical critiques. Akers (1998) argues that structural vari ables affect variation in crime only in that they provide contingencies of re inforcement and punishment for individual behavior. Structure serves as a distal c ause of crime, providing the individual learning environment that affects an individualÂ’s exposure to norm and normviolating contingencies (Akers, 1968, 1998) Microsocial theories offer proximate causes of crime (Akers, 1998), aggregat es of which provide group rates. An at first, seemingly condemning t heoretical criticism of the social structure-social learning m odel is that it treats all structural variables without
56 distinction. Sampson (1999), for example, characterizes social structure-social learning theory as an explanation for how social structural patterns influence individual variations in the exposure to social learning variables, notably delinquent definitions. He correctly summari zes the link from social structure to social learning as involving differing expos ure levels that affect the initiation, continuance, or desistance, along with t he frequency and versatility, of criminal behavior. Sampson (1999) characterizes the so cial structure-social learning statement, however, as a quest to list ma crosocial variables that influence exposure to learning patte rns conducive to crime. Sampson contends that such treatment puts social structure outside the scope of the theo ryÂ—all structural variables are exogenous to the m odel. Sampson questions this approach, suggesting that in doing so, social structur e-social learning theory inappropriately separates social mechanisms from theor izing, as the model includes any macrosocial variable that has an effect on the social learning process regardless of its origin. Sampson (1999) objects to the Â“every thing mattersÂ” approach, suggesting that a useful theory needs to make pres umptive falsifiable statements about the social structure, as do conflict, social di sorganization, and anomie/strain theories. He maintains that social structure-soci al learning theory is uninterested in the sources of social structur al arrangements, or their theoretical ordering. He
57 suggests that the social structure-soci al learning theory incorrectly divorces microsocial mechanisms from the rationale of structural or cultural sources. Sampson rejects the social structure-so cial learning model as unsatisfying and not useful. Krohn (1999) also suggests that the social structur e-social learning model does not adequately specify the links betwe en the macrosocial and social learning variables. He suggests that the model does not fully integrate levels of explanation because there ar e no propositions linking t he exogenous structural variables to the social learning process. Krohn sees potential in the model, but he believes the theory falls short. For Krohn (1999), an acceptable social structure-social learning statement, a useful cross-leve l integration of macrosocia l theoretical explanations for crime with social learning theory, must contain hypotheses explaining why certain social structural variables result in different levels of associations, definitions, imitation, and reinforcement. Krohn views social structure-social learning theory as currently unacceptabl e because it is not a propositional integration. Akers (1999) addressed SampsonÂ’s ( 1999) and KrohnÂ’s (1999) criticisms by noting that the theory does distinguish structural variables: The theory predicts that structural variables associated with crime rates will also relate to social learning variables. The model excludes structural variables that do not
58 empirically influence crime rates. Moreov er, Akers points out that the theory specifically presumes that variables fr om social disorganization, conflict, and anomie theories will have an effect in t he model. Akers (1998) admits the lack of linking propositions; however, he suggests that the theory instead conceptually attempts to Â“integrate across levels by linking the variables, causes, and explanations at the structur al/macro level (that account for different absolute and relative levels of crime) to probable effects on individual behavior through social learning variablesÂ” (p. 329). Although AkersÂ’ (1999) response is vague, perhaps unsatisfying to some, social structure-social lear ning is an elaboration of social learning theory and it is intentionally abstract. The theory is a cross-le vel end-to-end conceptual integration, not a propositi onal integration. The social structure-social learning model is concerned with how social lear ning theory mediates the influence of structural variables on crime rates, and therefore, individual behavior. Moreover, despite AkersÂ’ agreement that linking prop ositions are absent from the theory, and inviting others to help specify Â“the mo st underdeveloped part of the theoryÂ” (Akers, 1999, p. 491), social structur e-social learning does indeed make interrelated statements among its propositions. Sampson (1999) and Krohn (1999) may confuse AkersÂ’ (1999) vagueness in describing the theoretical linkages between social structural variables and social learning variables for inadequacy in doing so, perhaps overlooking
59 CresseyÂ’s (1960) warning t hat criticism not based on research is not a valid critique of a theory, rat her it is a proposal for new research. Akers (1998) specifies that variations in social stru cture explain variations in crime rates because of their influence on social learni ng variables. He expl ains further that this occurs because of the different ial learning environments produced by societal structure and culture. That is, structure provides individual learning environments that affect an individualÂ’s exposure to norm and norm violating contingencies. The issue may not be the absence of linking propositions; rather critics may disagree with the linking propositions as presented, or as Sampson (1999) notes, Â“I have a different theoretical inte rpretation of ultimate ly ambiguous dataÂ” (p. 448). Sampson (1999) and Krohn (1999) do not provide evid ence that the structural variables do not operate on the so cial learning variables as posited by Akers (1998, 1999), rather they suggest more preferable social structural explanations for crime (see Sampson, 1999), or better uses for t he theory if more fully specified (see Krohn, 1999). Samp son and Krohn do not refute social structure-social learning theor y; rather they present re search ideas that differ from AkersÂ’ interpretation of, perhaps even his interest in, ambiguous data and views on the role of theory. Sampson (1999) points out that the social structure-social learning structural variables are not importanceprioritized such as in Blau and BlauÂ’s
60 (1982) test of strain theory, nor are the propositions as a priori falsifiable as those offered by social disorganization theory. Sampson (1999) would like to see the theory better address the macr o-level concern with why society has the social systems (e.g., culture, age structure, cl ass and race systems) that it does. Krohn (1999) would like to see social struct ure-social learning theory better address macrosocial structure and dev elopmental processes. However, operationalizing the st ated propositions and explicating functional relationships is the role of research (Short, 1960). Disliking the social structure-social learning t heory as stated does not refute the theory; rather a compilation of studies finding no support for its propositions may do so (see Popper, 2002; Lakatos, 1978). Moreover, Krohn (1999), and to some extent Sampson (1999), use questionable exam ples to support their points. Krohn (1999) uses the aging out e ffect (see Akers & Lee, 1999; M. Gottfredson & Hirschi, 1990; Hirschi & Gottfredson, 1983; Sampson & Laub, 1993; Steffensmeier, Allan, Ha rer & Streifel, 1989; Warr, 1993) as an example of why social structure-social learning t heory falls short as an adequate explanation of crime and criminal behavior thr ough its lack of macrosocial linking propositions. In doing so, though, he incorre ctly asserts that social learning theory must incorporate development al perspectives (e.g., Moffitt, 1993; Sampson & Laub, 1993; Thornberry, 1987) to structurally explain the decreasing prevalence in crime as age increases.
61 Researchers have not fully explored t he social learning process as it relates to the aging out effect, but the micr o-level social learni ng theory implicitly explains the aging out effect as it is, and the social structural elaboration may address the issue even more so. Althoug h not expressly not ed by Akers and Lee (1999) in their longitudinal study of adolescent substance use and their subsequent discussion of the age and crime effect as a function of age-related changes in differential reinforcement, reinforcement schedules may contribute to the aging out explanation thr ough changing associations and the extinction of no longer reinforced behavior. For example, reinforcement occurs w hen there is a balance of anticipated or actual rewards over punishments. Reinforcement has three modalities: amount, frequency, and probability (Akers, 1998). Various reinforcement schedules control the emi tting of behavior (Akers, 1998). Generally, behavioral frequency corresponds with social rein forcement frequency (Hamblin, 1979; Herrnstein, 1974). Some social behavioral reinforcement occurs infrequently, however, so individuals seek behavioral choices that optimi ze reinforcement (Herrnstein & Loveland, 1975). Akers ( 1998), notes, Â“therefore, a given behavior must be seen in the context of all other concurrently available schedules and sources of reinforcementÂ” (p. 70). Much of what researchers know about reinforcement schedules comes from laboratory studies with animals such as pigeons and rats (Herrnstein &
62 Loveland, 1975; B.F. Skinner 1953); however, there are clear implications for social behavior (see Bandura, 1977). Behavior that is reinforced each time it is emitted is on a continuous schedule of reinforcement. Behavior that is not reinforced on each occurrence is on one of four intermittent schedules of reinforcement (B.F. Skinner, 1953). A fixed ratio schedule refers to reinforcement that occurs after a certain number of responses (e.g., every tenth response), whereas a variable ratio schedule characterizes reinforcement that occurs after a variable number of responses (e.g., af ter the fifth response on one occasion, after the second response on another occasion, etcÂ…). A fixed interval schedule depicts reinforcement that occurs after a certain amount of el apsed time (e.g., every ten minutes), and a variable interval schedule refers to reinforcement that occurs after a varying amount of elapsed time (e.g., after five minutes on one occasion, after two minutes on another occa sion, et cetera; B.F. Skinner, 1953). Reinforced behavior is more probable to occur again in the future (see Akers, 1998; B.F. Skinner, 1953), and behav ior that is not reinforced is extinguished (see B.F. Skinner, 1953) Behaviors that are on continuous schedules of reinforcement extinguish easily when not reinforced. Ratio schedules of reinforcement tend to produc e higher response rates than interval schedules. Variable schedules tend to be mo re difficult to extinguish than fixed schedules (B.F. Skinner, 1953). Social behav ior is generally on a variable interval schedule of reinforcement (Hamblin, 1979; Herrnstein, 1974; see Akers, 1998).
63 Following this line of thought, devi ant behavior that was previously reinforced but is no longer reinforced due to differential associations, or other changes in the social learning variables, would be expected to extinguish at a slow rate. Extinction would occur in the absence of rein forcement, but its effect would not be immediate due to the in termittent schedule of reinforcement inherent in soci al phenomenon. For example, an adolescent that previously received reinforcement for theft may, in the presence of changing a ssociations such as peer (Thornberry, 1987) or friendship (Haynie, 2002) net works, intermittently continue the response, fail to receive reinforcement, and discontinue the response over time. The amount of time to ex tinction would depend upon previous rates and intervals of reinforcement, producing a va riable rate of extinction. Although providing a more detaile d explanation of the underlying mechanism than previous researchers co mmenting on the observation, the aging out example is consistent with the fi ndings of Lanza-Kaduce, Akers, Krohn, and Radosevich (1984), who investigated social learning theoryÂ’s ability to account for the cessation of alcohol and marij uana use by adolescents. They found that differential associations played a role in substance desistance. Such rationale is further consistent with Winfree, Selle rs, and ClasonÂ’s (1993) conclusion that changing reference groups or associati ons with significant others may alter previous behavior, in their investi gation adolescent drug use, through new
64 definitions, reinforcements, and punishments. The described process of variable-interval microsocial reinforcement schedules extends to macrosocial structur e through the notion of sets and schedules of reinforcement contingencies (see Akers, 1998; Lee et al., 2004). Although the changing associations describ ed in the adolescent theft example result in variable individual reinforcem ent schedules, the associations provide schedules of reinforcement contingencies. No to low incidence of criminal behavior before age 6 for example, with a gradual increase during childhood until adolescence around age 12, turning into a sharp increase that peaks at age 17 or so, and continues its decline through y oung adulthood until finally tapering off in mid-adulthood around age 35-36, is not beyond the explanation of social learning theory, or social structuresocial learning theory by extension. The extension of microsocial reinfo rcement as an explanation for the aging out effect to the ma crosocial level through sc hedules of reinforcement contingencies may be better described by drawing on SampsonÂ’s (1999) discussion of differential associations, and his reference to Glueck and GlueckÂ’s (1950) birds of a feather char acterization. In that exam ple, Sampson attempts to reconcile the effect of delinquent peers on delinquency with WarrÂ’s (1998) account that marriage correlates with desis tance in crime. Sampson concludes, based in part on a summary of WarrÂ’s pos ition as conceding that the mechanism of transmitting behavior among delinquent s remains unknown, that social
65 learning theory cannot explain why ma rriage results in less time spent with delinquent peers, and thus, le ss individual delinquency. When the analysis remains at the indi vidual level, as in the earlier adolescent theft example, and Samps onÂ’s (1999) approach to the marriage example, various individual reinforcem ent schedules affect the emitting of individual behavior. However, peer asso ciations, friendship groups, and marriage are meso-level groups in which individuals are differentially located. Akers (1998) incorporates this depiction in his soci al structure-social learning model as differential social location in primary, secondary, and reference groups, as well indirectly, through the notion of congr egating with like other s, part of the differential location in the social structure dimension. Sampson (1999) asks why marriage affe cts individual association with delinquent peers and individual delinquency. As meso-level groups, delinquent peers and marriage may present conflicting contingencies of reinforcement. The social structure of friends hip groups and family groups provides the opportunities for an individual to receive reinforcemen t, or punishment, for social behavior. The emitting of individual delinquent b ehavior depends on the amount and frequency of reinforcement contingencies supportive of delinquency, versus non-supportive contingencies. In the marriage example, more frequent associations with a spouse who does not reward delinquency than delinq uent peers who do reward delinquency,
66 will lead to reductions in delinquency and ul timately, extinction of the delinquent behavior. Delinquency extinguishes when it is not reinforced. Upon extinction, as well as during the process, through t he notion of maximizing opportunities for reinforcement, association with the rewa rding spouse will replace associations with delinquent peers who reward behavior that is no longer emitted. As the delinquent behavior no longer occurs, t here is no longer an opportunity for reinforcement in such an environment, and indeed conformity may result in punishment, so the behavior of associati ng with deviant peers may extinguish as well. Although social learning theory is near silent on the importance and measurement of reinforcement schedules, and the social structural elaboration only briefly mentions social structural contingencies of reinforcement (see Akers, 1998, pp. 322-323), the concepts are undeniably present in the theory. Moreover, in contrast to KrohnÂ’s (1999) assertion that social struct ure-social learning theory does not offer suitable linking propositi ons to explain why the macrosocial variables might be expected to affect levels of social learning, such statements may be derived from the theory, at least as it relates to the example he used. At the individual level, social learning accounts for the aging out effect through reinforcement schedules. At the ma crosocial level, social structure accounts for differential reinforcement schedules through contingencies of reinforcement. Both refutable statements co me directly from the social structure-
67 social learning explication. Finding the question important, and developing the hypotheses, is the role of research. Likewise, SampsonÂ’s (1999) discussion of the role of theory and his desire to explain macrosocial stru cture, both advances a res earch question rather than offering a valid theoretical critique, and additionally misidentifies an implication present in AkersÂ’ (1998) expl ication of social structur e-social lear ning theory. First, contrary to SampsonÂ’s (1999) assertion, social structure-social learning theory does make presumptiv e falsifiable statements about social structure. Akers (1998) notes, The macroand meso-level variables determine the probabilities that an individual has been, is, or will be exposed to different levels of the social learning variables. The different leve ls of these variables determine the probability that the individua l will begin, persist, or desist from behavior, and at what frequency and degree of spec ialization or versatility. This behavior is translated into crime rates. (p. 335) The statements may not be to SampsonÂ’s satisfaction, but they nonetheless exist in the theory. Second, again contrary to SampsonÂ’s (1999) assertion, social structuresocial learning theory does not treat all macrosocial variables as equal, and although not emphasized, the theory does im ply, if not explicit theoretical ordering, importance-prioritiz ed structure. In his descrip tion of differential social location in primary, secondary, and refe rence groups, along with a reference to sex, race, and age, Akers (1998) implies t hat the meso-level social structural dimensions are the mechani sms through which the other two social structural
68 dimensions, more distal causes, direct ly affect individual behavior. Akers prioritizes differential social location in primary, secondary, and reference groups, along with differential location in the soci al structure, as more important than differential social organization and theor etically defined structural causes because of their role in providing contex t to the social learning process. In sum, Akers (1998) offered a theor y that organized pr opositions between macro-level and meso-level social arr angements and microsocial behavior. Akers viewed the social structure-social lear ning theory as a logical extension of previous research, and he offered a pos t hoc analysis of how previous macrolevel research findings, macrosocial facts, are consistent with the theory. Akers did not explicitly test the t heory at the time of its expl ication; however, neither did his critics. Moreover, the research avenues suggested by Sampson (1999) and Krohn (1999) do not go against the rati onale both expressed and implied by AkersÂ’ (1998) social structure-social learning theory; rat her, the research suggestions may merely fall outside of AkersÂ’ interests. Akers (1998) intentionally offered an abstract theoretical elaboration of social learning theory. He is more inte rested in explaining criminal behavior (Akers, 1998, 1999) than he is in explaining societal stru ctures. AkersÂ’ cross-level integration tries to explain how existing so cial structure explai ns crime through its effect on individual levels of social learning. There are obstacles to testing Aker sÂ’ (1998) social structure-social
69 learning model, however. Most notably, data allowing simult aneous examination of macrosocial and microsocial vari ables are uncommon (Lanza-Kaduce & Capece, 2003). Empirical validity Although testing the social structure-so cial learning model is difficult, there has been promising research in this area. In one study with lim ited structural measures, researchers conc luded that family well-be ing and social learning partially mediated the impact of occupat ional structure on adolescent violence (Bellair et al., 2003). Bellair and co lleagues modeled differential social organization through the variables l abor market opportunity, concentrated disadvantage, and urbanicity. They defi ned their structural boundaries by U.S. zip code. They assessed their model with hierarchical regression and once they added the mediating variables to the m odel, the effects on adolescent violence reduced, and concentrated disadvantage no longer directly affected violent attitudes. In another study, researchers conclude d that social learning partially mediated the relationship between st ructural variables and binge drinking (Lanza-Kaduce & Capece, 2003). The mo deled social structure variables included differential social organization (urban, suburb an, or rural university), differential location in social structure ( gender, race), differential social location in meso-level groups (Fraternity/Sorority in volvement, extracurricular involvement),
70 and two single-index theoretical variables: integration into academics (B or better grade point average) and conflic ting culture (opinion of whether alcohol is central to the groups male students, female students, faculty and staff, alumni, and athletes). Lastly, researchers concluded that so cial learning partially mediated the relationship between structural variables and adolescent substance use (Lee et al., 2004). Social structural variables included differential social organization (community size), differential location in social structure (gender, social class, age), and differential location in primar y groups (family structure). Lee and colleagues assessed direct and indirect e ffects in their models with structural equation modeling. The three social structure-social l earning studies show promise for the model, but each has limitations. Aside from their varying statistical sophistication and microsocial measures, none of t he tests extensiv ely measured the differential social organization and theor etically defined structural causes dimensions posited by Akers (1998). Lee and colleagues (2004) tested a model with community size (rural, urban, or suburban) as the sole indicator of differential social organization, and they excluded theoretically defined struct ural causes entirely. The Lanza-Kaduce and Capece (2003) model likewise measured differential social organization with one indicator (a dummy-coded university variable), and their two theoretically
71 defined structural causes measures (i ntegration into academ ics and cultural climate) did not tap strong theoretically defined macrolevel predictors (see Pratt & Cullen, 2005). Further, although Lanz a-Kaduce and Capece concluded that there was support for the partial mediati on hypothesis, they assessed their model with standardized coefficients (ordinary least squares [OLS] regression) to assess the change between full and partial models, a technique Baron & Kenny (1986) and James and Brett (1984) sugges t cannot be used to differentiate mediation because OLS does not allow for causal ordering. Although Bellair and colle agues (2003) modeled disadvantage, urbanicity, and family disruption measures that are popular in the literature (e.g., Bergesen & Herman, 1998; Curry & Spergel, 1988; Krivo & Peterson, 1996; Morenoff & Sampson, 1997; Sampson, 1986, 1987; Sampson & Groves, 1989; Sampson & Raudenbush, 1999; Sampson et al., 1997; D. A. Smith & Jarjour a, 1988; Warner & Pierce, 1993), they indexed AkersÂ’ (1998) differential social organization and theoretically defined structural causes dimensions with only four measures. Moreover, they added an additional interveni ng process between social structure and social learning, family well-being, and perhaps their most interesting finding, the mediation of concentrated disadvant age, involved mediat ion of attitudes (definitions), not their outcome meas ure. Although Bellai r and colleagues gave attention to the linking mechanisms between social structure and social learning, they mainly did so through the altered m odel that included the family well being
72 concept. Further distorting interpretation of t heir results as to the adequacy of the social structure-social learning model Bellair and colleag ues (2003) aggregated social structure at the zip code level. This is, somewhat removed from the notion of community advanced by social diso rganization theory and adopted by Akers as likely to influence individual learning environments. Census zip code tabulation is a statis tical entity created by the Census Bureau to represent an aggregation of t he predominant zip code in a census block (U.S. Census Bureau, 2000). Wher eas census blocks nest within block groups, and block groups nest within census tracts, the Census Bureau reports zip code tabulation areas as a subset of the nation. The Census Bureau does not specify its hierarchy, and they do not report its average size. Another study relevant to AkersÂ’ (1998) social structure-social learning model is that reported by Hoffmann (2002), who tested a contextual model that assessed the effects of community diso rganization and raci al segregation on a logged delinquency scale. Starting from the social structural tradition, Hoffmann measured social structur e at the zip code level, and he indexed community disorganization through the percent of female-headed hous eholds, the percent of unemployed or out of work, and the per cent below the poverty threshold. Hoffmann created a dissimilarity i ndex to measure segregation. Hoffmann (2002) did not explicitly test t he social structure-social learning
73 model, though he did draw on it in his res earch. Hoffmann was most interested in testing community structure as the cont ext for nested individual behavior through measures of social control, strain and differential association. He assessed his model with HLM, using conventional defin itions and peer expectations to index differential association and social lear ning, as well as interaction terms. Hoffmann (2002) reported that indicators of the percent of female-headed households, the percent of unemployed or out of work males, and the percent below the poverty threshold signific antly affected his logged delinquency measure, and that the relationship was not mediated or moderated by his social learning measures. In combination with hi s reported results of testing the social control and strain measures, Hoffm ann concluded that attempts to link macrosocial and microsocial theoretical explanations for cr ime and criminal behavior Â“may be slightly misdirectedÂ” (p. 779). Like the three specific tests of the so cial structure-social learning model, HoffmannÂ’s (2002) study has strengths and weak nesses in its inference to AkersÂ’ (1998) hypothesized relationships between so cial structure, social learning, and individual criminal behavior. Hoffmann corrected for the per ceived inadequacy of OLS regression to assess cross-level effect s by using HLM, a technique suited to individuals nested within a social struct ure. However, like Bellair and colleagues (2003), he aggregated social struct ure at the zip code level. Moreover, Hoffmann (2002) only used four measures of social structure,
74 whereas social structure-social learning theory identifies four social structural dimensions, two dedicated solely to ma crosocial correlates. Further, Hoffmann was only able to index one social lear ning concept directly: definitions. Hoffmann (2002) acknowledged that he had no measure of peer associations, and he did not address the conc ept of imitation. As to differential reinforcement, Hoffmann questionably co ncluded that peer expectations sufficiently indexed differential reinforc ement, as the survey instrument asked questions about friendsÂ’ expectations about life goals. However, the measure asked no direct questions regarding del inquency, the behavior under study, instead asking the respondent to report their friendsÂ’ attitudes toward conventional goals; specifically, w hether they view getting good grades, graduating from high school, education beyond high school, and studying as important. Hoffmann (2002) did not specifically set out to test social structure-social learning theory; rather he viewed social structure thr ough a contextual lens. In sum, it is questionable that his measures of both social structure and social learning adequately tested AkersÂ’ (1998) theory. However, HoffmannÂ’s research does question the social structure-soci al learning model specification with research, rather than pure reasoning such as employed by Sampson (1999) and Krohn (1999). Moreover, HoffmannÂ’s (2002) research s uggests that the social structure-
75 social learning model may indeed be in complete until it can more adequately explain how the social structural va riables impinge on the social learning process. Hoffmann may have taken social st ructure-social lear ning theory in a direction removed from its implied tenets, as perhaps did the Bellair and colleaguesÂ’ (2003) test; however, the theor y does not expressly speak to, let alone admonish, those research directi ons. It seems apparent that social structure-social learning t heory must address the macros ocial literature, despite AkersÂ’ (1998, 1999) implied lack of interest in the topic.
76 Chapter Three Crime Rate Determinants Criminal Behavior and Environment Akers (1998) suggests that social lear ning theory mediates the effects of social structure on crime and criminal behavior. The social structure-social learning model proposes that four social structural dim ensions affect crime rates, only in as much as they affect th e intervening social learning process and individual criminal and deviant behavio r. Social structure provides the environment by which social lear ning produces individual behavior. Two of the dimensions, di fferential social organization and theoretically defined structural causes, draw from t he domain of macrosocial theorists as Akers (1998) specifically incorporates known and unknown crime rate correlates and theoretically derived group crime rate explanations. Akers does not, however, fully explain how the two dim ensions impinge on t he social learning process. Akers is instead content on noting their importance and generally describing some of the indicators current ly known to correlate with crime (see Akers, 1998, 1999). In discussing differential social organi zation, for example, Akers (1998) notes that this social structural di mension aims to incorporate known and
77 unknown social structural correlates of crime, be they derived theoretically or merely identified through previous studies as having a relationship with crime, deviance, and criminal behavior. He de scribes the dimension in terms of Â“ecological, community, or geographica l differences across systemsÂ” (Akers, 1998, p. 332). Akers uses urbanicity and population size as two main examples. Akers appears, in this di mension, concerned only wit h whether the identified social structure associates with crim e, not the correlateÂ’s theoretical conceptualization. In relating the theoretically defined structural causes dimension, Akers (1998) attends to the notion that macr osocial researchers conceptually define social structural correlates in a certai n way, but he again leaves determination of the precise relevance to others (see Aker s, 1998, 1999). Akers groups theoretical social structural explanations into a category of social disorganization and conflict, remarking, Â“both vi ew social order, stability, and integration as conducive to conformity, and disorder and malin tegration as conducive to crime and devianceÂ” (p. 334). As with t he differential social organ ization dimension, Akers only vaguely identifies indicators of this dimension. Evidenced by the three reported tests of social structure-social learning theory (Bellair et al., 2003; Lanza-K aduce & Capece, 2003; Lee et al, 2004), researchers viewed the social structural dimensions differently, incorporating a wide range of indicators and explanations as to their relevance. More notably,
78 none of the researchers were able to us e AkersÂ’ (1998) explication of the theoretical dimensions to expressly rela te how their measures influence social learning and individual behavior. After three tests of AkersÂ’ (1998) soci al structural elaboration, theoretical questions remain. What indicators measur e differential social organization and theoretically defined structural causes ? How do these dimensions directly influence the social learning process? Social Structural Crime Correlates and Explanations. Background. There is much macrosocial literature re lating societal org anization to rates of crime. Research dates at least s poradically to Quetel et (1831/1984) who statistically examined official crime rate data in France. He advocated the examination of crim e through the calculation of av erages, rather than through examining individual characteristics. He was interested in constant causes of crime, determined through probabilities, as opposed to accidental causes, which he characterized as stemming from means and opportunities, if not free will. Quetelet (1831/1984) report ed that age was the most important cause of crime, with an aging out effect around age 25 years (peaking between 21 and 25). He further noted that sex (maleness) was a great influencer of crime (nearly threefold for males to females for all crim es in his sample), and that social class and poverty were additional leading correlates.
79 Quetelet (1831/1984) conc luded that natural forces beyond free will contributed to crime, and that age, sex, poverty, and education, for example, were crime propensities As Quetelet observed that the same crimes were Â“reproducedÂ” year after year in the sa me proportions (1826-1829), he viewed crime as a Â“sad condition of the hum an speciesÂ” (Quetelet, 1831/1984, p. 69). Quetelet viewed crime as a scientific law, terming his observation Â“physical factsÂ” or Â“general facts, Â” and he noted that one could not under stand crime until one understood the general facts upon which soci ety existed. As such, Quetelet believed that society caused crime by a ffecting the social masses through its social system. Empirical research. Three prominent studies have tried to make sense of modern macrosocial literature, varying in their degrees of broadness. Chiricos (1987) reviewed the findings from 63 studies regarding unem ployment and crime rates. Although comprehensive, the topic was narrow and the methodology was descriptive. He categorized the studies by type, cross-se ctional or longitudinal, and concluded that the unemployment-crime relationshi p was more consistent and stronger in the cross-sectional studies. Although ma king few firm conclusions, Chiricos noted that unemployment a ffected crime differently based on the level of aggregation: unemployment had stronger effects on the cr ime rate at smaller units of aggregation (e.g., SMSA versus State).
80 Land, McCall, and Cohen (1990) summari zed the results of 21 studies regarding the structural covariates of homic ide. Although restrict ed substantively, Land and colleagues, in contrast to Chiric os (1987), examined a broad range of presumed social structural correlates. Revi ewing the literature, they started with the notion that such measures as popul ation size, population density, racial heterogeneity, and age structur e were not stable predictors of homicide. In regard to all of the variables under anal ysis, which included the other measures percentage divorced, percent age of children under aged 18 years or younger not living with both parents, percentage of families in pov erty, median family income, percent unemployed, the Gini index of i nequality, and living in the South, they concluded that only one meas ure was statistically significant, and moving in the same direction, across all studies: t he percentage of children under aged 18 not living with both parents. Having analyzed the literature, La nd and colleagues (1990) estimated a baseline model of the 11 predictors using OLS regression at the SMSA, city, and state level. Their years under analysi s were 1960, 1970, and 1980, and they replicated their model on 1950 data. Land and colleagues (1990) concluded first t hat the problem of invariance across time and homicide studies was due to structural covariate multicollinearity. They c autioned that future studies should attend that issue. Secondly, they concluded that the most stable predictor of homicide was a
81 resource-deprivation/affluence index. T hat measure derived from principalcomponents analysis and it expanded Loftin and HillÂ’s (1974) structural poverty index, as it comprised median family income, the percentage of families below the poverty line, the Gini index of inequality, percent Black, and the percentage of children aged 18 years or younger not liv ing with both parents. Finally, they concluded that the population and perc entage divorced measures were strong covariates of homicide, and that the unem ployment rate and age structure were less consistent predi ctors. The third prominent study that has organized the macrosocial crime rate literature is the most comprehensive review to date, as well as the most recent. Pratt and Cullen (2005) examined social st ructural predictors far more generally than previous efforts, and their study is the most st atistically rigorous review as they utilized a meta-analytic procedur e that controlled for measurement technique conditioning effects. Pratt and Cullen (2005) examined 31 soci al structural crime predictors across 214 empirical studies (509 stat istical models) published between 1960 and 1999. They looked both at studies t hat used aggregate measures to predict crime rates without specifying a theoretical rationale, as well as those utilizing a theoretical framework. The seven spec ified theories included in the study are social disorganization, anomie/strain, re source/economic deprivation, routine activity, deterrence/rational choice, soci al altruism, and subcultural. Pratt and
82 CullenÂ’s main findings both rank-order the efficacy of specific macrosocial predictors and identify the macrosoc ial theories that have been adequately tested, along with a conclusion of the t heoryÂ’s overall empirical support (weak, moderate, high). Pratt and Cullen (2005) estimated an independence-adjusted mean effect size in order to control for the type of measurement used by a particular study. Rank-ordered by the adjusted effect size the 31 crime predict ors they examined (p. 399) are (1) strength of economic instit utions, (2) length of unemployment, (3) firearms ownership, (4) percent nonWhite, (5 ) incarceration effects, (6) collective efficacy, (7) percent Black, (8) religion e ffect, (9) family disr uption, (10) poverty, (11) unsupervised local peer groups, (12) h ousehold activity ratio, (13) social support/altruism, (14) inequality, (15) ra cial heterogeneity index, (16) urbanism, (17) residential mobility, (18) unemployment with age restriction, (19) age effects, (20) southern effect, (21) unemployment with no length cons ideration, (22) socioeconomic status, (23) arrest ra tio, (24) unemployment with no age restriction, (25) sex ratio, (26) structural density, ( 27) police expenditures, (28) get-tough policy, (29) educati on effects, (30) police per capita, and (31) police size. Pratt and Cullen (2005) found four consis tently robust social structural factors: racial composition (both perc ent nonWhite and percent Black), economic deprivation, and family disruption. These factors were strong and stable
83 predictors across studies that used them to index theoretical concepts such as the racial heterogeneity, poverty, and family disruption measures used to test social disorganization theory, as well as when they were viewed as a composite concentrated disadvantage (e.g., Sampson et al., 1997) measure. Pratt and Cullen (2005) concluded t hat social disorganization and resource/economic deprivation theor ies received high empirical support, anomie/strain, social support/altruism, and routine activity theories received moderate support, and rational choice/det errence, and subcultural theories received only modest support. They further concluded that each of the theories except anomie/strain and social support/a ltruism have been adequately tested, and that routine activity, rational choice/ deterrence, and subcul tural theory results are conditioned by their methodologies. Pratt and CullenÂ’s (2005) use of the term resource/economic deprivation theory refers mainly to conflict perspecti ves that emphasize poverty either from absolute or relative positions. Such char acterization does not distinguish whether poverty and economic deprivation were pi tted against one another or viewed as a construct. Pratt and Cullen do not seem to intend this theoretical grouping as a clean theoretical distinction, as they a ssessed poverty and inequality separately, grouped them together for the purposes of description, and warned that their study cannot distinguish the absol ute and deprivati on paradigms. The substantive conclusion to be drawn from this grouping is that both poverty and
84 relative deprivation were two of the stronger macrosocial predictors of crime rates. Pratt and Cullen (2005) use the term so cial disorganization theory to represent the tradition of Shaw and Mc Kay (1942), who, drawing on DurkheimÂ’s (1897/2002) notion of rapid societal change, sought an explanation for the spatial distribution of Chicago delinquency rates in neighborhood communities. Shaw and McKay (1942; Shaw et al., 1929) at first examined Chicago juvenile delinquency rates that spanned several decades in the early 1900s. They later added more decades, accumulating Chicago delinquency data for a period of 65 years, and more cities to include Philade lphia, Boston, Cinci nnati, Cleveland, and Richmond, Virginia (S haw & McKay, 1969). Before sharing their conclusions, Shaw and McKay (1969) stated their questions. They asked, 1. To what extent do the rates of delinquents and criminals show similar variations among the local communities in different types of American cities? 2. Does recidivism among delinque nts vary from community to community in accordance with rates of delinquency? 3. To what extent do variations in rates of delinquents correspond to demonstrate differences in the economic, social, and cultural characteristics of local communities in different types of cities? 4. How are the rates of delinquents in particular areas affected over a period of time by successive cha nges in the nativity and nationality composition of the population? 5. To what extent are t he observed differences in the rates of delinquents between children of foreign and native parentage due to a differential geographic distribution of thes e two groups in the city? 6. Under what economic and social conditions does crime develop as a social tradition and become embodied in a system of criminal values. 7. What do the rates of delinquents, when computed by local areas for
85 successive periods of time, reveal wit h respect to the effectiveness of traditional methods of treatment and pr evention, of wide variations in rates of delinquents in different types of communities? (Shaw & McKay, 1969) Shaw and McKay (1969) qualified their conclusions by acknowledging that others may interpret their results different ly. Shaw and McKay first concluded that there is a relationship between local comm unity conditions and rates of juvenile delinquency. They noted that communi ties with high rates of delinquency exhibited different social and economic conditions than communities with low delinquency rates. They remarked, [The] high degree of consistency in the association between delinquency and other characteristics of the community not only sustains the conclusion that delinquent behavior is related dynamically to the community but also appears to es tablish that all community characteristics, including delinquency, are products of the operation of general processes more or less common to American cities. Shaw & McKay, 1969) Referring to the Chicago data, S haw and McKay (1942, 1969) further noted that delinquency rates remai ned stable during the years under examination, regardless of the neighborhoodsÂ’ racial or ethnic composition. The populations of neighborhoods with high delinquency rates were mainly comprised of immigrants. Further, they found that delinquency rates increased the further away from the central co re of the city. They reasoned that delinquency must be related to inherent communi ty characteristics. Taking a different approach to rapid growth than Shaw and McKay (1942, 1969), Wirth (1938) observed that a large city represents many people that have
86 little in common. He concluded that ur banism, the rapid growth associated with the development of cities, resulted in s uperficial social relations. According to Wirth, such heterogeneity may result in Â“personal disorganization, mental breakdown, suicide, delinquency, crime, co rruption, and disorder. . (p. 230).Â” Early research derived from Wirth (1938) tended to look at a cityÂ’s population density, the number of peopl e packed into a geographical area, and the various stratifications that resulted from masses of people that knew larger groups only superficially, such as race composition, sex composition, age composition, and poverty. As gleaned from Pratt And Cullen ( 2005), researchers often use urbanicity or population density variables either as it ems of interest or as a statistical controls (Allison, 1972; Archer, Gar dner, Akert & Lockwood, 1978; Bursik & Webb, 1982; Byrne, 1986; Copes, 1999; Gi bbs & Erickson, 1976; Jackson, 1984; Krohn et al., 1984; Mencken & Barnett, 1999; Mladenka & Hill, 1976; Morenoff & Sampson, 1997; Osborn, Trickett & Elder, 1992; Pressman & Carol, 1971; Sampson, 1985; Sampson & Groves, 1989; M.D. Smith & Brew er, 1992; Stafford & Gibbs, 1980; Warner & Pierce, 1993; We bb, 1972). As to e fficacy, Pratt and Cullen (2005) concluded that urbanicity has high strength (an effect size estimate two standard errors above the pooled me an across studies with various methodological specifications) and high st ability (degree in change of effect size when accounting for model methodology) and structural density has moderate
87 strength (an effect size estimate wit hin two standard errors above the pooled mean) and moderate stability as a predictors of crime rates. The literature reports fr equent examinations of racial composition as a correlate of crime rates, measured either as the percent or proportion of a given population that is nonWhite or Bla ck (Chamlin, 1989; Liska, Logan & Bellair, 1998; Neapolitan, 1998; Sampson, 1985, 1986; M.D. Sm ith & Bennett, 1985; D.A. Smith & Parker, 1980; Stafford & Gibbs, 1980; Williams, 1984; Williams & Flewelling, 1988), as well as numerous studies with age, sex, and poverty measures (e.g., Allison, 1972; Bailey, 1984, 1999; Baum, 1999; Blau & Blau, 1982; Britt, 1992; L. Cohen & Land, 1987; Copes, 1999; Curry & Spergel, 1988; Gartner, Baker & Pampel, 1990; Gauthier & Bankston, 1997; Glaser & Rice, 1959; Greenberg, 1985; Kapuskinski, Brai thwaite & Chapman, 1998; Messner, 1982; Messner & Sampson, 1991; OÂ’Brien, 1991; Osborn et al., 1992; Patterson, 1991; R.D. Peterson & Bailey, 1988; Ph illips & Votey, 1972; Sampson, 1985, 1987; D.A. Smith & Jarjour a, 1988; Steffensmeier, Streifel & Harer, 1987; Steffensmeier, Streifel & Shihadeh, 1992; Warner & Pier ce, 1993; Warner & Roundtree, 1997). Pratt and Cullen (2005) found percent Black, percent nonWhite, and poverty measures to have high strength and high stability as crime rate predictors, age structure to have moderate strength and high stability, and sex structure to have m oderate strength and stability. Some researchers have suggested, howev er, that WirthÂ’s (1938) view of
88 urbanism, particularly as it relates to the importance of pop ulation density, does not recognize that other factors may moderate the e ffect of population density on crime, or that the relationship may be spurious (Kasarda & Janowitz, 1974). Rather than forming an attachment to the community, or lack of attachment because of dense populations and superfici al relations, individuals may instead assimilate to a community system of fr iendship and kinship networks over time (Park & Burgess, 1925). Although Wirth (1938) discussed many urban factors beyond population density, such as residential mobility, he viewed density, the accumulation of large numbers in a small area, as mainly produc ing the other characteristics through the absence of intimate contacts and t he loss of formal control. He viewed urbanicity as creating DurkheimÂ’s (1897/2002) anomie through an interplay among a populationÂ’s number, it s density, and heterogeneity. Some researchers, however, sugges t that an individualÂ’s length of residence (Kasarda & Janowitz, 1974), an indi vidualÂ’s low residential stability or high residential mobility (Sampson & Grov es, 1989), operates more in line with Shaw and McKayÂ’s (1942, 1969) rationale; that high residential mobility, low residential stability, in part produces the lack of cohesiveness found in a community, and that population density is not important when residential mobility is controlled (Kasarda & Janowitz, 1974). Sampson and Groves (1989) characterized Shaw and McKayÂ’s theory as
89 specifying that disruptions in comm unity organization stemming from low economic status, ethnic heterogeneity, and residential mobility, influence variations in rates of delinquency. They noted that although macrosocial researchers frequently examined measur es derived from Shaw and McKayÂ’s (1942, 1969) findings, such as the effe cts of residential mobility, racial composition, and poverty meas ures on crime rates, ther e had been no direct test of Shaw and McKayÂ’s social disorganization theory. Arguing that the prime r eason social disorganization theory had never been tested was mainly a matter of suitable data, as opposed to theoretical shortcomings, Sampson and Groves ( 1989) examined the theory with Great Britain community-level and aggregated se lf-report crime and vi ctimization data. First, they defined social disorganization as Â“the inability of a community structure to realize the common values of its residents and maintain effective social controls (Kornhauser 1978, p. 120; Burs ik 1984, p.12 )Â” (Sampson & Groves, 1989, p. 777). Next, Sampson and Groves (1989) expl ained that social disorganization should be measured by the effectiv eness of those controls. Social disorganization results from a communityÂ’s inability to formally or informally supervise its residents, so it can be indexed by the communityÂ’s number and types of social networks. They measur ed social disorganization as sparse friendship networks, unsupervised gr oups of juveniles (teens), and low
90 participation in community organizations. Additionally, Sampson and Groves (1989) gave attention to the types of social structure that might be expe cted to impact delinquency. Drawing on Kornhauser (1978), Kasarda and Janowit z (1974), Krohn (1986), and Sampson (1987), they identified socioeconomic stat us (SES), residential mobility, racial and ethnic heterogeneity, fa mily disruption, and urbanization as the five exogenous processes to social disor ganizationÂ’s effect on delinquency. Sampson and Groves (1989) explai ned that SES was hypothesized by Shaw and McKay (1942, 1969) to affect delinquency through the mediation of social disorganization. Low community SES represents a dearth of the resources necessary to result in a strong organizational community base. Referencing Kornhauser (1978) and Byrne and Samp son (1986), Sampson and Groves (1989) noted that previous res earch that failed to find direct SES effects on crime rates inadequately measured t he intervening process. Sampson and Groves (1989) observed that residential mobility was in Shaw and McKayÂ’s (1942, 1969) original m odel as a disruptor of social networks that might otherwise be form ed if not for the lack of kinship to the community. Temporary, transient residents do not form strong friendship bonds and ties (Sampson & Groves, 1989). There is much research on residential mobility or residential instability (Lewis & Salem, 1986; Sampson, 1988; Tittle, 1989) in the literature (e.g., Baum, 1999; Bellair, 1997; Bursik & Grasmick, 1992; Crutchfield,
91 Garken & Grove, 1982; Fleisher, 1966; Heitgard & Bursik, 1987; Krivo & Peterson, 1996; Miethe, Hughes & McDowall, 1991; Sampson, 1986; Sampson & Raudenbush, 1999; D.A. Sm ith & Jarjoura, 1997; Veysey & Messner, 1999; Warner & Pierce, 1993; Warner & Roundtree, 1997; Weicher, 1970). Sampson and Groves ( 1989) likewise observed that Shaw and McKay (1942, 1969) identified racial and ethnic heterogeneity as important to the model. Shaw and McKay argued that heterogeneity affected the ability of community residents to achieve consensus, and Sampson and Groves noted that previous research that tested the direct effect s of heterogeneity on cr ime, like SES, failed to properly account for social diso rganizationÂ’s intervening process. Sampson and Groves (1989) derived t heir measure of family disruption from SampsonÂ’s (1987) argument that community controls are negatively impacted in communities having low leve ls of two-parent households. Sampson and Groves explained that two-parent households offered better networks of control both for their own children, and for other children within the community network. Lastly, Sampson and Groves (1989) explained that urbanization was implied by Shaw and McKayÂ’s (1942, 1969) intracity theory as contributing to the capacity to establish effective community controls. Sampson and Groves incorporated the level of urbanicity into their model so that they could rule out between-community urbanization effects.
92 Sampson and Groves (1989) concluded t hat there was over all support for their model. They found that socially disorganized communities had disproportionately high rates of delinquen cy, and that social disorganization (sparse friendship groups, unsupervised t eens, low organizational participation) partially mediated the effects of SES, re sidential mobility, ethnic heterogeneity, and family disruption (communi ty structural characterist ics) on their delinquency measures. Other researchers have since tested social disorganization theory with mixed results. Veysey and Messner ( 1999) reexamined Sampson and GrovesÂ’ (1989) data using structural equation modeling, finding only partial support for the social disorganization mediation hypothesis. Instead, they suggested that social disorganization represents more t han one mechanism, and that its operation supports additional theories of crime than so cial disorganization theory, including peer affiliation theories. First, Veysey and Messner (1999) argued that SEM analyses revealed that social disorganization as measured by Sampson and Groves (1989) did not comprise a single construct. The indica tors instead measured separate social processes. Veysey and Messner suggest ed that although the construct did not measure one distinct dimens ion, and although it was not a mediator of each of the community-level variables, it c ould be that the construct works as hypothesized but was measured poorly.
93 Further, Veysey and Messner (1999) observed that the strongest mediation of community effects came from the communityÂ’s perception of unsupervised teens. As analyses revealed it was a distinct intervening dimension, Veysey and Me ssner concluded that Sampson and GrovesÂ’ (1989) conclusion of clear support for social di sorganization theory was overstated. Veysey and Messner instead likened the peer group meas ure more to Akers and colleaguesÂ’ (1979) social lear ning theory than social diso rganization theory. They found the test of social disorganizat ion theory to be important, but they suggested that future studies seek stronger theoret ical measures. Lowenkamp, Cullen & Pratt (2003) a ttempted to replicate Sampson and GrovesÂ’ (1989) findings on BCS data 10 years newer than the data used by Sampson and Groves, thus examining t he stability of the findings. Lowenkamp and colleagues used a similar dataset and measures to those used by Sampson and Groves, but they exam ined a different time and place. Lowenkamp and colleagues concluded that t heir results were generally consistent with those of Sampson and Groves, and that the general propositions of social disorganization theory were supported. Lowenkamp and colleagues (2003) addr essed Veysey and MessnerÂ’s (1999) characterization of Sampson and GrovesÂ’ (1989) study as supporting multiple theoretical explanations as one worthy of future research. They suggested that future research explore the mechanisms as to why the variables
94 have the effects that they do. D. Gottfredson, McNeil, and Got tfredson (1991) investigated the mechanisms by which characteristics of a social area affect individual delinquency. Although they used social disorganization measures, they expanded on some of Sampson and Groves Â’ (1991) measures, and they did not aggregate the individual level survey data as did Sampson and Groves. D. Gottfredson and colleagues instead examined the effects of social structure directly on individual level delinquency. D. Gottfredson and colleagues (1991) argued that researchers had long been interested in the mechanisms by whic h social structure impacts individual behavior, but that no previous study had suit ably looked at the issue in light of ecological research such as that by Shaw and McKay (1942) and Sampson and Groves (1989). They further argued t hat two (Reiss & Rhodes, 1961; Johnstone, 1978) of the three published articles t hat had drawn conclusions regarding the effects of area characteristics on individual level crime used unsound methodologies: They violated HauserÂ’s (1970) caution against a contextual fallacy, misinterpreting groups effect s when shifting conclusions from an individual level of analysis. The third study, D. Gottfredson and colleagues (1991) argued, was methodologically sound, and it offered a mo re complete multi-level test of the effects of social structure on individua l delinquency, but its lack of broad social
95 structural measures failed to shed more light on how the macrosocial process affected individual level behavior. Si mcha-Fagan and Schwartz (1986) assessed the contextual effects of community economic level, community disorder, community organizational base, and comm unity residential stability on selfreported and officially recorded delinquen cy through the intervening mechanisms of bonds to conventional social roles and bonds to deviant social groups in a sample of 12 New York City neighborhoods. They advanced their model as representing portions of social disorganizati on, subcultural, and labeling theories. Simcha-Fagan and Schwartz (1986) concluded that one community level construct representing social disor ganization theory and another construct representing the subcultural perspective found strong empirical support. SimchaFagan and Schwartz reported that both constructs im pacted a communityÂ’s ability to sustain organiz ational participation, and that the variance between group effects on their delinquency measures was much reduced by the addition of individual-level variables. They su mmed their findings, in part, commenting, Â“[The study] indicates that when the reduced-form equation is more fully specified, community effects on delinquency are to a large extent mediated by socialization processes. The considerat ion of direct effects of community characteristics on delinquency thus involv es an oversimplificationÂ” (p. 695). D. Gottfredson and colleagues (1991) utilized a design strategy similar to Simcha-Fagan and Schwartz (1986) but th ey broadened the sample of social
96 areas by examining a convenience sample of 10 middle or high schools across 4 U.S. cities. They measured self-re ported delinquency, which comprised aggression, theft, property damage, and drug involvement measures. At the individual level, they measured parent al education, negative peer influence, parental attachment and supervision, school attachment and commitment, involvement, and belief in conventional rules. D. Gottfredson and colleagues (1991) i ndexed their social area measures with U.S. Census block group data, conduc ting factor analysis on the variables female-headed households, welfare, pover ty, divorced, male unemployment, female unemployment, male employment, female employment, professional or managerial employment, family income, education, farm income, and nonpublic school enrollment. They extracted variabl es representing two factors, labeling female-headed households, high welfare, high poverty, high divorce rate, and low male employment disorganization They called their second factor affluence and education which comprised incomes above the median level, high professional or managerial employment, completion of high school, employed females, and a low farm income to wages and salaries ratio. D. Gottfredson and colleagues (1991) concluded that their study provided only slight support for the notion, follo wing the rationale of Shaw and McKay (1942), that weak family st ructure reduces the contro l that is exerted over children, thereby resulting in increased interpersonal, aggressive delinquency. In
97 such areas, they concluded that children bonded less with controlling institutions and reported more negative peer influences than more organized areas. They also found that SES contributed to delin quency, though they concluded that the mechanism was not community control, as there was no effect on the bonding and peer association variables, and rather than affecting interpersonal violence, SES only impacted delinquencies such as theft and vandalism. Although measuring some concepts similar to Sampson and Groves (1989), and finding some support for some of the hypothesized relationships, D. Gottfredson and colleagues (1991) concluded t hat differences in social areas do not greatly influence in dividual delinquency. They commented, All [the limitations of the study] notwithstanding, the assumption that community characteristics explain much of the differences among individuals in criminal behavior no longer seems tenable. A maximum of 2% of the variance in individual delinquency is accounted for by area factors in any of the mu lti-level studies exami nedÂ—and a more reasonable estimate is less than 1%. The result s of every multilevel study relating individual delinquency to measures of ar ea characteristics imply that most of the variability among individuals must have sources other than differences in the communities they i nhabit. (D. Gottfredson et al., 1991, p. 221) Although D. Gottfredson and colleagues (1991) and Simcha-Fagan and Schwartz (1986) were interested in the ques tion of social disorganization, both studies, unlike Sampson and Groves (1989) examined the effects of aggregate community measures directly on individual delinquency. Both studies argued that some type of social process inte rvened between social structure and delinquency. The studies furt her distinguished themse lves from Sampson and
98 Groves (1989) as they used U.S. Census data to measure community structure. Further, D. Gottfredson and co lleagues (1991) suggested t hat better measures of social disorganization by community might have yiel ded different results. Sun, Triplett, and Gainey (2004) a ttempted to replicate Sampson and GrovesÂ’ (1989) tests of social disorgani zation theory, returning the level of analysis to the aggregate level, examin ing the impact of community on crime rates, but using U.S. Census data and in corporating broader measures of some of the theoretical constructs. They analyzed a sample ( N = 8155) that comprised 36 neighborhoods across 7 U.S. cities. Sun and colleagues (2004) operationalized SES as a scale comprised of the percentage of the co mmunity with an income above $20,000, percent employed, and the percent age of college graduates. They measured residential mobility as the percentage of residents t hat had resided in the community less than five years. They used BlauÂ’s (1977) index of intergroup relations to measure racial heterogeneity, and t hey measured family disrupt ion as the percentage of community residents that were divorc ed or separated. They held urbanicity constant, as all communities in the sample were considered urban. Sun and colleagues (2004) measured the intervening construct local social ties as the percentage of neighb ors who reported doing things together, and they measured organizational participati on as the percent of residents who attended community meetings during the previous 6-12 months relating to area
99 drug problems. Sun and colleagues m easured unsupervised teens as the percent of residents who considered di sruptions around schools as a problem. Their dependent variables were robbery and assault rates. Sun and colleagues (2004) modeled pat hs that accounted for those reported by Veysey and MessnerÂ’s (1999) replication of Sampson and GrovesÂ’ (1989) study, concluding that social diso rganizationÂ’s mediation of community effects on crime found only partial support. They found that each of the social disorganization measures di d not mediate the community -level effects; rather only the local social ties measure did so effectively. They, like the other tests of social disorganization theory, suggested that future research employ better measures of the theorized constructs. Applicability to social st ructure-social learning. Akers (1998) suggests that the soci al learning process mediates the effects of social structure on crime and criminal behavior. Although he proposes four social structural dimensions, two of the dimensionÂ’s indicators overlap as they both seek empirically sound macrosoc ial correlates of crime rates, one from the angle of incorporating known correlates, be they atheoretical or theoretically derived, and the other focusi ng specifically on theoret ical explanations. Akers appears mainly unconcerned with the source of the social structural variables, beyond their empirical relationship with crime. Akers likewise is not concerned with theoretically derived rationales, bey ond noting that the most promising
100 theories are anomie, social di sorganization, and conflict. Pratt and Cullen (2005) provided the most comprehensive and recent examination of macrosocia l predictors of crime ra tes. Their meta-analysis suggested that social disorganization and the conflict notions of resource or economic deprivation prov ide adequately tested and highl y supported theoretical macro-level explanations for crime. Pratt and Cullen found that racial composition, family disr uption, and poverty were t he most robust macrosocial crime rate predictors, and they suggested that macros ocial theoretical tests would be misspecified without their inclusion. In additi on, they identified other moderate or highly strong and stable macros ocial predictors such as urbanism, structural density, age, and sex, among others. Sampson and Groves (1989) demonstr ated how to measure and test social disorganization theory, a rationale that was adapted to U.S. Census data by Sun and colleagues (2004). D. Gottfredson and colleagues (1991) and Simcha-Fagan and Schwartz (1986) showed how the effects of macrosocial variables could be tested on individual delin quency directly, though both studies modeled intervening variables that in par t contained social learning (deviant peers) measures. Although not testing so cial disorganization theory, per se, Hoffmann (2002), discussed in the previous chapter, likewise examined the direct effects of social structure on individua l delinquency including various intervening measures, some of which were intended to represent social learning variables.
101 Some of the macrosocial research f ound weak social structural effects, suggesting that future research should seek better theoretical measures (e.g., D. Gottfredson et al., 1991; Lowenkamp et al., 2003; Sun et al., 2004; Veysey and Messner, 1999). Although working from a fram ework different than that of social disorganization, and examining a narrow outcome measure, Land and colleagues (1990) warned that in addition to measuring structural covariates consistently, researchers must make su re that the intercorrelation between predictors does not interfer e with the power of the statistical examination. Although the macrosocial literature approaches the problem of crime from a position differently than that of Ak ers (1998), none of the reviewed literature convincingly refutes his viewpoint. Instead, much of the literature supports AkersÂ’ notion that social disorganization and conf lict theories are important macrosocial correlates, and three studies showed how their indicators, as well as other macrosocial crime covariates might be tested on individual level data.
102 Chapter Four Rationale for the Present Study Overview The present research contributes to t he theoretical body of literature in two major ways. First, this study distinguishes itself from previous research on AkersÂ’ (1998) social structure-social learning theory by incorporating more complete measures of the differential social organization and theoretically defined structural causes dimensions, and it se condly explores how the dimensions may impinge on the social learning process. It responds to AkersÂ’ (1999) call to help specify the most underdevelo ped portion of the social st ructure-social learning model. Sutherland (1939) began with an interest in explaining both crime and criminal behavior, which led to a t heory that discussed bot h macrosocial and microsocial structures and processes. Sutherland (1947) revised the theory, however, such that its final version cons trained itself to microsocial processes. What began as a broad, general theory of both crime and criminal behavior ended up as a delimited explanation of t he general processes that influence deviant and conforming behav ior at the individual le vel of explanation. Sutherland (1947) retained the notion that social disorganization and
103 normative conflict are involved in the fo rmation of individual criminal behavior, that differential social organization pr ovides the opportunity for differential associations to occur, but his final version of the t heory did not specify the links between social structure and criminal behav ior. Sutherland remained interested in both an epidemiological an d etiological explanation for crime and criminal behavior, but his formal theoretical statements excluded macrosocial considerations. Burgess and Akers (1966) revised Su therlandÂ’s processual theory to better specify the learning process, keeping the theory focused on the microsocial level. Akers (1998) later elaborat ed social learning theory such that it attempts to explain both the macrosocial structure and microsocial processes that lead to deviant or conforming behav ior, and ultimately crime rates, by viewing social structure as the l earning environment for individual behavior (Akers, 1968). Akers (1998) revisited the form al cross-level specification of crime causality abandoned by Su therland (1939, 1947). Akers (1998) referenced SutherlandÂ’s (1947) earlier lack of macrosocial linking propositions as an impetus for hi s explicating social structure-social learning theory. Although Sampson (1999) and Krohn (1999) have suggested that Akers (1998) likewise fails to prov ide suitable linking propositions between social structure and social learning, Akers (1999) suggests that the model specifies relationships enough for empirical testing.
104 Although acknowledging the concerns about macrosocial variables, Akers (1999) concludes that social structuresocial learning theory requires better empirical testing with cross-level data, not further theorizi ng. Akers (1998) and colleagues (Lee et al., 2004) sugg est that research in this area should test more comprehensive models that include broader indicators of social structure, especially those derived from macrosocial t heories of crime. The point of the social structure-social learning specification is that social structure only influences individual behavior through its influence on social learning variables. The theory hypothesiz es that theoretical concepts already known to influence crime rates do so through their influence on reinforcement contingencies. Therefore, the social stru cture-social learning model does account for theoretically derived macrosocial determinants. Study Objectives Although AkersÂ’ (1998) social structur e-social learning model is testable without further theoretical work linking t he structural variables to the social learning process, theoretically derived macrosocial measures need better attention. Past empirical tests have not fully captured the dimensions described by Akers, and researchers (Bellair et al., 2003; Lanza-Kaduce & Capece, 2003; Lee et al, 2004) have been unable to suitably explain why social structure might be expected to influence the social learning process. The present research draws on the macr osocial literature to measure both
105 the differential social organization dimens ion, and the dimension that represents theoretically defined causes, notably measur es endorsed in previous research by Sampson and Groves (1989), D. Go ttfredson and colleagues (1991), and Sun and colleagues (2004), among others (see Pra tt & Cullen, 2005). Its major goal is to operationalize AkersÂ’ (1998) stated propositions and explicate functional relationships between suitable measuresÂ—t o state the hypotheses requested by Krohn (1999) that explain why certain structural variables result in different levels of the social learning variables, in a manner that gives attention to social structural explanations consistent wit h the expectations of Sampson (1999). Another major aim of the present resear ch is to critically examine AkersÂ’ (1998) statement that social learning theory mediates the effect of macrosocial variables on criminal behavior. Beyond whether the model is measured correctly, or finds statistical support, AkersÂ’ use of the term mediation warrants scrutiny. As stand-alone theories, macrosocial ex planations typically compete with microsocial explanations (Akers, 1998), though they operate at different levels of explanation. Figure 2 present s these theoretical models us ing social structure as a macro-level explanation for crime rate s and social learning as a micro-level explanation for criminal behavior.
106 In a cross-level integrated approach, proximate microsocial processes intervene between distal macrosocial causes of behavior and group rates. Social structure affects group rates only th rough their effect on the microsocial processes; rather, social structure has no effect on group rates independent of microsocial processes. In a social learni ng framework, social structure provides learning contingencies for individual behavio r, ultimately influencing crime rates. Earlier, Figure 1 depicted the social struct ure-social learning model as devised by Akers (1998), showing the indicators of each dimension. Figure 3 depicts the theoretical model of all relationships, representing each dimension as a latent construct. Social Structure-Social Learning Theoretical Model Figure 3 Criminal Behavior Social Structure Social Learning Crime Rates Social Structure and Social Learning Theoretical Models Figure 2 Crime Rate Social Structure Social Learning Criminal Behavior
107 Akers (1998) suggests that although Figure 3 describes the explanation for criminal behavior and crime, st atistical models cannot adequately demonstrate such due to biased sampling, measurement error, and an inability to control for all factors. Beyond the stat istical issues, Akers points out that researchers should not expect to m odel human behavior perfectly, and that researchers should not seek full dete rministic models. Akers consequently expects imperfect social lear ning mediation, commenting, The [social structure-social learning model] is depicted in the way it is to show that it can be tested with empirica l data in a multivariate statistical model. What kind of empirical findings what magnitude of coefficients from such a statistical analysis, will be taken as confirming or disconfirming the theory? It depends on how strongly or unequivocally the expected relationships are stated. The strongest expectation is that vari ations and stabilities in the behavioral and cognitive variables in the soci al learning process account for all variations and stabilities in cr iminal behavior and thereby mediate all of the significant relationships between the structural variables and crime. The more realistic statement is that vari ations and stabilities in the behavioral and cognitive variables specified in the social learning process account for a substantial portion of individual variations and stabilities in crime and deviance and mediate a substantial portion of the relationship between most of the structural variables in t he model and crime. A weak statement of the theory is that the social learning process accounts for some portion of the variation and stability in criminal behavior and mediates some portion of the relationship between the correlates and crime. (Akers, 1998, p. 340) Although a full mediating model is i deal (no direct path between social structure and crime rates), Akers (1998) suggests that social structure-social learning theory strives for substantial mediat ion (a weaker direct path from social
108 structure to delinquency t han that through social le arning to delinquency). However, Akers does not specify what qualif ies as substantial mediation, beyond noting that, The more closely the results of the analysis show relationships as predicted by the model, the more one can conclude that the theory has been supported. . If subs tantial portions of the variations (by normally accepted standards in social science) are accounted for by the variables in the theory, then it is c onfirmed. (Akers, 1998, p. 341) What are the normally accepted social science standards for substantial mediation? Akers (1998) does not say. The present research seeks a better specification of mediati on generally, and substantial mediation particularly. Mediation and Substantial M ediation versus Moderation Similar to the present research, n one of the reported tests of social structure-social learning t heory has incorporated crime rates into the empirical test of the model. Each previous test has treated structure si milarly: Structure serves as that which influences microsoc ial behavior, whether that structure is occupational, university association, or some other community aggregate. Although not making strong statements on the issue, each of the previous researchers has evaluated test results according to a partial or substantial mediation standard. Figure 4 depicts generally the theor etical model tested in previous research, as well as the pres ent study (using delinquency as a proxy for criminal behavior).
109 As depicted, social structure both di rectly and indirectly (through social learning) affects delinquency. Akers (1998) suggests that partial mediation is present when the path from social structur e to delinquency is weaker with social learning in the model than it would be with social learning not in the model. In specifying the social structuresocial learning model, Akers (1998) points out that he considers his effort to be theory elaboration along the lines of that proposed by Thornberry (1989). This does not seem to be an inconsequential point. Althoug h the social structure-soci al learning model has its roots in SutherlandÂ’s (1939, 1947) work, AkersÂ’ theory elaboration expands out from social learning theor y, attempting to see how far the theory will extend, rather than down from SutherlandÂ’s concept of differ ential social organization. By labeling his social structural extension of social learning t heory an elaboration, Akers appears to be both taking a positi on on the theory competition versus theory integration debate, and he seems be rejecting the views of critics that Model of the Relationship Be tween Social Structure and Figure 4 Delinquency with Social Learni ng as a Substa ntial Mediator Delinquency Social Structure Social Learning
110 expect theoretical propositions linking ma crosocial explanations to the modelÂ’s microsocial processes. AkersÂ’ (1998) approach may be adequate if social learning does indeed mediate social structural effects on delinquency, if adding the social learning process that explains indi vidual delinquency into the model eliminates the effects of social structure on crimi nal behavior, the aggregate of which form crime rates. However, AkersÂ’ specification is less satisfying when full mediation does not occur. Noted earlier, Akers explains that expecting full mediation from a statistical model is unrealistic, as sampling bias and measurement erro r affect results. Instead, Akers suggests that the t heory finds satisfactory support when substantial mediation is evident. However, Akers does not explicate this term. He does not sufficiently define substantial mediation. Moreover, AkersÂ’ (1998) use of t he terms mediation and substantial mediation may be inconsistent with hi s and SutherlandÂ’s (1939, 1947) various explanations of the rela tionship between social structure and the microsocial processes that affect criminal behav ior. For example, Sutherland (1947) suggests that crime is r ooted in social structure, as differential social organizations provide the opportunity for di fferential associations. One concludes that groups organize for or against crimi nal behavior. Social disorganization and culture conflict affect the formation of individual criminal behavior. Akers (1968, 1973, 1977, 1985, 1992, 1998) continually describes the
111 social structure and social learning re lationship in a similar manner, suggesting that social structure provides the conti ngencies for social learning to occur. One concludes that social structure provides the environment that shapes individual behavior through the process of social learni ng. Social structural situations shape individual behavior. The contexts of social interaction produce learning environments conducive to conformity or nonconformity. AkersÂ’ (1998) social structure-social learning model describes mediation, yet his narratives explaining the proc ess may describe a contextual, or moderating effect. As described, social structure may affect individual behavior through its interaction with social learni ng. Although Akers is clear that social learning intervenes between social stru cture and criminal and deviant behavior, his use of partial mediation as an acceptable standard seemingly clouds the distinction between mediat ion and moderation. For example, the theoret ical model described earlie r (Figure 4) as that which has been tested in t he literature (Bellair et al., 2003; Lanza-Kaduce & Capece, 2003; Lee et al., 2004) derives from AkersÂ’ (1998) use of the term mediation and his supposition that partial mediation is that by which the theory should be judged. Recall, however that the social struct ure-social learning model advanced by Akers (Figure 1) has no dire ct path from social structure to individual behavior. Social structure-social learning theory suggests that the social learning process leading to criminal behavior fully mediates the effects of
112 social structure on crime rates. AkersÂ’ (1998) suggests that the theor etical model should be relaxed for purposes of testing its va lidity, and he introduces t he notion of substantial mediation for that purpose. The model depic ted in Figure 4 is the tested model. It excludes crime rates from consideration, but more importantly, it allows a direct path from social structure to deviant, criminal, and delinquent behavior, as well as an indirect path to delinquent behavior through the social learning process. The tested model derives from AkersÂ’ de scription of the model through use of the term mediation, serving as a relaxed mode l that depicts statis tical mediation of AkersÂ’ theoretical concepts. Although Figure 4 correctly depicts stat istical mediation (Rozeboom, 1956; see Baron & Kenny, 1986; James & Bret t, 1984; Judd, Kenny & McClelland, 2001; Kraemer, Stice, Kazdin, Offord & Kupfer, 2001), researchers often incorrectly use mediation and moderati on as synonyms (see Baron & Kenny, 1986), sometimes in the same article (e.g., Findley & Cooper, 1983; Harkins, Latane & Williams, 1980). Holmbe ck (1997), for example, not ed that a researcher verbally described moderation, visually i llustrated mediation, and tested neither. Researching tests and reports of inte raction in nonlinear models, Chunrong & Norton (2003) examined 72 articles published between 1980 and 1999 in the econometrics literature and c oncluded that none of them reported the results correctly.
113 Adding to the confusion, methodol ogists note that both mediators and moderators sometimes produce incomplete statistical reduction in bivariate effects when added to a model (Baron & Kenny, 1986). For ex ample, if the bivariate effect between social structur e and delinquency is reduced but not fully accounted for by the addition of social l earning variables, the resulting indirect effects between social structure and delin quency may be the result of social learning intervening between the variables (statistical mediation). However, the weaker but still present indirect effect s of social structure and delinquency may have to do with the way social structure in teracts with social learning (statistical moderation). Mediation accounts for the relations hip between an independent variable and a dependent variable, whereas moderati on describes the circumstances in which the relationship exists, or when the effects will hold (Baron & Kenny, 1986). Mediation relates to the proce ss that produces t he dependent variable, whereas moderation relates to the magnitude of its effect (Judd et al., 2001). An identified independent variable di rectly influences a mediator variable, whereas a moderator variable influences the rela tionship between the independent variable and a dependent variable (Baron & Kenny 1986; Kraemer et al., 2001). A mediator variable is consistent with a general explanation, whereas a moderator variable implies a conditional relations hip (Friedrich, 1982). Baron and Kenny (1986) summarize the diffe rence between mediators an d moderators by noting,
114 Â“[Whereas] mediator-oriented research is more interested in the mechanism than in the exogenous variable itself. . moder ator research typically has greater interest in the predictor va riable per seÂ” (p. 1178). As it relates to social structure-so cial learning, keeping delinquency as a proxy for criminal behavior, mediat ion suggests that there would be no relationship between social structure and cr ime rates if not for social learning and delinquency. Social learning is the proce ss by which social structure affects delinquency and ultimately crime rates. Social structure directly influences social learning. If moderation is at work, soci al learning and delinquency are the circumstances by which the relationship between social struct ure and crime rates exists. The effects of social structure on crime rates hold when social learning is considered. Social learning influences the magnitude of social structureÂ’s effect on crime rates. In sum, m oderation implies that the c ausal relationship between social structure and delinquency changes as a function of social learning (see Baron & Kenny, 1986). Social learning cond itions social structureÂ’s effect on delinquency (see Friedrich, 1982; Hoffmann, 2002). Although substantial medi ation, the standard advoc ated by Akers (1998) as suitably testing the social structur e-social learning model, may be suggestive that mediation is in play, the approac h leaves open the possibility of social learning as a moderator. If full mediat ion does not occur, and researchers have
115 not ruled out moderation beforehand, researc hers may misinterpret the results. Akers (1998) suggests that social lear ning variables relate to social structural variables as a mediator. Ma crosocial critics of that position would suggest that if social learning relates to the macrosocial variables at all, it is as a moderator. Both Akers and the social stru ctural critics might agree that social learning and social structure relate to one another, but they would disagree on the type of relationship. If a researcher te sts a model of social structural effects on delinquency first without social learning variables in the model and then later with social learning variables in the model, and the variables are expected to relate with one another, one would expect the macrosocial coefficients to be different. There are circumstances, how ever, in which both mediation and moderation may result in the reduction of the social structural coefficients. Researchers can only test mediation through techniques that allow causal modeling, however, and medi ation should only be tested after moderation has been ruled out (see Baron & Kenn y, 1986; James & Brett, 1984). Each of the social learning tests in the literature report finding evidence of mediation, but none report having tested moderation (Bellair et al., 2003; LanzaKaduce & Capece, 2003; Lee et al., 2004). Hoffmann (2002) reports testing moderation, but he found no e ffects between social structure and social learning, be it moderating or mediating. In additi on to not sufficiently indexing the differential social organization and theor etically defined structural causes
116 dimensions theorized by Akers (1998), social structure-social learning tests have also not adequately accounted for the po ssible alternative explanation of moderation. If social learning is a mediator, the relationship between social structure and delinquency is spurious as social stru ctural effects on delinquency only occur through their effects on the social learni ng process. If social learning is a moderator, social structur e affects delinquency through an interaction with the social learning process. Figure 5 illust rates these two testable hypotheses.
117 Unlike the mediation model, with moder ation, social structure and social learning occupy the same level of antecedence to delinquency (see Baron & Kenny, 1986). The additional variable repr esents the product of the independent variable and the presumed moderator (B aron & Kenny, 1986; Friedrich, 1982; James & Brett, 1984; Judd et al., 2002; Kramer et al, 2001). Discussing moderation, Baron and Kenny (1986) comment, Social StructureDelinquency Social Learning Social Structure (SS) Social Learning (SL) Product Term (SS) X (SL) a b c ab c Mediator Hypothesis (substantial mediation) DelinquencyPath Diagram of Hypothes es Depicting Social Le arning as a Moderator Figure 5 and a Mediator of the Social St ructural Effects on DelinquencyModerator Hypothesis
118 The diagrammed [model] has three c ausal paths that feed into the outcome variable . .The moderator hypothesis is supported if the interaction (Path c ) is significant. There may also be significant main effects for the predictor and the moderator (Paths a and b ), but these are not relevant conceptually to testi ng the moderator hypothesis. (p. 1174) Researchers find support for moder ation when the path between the interaction term and the dependent variable is significant, regardless of the significance of the independent and moderat ing variable paths (Baron & Kenny, 1986). If moderation is present, if the path between the social structure and social learning interaction term and delin quency is significan t, the rest of the model need not be interpreted. Describing mediation, Bar on and Kenny (1986) note, A variable functions as a mediator when it meets the following conditions: (1) variations in levels of the inde pendent variable significantly account for variations in the presumed mediator (i.e. Path a ), (b) variations in the mediator significantly account for vari ations in the dependent variable (i.e., Path b ), and (c) when Paths a and b are controlled, a previously significant relation between the independent and dependent variables is no longer significant, with the strongest demonstr ation of mediation occurring when Path c is zero. In regard to the la st condition we may envisage a continuum. When Path c is reduced to zero, we have strong evidence for a single, dominant mediator If the residual Path c is not zero, this indicates the operation of multiple m ediating factors. Because most areas of psychology, including social treat phenomena that have multiple causes, a more realistic goal may be to seek mediators that significantly decrease Path c rather than eliminating the relation between the independent and dependent variables alto gether. From a theoretical perspective, a significant reduction demons trates that a given mediator is indeed potent, albeit not both a necessa ry and sufficient condition for an effect to occur. (p. 1176) Referring back to the debate between Akers (1999), Sampson (1999) and Krohn (1999), Akers may be insistent on a mediation relationship because he
119 has started with the social learning proc ess, expanded out, and is trying to see how far the explanation goes (Akers, 1998) Sampson, however, starts with a macrosocial perspective and although he ma y buy a moderating effect, that is unclear, he does not accept AkersÂ’ implicat ion that social stru cture is important only to the extent that it provides the opportunity for social learning to occur. Krohn starts with the life-c ourse perspective, in the example given, and he thinks social structure-social learning is inte resting, perhaps helpfu l, if it can help explain the various macrosocial processes that impact crime over the life-course. He may be expecting a moderating effect. Based on AkersÂ’ (1968, 1973, 1977, 1985, 1992, 1998) description of the relationship between social structure and social learning, as well as the definitions of mediation or moderation, both explanations seem plausible. Social structure may influence crime rates only because it sets the opportunities for various individual level reinforcement schedul es to occur, resulting in criminal behavior that aggregates to the group level. Social structure may affect crime rates both inherently, or in combination with various individua l social learning components. In sum, the present research contri butes to the theoretical body of literature through its examin ation of social learning theoryÂ’s generalizability across levels of explanation. The re search specifically models strong macrosocial measures that index AkersÂ’ (1998) differential social organization
120 and theoretically defined struct ural causes dimensions, attempting to explain why social structure should influence social l earning. Further, the research attempts to clarify whether social learning in tervenes between social structure and delinquency as a mediator, or if it interact s with social structur e as a moderator, if there is any relationship at all. Functional Relationships Recall that the social structure-soci al learning model makes predictions about social structure, soci al learning, individual criminal behavior, and crime rates (see Figure 1). Akers ( 1998) justifies inclusion of crime rates in the model as that which traditionally correlates wit h social structure. Akers views the insertion of social learning theory between social structure and crime rates as the answer to the question, by what process does social structure affect crime rates? Akers (1998) contends that his crosslevel integration of theoretical explanations for crime and cr iminal behavior is logically consistent because both levels of explanations seek answers to the same question. Akers characterizes crime rates as the sums of individual crimes committed by those individuals falling within the system. Aker s argues that crime rate s are little more than an aggregate of criminal behaviors. Although researchers generally use soci al structural theories, along with atheoretical macrosocial crime correlates, to make predictions about crime rates, as noted earlier, some researchers have re lated social structural factors to
121 individual behavior (Simcha-Fagan & Schwartz, 1986; D. Gottfredson & colleagues, 1991). Such studies have followed the rationale that adequate evaluations of contextual effects must simultaneously index soci al structural and individual-level measures (Blau, 1960; Simcha-Fagan & Schwartz, 1986). Although such literatur e adequately addresses that portion of AkersÂ’ (1998) model that connects social struct ure to individual behavior, there is no support in the literature for aggregating the micro-level behaviors back to the aggregate rate level as advanced by Akers. In contrast, the literature suggests that such theoretical formulation ma y create an aggregation inconsistency (see Blalock, 1984; Bursik & Grasmick, 1996; Hannan, 1971). Moreover, although Akers has advanced crime rates as part of the theoretical model, researchers have excluded that link from each test of the model. Consistent with AkersÂ’ (Lee et al., 2004) test of social structure-social learning theory, as well as the other two r eported tests in the lit erature (Bellair et al., 2003; Lanza-Kaduce & Capece, 2003), t he present study does not examine that portion of AkersÂ’ (1998) model that makes predictions about crime rates from the observation of individua l criminal behavior. The present research instead holds that portion of the model as inconsis tent with the literatur e, and it examines the relationship solely among social st ructure, social learning, and individual delinquency. Both Sutherland (1947) and Akers (1998) contend that crime is an
122 expression of social organization. Akers elaborates that so cial structure and culture provide differential learning environments that infl uence an individualÂ’s learning contingencies. Akers suggests that social struct ure affects delinquency only through its effect on the social learning process. Aker s posits that four social structural dimensions pr oduce social learning, which in turn accounts for individual criminal behavior, but he does not explain how the social structure variables actually operate to create vari ations in associations, definitions, reinforcements, and models. One way that the social st ructural dimensions may re late to social learning and individual delinquency antecedent to group crime rates is through reinforcement contingencies, discussed earlier in the section that relates SampsonÂ’s (1999) and KrohnÂ’s (1999) conc erns about the social structural elaboration of social learning theory. Re call that individual reinforcement for social behavior occurs when there is a balance of actual or anticipated rewards over punishment. Individual reinforcem ent schedules derive from sets of reinforcement contingencies. Individual behavior that is not emitted is not eligible for reinforcement (or punishment). Social learning theory sugges ts that individual behavior is unlikely to be emitted when reinforcement is un likely. Reinforcement operates through amount, frequency, and probability modalit ies, and individual behavior is not actually reinforced all of the time. Ra ther, individual behavior is generally
123 reinforced on variable interval schedules. Because behavior is intermittently reinforced, because there is always a chance for reinforcement, individ uals continue learned behavior until reinforcement stops. Individuals conti nue social behavior until the balance of anticipated rewards no longer exceeds that of punishment (ext inguishment). At the macrosocial level, social stru cture provides arrangements of various sets of reinforcement contingencies (Akers, 1998). Structure provides the occasion for reinforcement contingencies to occur, thereby affecting individual reinforcement schedules. Individual behavio r cannot be reinforced if it is not emitted, and its emittance is dependent on both the reinforcement schedules and the reinforcement contingency. The linking mechanism requested by Krohn (1999), therefore, is that soci al structural variables infl uence variations in social learning variables by providing the environmental setting for contingencies of reinforcement. Social learning variabl es then produce various reinforcement schedules that lead to the onset, continuanc e, or desistance of individual deviant behavior. Akers (1998) suggests that social struct ure affects crime through its effect on social learning. The macrosocial literat ure review suggested that indicators of social disorganization theoryÂ’s antecedent macro-level variables (SES, ethnic heterogeneity, residential mobi lity, family disruption), along with other various social structural measures such as population density, race, sex, age, and
124 poverty, find moderate to high strength and stability as predictors of crime across a wide range of empirical tests. These known social structural correlates and theoretically derived composite measures may affect social learning variables and individual delinquency directly through their various sets of reinforcement contingencies. Population density, for example, may affect delinquency through the inability of highly dense communities to provide social structural learning contingencies of individual reinforcement that are conducive to law conformity. Smaller communities, less dense populati ons, are better able to exert more control over community members than more densely populated areas (see Bursik & Grasmick, 1993). Smaller co mmunities may offer more homogeneous reinforcement schedules. Because behavior emittance corresponds with social reinforcement frequency (Hamblin, 1979), and because individuals seek opportunities to maximize social reinforcement for indi vidual behavior (Herrnstein & Leveland, 1975), homogeneous populations (e.g., less population density) may exert more influence over individual behavior. Various individual reinforcement schedules control the emitting of behavior. Social structure, in this case homogeneous populations, controls the reinforcement contingencies. Large population densities may produce more delinquency than low population densities because such societal makeup provides more opportunities
125 for reinforcement of delinquent behavior. Heterogeneous popula tions offer more behavioral choices, and thus more pl entiful and differing reinforcement contingencies. When smaller groups hold di ffering views of mores than those of the larger community (see Sutherland, 1939), contingencies for reinforcement of those differing views will occur. The same logic equally applies to other social structures. Individuals that have little in common with their larger group, individuals that have superficial group and community relations such as thos e stratified by race, sex, age, or poverty, for example, may be less likely to be controlled by larger groupings (e.g., the community). The individuals instead may be more likely to engage in behavior learned in their smaller groupi ngs, and because of the process of maximizing social reinforcement, indi viduals may emit the behavior even when such behavior goes against societal norms. Such societal makeup, a high populatio n density of people with superficial relations, small groups stratified by race sex, age, or poverty, may result in varying levels of differential associati ons, definitions, imitation, and differential reinforcement. The social structure provi des different sets of contingencies of reinforcement, differential behavioral re wards, thus producing individual reinforcement schedules that lead to diffe rential patterns of delinquent behavior.
126 Hypotheses The present research tests hypotheses derived from three of AkersÂ’ (1998) four social structure-social l earning dimensions: differential social organization, differential location in the social structure, and theoretically defined structural causes. Figures 6-10 depict the studyÂ’s social structure-social learning moderator and mediator hypotheses for each differential social organization indicator, and Figure 11 portrays the dimensionÂ’s hypothesis. Figures 12-14 represent the indicator hypotheses and Fi gure 15 the dimension hypothesis for differential location in the structure. Fi gures 16-19 portray the hypotheses for the theoretically derived structural causes dimension, and Figure 20 depicts its dimension hypothesis. Figure 21 depicts t he hypothesized structural model of each of the three measured dimensions.
127 Moderator Hypothesis Population Density Differential Association Product Term: (PD X DA) Mediator Hypothesis Population Density Differential Association Moderator Hypothesis Population Density Definitions Product Term: (PD X D) Mediator Hypothesis Population Density Definitions Moderator Hypothesis Population Density Rewards Product Term: (PD X R) Mediator Hypothesis Population Density Rewards Moderator Hypothesis Population Density Costs Product Term: (PD X C) Mediator Hypothesis Population Density Costs Population Density (PD) and Costs (C) Hypotheses Delinquency Delinquency Delinquency Population Density (PD) and Rewards (R) Hypotheses Delinquency Delinquency Delinquency Figure 6 Path Diagram for SSSL Dimension I (Population Density) Hypotheses Population Density (PD) and Differe ntial Associations (DA) Hypotheses Population Density (PD) a nd Definitions (D) Hypotheses Delinquency Delinquency
128 Moderator Hypothesis Race Composition Differential Association Product Term: (RC X DA) Mediator Hypothesis Race Composition Differential Association Moderator Hypothesis Race Composition Definitions Product Term: (RC X D) Mediator Hypothesis Race Composition Definitions Moderator Hypothesis Race Composition Rewards Product Term: (RC X R) Mediator Hypothesis Race Composition Rewards Moderator Hypothesis Race Composition Costs Product Term: (RC X C) Mediator Hypothesis Race Composition Costs Race Composition (RC) and Definitions (D) Hypotheses Delinquency Delinquency Delinquency Figure 7 Path Diagram for SSSL Dimension I (Race Composition) Hypotheses Race Composition (RC) and Differential Associations (DA) Hypotheses Delinquency Race Composition (RC) and Rewards (R) Hypotheses Delinquency Delinquency Race Composition (RC) and Costs (C) Hypotheses Delinquency Delinquency
129 Moderator Hypothesis Sex Composition Differential Association Product Term: (SC X DA) Mediator Hypothesis Sex Composition Differential Association Moderator Hypothesis Sex Composition Definitions Product Term: (SC X D) Mediator Hypothesis Sex Composition Definitions Moderator Hypothesis Sex Composition Rewards Product Term: (SC X R) Mediator Hypothesis Sex Composition Rewards Moderator Hypothesis Sex Composition Costs Product Term: (SC X C) Mediator Hypothesis Sex Composition Costs Sex Composition (SC) and Costs (C) Hypotheses Delinquency Delinquency Delinquency Sex Composition (SC) and Rewards (R) Hypotheses Delinquency Delinquency Delinquency Figure 8 Path Diagram for SSSL Dimension I (Sex Composition) Hypotheses Sex Composition (SC) and Differential Associations (DA) Hypotheses Sex Composition (SC) and Definitions (D) Hypotheses Delinquency Delinquency
130 Moderator Hypothesis Age Composition Differential Association Product Term: (AC X DA) Mediator Hypothesis Age Composition Differential Association Moderator Hypothesis Age Composition Definitions Product Term: (AC X D) Mediator Hypothesis Age Composition Definitions Moderator Hypothesis Age Composition Rewards Product Term: (AC X R) Mediator Hypothesis Age Composition Rewards Moderator Hypothesis Age Composition Costs Product Term: (AC X C) Mediator Hypothesis Age Composition Costs Age Composition (AC) and Definitions (D) Hypotheses Delinquency Delinquency Delinquency Figure 9 Path Diagram for SSSL Dimension I (Age Composition) Hypotheses Age Composition (AC) and Differential Associations (DA) Hypotheses Delinquency Age Composition (AC) and Rewards (R) Hypotheses Delinquency Delinquency Age Composition (AC) and Costs (C) Hypotheses Delinquency Delinquency
131 Moderator Hypothesis Poverty Differential Association Product Term: (P X DA) Mediator Hypothesis Poverty Differential Association Moderator Hypothesis Poverty Definitions Product Term: (P X DA) Mediator Hypothesis Poverty Definitions Moderator Hypothesis Poverty Rewards Product Term: (P X DA) Mediator Hypothesis Poverty Rewards Moderator Hypothesis Poverty Costs Product Term: (P X DA) Mediator Hypothesis Poverty Costs Poverty (P) and Costs (C) Hypotheses Delinquency Delinquency Delinquency Poverty (P) and Rewards (R) Hypotheses Delinquency Delinquency Delinquency Figure 10 Path Diagram for SSSL Dimension I (Poverty) Hypotheses Poverty (P) and Differential Associations (DA) Hypotheses Poverty (P) and Definitions (D) Hypotheses Delinquency Delinquency
132 Path Diagram for the Social Structure-Social Learning Dimension I Figure 11 Hypothesis that Social Learning Mediates the Effect of Differential Social Organization on Delinquency Delinquency Social Organization Social Learning
133 Moderator Hypothesis Individual Sex Differential Association Product Term: (IS X DA) Mediator Hypothesis Individual Sex Differential Association Moderator Hypothesis Individual Sex Definitions Product Term: (IS X DA) Mediator Hypothesis Individual Sex Definitions Moderator Hypothesis Individual Sex Rewards Product Term: (IS X DA) Mediator Hypothesis Individual Sex Rewards Moderator Hypothesis Individual Sex Costs Product Term: (IS X DA) Mediator Hypothesis Individual Sex Costs Individual Sex (IS) and Definitions (D) Hypotheses Delinquency Delinquency Delinquency Figure 12 Path Diagram for SSSL Dimension II (Individual Sex) Hypotheses Individual Sex (IS) and Differential Associations (DA) Hypotheses Delinquency Individual Sex (IS) and Rewards (R) Hypotheses Delinquency Delinquency Individual Sex (IS) and Costs (C) Hypotheses Delinquency Delinquency
134 Moderator Hypothesis Individual Race Differential Association Product Term: (IR X DA) Mediator Hypothesis Individual Race Differential Association Moderator Hypothesis Individual Race Definitions Product Term: (IR X DA) Mediator Hypothesis Individual Race Definitions Moderator Hypothesis Individual Race Rewards Product Term: (IR X DA) Mediator Hypothesis Individual Race Rewards Moderator Hypothesis Individual Race Costs Product Term: (IR X DA) Mediator Hypothesis Individual Race Costs Individual Race (IR) and Costs (C) Hypotheses Delinquency Delinquency Delinquency Individual Race (IR) and Rewards (R) Hypotheses Delinquency Delinquency Delinquency Figure 13 Path Diagram for SSSL Dimension II (Individual Race) Hypotheses Individual Race (IR) and Differential Associations (DA) Hypotheses Individual Race (IR) and Definitions (D) Hypotheses Delinquency Delinquency
135 Moderator Hypothesis Individual Age Differential Association Product Term: (IA X DA) Mediator Hypothesis Individual Age Differential Association Moderator Hypothesis Individual Age Definitions Product Term: (IA X DA) Mediator Hypothesis Individual Age Definitions Moderator Hypothesis Individual Age Rewards Product Term: (IA X DA) Mediator Hypothesis Individual Age Rewards Moderator Hypothesis Individual Age Costs Product Term: (IA X DA) Mediator Hypothesis Individual Age Costs Individual Age (IA) and Definitions (D) Hypotheses Delinquency Delinquency Delinquency Figure 14 Path Diagram for SSSL Dimension II (Individual Age) Hypotheses Individual Age (IA) and Differential Associations (DA) Hypotheses Delinquency Individual Age (IA) and Rewards (R) Hypotheses Delinquency Delinquency Individual Age (IA) and Costs (C) Hypotheses Delinquency Delinquency
136 Path Diagram for the Social Structure-Social Learning Dimension II Figure 15 Hypothesis that Social Learning Mediates the Effect of Differential Location in the Social Structure on Delinquency Delinquency Location in Social Structure Social Learning
137 Moderator Hypothesis SES Differential Association Product Term: (SES X DA)Mediator Hypothesis SES Differential Association Moderator Hypothesis SES Definitions Product Term: (SES X DA)Mediator Hypothesis SES Definitions Moderator Hypothesis SES Rewards Product Term: (SES X DA)Mediator Hypothesis SES Rewards Moderator Hypothesis SES Costs Product Term: (SES X DA)Mediator Hypothesis SES Costs SES (SES) and Costs (C) Hypotheses Delinquency Delinquency Delinquency SES (SES) and Rewards (R) Hypotheses Delinquency Delinquency Delinquency Figure 16 Path Diagram for SSSL Dimension III (SES) Hypotheses SES (SES) and Differential Associations (DA) Hypotheses SES (SES) and Definitions (D) Hypotheses Delinquency Delinquency
138 Moderator Hypothesis Ethnic Heterogeneity Differential Association Product Term: (EH X DA) Mediator Hypothesis Ethnic Heterogeneity Differential Association Moderator Hypothesis Ethnic Heterogeneity Definitions Product Term: (EH X DA) Mediator Hypothesis Ethnic Heterogeneity Definitions Moderator Hypothesis Ethnic Heterogeneity Rewards Product Term: (EH X DA) Mediator Hypothesis Ethnic Heterogeneity Rewards Moderator Hypothesis Ethnic Heterogeneity Costs Product Term: (EH X DA) Mediator Hypothesis Ethnic Heterogeneity Costs Ethnic Heterogeneity (EH) and Definitions (D) Hypotheses Delinquency Delinquency Delinquency Figure 17 Path Diagram for SSSL Dimension III (Ethnic Heterogeneity) Hypotheses Ethnic Heterogeneity (EH) and Differential Associations (DA) Hypotheses Delinquency Ethnic Heterogeneity (EH) and Rewards (R) Hypotheses Delinquency Delinquency Ethnic Heterogeneity (EH) and Costs (C) Hypotheses Delinquency Delinquency
139 Moderator Hypothesis Residential Mobility Differential Association Product Term: (RM X DA) Mediator Hypothesis Residential Mobility Differential Association Moderator Hypothesis Residential Mobility Definitions Product Term: (RM X DA) Mediator Hypothesis Residential Mobility Definitions Moderator Hypothesis Residential Mobility Rewards Product Term: (RM X DA) Mediator Hypothesis Residential Mobility Rewards Moderator Hypothesis Residential Mobility Costs Product Term: (RM X DA) Mediator Hypothesis Residential Mobility Costs Residential Mobility (RM) and Costs (C) Hypotheses Delinquency Delinquency Delinquency Residential Mobility (RM) and Rewards (R) Hypotheses Delinquency Delinquency Delinquency Figure 18 Path Diagram for SSSL Dimension III (Residential Mobility) Hypotheses Residential Mobility (RM) and Differe ntial Associations (DA) Hypotheses Residential Mobility (RM) a nd Definitions (D) Hypotheses Delinquency Delinquency
140 Moderator Hypothesis Family Disruption Differential Association Product Term: (FD X DA) Mediator Hypothesis Family Disruption Differential Association Moderator Hypothesis Family Disruption Definitions Product Term: (FD X DA) Mediator Hypothesis Family Disruption Definitions Moderator Hypothesis Family Disruption Rewards Product Term: (FD X DA) Mediator Hypothesis Family Disruption Rewards Moderator Hypothesis Family Disruption Costs Product Term: (FD X DA) Mediator Hypothesis Family Disruption Costs Family Disruption (FD) and Definitions (D) Hypotheses Delinquency Delinquency Delinquency Figure 19 Path Diagram for SSSL Dimension III (Family Disruption) Hypotheses Family Disruption (FD) and Differential Associations (DA) Hypotheses Delinquency Family Disruption (FD) and Rewards (R) Hypotheses Delinquency Delinquency Family Disruption (FD) and Costs (C) Hypotheses Delinquency Delinquency
141 Path Diagram for the Social Structure-Social Learning Dimension III Figure 20 Hypothesis that Social Learning Medi ates the Effect of Theoretically Defined Structural Causes on Delinquency Delinquency Theoretically Defined Structural Causes Social Learning Figure 21 Path Diagram for the Social Structure-Social Learning Hypothesis that Social Learning Mediates the Effect of Social Structure on Delinquenc y Theoretically Defined Structural Causes Social Learning Delinquency Location in Social Structure Social Organization
142 Chapter Five Research Design and Analytic Strategy Sample The present research conducts anal yses of microsocial data obtained from an existing dataset, merged with macros ocial data. The individual-level data for this study come from a 1998 cross-se ctional survey of Largo, Florida high school and middle school students (see War eham, Cochran, Dembo, & Sellers, 2005). Largo is a metropolitan area comprising 15.41 square miles in west central Florida. Its population during the 1990s was around 69,000 people: 47% male, 92% White, 9% foreign-born, 20% never married, and 16% aged younger than 18 years (U.S. Census Bureau, 1990, 2000). Roughly 6% of LargoÂ’s families had income below the poverty level, and the cityÂ’s 1998 median adjusted household income was $42,000 (Largo Chamber of Co mmerce, 1998; U.S. Census Bureau, 1990, 2000). The 1998 City of Largo official crime rate (per 100,000) was 5,019: 3 murders, 24 forcible rapes, 65 r obberies, 347 aggravated assaults, 642 burglaries, 2,159 larcenies, and 185 motor vehicle thefts (Florida Department of Law Enforcement, 1999). The Largo public high school, one of seve ral high schools in the area, had
143 1,948 enrolled students (gr ades 9-12) during the 19981999 school year, with an average class size of 31 students. There were 150 school-related reports of crime or violence that year: 18 violent ac ts against people; 25 incidents of fighting or harassment; 9 possession of weapon in cidents; 3 incidents of property damage; 83 alcohol, tobacco, and other drug incidents; and 12 other nonviolent or disorderly incidents (Florida Department of Education, 2003). The Largo middle school, one of two area middle schools, had 1,294 enrolled students (grades 6-8) duri ng the 1998-1999 school year, with an average class size of 25 st udents. There were 61 school-re lated reports of crime or violence that year: 18 violent acts against people; 6 incidents of fighting or harassment; 10 possession of weapon incident s; 4 incidents of property damage; 13 alcohol, tobacco, and other drug inci dents; and 10 other nonviolent or disorderly incidents (Florida Depa rtment of Education, 2003). In December 1998, students from a rando m sample of 30 third-period high school classes and all middle school Soci al Studies classes completed a 239item questionnaire (see Wareham et al., 2005). The study employed passive parental consent procedures that were approved by the university Institutional Review Board (IRB). All survey in formation was anonymous, and researchers kept the street intersection nearest to the respondentÂ’s home address (asked in order to link the respondent to a C ensus block group) confidential. Although researchers adv ised students that parti cipation was voluntary
144 (Wareham et al., 2005), consistent with t he tenets of inform ed consent (see APA, 1992; D. Smith, 2003), passive parent al consent for juveniles has been controversial. In active parental consen t, parents receive written notification of the study and signify permission for the incl usion of their child in writing. With passive parental consent, researchers in form parents of the intended research, and interpret a lack of objection as permissi on to include the child in the study (Pokorny, Jason, Schoeny, Townsend & Curie, 2001). Researchers use informed consent pr ocedures to ensure that individual participation is voluntary (D. Smith, 2003). Legal and ethical considerations generally require parental permission to include juveniles in research (APA, 1992; D. Smith, 2003), but participation fr om active parental consent is often lower than that of passive parental consent (Pokorny et al., 2001), so researchers simultaneously consider se lection bias (see Anderman, Cheadle, Curry, Diehr, Shultz & Wagner, 1995). In the Largo study, however, the re searchers were especially concerned with the ethical considerat ion of confidentiality. T he Largo police department funded the research with a Community Or iented Policing grant (see Wareham et al., 2005). As the researchers solici ted sensitive information from the respondents such as involvement in illegal behavior and the intersection of streets closest to their residence, the researchers decided, and the IRB concurred, that passive parental consent best protected the identity and privacy
145 of the respondents. The researchers did no t want the police department to have access to the names, block groups, and se lf-reported illicit behaviors of the study respondents. On the day of survey administration, a researcher described the purpose of the study, explained that participati on was voluntary, and remained available to answer questions (Wareham et al, 2005). The survey response rate was 79% ( N = 625) for the high school and 81% ( N =1,049) for the middle school. The community-level data for the pres ent study come from the 2000 U.S. Census of population and Housing Summary File 3, aggregated at the Pinellas County block group level (U.S. Census Bureau, 2000), and from information collected in the Largo survey. The pr esent study adopts the approach of including block-groups for which at least one respondent resided (see D. Gottfredson et al., 1991; see also, Rount ree, Land & Miethe, 1994; Sampson et al., 1997). The Census 2000 aggregates reporting areas hierarchically. A census tract is a geographic statistical subdivision of a county. Tracts average about 4,000 people and the Census Bureau in tends tracts to be relatively homogeneous across population, economic status, and living condition characteristics. The Census Bureau defines tracts with input from local officials, and they characterize a tract as repr esenting a neighborhood. Census 2000 was the first decennial census that covered t he entire country by tract (U.S. Census
146 Bureau, 2000). Census blocks are smaller aggregates in area, such as a block bounded by city streets, and they average about 85 people (Myers, 1992). The Census 2000 identifies blocks through a four-digit numbering system, one different than that used in previous censuses (U.S. Census Bureau, 2000). A block group is a cluster of census blocks whose number begins with the same first digit as other blocks within t he tract. Census block groups typically contain between 600 and 3, 000 people depending on the urbanicity of the measured area, with an ideal size of 1, 500 people (U.S. Census Bureau, 2000). In the Census 2000, blocks nest wit hin block groups, which nest within census tracts, which nest within counties of the 50 states and the District of Columbia. Before the state level, the C ensus 2000 subdivides the United States first into four regions and then into nine divisions. Although the census collects information from blocks, the smallest geographic subdivision for which the Census Bureau publicly reports, the block group is the lowest level of aggregated data provided in summary file 3 (U.S. Census Bureau, 2000). The U.S. Census Bureau divides repor ting areas hierarchically, and it treats the detail of information similarly. The Census Bureau typically reports broader characteristics for the political and statistical subdivisions that are closer to the top of the reporting hierarchy (Myers, 1992). Summa ry file 3 details social, economic, and housing characteristics (e.g ., marital status, 1999 income, year
147 moved into residence) from a generally 1 in 6 sample (long-form) of roughly 19 million housing units, as well as 100 percent (short-form) char acteristics (e.g., household relationship, sex, age, race). There is no sampling error associ ated with the 100-percent data (U.S. Census Bureau, 2000). There is sampling error associated with the short-form data collection method, however, as the Census 2000 asks a portion of the population more questions than it does the entire population. A fter collecting all data, the Census Bureau weights the sa mple responses upward so that they estimate the responses of the census population (U.S. Census Bureau, 2000; see Myers, 1992). Sampling error varies across Census 2000 tables, but many researchers consider the error ignorable (Myers, 1992). The present studyÂ’s merged sample size ( N = 1,674) first decreased during the coding process that linked respondents to a census block group. Students provided the street names of intersections nearest where they lived. The response rate was 83.6% ( N = 1400). Researchers geocoded usable responses ( N = 1,188) and assigned them a 2000 Census identification number (Wareham et al., 2005). The sample further decreased for the present analysis during listwise deletion (the method preferr ed in SEM analysis; Kline, 1998; see also discussion in D. Kaplan, 2000) to account for missing questionnaire responses ( N = 1062). The resultant sample size meets rules of thumb in the literat ure suggesting that
148 SEM analyses employ samples of at leas t 200 cases when there are ten or more variables (Loehlin, 1992), at least 15 cases for each measured variable or indicator (Stevens, 2002), or at least 5 cases for each parameter estimator including error terms and path coe fficients (Bentler & Chou, 1987). One way researchers deal with missing cases is to impute values for missing data. The idea is that missing dat a may bias the sample, and estimating the value of the absent responses allo ws analysis to continue as if the information were complete (Brick & Kalton, 1986). Although the approach may reduce sample bias (Kalton & Kasprzyk, 1986), researchers do not recommend imputation with path model ing because the substituted means may distort variance and covariance information (see Brick & Kalton, 1986; Kalton & Kasprzyk, 1986), a key component to structural equation modeling. In the present research, the number of missing cases ( n = 126) exceeds the 5% rule of thumb re searchers generally use to assume randomness (Kalton & Kasprzyk, 1986; Kline, 1998; Tabachnick & Fidell, 2001). If data are missing completely at random, the sample remain s unbiased. The individual sample, the census-coded sample, and the sample under analysis compare, however, on demographic characteristics (see Table 1), and t -tests showed no statistical differences among their means ( p >.05).
149 The present study considers the re sponses not included in the sample under analysis as missing completely at ra ndom and therefore ig norable (see P. Allison, 2001; Kalton & Kasprzyk, 1986; Kline, 1998; Tabachnick & Fidell, 2001). Respondents in the Largo sample are 47% male, 80% White, and they average 14 years of age. Measures Dependent variable. Self-reported delinquency is the dependent variable. Its measurement is consistent with that reported in the literature (see Akers et al., 1979; Elliott et al., 1979; Elliott et al., 1985; Farrington, Loeber, Stouthamer-Loeber, Van Kammen & Schmidt 1996; Huizinga & Elliott, 1986; Piquero, MacIntosh & Hickman, 2002; Regnerus, 2002), given the constraints of secondary data analysis (Riedel, 2000). The present studyÂ’s SEM analyses in terpret self-reported delinquency as InitialCensusFinalInitialCensusFinalInitialCensusFinal Sample Coded Sample Sample Coded Sample Sample Coded Sample Mean1.501.481.422.214.171.1243.7913.8313.87 SD .126.96.36.199.40.401.991.981.97 N 166211821062161711561062165211781062 (2 = Male)(2 = nonWhite) RaceAge Sex (in years) Table 1 Missing Values Analysis
150 a latent construct with one indicator, whereas the correlation and OLS regression analyses characterize the variable as a summed index. The questionnaire asked, 1) Â“Have you ever skipped classes without an excuse?Â” 2) Â“Have you ever stolen things worth $50 or less?Â” 3) Â“Have you ever stolen something worth more than $50?Â” 4) Â“Have you ever hit someone with the idea of hurting them?Â” 5) Â“Have you ever attacked someone with a weapon?Â” 6) Â“Have you ever used marijuana?Â” Respondents chose one of three respons es: no, never; yes, but the last time was more than a year ago; and yes, in the past 12 months. Respondents that reported delinquency in the previous year furt her marked the number of instances. The study equates observations more frequent than once weekly (52 or more instances) to eliminate unnecessa ry outliers, creating a linear composite (0-312). As intuitively obvious from t he distribution of frequencies in Table 2, however, normality indices suggest the possi bility of skew (4.77) and kurtosis (32.84).
151 Statistical analyses for this research assume normality. Skew and kurtosis are absent when their indices equal zero, and a rule of thumb is there may be cause for concern when skewness is greater than 2 and kurtosis is greater than 7 (Curran, West & Finch, 1996; Muthen & Kaplan, 1992), though kurtosis is usually the most problematic for variance and covariance techniques that assume a multivariate normal distribution (Browne, 1984; Finch, West & MacKinnon, 1997; DeCarlo, 1997; Mardia, Kent & Bibby, 1979). A nonnormal distribution may result in bi ased correlation coefficients that may affect interpretation of the null hypothesis (Hatcher, 1994; West, Finch & Curran, 1995). Positive skew such as that which may be present in these data 053750.56 18158.19 25263.09 34667.42 44271.37 54775.80 6-108083.33 11-205888.79 21-303892.37 31-401393.60 41-521895.29 53-1044399.34 105-2347100.00 Table 2 Frequency Distribution and Cumulative Percentages for Self-Reported Delinquency (N = 1121) Delinquency Count Frequency Cumulative Percent
152 produces negatively biased estimator standard errors that may re sult in a lack of statistical power and an erroneous accept ance of the null hypothesis (Hatcher, 1994; Jaccard & Wan, 1996; West et al., 1995). Although the literat ure provides guidance in testing for multivariate normality in SEM (e.g., West et al., 1995), some researchers suggest that univariate normality is a necessary but not sufficient requirement for multivariate normality (Jaccard & Wan, 1996). Some researchers further suggest that univariate skew and kurtosis must be less than the absolute value of 1 to assure multivariate normality (D. Kaplan, 2000). Oth ers suggest that such a strategy is too conservative (Jaccard & Wan, 1996). Instead, some researchers address nonnormality through the consideration of statistica l tests that do not assume normality. For example, the self-reported delinquency va riable represents the numbe r of times a respondent committed a specific delinquent act in t he previous year. The responses range from zero to 234. Although researchers ty pically treat such data as continuous, as they view such questions as indexi ng a continuous measure of involvement in crime or delinquency (e.g., Hoffmann, 2002) potential responses must be above zero, and in this study, they are capped at 312. Zero is the most frequent response (52%), and high counts of self -reported delinquency are somewhat rare in these data (17% > 11). Accordingly, so me researchers might view statistical techniques designed for count data as appropriate.
153 The notion of count data refers to t he number of times an event occurs. Rather than a continuous response, a count is always a non-negative discrete number (e.g., 0, 1, 2, 3, etcÂ…). This type of response variable is common in event history analysis (DeMaris, 2004). Ev ent count observations comprise a fixed domain (King, 1988) that can be tem poral or spatial (DeMaris, 2004). For example, the delinquency responses in the present study em body the event of delinquency and the domain of one year. Res earchers might reasonably consider the respondentÂ’s self repor ted delinquency during the previous year an event count. OLS regression, along with SEM, relie s on the assumption of a normal distribution, and count data may violate t hat assumption; particularly when zero responses are overrepresented and high in tegers are rare. Some researchers (Cameron & Trivedi, 1998; DeMaris, 2004; Gardener, Mulvey & Shaw, 1995), including criminologists (Osgood, 2000), s uggest that OLS regression models are inappropriate for count data. OLS regression assumes a normal distribution, and a large positive skew may violate that assumption. Instead, some researchers (Camer on & Trivedi, 1998; DeMaris, 2004; Gardener et al., 1995; Osgood, 2000) advocate Poisson-based regression analyses, as the Poisson distribution does not assume normality. The Poisson distributionÂ’s variance is equal to it s mean, however, and overdispersed (variance exceeding its mean) data such as those in the present study, although
154 not violating Poisson assumptions of a skewed non-negative distribution, do violate the PoissonÂ’s equidispersion pr operty (see DeMaris, 2004; Long, 1997). Still Poisson-based, researchers may tu rn to negative binom ial regression or zero modified models when equidispersion is violated as they allow a variance greater than the mean (C ameron & Trivedi, 1998; De Maris, 2004; Long, 1997; Gardener et al., 1995; Osgood, 2000). OLS regression is not the main analytic al technique in the present study, however. The present study uses path analysis and SEM to examine possible mediation effects of soci al learning on social st ructure and delinquency as hypothesized by Akers (1998). SEM is a cross-level alternative to OLS regression when both direct and indire ct effects are of interest. Poisson regression is an alternative to OLS regression when assumptions of normality are doubtful. Binomial regressi on, along with various zero modified models, is an alternative to Poisson r egression when the conditional variance is greater than the conditional mean. Much as researchers use alternative analytic techniques with nonnormal regression distri butions, researchers likewise make use of multi-level tools that relax normality assumptions. Researchers use hierarchical gener alized linear models (HGLM), for example, as an alternative to HLM for binar y, multinomial, ordinal, and count data (Raudenbush, Bryk, Cheong & Congdon, 2001). However, Raudenbush and colleagues note that for most nonnormal data, a simple transformation suitably
155 norms the distribution and that researchers typically do not have to resort to a generalized multi-level model. Land and coll eagues (1990), as well as Jaccard & Wan (1996), likewise note that researc hers may appropriately transform either independent or dependent variables for reasons of linearity. Researchers have used generalized esti mating equations (GEE) to model count data in SEM (Zeger & Liang, 1986), but the technique is complicated, only produces quasi-likelihood resu lts, and it does not derive correlation structures. The approach instead focuses on mean stru cture, and it atte mpts a Â“workingÂ” correlation matrix (Skrondal & Rabe-Heske th, 2004). Researchers alternatively tend to use weighted least squares (WLS ), an asymptotically distribution free estimator (Browne, 1984), as alternat ives to maximum likelihood (ML) or generalized least squares (GLS) estimations (see Bollen, 1989) when assumptions of normality are not met. Much like zero modified models acc ount for the overrepresentation of zeros predicted by negative binomial regr ession by modeling the predicted zeros (Long, 1997), WLS accounts for nonnormality by weighting covariance matrices. Although the technique produces unbiased parameter estimates, standard error estimates, and chi-square goodn ess-of-fit estimates in large samples, it is computationally demanding (West et al., 1995). Olsson, Foss, Troye, and Howell (2000) conducted a simulation study derived from recommendations in the lit erature to use WLS for nonnomally
156 distributed data, contrast ing it with ML and GLS estimation methods. They modeled 11 conditions of kurtosis (ranging fr om Â–1.2 to +25.45, mild to severe), 4 models (3 containing misspecification), and 5 sample sizes. Olsson and colleagues (2000) concluded, The results can be summarized as follows: The performance in terms of empirical and theoretical fi t of the three estimation methods is differentially affected by sample size, specificati on error, and kurtosis. Of these three methods, ML is considerably more insensitive than the other two variations in sample size and kurtosis Only empirical fit is affected by specification errorÂ—as it should be. Moreover, ML tends in general not only to be more stable, but also demons trates higher accuracy in terms of empirical and theoretical fit compar ed to the other estimators. (pp. 577578) Olsson and colleagueÂ’s (2000) findings are consistent with Lei and Lomax (2005), who specifically tested the effects of SEM nonnormality through simulation and concluded, Â“nonnormality condi tions have almost no effect on the standard errors of parameter estimates regardless of the sample size and estimation methodsÂ” (p. 16). Although other researchers have likewise concluded that the assumption of SEM normality is robust in its estimation of parameters (Fan & Wang, 1998), Lei and Lomax (2005) fu rther sought identification of the more robust goodness-of-fit indices. T hey concluded that nonnormality should not prevent researchers from interpreti ng parameter estimates as usual, and that the normed fit index (NFI), the non-norm ed-fit index (NNFI), and the comparative fit index (CFI) are more appropriate indexes than the chi-square test statistic. West and colleagues (1995) likewise s uggest that SEM is robust to SEM
157 violations of normality, and they further argue that SEM is robust to scaling assumptions. West and colleagues obs erve that although SEM assumes continuous variables with a multivariate normal distribution, r eal data often do not satisfy the assumptions. They cite measur es of the amount of substance use as an example. To address potential mu ltivariate nonnormality, West and colleagues recommend linear data tr ansformation. Transformation preserves the order of observations and the broad meaning of a variable, but it alters t he distance between observations (West et al., 1995), thus stabilizing its variance (Stone & Hollenbeck, 1989). Transformation is possible when a variableÂ’ s scale has no inherent meaning, and the point is to reexpress variables so that their distribution looks like a normal distribution (Jaccard & Wan, 1996). So me researchers recommend transforming all variables to remedy normality, unless doing so would hinder interpretation, as transformations generally improve re sults (Tabachnick & Fidell, 2001). The transformation suggested by moderat e to substantial positive skew is a logarithm (log10; Tabachnick & Fidell, 2001). Only positive numbers can have a logarithm, and as the present research dependent variable contained zeros, the constant .50 was added to each value before the log10 transformation (see Tabachnick & Fidell, 2001; West et al., 1995). Transform ing the study dependent variable dramatically reduced univariate skewness (.84) and kurtosis (-.507), bringing both indexes under Curran and colleagues (1996) and Muthen and
158 KaplanÂ’s (1992) rule of thumb, thus allowing im proved evaluation of the distribution. The present research assessed the cons truct validity of the theoretically reasoned delinquency scale through princi pal-components analysis, using the eigenvalue-one criterion for prior comm unality estimates (Kaiser, 1960; see Hatcher, 1994; Mulaik, 1987; Stevens, 2002). The Kaiser criterion suggests that there is only one dimension present am ongst variables when the eigenvalue (its contribution to the variance) is lower than 1.00 (Hatcher, 1994). The goal was to assess whether the six variables r epresented one underlying dimension (see Tinsley & Tinsley, 1987); to see if they measure what they purport to measure (Farrington et al., 1996; Huizin ga, Esbensen & Weiher, 1991). The methodological literatu re reports two approaches, principalcomponents (uses a correlation matrix diagonal) and common factor (estimates reliability through an iterative process) analysis. There is no consensus as to which approach is more appropriate under what circumstances (see Comrey, 1978; Ford, MacCallum, and Tait, 1986; St ewart, 1981; Tinsley & Tinsley, 1987), but Snook and Gorsuch (1989) conducted a simulation study and found that both methods yield similar results as the num ber of items increase. In an exhaustive literature review, Guadagnol i and Velicer (1988) likew ise found no substantive differences in drawn conclusions between the two techniques, and Thompson and Daniel (1996) further concluded that either factor analysis approach is
159 suitable as long as the researcher reports the utilized technique. R.A. Peterson (2000) reported meta-anal ytic results, indicating that in addition to which technique to use, there is also no consensus on what constitutes a low or high factor loadi ng or how much explained variance is acceptable. He found, however, that m any researchers judge factor loadings similar to that explained by Hair Anderson, Tatham, and Black (1998): .30, minimally acceptable; .40 and larger, important; .50 and larger, practically significant. R.A. Peterson indicated that in his study, the average factor loading was .32 and the average explained va riance was 56.6%. R.A. Peterson concluded, in concurrence with Thompson and Daniel (1996), that regardless of which variable variance is analyzed, uni ties in principal-components analysis and communality in common factor analysis, neither differs on derived substantive conclusions. In the present study, analys is of the six variables used to construct the delinquency scale suggests that there is one underlying construct (eigenvalue = 2.42). Each of the variables loaded in t he practically significant range (Hair et al., 1998), higher than .50, (skip class = .61, stolen < $50 = .69, stolen > $50 = .67, hit = .60, weapon = .62, marijuana = .62), accounting for 40.44% of the variance. Microsocial independent variables. The individual-level independent variabl es comprise measures of each of the social learning concept s except imitation, whic h the questionnaire did not
160 index. Analysis of the variables used to construct the scales revealed that the skewness and kurtosis index for each variab le satisfies the adopted rule of thumb for univariate normality (skewness < 2; kurtosis < 7). The study assesses internal consist ency of the scales through CronbachÂ’s (1951) coefficient alpha ( ). Coefficient alpha seeks to assess research generalizability by evaluatin g whether measures are re liable; whether repeated measures yield similar results (Nunnal ly, 1978). CronbachÂ’s alpha is a widely used and accepted scale-construction relia bility statistic, with researchers generally accepting a scaleÂ’s reliability when > .70 (Nunnally, 1978; see Hatcher, 1994). Cortina (1993) warns, how ever, that CronbachÂ’s alpha can only confirm unidimensionality after unidimens ionality has been established, and it should be used in conjunction with prin cipal-components or common factor analysis. Differential associations is measured similar to that of Akers and colleagues (1979) and Elliott and colleagues (1985). The index is a 4-item summated scale of the number of res pondent friends who have skipped school, stolen something worth $50 or less, hit someone with the idea of hurting them, or used marijuana (see Table 3 following this section). Unidimensionality analyses for the scale suggested one underlying construct (eigenvalue = 2.46; = .78). The variables loaded in the practically sign ificant range (skip class = .83, steal = .80, fight = .72, marij uana = .80), accounting for 61. 55% of the variance.
161 Definitions is an 8-item summated scale comprised of four questions asking whether the respondent ag reed it is okay to skip school, steal little things, get into a fight, and use marijuana under certain conditions, and four questions asking the respondent if they would feel any guilt if they engaged in the described behaviors (see Table 4 following this section). The techniques of neutralization measures derive from Sykes and Matza (1957) and Akers and colleagues (1979). The guilt measures derive from Winfree and Bernat (1998). The scale measures loaded on one dimension (eigenvalue = 4.09; = .86), with each variable in the practically signific ant range (skip class neutralization = .71, steal neutralization = .63, fight neutraliz ation = .60, marij uana neutralization = .75, skip class guilt = .78, steal guilt = .77, fight guilt = .68, ma rijuana guilt = .77), accounting for 51.14% of the variance. Two scales measure differential reinfo rcements, both derived from Akers and colleagues (1979). Rewards is 4-item summated scal e of the degree of fun the respondent would experience from ski pping school, stealing something worth $50 or less, hitting someone with the idea of hurting them, or using marijuana (see Table 5 following this section) The items loaded on one dimension (eigenvalue = 2.24; = .74), with each variable in t he practically significant range (skip class = .75, steal = 79, hit = .74, marijuana = .72) accounting for 56.06% of the variance. Costs is a 4-item summated scale of whether parents would lose respect
162 for the respondent skipping school, steali ng something worth $50 or less, hitting someone with the idea of hurting them, or using marijuana (see Table 6 following this section). The scale items loaded on one dimension (eigenvalue = 2.51; = .80), with each variable in the practically significant range (skip class = .82, steal = .83, hit = .75, marij uana = .77), accounting for 62.77% of the variance.
163 n % 30025.6 48941.7 1159.8 18415.7 867.3 1174100.0 75064.3 30726.3 554.7 342.9 211.8 1167100.0 56748.2 42436.0 74603 574.8 554.7 1177100.0 60551.7 27423.4 867.3 1099.3 978.3 1171100.0 5. All of them. "How Many of Your Current Friends Have:" 1. None of them. 2. A few of them. 3. Half of them. 4. Most of them. 2) Stolen something worth $50 or less? 3. Half of them. 4. Most of them. Table 3 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Differential Associations Index (Range 2-20) 5. All of them. Questions and Responses 1) Skipped school? 1. None of them. 2. A few of them. 3) Hit someone with the idea of hurting them? 1. None of them. 2. A few of them. 3. Half of them. 4. Most of them. 5. All of them. 4) Used marijuana? 1. None of them. 2. A few of them. 3. Half of them. 4. Most of them. 5. All of them.
164 n % 36130.8 37532.0 29925.5 13811.8 1173100.0 58750.0 33828.8 17615.0 726.1 1173100.0 26222.5 32227.6 41735.7 16614.2 1167100.0 69459.3 25021.4 13811.8 887.5 1170100.0 42135.7 25321.4 24720.9 25921.9 1180100.0 67257.2 25421.6 16514.0 847.1 1175100.0 35530.2 26822.8 23319.8 31827.1 1174100.0 58049.5 16213.8 15213.0 27823.7 1172100.0 2. Disagree 3. Agree 4. Strongly agree 5) How guilty would you feel if you skipped school? 1. Very guilty 2. Fairly guilty 3. A little guilty 4. Not very guilty at all 6) How guilty would you feel if you stole something worth $50 or less? 4. Strongly agree 4) It's okay to use marijuana since it's not really harmful. 1. Strongly disagree 3) It's okay to get into a physical fight with someone if they insult or hit you first. 1. Strongly disagree 2. Disagree 3. Agree Table 4 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Costs Index (Range 4-32) 4. Strongly agree Questions and Responses 1. Strongly disagree 2. Disagree 4. Strongly agree 1) It's okay to skip school if nothing important is going on in class. 1. Strongly disagree 2. Disagree 3. Agree 3. Agree 2) It's okay to steal little things from a store since they make so much money it wont hurt them. 1. Very guilty 2. Fairly guilty 3. A little guilty 4. Not very guilty at all 7) How guilty would you feel if you hit someone with the idea of hurting them? 1. Very guilty 2. Fairly guilty 3. A little guilty 3. A little guilty 4. Not very guilty at all 4. Not very guilty at all 8) How guilty would you feel if you used marijuana? 1. Very guilty 2. Fairly guilty
165 n % 43637.1 30125.6 24821.1 19116.2 1176100.0 65555.7 25021.3 16414.0 1069.0 1175100.0 54546.4 26222.3 20417.4 16414.0 1175100.0 69659.3 16113.7 12810.9 18816.0 1173100.0 4. A lot 1) How much fun or Â‘kickÂ’ would you get if you got away with skipping school? 1. None at all 2. A little 3. Some 2) How much fun or Â‘kickÂ’ would you get if you got away with stealing something worth $50 or less? 3. Some Table 5 Frequency Distribution and Percentages for the Questionnaire Responses that Comprise the Rewards Index (Range 4-32) 4. A lot Questions and Responses 1. None at all 2. A little 3) How much fun or Â‘kickÂ’ would you get if you got away with hitting someone with the idea of hurting them? 1. None at all 2. A little 3. Some 2. A little 3. Some 4. A lot 4. A lot 4) How much fun or Â‘kickÂ’ would you get if you got away with using marijuana? 1. None at all
166 Macrosocial independent variables. The community-level independent vari ables comprise several measured variables or latent constructs (view ed as summated or averaged scales in correlation and OLS regression analyses) corresponding with three of AkersÂ’ (1998) four social structural dimens ions. The Largo questionnaire did not index the differential social location in pr imary, secondary, and reference groups dimension. n % 36130.8 37532.0 29925.5 13811.8 1173100.0 58750.0 33828.8 17615.0 726.1 1173100.0 26222.5 32227.6 41735.7 16614.2 1167100.0 69459.3 25021.4 13811.8 887.5 1170100.0 2. Probably would 3. Probably would not 4. Definitely would not 4. Definitely would not 4) Would your parents lose respect for you if you used marijuana? 1. Definitely would 3) Would your parents lose respect for you if you hit someone with the idea of hurting them? 1. Definitely would 2. Probably would 3. Probably would not Table 6 Frequency Distribution and Percentages for the Questionna ire Responses that Comprise the Costs Index (Range 4-32) 4. Definitely would not Questions and Responses 1. Definitely would 2. Probably would 4. Definitely would not 1) Would your parents lose respect for you if you skipped school? 1. Definitely would 2. Probably would 3. Probably would not 2) Would your parents lose respect for you if you stole something worth $50 or less? 3. Probably would not
167 In describing the differential social organization and theoretically defined structural causes dimensions, Akers ( 1998) noted that there is some conceptual overlap based on the way different res earchers view theoretical constructs. Although such is perhaps adequate concept ually, it presents the potential for multicollinearity when operat ionalizing and simultaneousl y modeling measures in each structural dimension. Recall that Land and colleagues (1990) concluded in part that the invariance of previously reported macros ocial covariates of homicide may have been influenced by multicollinearity am ong the structural variables. They recommended that future research use standard definitions for structural variables and consider multicollinearity among variables. Also, recall that the three macrosocia l constructs Pratt and Cullen (2005) found most efficacious in predicting cr ime could be conceptualized either as indicators of social disorganization or as a composite concentrated disadvantage measure. Lastly, recall that Pra tt and Cullen concluded that social disorganization and resource/economic depr ivation theories (both sharing some measures) found the most em pirical support, the only tw o theories of the seven evaluated that were found to be highly supported. The present research operationalizes m easures that indicate three of the four social structure-soci al learning dimensions by balancing AkersÂ’ (1998) theoretical descriptions, SampsonÂ’ s (1999) and KrohnÂ’s (1999) theoretical
168 concerns about the social structure-soci al learning model, Land and colleagueÂ’s (1990) methodological concerns for multicollinearity among macrosocial variables, in their case covariates of homicide rates, and Pratt and CullenÂ’s (2005) identification of important social st ructural covariates of crime generally, along with measurement specifications from Sampson and Groves (1989), D. Gottfredson and colleagues (1991), and S un and colleagues (2004). Univariate analysis of each variable suggested that each satisfied the rule of thumb for normality (skewness <2; kurtosis <7), ex cept for the race composition and ethnic heterogeneity measures, which did so after a log10 transformation. Five measures index the social struct ural correlates/differential social organization dimension. Population density measures the census block-group population divided by its squar e miles of land area. Aker s (1998) specifies this variable as indexing the di mension, and it further derives from Sampson and Raudenbush (1999), among others (e.g., Roncek & Maier, 1991; Warner and Pierce, 1993). Race composition measures the log10 proportion of census block-group residents who are Black (e.g., Liska et al., 1998; Sampson, 1986). As several proportions equaled zero, the constant 00001 was added to the variable before transformation, bringing the skewness and kurtosis indexes within range of the normality rule of thumb. Sex composition measures the proportion of c ensus block-group residents
169 who are male. This measure follows that of Glaser and Rice (1959). Age composition measures the proportion of census block-group residents aged 16-24 years. This measure is like wise consistent with Glaser and Rice (1959), among others (e.g., L. Cohen & Land, 1987; Land et al., 1990). Near poverty measures the proportion of ce nsus block-group residents aged 15 years and older with a ratio of inco me to poverty lower than 1.25 times the poverty threshold. The index measures relative ra ther than absolute poverty, in order to capture deprivation (e.g ., Brady, 2003; Gor don, 1972; Hagenaars, 1991). It taps that portion of the popul ation thought to be Â“underemployed.Â” Three measures index the differentia l location in social structure dimension. Individual sex measures the sex of the Largo survey respondents (2 = male). Individual race measures the race of the Largo survey respondents (2 = nonWhite). Individual age measures the age in y ears of the Largo survey respondents. Akers (1998) specifies each of these measures as indexing the dimension. Sex and age fu rther derive from Lee and colleagues (2004) and sex and race from Lanza-Kaduce and Capece (2003). Four measures index t he theoretically derived structural causes dimension. Each of the measures operationalizes Sa mpson and GrovesÂ’ (1989) conceptualization of the social disor ganization theory exogenous variables, as adapted to U.S. census data by Sun and colleagues (2004). The present study adopts the terminology of Sun and collea gues, and like their model, Sampson
170 and GrovesÂ’ concept of urbanization is hel d constant, as each of the sample census block-groups are located in an urban area. Although Sun and colleagues approximated Sampson and Grov esÂ’ measure of friendship ties, the Largo data did not capture such data. This is not pr oblematic to the present study, however. Sampson and Groves (1989) used thei r intervening variables to index social disorganization. AkersÂ’ (1998) social structure-social learning theory relies, as the operationalization of this dimension pertains to his theory, on the same types of exogenous variables used by Sampson and Groves. However, Akers advances a different intervening mechanism. Moreover, had measures of friendshi p ties been available in the Largo data, they would have most likely represented AkersÂ’ (1998) differential social location in primary, secondary, and refer ence groups dimension. That dimension is not modeled in this research; howeve r, Akers observes that the meso-level dimension indicators interplay with the microsocial learning variables closely. This research tests whether social lear ning variables mediate social structural variables, the effective, though not conceptual role that social ties play in the social disorganization model. The strict m easurement of the t heoretically derived dimension is not deemed weakened by the exclusion of the friendship ties measurement, or Sampson and GrovesÂ’ (1989) other two intervening measures. Socioeconomic status (SES) is a scale comprised of the mean z -scores of four indicators. Three measures deriv e from Sampson and Groves (1989): the
171 proportion of census block-group resi dents with an income greater than $20,000 (also used by Sun et al, 2004), the pr oportion of census block-group residents with professional jobs (also used by D. Gottfredson et al., 1991), and the proportion of census block-group residents that are college graduates (also used by Sun et al., 2004). The fourth measure, the proportion of census block-group residents that are employed, derives from Sun and colleagues (2004). Unidimensionality analyses for the scale suggested one underlying construct (eigenvalue = 2.60; = .81). The variables loaded in the practically significant range (income $20,000+ = .79, employed = .67, college graduates = .93, professional job = .82), accounting for 65.01% of the variance. Ethnic heterogeneity is a measure similar to that of BlauÂ’s (1977) index of intergroup relations. Researchers (e.g., Sampson & Groves, 1989; Sun et al., 2004) indexing racial heterogene ity use the Blau index as opposed to the percent of the population that is Black in order to examine spatial distributions that approximate segregation. Conceptually, BlauÂ’s (1977) measure a sks, what proporti on of the group would have to change residence in order to have an even distribution of groups in each neighborhood. Although the measure is able to capture more than one race, recent measures have been created that attempt to examine ethnicity. Moreover, recent measures give attention to relative diversity (taking the larger group into account), as opposed to absolute diversity (merely t he proportion of
172 each group). Ethnic heterogeneity is measured in this research through MalyÂ’s (2000) neighborhood diversity index (NDI). The spatial differentiation formula is NDI = .5( CW CBGW + CB CBGB + CH CBGH + CA CBGA ) The logic of the formula is such that census block-group (CBG) populations for White (W), Black (B), Hispanic (H), and Asian (A) are compared to the respective city (C) populations. The White, Black, and Asian categories only include those who did not additionally identify themselves as Hispanic. The index ranges from 0-1 and the higher the score, the more segregated, less diverse the neighborhood (Maly, 2000). Simila r to the race composition measure that indexes the differential social organization dimension, the ethnic heterogeneity measure represents its log10 transformation, satisfying the normality skewness and kurtosis rule of thumb. Residential mobility is measured similar to th at of Sun and colleagues (2004). It represents the proportion of cens us block-group residents who lived in a different home four years earlier. Lastly, family disruption is a scale comprised of the mean z-scores of two indicators. The proportion of census bl ock-group residents who are divorced or separated derives from Sampson and Gr oves (1989) and Sun and colleagues (2004). The proportion of female-headed hous eholds with children derives from D. Gottfredson and colleagues (1991), an es timation of the single parents with
173 children measure used by Sampson and Groves. Unidimensionality analyses for the scale suggested one underlying construct (eigenvalue = 1.39; = .52). The variables loaded in the practically signific ant range (divorced or separated = .83, female headed household with kids = .83), accounting for 69.34% of the variance. Table 7 summarizes the descriptive properties of all variables under analysis. Table 8 reports the inter-co rrelations among the variables. Although there are many significant inter-correlati ons, as is to be expe cted with variables such as poverty, race, and SES, as well as among the social learning variables, none of the coefficients exceeds .90 (the highest being -.82), a rule of thumb for redundancy (Tabachnick & Fidel, 2001). Moreover, those with the highest correlation coefficients tend to index differ ent social structure-social learning dimensions, an expectation explained by Akers (1998).
174 MinMax MSD 105.807729.273811.811446.95 -5.00-.02-2.071.34 0.360.540.470.03 0.000.240.080.03 0.010.650.140.08 1.002.001.470.50 1.002.001.200.40 11.0019.0013.871.97 -4.381.760.000.79 -2.10-.03-1.300.42 0.210.790.490.10 -2.053.570.000.82 4.0020.007.733.51 8.0032.0016.545.86 4.0016.007.763.25 4.0016.008.043.06 -.302.370.280.70Note. *log10 transformation **scores based on mean z -scoresDependent Costs Differential Associations Definitions Rewards Delinquency* SSSL III: Family Disruption** Intervening SSSL II: Individual Age SSSL III: SES** SSSL III: Ethnic Heterogeneity* SSSL III: Residential Mobility SSSL I: Age Composition (16-24) SSSL I: Near Poverty SSSL II: Individual Sex (Male) SSSL II: Individual Race (nonWhite) SSSL I: Race Composition (Black)* SSSL I: Sex Composition (Male) Table 7 Descriptive Statistics for Vari ables Under Analysis (N = 1062) Exogenous Variable SSSL I: Population Density
175 Procedure General issues and moderation. The present study tests a portion of AkersÂ’ (1998) so cial structure-social learning cross-level elaboration. The re search employs correlation, multiple regression, and SEM analyses. Researchers may not make statem ents about individual behavior from analysis of aggregate behavior. Doing so re sults in an ecological fallacy because the statistical properties of groups of peopl e do not substitute for the descriptive properties of its individuals (Robinson, 1950 ). Also, an atomistic or individualistic fallacy occurs when drawing inferences abo ut groups from examining individual behavior (Diez-Roux, 1998; Hannan, 1971, 1985; see the contextual fallacy 12345678910111213141516 Â—.34*.00 .25*.28*-.02 .01 -.07*-. 40*-.08*.13*.09*.01 .00 .04 .04 Â—.24*.47*.51*.01 .18*-.08*-.58*.36*.16*.49*-.02 -.02 .03 .08* Â—.32*.14*-.02 .07*-.03 -.02 .20*.19*.25*-.03 -.01 -.01 -.02 Â—.43*-.05 .20*-.03 -.36*.52*.08*.33*-.03 -.05 -.04 .04 Â—.00 .18*-.11*-.82*.61*.36*.75*-.04 -.01 .04 .05* Â—-.08*.04 -.02 -.01 -.02 .01 .12*.22*.14*.08* Â—-.03 -.22*.30*-.01.17*.01 .00 .06*.07* Â—.07*-.07*-.05*-.08*.24*.24*.01 .01 Â—-.46*-.37*-.71*-.00 -.03 -.08*-.08* Â—.05*.45*-.06*-.03 .01 .03 Â—.41*-.02 -.02 .01 .03 Â—-.02 .01 .04 .06* Â—.67*.51*.24* Â—.65*.37* Â—.25* Â— Note: p < .05 (one-tailed t -test) 14. Definitions 15. Rewards 16. Costs Table 8 Inter-correlations Among Explanatory Variables (N = 1062) Variable 3. SSSL I: Sex Composition 1. SSSL I: Population Density 2. SSSL I: Log10 Race Composition 8. SSSL II: Individual Age 12. SSSL III: Family Disruption 13. Differential Associations 6. SSSL II: Individual Sex 7. SSSL II: Individual Race 9. SSSL III: SES 10. SSSL III: Log10 Ethnic Heterogeneity 11. SSSL III: Resdiential Mobility 4. SSSL I: Age Composition 5. SSSL I: Near Poverty
176 discussion in Hauser, 1970). Researcher s may, however, examine social structure exogenous to individual behavior. Such an approach views aggregates as microsocial antecedents (B lalock, 1984; Diez-Roux, 2003). The implication of Robinson (1950) is that researchers may not examine the effects of social stru cture on crime rates and make inferences about criminal behavior. The implication of Hannon (1971, 1985) is that researchers may not examine the effects of social learning on criminal behavior and make inferences about crime rates. The imp lication of Blalock (1984) is that researchers may make inferences from the examination of the effects of social structure on criminal behavior. Akers (1998) may not provide suitable li nking propositions as to why social structure influences criminal behavior (e.g., Krohn, 1999), but Blalock (1984) provides the statistical justification to examine the relationship. Much like the confusion over whether a variable is a moderator or a medi ator (Saunders, 1956; Velicer, 1972; Zedeck, 1971), however, resear chers likewise tend to disagree on suitable test procedures (e.g., Arnold 1982, 1984; Baron & Kenny, 1984; Findley & Cooper, 1983; Harkins et al., 1980; Jaccard & Wan, 1995, 1996; Saunders, 1956; Stone & Hollenbeck, 1984, 1989). The methodological literature suggests five basic approaches (Bollen & Paxton, 1998; Jaccard & Wan, 1996; Joreskog & Yang, 1996; Klein & Moosbrugger, 2000; Ping, 1996), varying in their statistical sophistication and
177 agreement as to the statistical power of OLS regression models (see Baron & Kenny, 1986; Jaccard & Wan, 1995, 1996; Stone & Hollenbeck, 1989). The choice mainly rests between OLS regression models versus complicated SEM models that vary in their abili ty to account for the correlat ion of variable indicators with their multiplicative terms, as well as the degr ee to which they address (ignore; focal point) OLS regression power. The present research uses path analytic techniques to test AkersÂ’ (1998) assertion that that the so cial learning process mediates the effect of social structural variables on delinquency. The study is interested in testing AkersÂ’ assertion of mediation, but for the reasons described earlie r, it must first examine potential moderation. The present study adopts the noti on that SEM latent modeling is inappropriate for interactions without sophisticated variable construction corrections (Jaccard & Wan, 1996), and that the OLS methodology (the Figure 8 moderator hypothesis) suffi ciently addresses the question of moderation (see Baron & Kenny, 1986; Stone & Hollenbeck, 1989). Moreover, Jaccard & Wan (1996) note that OLS regressi on is a special case of st ructural equation modeling and that measuring an indicator with no e rror, such as through OLS regression, is effectively equivalent to constraini ng a SEM path to zero, thereby producing similar results. Likewise, Friedrich ( 1982) advocates OLS regression to test moderation. He systematically addressed each criticism of the approach in the
178 literature, and conclud ed that modeling conditi onal rather than general relationships is not complicated with OLS r egression, and that it provides a much better detailed depiction of the rela tionship between dependen t and independent variables. Mediation. After examining moderation, the pres ent research tests mediation. The analytic approach balances sophistica tion and parsimony to address the research question: How does the social l earning process interact with the effects of social structure on delinquency? Do di fferential associations, definitions, and differential reinforcement mediate social structureÂ’s effects? Do the social learning elements interact with social structure in some way that produces delinquency? Hierarchical social structures are common (Galtung, 1969; Lazarsfeld & Menzel, 1961), and as noted in the social l earning literature, individuals typically nest within various groups. Although res earchers have long understood the need for statistically separating group and individu al effects (Blau, 1960; Davis, Spaeth & Huson, 1961), there is little consensus on proper statistical techniques (see discussion in Bursik & Grasmick, 1996). Some previous tests of the social st ructure-social learning model have employed OLS regression. This proc edure pools individual and structural explanatory variables, regressing t he individual level dependent variable
179 simultaneously. Researchers assess cross-level effects by analyzing standardized coefficients (e.g., Lanza-Kaduce & Capece, 2003). However, OLS regression does no t adequately allow assessment of mediating effects. Because the method pools all of the variables, the linear, additive approach cannot discern causal terms, a requisite of mediation (James & Brett, 1984). Additionally, if the social learning mediator is measured with error, a likely occurrence, OLS regression may underestimate the effect of social learning and overestimate the effect of social structure, possibly overlooking successful mediation (see Baron & Kenny, 1986; Judd & Kenny, 1981). Likewise, the attenuated measures and overestimation of social structural effects may lead to incorrect conclusions that social stru cture causes social learning and social learning causes delinquency, the effe ct expected when mediation is present (Baron & Kenny, 1986). As such, when using OLS regression to assess a mediating effect, variable measurement error may result in a successful mediation going unnoticed, as well as conclusions that mediation exists when it does not. Type I error and Type II error are both possible concerns. OLS regression is not a suitable method for testing mediation (B aron & Kenny, 1986; James & Brett, 1984; Judd & Kenny, 1981). In addition, the pooled OLS regr ession approach ignores presumed multilevel methodological problems of nes ted data (Hox & Kreft, 1994). Ordinary significance tests assume explanatory variable independence. Tests that violate
180 the assumption, a possibility when using ne sted data, risk inflating Type II error. Suitable designs require analytic models t hat can handle two sources of variation (within and between), as well as unequa l group sizes. Further, suitable techniques must attend to effects that ar e random rather than fixed, and potential cross-level interaction (Hox & Kreft, 1994). Hierarchical linear modeling (HLM) is common in the psychological and educational field, whose researchers co mmonly use the technique to disentangle the cross-level effects of nested vari ablesÂ—to isolate individual effects independent of group effect s (Hox & Kreft, 1994). The technique handles unequal sample sizes, assumes intraclass correlation, rather than independent observations, and models random effects. Education researchers typically wish to assess the effects of a treatment tested in a classroom. However, res earchers interested in assessing the advantages of a particular assessment tool for example, must, when testing the effects, first account fo r classroom characteristi cs. Before assessing test differences (within), researchers account for classroom differences (between). Some researchers assuming cross level interaction (Bryk & Raudenbush, 1992) have applied the same reasoning to social problems generally (Hox & Kreft, 1994), as well as the examinati on of characteristics and crime (e.g., Hoffmann, 2002; Sampson et. al., 1997; R ountree et al., 1994; Silver & Miller, 2004; Wooldredge, 2002). To acc ount for the possibility that individual regression
181 residuals correlate with regression residuals within a neighborhood, HLM separates residual variance into two co mponents: individual-level variance and random neighborhood variance (Bryk & Raudenbus h, 1992). HLM tests statistical significance at both levels. Although HLM may be appropriate for examining the nested structure inherent to individuals and their neighbor hood, the more pr essing aim of the present study is to examine social learning as a mediator of social structure. The cross-level effect is the item of intere st. Moreover, some educ ational simulation studies found equally unbiased estimates between OLS regression and HLM (see Kreft, 1996). Researchers conduct simulation studies to compare the results from one statistical technique against another (C onway & McClain, 2003). In the case of the OLS regression versus HLM study, the author (Kreft, 1996) likely started with the question of whether HLM was necessary under certain circumstances. Researchers may conduct simulations with empirical data, or they may build a testable model with hypothetical data, test ing validity through any of a number of simulation software programs (Conway & McClain, 2003). The OLS regression versus HLM fi nding is important to educational researchers because if not for the possibi lity of unwanted structural influences, they would typically employ analysis of variance (ANOVA), or multivariate analysis of variance (MANOVA) to test t heir hypotheses, techniques that work
182 from a similar set of assumptions as OLS regression. Educational researchers such as those depicted in the example mainly wish to assess whether the exam procedure works, and HLM is merely a technique used to account for other explanations. Similarly, criminologists examining mu ltilevel problems might, if not for the possibility of assumption violations, use OLS regression. If HLM and OLS regression produce similarly unbiased estimate s, the researcher may not want to use the more sophisticated technique. As noted earlier, however, OLS r egression may be inappropriate for testing mediation. James and Brett (1984) suggest that researchers must use path analytic techniques to assess medi ation. Baron and Kenny (1986) likewise recommend path modeling to test mediati on, noting that the method allows simultaneous testing of all relevant paths. Structural equat ion modeling. Structural equation modeling (SEM) is a family of sophi sticated algebraic techniques that extends t he OLS regression methodology through the analysis of correlation matrices (Anderson & Gerbing, 1988; King & King, 19 97; Kline, 1998, 2005; McDonald & Ho, 2002; Raykov & Marcoulides, 2000). SEM uses the general linear model like OLS regression, but it has a more relaxed set of assumptions. SEM comprises path analysis models of observed variables, confirmatory
183 factor analysis models that examine the non-causal pattern of relationships among latent constructs, structural r egression models that specify causal relationships of regression constructs and latent change models that examine effects over time (Kline, 1998, 2005; Raykov & Marcoulides, 2000). Factor analysis comprises models of latent variabl es that have multiple indicators but no hypothesized direct effects between one another. Factor analysis models the correlation of latent variables (Raykov & Marcoulides, 2000). Researchers use path analysis to spec ify causal relationships and test theoretical models among mani fest (observed) variables (Hatcher, 1994; Kline, 1998, 2005; Raykov & Marcoulides, 2000). Path analysis tests hypothesized paths among variables, but like OLS regre ssion, it cannot estimate measurement error. Each path produces coefficients t hat equate to the par tial correlations calculated in OLS regression. Although the path analysis produces both raw and standardized coefficients, researchers typi cally report the standard (beta weights) scores (McDonald & Ho, 2002). Although SEM is an umbrella of tec hniques, researchers generally reserve the term SEM for models that examine the causal ordering of latent constructs, which use several manifest variables as indicators (Raykov & Marcoulides, 2000). The SEM approach allows resear chers to examine the underlying structure among variables (King & King, 1997) based on a proposed theoretical relationship (Raykov & Marcoulides, 2000). SEM tests models, not builds them.
184 Researchers typically represent manifest and latent variables visually in a path diagram with different symbols (Ra ykov & Marcoulides, 2000). As variables might simultaneously be the outcome of one variable and the predictor of another, both dependent and independent, res earchers instead refer to path analytic variables as exogenous and endogenous (Hatcher, 1994). Exogenous variables have no paths coming into them but paths going out. They are antecedent variables whose causes lay outside the model. Endogenous variables have at least one path coming in (consequent variable) and they may have paths going out (mediating variable). Figure 22 illustrates two mediating models: path analysis with manifest variables and path analysis with latent va riables. The path diagrams depict latent variables as oval, observed variables as rectangle, latent variable error (disturbance) as circles containing a Â“d,Â” measured variable error by an Â“e,Â” exogenous variable correlation by a twoarrowed curved connector, and path direction by a one-arrowed straight line (see Hatcher, 1984).
185 Referring back to the OLS regressi on versus HLM simulation studies, researchers have conducted similar anal yses comparing HLM with SEM. Julian e eeeeee eee eee Path Analysis with Manifest Variables Path Analysis with Latent Variables (SEM) Figure 22 Path Analysis Illustrations with Manifest and Latent Variables Predictor Factor Mediating Factor Indicator Variable Indicator Variable Indicator Variable Indicator Variable Indicator Variable d Predictor Variable Mediating Variable Outcome Variable Predictor Variable Predictor Factor Indicator Variable Indicator Variable Indicator Variable Indicator Variable Outcome Factor Indicator Variable Indicator Variable Indicator Variable d
186 (2001) employed simulation models to assess the consequences of using SEM instead of HLM with nested data. He starte d with the statistica l and logical crosslevel question of how best to discern the most appropriate level of analysis, given certain testable hypotheses. Working from an educational framew ork, Julian (2001) began with CronbachÂ’s (1976) argument that the hierarchical nature of educational data confounds individual assessment. Julian (2001) noted that multilevel SEM software exists, but that the technique is advanced and behavioral science researchers are not likely to be trained in assessing multilevel data structures. Julian suggested that researchers alte rnatively collect dat a with Â“conveniently organized groups of individualsÂ” (p. 330), and either overlook dependence among variables in order to examine the underlying structure, or conclude that any dependence is likely to impact the data minimally. Julian (2001) tested four di fferent group to member configurations (100/5, 50/10, 25/20, 10/50), maintaining a consistent sample size ( n =500). His models contained three varying intraclass correlations (.05., .15, .45), representing low, moderate, and high correlati on. Julian assessed the models with confirmatory factor analysis, and he concl uded that the low intracla ss correlation chi-square model fit statistic is relatively unb iased in SEM, along with parameter and standard error estimators. Julian was less enthusiastic about the implications when the intraclass correlations are above .05 or for decreasing group to
187 member ratios, suggesting that researchers consider alternative strategies under such conditions to avoid estimation problems. The implications of JulianÂ’s (2001) findings to the present study are unclear. Julian examined a simple data st ructure, designed to hypothetically examine the consequences of sampling groups of individuals to obtain a suitable size of individual responses for as low co st as possible, convenience, or some similarly minded rationale. JulianÂ’s group to individual ratios imply completely nested individuals, individual s only belonging to one group. Also, JulianÂ’s groups to members ratios may not generalize to the types of social situations under analysis in the present study, as the present study comprises relatively few social structures (neighborho ods) compared to the num ber of individuals. Further, although the chi-square test statistic may be the most popular SEM goodness-of-fit indicator (Lei & Lom ax, 2005), some researchers (Bentler & Bonett, 1980; Specht, 1975) question it as an appropriate measure of SEM empirical fit, and SAS PROC CALIS, for ex ample, offers more than 20 goodnessof-fit indices (SAS Institute, 1999). Ol sson and colleagues (2000) concluded from their simulation study that the maximu m likelihood SEM root mean square error of approximation (RMSEA) model fit index is relatively insensitive to sample size and kurtosis, and relatively stable with mi sspecification of a nested structure. Moreover, Wendorf (2002) found nearly i dentical results between SEM and HLM in an examination of matched-pairs (h ierarchical dyad). Lastly, Krull and
188 MacKinnon (2001) conducted a simulation study of SEM compared to a multilevel mediational model, and they r eported no bias in the estimators or standard error. In sum, researchers use SEM to model causal paths and test theoretical relationships among latent variables (Hatcher, 1994). SEM models generally have multiple indicators, though the tec hnique can handle single-item measures (modeled without error) as well. However, SEM models with m any single-item measures may have identific ation problems (Hatcher, 1994, Kline, 2005). In that case, some researchers suggest path modeling as an alternative (Kline, 2005). Path analysis falls under the umbrella of SEM, but the technique only models measured variables. AkersÂ’ (1998) social structure-social learning model presumes that the community-level characteristics have an effect on individual delinquency, but hypothesizes that individual learning subs tantially mediates its effect. AkersÂ’ question is both one of mediat ion and theory. Path analytic techniques are well suited to examining theoretical causal st ructures generally, as well as assessing the direct and indirect effects advanc ed by Akers (see Baron & Kenny, 1986; James & Brett, 1984; D. Kaplan, 2000; Muthen, 1989; Tabachnick & Fidell, 2001). Although the implicatio ns of using SEM instead of HLM when the possibility of cross-level interaction seem mixed in the methodological literature
189 (Julian, 2001; Krull & MacKinnon, 2001; Olsson et al., 2000, Wendorf, 2002; see generally Kreft, 1996; Kreft & de Leeuw, 19 98), structural equation modeling is more appropriate to testing hypot heses and assessing mediation than hierarchical linear modeling (see Hatc her, 1994; Raykov & Marcoulides, 2000). Further, one study in the literature has used SEM to assess the social structuresocial learning model (Lee et al., 2004). The present research adopts the notion that SEM is the most appropriate techni que to test AkersÂ’ (1998) theoretically derived mediation statement. A priori measures. Although selecting SEM over HLM as the most suitable procedure to test the theoretical question, the present res earch does not ignore the possibility of a nested structure. The study addresses the nested individuals possibility, the main reason for using HLM instead of SEM, by ex amining the possibility of interaction between the social structural and social learning variables. Toward that end, the present research adopts FriedrichÂ’s ( 1982) view, supported by Baron and Kenny (1986) and James and Brett (1984), that OLS regression suitably assesses moderation through the incorporation of a multiplicative term. The present study proceeds to SEM analyses after assessing the possibility of moderation. SEM is usually a confirmatory rather than exploratory procedure that consists of two steps: deriving a measurement model and validating the model (Ander son & Gerbing, 1984).
190 In SEM path analysis with latent va riables, the measurement model describes the nature of the relationship between a number of latent variables, or factors, and manifest indicator variables that measure those latent variables (Hatcher, 1994). At this stage, the goal is to use confi rmatory factor analysis to develop the measurement model. First, the present research tackl es Lee and colleaguesÂ’ (2004) little explained assertion that social learning is a construct comprising, in this study, differential associations, definitions, rewards, and costs. The theoretical implications were discussed earlier; this portion of the study tests its construct validity. Still part of establishing the measurement model, the present research next examines AkersÂ’ (1998) theoret ical model. The measurement model identifies the latent constructs and m anifest indicators, but does not specify causal paths: Each latent variable is allowed to correlate with one another (Hatcher, 1994). SEM is a system of functional equat ions, and model identification is important. An underidentifi ed estimation, including fe wer linearly independent equations than unknowns (Asher, 1988), result s in an infinite number of possible solutions and, therefore, meaningless re sults. A saturated or just-identified estimation, a model that contains exac tly as many linearly independent equations as unknowns, provides unique identifiers, but the model always fits perfectly thus
191 invalidation becomes impossible. Research ers using SEM seek an overidentified modelÂ—a model that includes more linearly independent equations than unknowns (Hatcher, 1994). The next measurement model step is to test the model with goodness of fit measures. Goodness of fit tests do not es tablish which paths in a model are significant, rather they assist res earchers in deciding whether the model generally should be accepted or rejected. As mentioned earlier, there are many such measures in the liter ature, yet there is little consensus on which ones are best. One common approach requires the research er to a priori identify several fit assessment measures that reflect diverse criteria (see Jaccard & Wan, 1996). The idea is to use enough measures to assi st in determining measure fit, yet not so many as to imply a Â“shotgun approac h.Â” Kline (1998) recommends that researchers use at least four tests. The present research addresses the possibility of nonnormal data affecting st atistical power by adopting Lei and Lomax (2005) and Olsson and colleaguesÂ’ ( 2000) specifications for assessing model fit. The study sets SteigerÂ’s (1990) root mean square error of approximation (RMSEA), Bentler and Bonet Â’s (1980) normed-fit index (NFI), Bentler and BonetÂ’s non-normed fit index (NNFI), and BentlerÂ’s (1989) comparative fit index (CFI) as a priori indicators of model fit. The maximum likelihood function used by SEM reflects the difference
192 between the observed covariance matrix and the one predicted by the model. Instead of a perfect fit, research ers more pragmatically seek an acceptable fit. The RMSEA compares the es timated model with a satura ted model. A perfect fit has a value of zero (Olsson et al, 2000) This research adopts Hu and BentlerÂ’s (1998, 1999) RMSEA cutoff value of .06 as suggesting a good fit. The NFI estimates fit by examining t he chi-square of the estimated model against the chi-square of an independent ( null) model. NFI values range from zero to one. This research adopts Hu and BentlerÂ’s (1998) conclusion that values > .90 indicate a good fit. The NNFI adjusts the NFI to account for the possibility of large sample sizes unduly influencing the results (Type I error). The NNFI evaluates the modelÂ’s degrees of freedom. The present research a priori adopts BentlerÂ’s (1989) conclusion that values > .90 represent a good fit. The CFI compares the predicted co variance matrix with the observed covariance matrix, and like the NNFI, it acco unts for sample size (Bentler, 1989). The CFI also ranges between zero and one. Many researchers use a cutoff for this measure of .90 (see Hatcher, 1994; Tabachnick & Fidell, 2001). Hu and Bentler (1998), aware of the convention, tested the measure in a simulation study and concluded that .95 is a more appropria te cutoff. This research adopts Hu and BentlerÂ’s (1998, 1999) notion that va lues > .95 suggest a good fit. If the goodness of fit indexes sugges t that the meas urement model
193 reasonably fits the data, t he study proceeds to the se cond step in the two-step Anderson and Gerbing (1988) approach, s pecifying the structural model. The present research uses an alpha of .05 fo r all statistical analyses: correlation, regression, and SEM. The research addresse s the possibility of partial mediation in two ways. First, recall that Akers (1998) suggests that varying degrees of mediation show varying degrees of support for the theory, but that substantial mediation shows the strongest support. Akers does not define substantial mediation, however, nor does t he methodological literature. MacKinnon, Lockwood, Ho ffman, West and Sheets ( 2002) note that Baron and Kenny (1986) set the standard for under standing the full implications of mediation and moderation, co mmenting that a check of the social sciences index showed that their article has been cited more than 2,000 times. Although Baron and Kenny allow that a Â“significant r eductionÂ” in the effects of an independent variable on a dependent variable when adding a new variable to a model demonstrates mediational potency, they do not address how much of a reduction is important. Shrout and Bolger (2002) addressed that issue by commenting that researchers may examine an effect rati o. The effect ratio is computed by summing the indirect effects (paths Â“aÂ” and Â“bÂ” in the Figure 5 mediation hypothesis) and dividing by the direct effe cts (path Â“cÂ”). T he present research incorporates the use of Shrout and Bolger Â’s effect ratio to summarize mediational
194 effects. Although the effect ratio puts a standardized number to the mediational effects, it still does not define substant ial mediation, AkersÂ’ (1998) standard for assessing his theory. Toward that end, the present rese arch adopts the notion that AkersÂ’ (1998) substantial mediation hinges on t he degree to which the mediator variable reduces the correlation between t he independent and dependent variables. Substantial mediation means that the paths between the two variables substantially reduce when the social lear ning variables are added to the model. Although there is no universal standar d for researchers to assess the strength of statistically significant zero-or der correlates, one rule of thumb is that a coefficient absolute value between zero and .20 represents no or negligible correlation, .20 to .40 represents low correlation, .40 to .60 suggests moderate correlation, .60 to .80 suggests marked correlation, and .80 to 1.00 suggests high correlation (Franzblau, 1958; see Hinkle, Wiersma & Jurs, 1988). Note that the range of each category is .20 and that as one moves up the continuum from negligible correlation to high correlati on, the percent of change between categories decreases. The difference between the ceiling of low correlation (.40) and the ceiling of negligible correlation (.20) is 50 perc ent. The difference between the moderate (.60) and low (.20) ceilings is 33%, 25% fo r the differences between the marked (.80) and moderate (.60) ceilings, and 20% between marked (.80) and perfect
195 correlation (1.00). One way to view AkersÂ’ (1998) term substantial mediation is to assess whether mediational effects lower bivariate correlations from one zeroorder rule of thumb summary categorization to another. Adopting the zero-order categorization rule of thumb to path analytic mediational analysis is conceptually stra ightforward. Ignoring the different definitions for the coefficients, the diffe rent inherent meanings the categorization reduction standard suggests a relative reduc tion. Does the incorporation of a mediator reduce the relative strength of the previously thought association between an independent and dependent variable from high to marked, marked to moderate, moderate to low, or from low to negligib le? Selecting the appropriate reduction percentage that indicates subst antial mediation is less intuitive. With the explicated rule of thumb, the range for identifying substantial mediation is between 20% and 50%, depending on the characterization of the starting correlation. However, is subs tantially reducing a low correlation to a negligible correlation a substantial mediat ion? Can substantial mediation occur within a range? The present research adopts the view t hat substantial mediation occurs at the higher end of the ranges, as the intent of AkersÂ’ (1998) term is to show that a relationship between two variables is substantially weaker than previously thought. A substantial reduction in an already poorly regarded model is less meaningful than the reduction observed in a more moderately, markedly, or
196 highly regarded model. As su ch, the present study sets the a priori level of substantial mediation as reducing the otherwise not ed path between social structure and delinquency by 20 percent.
197 Chapter Six Results Preliminary Evidenc e on Relationships Bivariate correlations. Table 9 reports the zero-order correl ations between the social structuresocial learning variables and log10 delinquency (the explanat ory variable intercorrelations were depicted in Table 8). Ten of the 16 variables predicted to affect delinquency are statistically signifi cant bivariate correlates.
198 As noted earlier in a different contex t, one way to view the strength of a statistically significant zero-order corre late is through a continuum described by Franzblau (1958) and Hinkle and colleagues (1988). A coe fficient absolute value between zero and .20 suggests no or negligible correlation, .20 to .40 suggests low correlation, .40 to .60 suggests moderate correlation, .60 to .80 suggests marked correlation, and .80 to 1.00 suggests high correlation. Three of the five social structuresocial learning differential social organization dimension variables are bivariate correlates of log10 delinquency: population density, log10 race composition, and age co mposition. However, each correlation is negligible; moreover, all th ree correlations are in the direction Coefficient -.06* -.07* -.05 -.06* -.04 .14* -.06* .27* .03 -.05 .01 -.04 .58* .61* .38* .22* Note: p < .05 (one-tailed t -test) 14. Definitions 15. Rewards 16. Costs Table 9 Zero-Order Correlations for the Explanatory Variable Variables and Log10 Delinquency (N = 1062) 12. SSSL III: Family Disruption 13. Differential Associations 6. SSSL II: Individual Sex 7. SSSL II: Individual Race 3. SSSL I: Sex Composition 1. SSSL I: Population Density 2. SSSL I: Log10 Race Composition 8. SSSL II: Individual Age 9. SSSL III: SES 10. SSSL III: Log10 Ethnic Heterogene i 11. SSSL III: Resdiential Mobility 4. SSSL I: Age Composition 5. SSSL I: Near Poverty
199 opposite of that hypothesized. Each of the th ree differential location in the social structure variables are bivariate corre lates of the delinquency measure, though individual sex and individual race are so negligibly, and race is in the direction opposite of that hypothesized. Individual age correlates weakly in the direction expected. All of the theor etically defined structural causes variables are statistically non-significant as bivariate correlates of log10 delinquency. At the microsocial level, differential associations, rewards, and costs each correlate in the direction hypothesized with log10 delinquency moderately. Definitions do so markedly. OLS regression models. Following the procedures of Friedrich (1982), cons istent with Baron and Kenny (1986), Braumoeller (2004), Cl early and Kessler (1982), J. Cohen and Cohen (1983), James and Brett (1984) and Judd and colleagues (2001), the present research examines moderati on through OLS regression. The analyses incorporate a multiplicative term in a regr ession model that contains both a social structure-social learning dimension predictor and a su spected social learning moderator. The SES and family disruption models do not report standardized coefficients because those scales are comprised of z -scores. Such measurements are already standardized, and Friedrich (1982) recommends not reporting the standardized coefficients produced by OLS regression because the
200 interpretation is not the same as that normally implied. Tables 10-21 report the results of the moderator r egression models for each soci al structural dimension indicator and each social learning measure. se ( b ) B .00.09 .01.77* .00 -.26* .12 R 2 .35 F ( p < .05 ) 186.65 .00.11 .01.78* .00 -.25* .14 R 2 .39 F ( p < .05 ) 220.72 .00-.02 .02.78* .00-.09 .14 R 2 .15 F ( p < .05 ) 64.13 .00.03 .02.33* .00-.15 .16 R 2 .06 F ( p < .05 ) 21.07 -1.16E-05 .10 -3.62E-06 .15 -1.00E-05 b 4.32E-05 Population Density Rewards Independent Variables (Population Density) X (Definitions) (Population Density) X (Costs) Intercept -6.13E-06 1.38E-05 .08 Population Density Differential Association (Population Density) X (Differential Association) Intercept-.77* Intercept Population Density Definitions .09 -5.23E-06 5.35E-05 Table 10 OLS Regression Dimension I (Population Density) Moderator Models (N = 1062) Model *p < .05 (one-tailed tests); significant interactions in bold -1.12* -.32* -.32 Intercept (Population Density) X (Rewards) Population Density Costs
201 se ( b ) B .03.04 .01.52* .00-.12 .08 R 2 .34 F ( p < .05 ) 182.02 .04.05 .01.55* .00-.13 .09 R 2 .38 F ( p < .05 ) 217.13 .04-.12 .01.42* .01.06 .09 R 2 .15 F ( p < .05 ) 63.68 .04.08 .01.13* .01 -.20* .11 R 2 .06 F ( p < .05 ) 22.79 *p < .05 (one-tailed tests); sign ificant interactions in bold -.87* -.49* -.05 Intercept (Log10 Race Composition) X (Rewards) Log10 Race Composition Costs Table 11 OLS Regression Dimension I (Log10Race Composition) Moderator Models (N = 1062) Model Intercept-.57* Intercept Log10 Race Composition Definitions .07 -.00 .03 (Log10 Race Composition) X (Costs) Intercept -.01 .04 .03 Log10 Race Composition Rewards Independent Variables (Log10 Race Composition) X (Definitions) Log10 Race Composition Differential Association (Log10 Race Composition) X (Differential Association) .10 -.01 b .02 -.06 .09 .00
202 se ( b ) B 1.46 -.01 .08.73 .17-.15 .69 R 2 .34 F ( p < .05 ) 179.51 1.72 -.03 .05.68 .10-.07 .82 R 2 .38 F ( p < .05 ) 214.60 1.72 -.11 .10-.05 .21.44 .82 R 2 .15 F ( p < .05 ) 62.02 2.00 .01 .11.56 .24-.35 .95 R 2 .05 F ( p < .05 ) 19.27 -2.56 -.01 .19 .15 -.06 b -.31 Sex Composition Rewards Independent Variables (Sex Composition) X (Definitions) Sex Composition Differential Association (Sex Composition) X (Differential Association) (Sex Composition) X (Costs) Intercept -.17 .28 .13 Intercept-.46 Intercept Sex Composition Definitions .08 -.02 -.70 Table 12 OLS Regression Dimension I (Sex Composition) Moderator Models (N = 1062) Model *p < .05 (one-tailed tests); sign ificant interactions in bold -.59 .86 -.26 Intercept (Sex Composition) X (Rewards) Sex Composition Costs
203 se ( b ) B 1.24 .10 .01.75* .14 -.25* .11 R 2 .35 F ( p < .05 ) 185.84 1.60 .10 .01.73* .09-.18 .13 R 2 .38 F ( p < .05 ) 215.65 1.65 -.09 .02.33* .20.07 .14 R 2 .15 F ( p < .05 ) 61.69 1.89 .03 .02.32* .22-.14 .16 R 2 .06 F ( p < .05 ) 20.49 *p < .05 (one-tailed tests); sign ificant interactions in bold -.1.10* -.21 -.19 Intercept (Age Composition) X (Rewards) Age Composition Costs Table 13 OLS Regression Dimension I (Age Composition) Moderator Models (N = 1062) Model Intercept-.79* Intercept Age Composition Definitions .09 -.18 2.27 (Age Composition) X (Costs) Intercept -.27 .75 .07 Age Composition Rewards Independent Variables (Age Composition) X (Definitions) Age Composition Differential Association (Age Composition) X (D ifferential Association) .15 -.44 b 2.28 -1.93 .07 .14
204 se ( b ) B .51.07 .01.65* .06-.12 .08 R 2 .34 F ( p < .05 ) 180.27 .63.06 .01.68* .04-.11 .10 R 2 .38 F ( p < .05 ) 215.15 .62-.07 .01.38* .07.01 .10 R 2 .15 F ( p < .05 ) 62.41 .72.03 .01.28* .08-.10 .12 R 2 .05 F ( p < .05 ) 19.87 *p < .05 (one-tailed tests); sign ificant interactions in bold -.99* -.28* -.16 Intercept (Near Poverty) X (Rewards) Near Poverty Costs Table 14 OLS Regression Dimension I (Near Poverty) Moderator Models (N = 1062) Model Intercept-.69* Intercept Near Poverty Definitions .08 -.05 .50 (Near Poverty) X (Costs) Intercept -.08 .21 .06 Near Poverty Rewards Independent Variables (Near Poverty) X (Definitions) Near Poverty Differential Association (Near Poverty) X (Differential Association) .13 -.10 b .60 -.55 .08 .01
205 se ( b ) B .09-.08 .02.37* .01 .27* .14 R 2 .35 F ( p < .05 ) 185.39 .10-.19* .01.42* .01 .31* .16 R 2 .38 F ( p < .05 ) 216.76 .10.00 .02.27* .01.14 .16 R 2 .16 F ( p < .05 ) 64.62 .12-.04 .02.04 .01 .25* .18 R 2 .07 F ( p < .05 ) 25.10 *p < .05 (one-tailed tests); sign ificant interactions in bold -.54* -.34* -.04 Intercept (Individual Sex) X (Rewards) Individual Sex Costs Table 15 OLS Regression Dimension II (Individual Sex) Moderator Models (N = 1062) Model Intercept-.43* Intercept Individual Sex Definitions .05 .02 -.26 (Individual Sex) X (Costs) Intercept .03 -.05 .01 Individual Sex Rewards Independent Variables (Individual Sex) X (Definitions) Individual Sex Differential Association (Individual Sex) X (Differential Association) .07 .03 b -.11 .01 .06 .01
206 se ( b ) B .11.04 .02.72* .01-.18 .13 R 2 .34 F ( p < .05 ) 183.51 .13.06 .01.74* .01-.17 .16 R 2 .38 F ( p < .05 ) 217.68 .13-.07 .02.40* .02-.01 .16 R 2 .15 F ( p < .05 ) 63.87 .14-.02 .02.29* .02-.09 .18 R 2 .06 F ( p < .05 ) 20.54 -.13 .09 -.02 .14 -.02 b .07 Individual Race Rewards Independent Variables (Individual Race) X (Definitions) Individual Race Differential Association (Individual Race) X (Differential Association) (Individual Race) X (Costs) Intercept -.01 -.03 .07 Intercept-.69* Intercept Individual Race Definitions .09 -.01 .10 Table 16 OLS Regression Dimension II (Individual Race) Moderator Models (N = 1062) Model *p < .05 (one-tailed tests); sign ificant interactions in bold -.1.04* -.21 -.10 Intercept (Individual Race) X (Rewards) Individual Race Costs
207 se ( b ) B .02.14* .04.56* .00-.02 .31 R 2 .35 F ( p < .05 ) 192.95 .03.06 .02.40* .00.21 .37 R 2 .39 F ( p < .05 ) 227.75 .03.33* .04.59* .00-.23 .34 R 2 .22 F ( p < .05 ) 98.27 .03.19* .05.02 .00.22 .40 R 2 .12 F ( p < .05 ) 48.62 *p < .05 (one-tailed tests); sign ificant interactions in bold -1.14* -1.98* -1.06* Intercept (Individual Age) X (Rewards) Individual Age Costs Table 17 OLS Regression Dimension II (Individual Age) Moderator Models (N = 1062) Model Intercept-1.25* Intercept Individual Age Definitions .05 .00 .02 (Individual Age) X (Costs) Intercept .00 .07 .00 Individual Age Rewards Independent Variables (Individual Age) X (Definitions) Individual Age Differential Association (Individual Age) X (Differential Association) .11 .00 b .05 .12 .13 -.00
208 se ( b ) .06 .01 .01 .04 R2 .34 F ( p < .05 ) 180.68 .07 .00 .00 .05 R2 .38 F ( p < .05 ) 218.58 .07 .01 .01 .05 R2 .15 F ( p < .05 ) 63.01 .08 .01 .01 .06 R2 .06 F ( p < .05 ) 9.45 1 SES is a scale comprised of z-scores. Unstandardized coefficients are reported as the variables are already standardized. *p < .05 (one-tailed tests); significant interactions in bold -.13* Intercept (SES) X (Rewards) SES Costs (SES) X (Costs) Table 18 OLS Regression Dimension III (SES 1 ) Moderator Models (N = 1062) Model Intercept-.61* Intercept SES Definitions .07* .01* -.12 -.02* Intercept .02 -.10 .05* -.36* SES Rewards Independent Variables (SES) X (Definitions) SES Differential Association (SES) X (Differential Association) .12* .01 b -.06 -.02 .08* .01
209 se ( b ) B .10.06 .02.47* .01-.15 .14 R 2 .34 F ( p < .05 ) 179.72 .13.11 .01.45* .01 -.22* .17 R 2 .38 F ( p < .05 ) 216.09 .16-.05 .12.38* .02-.00 .01 R 2 .15 F ( p < .05 ) 62.09 .14.09 .02.06 .02-.22 .19 R 2 .06 F ( p < .05 ) 20.81 -.08 .08 -.00 .09 -.02 b .10 Log10 Ethnic Heterogeneity Rewards Independent Variables (Log10 Ethnic Heterogeneity) X (Definitions) Log10 Ethnic Heterogeneity Differential Association (Log10 Ethnic Heterogeneity) X (Differential Association) (Log10 Ethnic Heterogeneity) X (Costs) Intercept -.03 .15 .01 Intercept-.47* Intercept Log10 Ethnic Heterogeneity Definitions .05 -.01 .18 Table 19 OLS Regression Dimension III (Lo g 10Ethnic Heterogeneity) Moderator Models (N = 1062) Model *p < .05 (one-tailed tests); significant interactions in bold -.69* -.47* .07 Intercept (Log10 Ethnic Heterogeneity) X (Rewards) Log10 Ethnic Heterogeneity Costs
210 se ( b ) B .44-.05 .03.46* .05.14 .22 R 2 .34 F ( p < .05 ) 179.02 .53.06 .02.69* .03-.09 .26 R 2 .38 F ( p < .05 ) 213.45 .54.08 .03.53* .06-.17 .27 R 2 .15 F ( p < .05 ) 61.16 .62-.15 .04-.07 .07.34 .31 R 2 .05 F ( p < .05 ) 19.55 *p < .05 (one-tailed tests); sign ificant interactions in bold -1.13* -.62* .42 Intercept (Residential Mobility) X (Rewards) Residential Mobility Costs Table 20 OLS Regression Dimension III (Residential Mobility) Moderator Models (N = 1062) Model Intercept-.45* Intercept Residential Mobility Definitions .08 -.02 .43 (Residential Mobility) X (Costs) Intercept .14 -1.11 -.02 Residential Mobility Rewards Independent Variables (Residential Mobility) X (Definitions) Residential Mobility Differential Association (Residential Mobility) X (Differential Association) .09 .05 b -.34 .54 .11 -.07
211 Despite the inclusion of coefficients and R -squared in each model, these analyses only test for moderation. If the interaction path is significant, a se ( b ) .05 .01 .01 .04 R2 .34 F ( p < .05 ) 178.97 .06 .00 .00 .05 R2 .38 F ( p < .05 ) 214.82 .06 .01 .01 .05 R2 .15 F ( p < .05 ) 61.98 .07 .01 .01 .06 R2 .05 F ( p < .05 ) 19.26 1 Family disruption is a scale comprised of z-scores. Unstandardized coefficients are reported as the variables are already standardized. *p < .05 (one-tailed tests); significant interactions in bold -.13* Intercept (Family Disruption) X (Rewards) Family Disruption Costs (Family Disruption) X (Costs) Table 21 OLS Regression Dimension III (Family Disruption 1 ) Moderator Models (N = 1062) Model Intercept-.61* Intercept Family Disruption Definitions .07* -.00 .02 -.93* Intercept .00 -.07 .05* -.36* Family Disruption Rewards Independent Variables (Family Disruption) X (Definitions) Family Disruption Differential Association (Family Disruption) X (Differential Association) .12* -.00 b -.01 -.02 .08* -.00
212 moderator relationship is s upported, regardless of the si gnificance, or not, of the other two paths (Baron & Kenny, 1986). Mo reover, the paths between individual social structure and social learning vari ables are not interpreted the same way that they would be in a traditional OL S model meant to assess random effects (see Baron & Kenny, 1986; Braumoeller, 2004). In the OLS moderation models, the general equation is Y = 0 + X1 + 2X2 + 3X1X2 + In this type of model, 3 represents the impact of a joint increase in X1 and X2 on Y and 2 are lower order terms in the model, and their coefficients do not represent the impacts of X1 on Y or X2 on Y generally. Instead, the coefficients represent the impact of X1 on Y when X2 = 0 or X2 on Y when X1 = 0 (see Braumoeller, 2004). Consequently, it is incorrect to think of X1 and 2X2 as the main effects of the model, compared to 3X1X2 as the interaction effects of the model (Friedrich, 1982). Instead, the X1 and X2 equations in the model are useless to the moderation hypothesis (s ee Baron & Kenny, 1986; Braumoeller, 2004; Friedrich, 1982). Each social structure-social learni ng dimension has at least one indicator with a statistically significant multiplic ative term. In the differential social organization dimension, population density st atistically interacts with differential associations and with definitions to jointly reduce log10 delinquency; race composition statistically interact s with costs to jointly reduce log10 delinquency;
213 and age composition statistically interacts wit h differential associations to jointly reduce log10 delinquency. One differential location in social structure indicator, individual sex, statistically interacts separately with di fferential associations, definitions, and costs to jointly increase the delinquency measure. The theoretically defined SES structural causes measure statistically interacts with the social learning measure of definitions to jointly increase log10 delinquency, whereas the statistical interaction between ethnic heterogeneity and definitions jointly decrease the delinquency measure. Direct and Indirect Effects Initial and revised measurement models. The implications of the moderati on analyses are not straightforward. Although the OLS regression models lend support to several of the moderator hypotheses, albeit some in directions differently than that expected, some variables in each dimension have statistica lly non-significant multiplicative terms, indicating that tests of the m ediational model are warranted. Following the procedures of James and Brett (1984), consistent with Baron and Kenny (1986), MacKinnon and colleagues (2002), and Shrout and Bolger (2002), the present research ex amines mediation through path analytic techniques. The study follows Anderson and GerbingÂ’s (1988) two-step approach of trying to establish a m easurement model before examin ing a structural model.
214 As mentioned earlier, SEM is sensitive to one-indicator models, and further, a fully saturated model has an infi nite number of possi ble solutions that do not allow fit assessment. One way to address the issue of numerous oneindicator measures is to assess a path model of manifest variables. Figure 23 depicts an example using population dens ity as the exogenous variable and differential associations as the intervening variable. Two problems occur from this approach. First, the model is fully saturated, thus not allowing for an assessment of fit. Second, the model assumes no measurement error, thereby not distingu ishing itself meaningfully from OLS regression. Lee and colleagues (2004) presumably addr essed these issues in their test of AkersÂ’ (1998) social structur e-social learning model through their Path Diagram for Social Structure-Social Learning Dimension I Figure 23 (Population Density), Social Lear ning (Differential Associations), and Delinquency Delinquency Poulation Density Differential Associations
215 parsimonious inclusion of a latent social learning construct. The logic of such a measure is that as social learning vari ables tend to correlate with one another (see discussions in Akers, 1998, 1999), they represent a higher social learning factor. By incorporating the construct social learning in their SEM model and testing the mediation of factor s, Lee and colleagues avoided having an intervening one-indicator variable, a sit uation problematic to SEM analysis (see Hatcher, 1994), and they were able to attended to the issue of saturation by constraining an index path in each latent variable. The present research follows Lee and colleaguesÂ’ (2004) example by constructing a latent social learning variabl e. Its construct validity is assessed by factor analysis. As mentioned earlier, pr incipal-components analysis and factor analysis are similar techniques that tend to produce similar results, though differing in their conceptualization of the underlying causal structure (see Hatcher, 1994). Principal-components analysis was used earlier to assess the survey and social structural scales because the measur es were viewed as additively creating a higher factor. In contrast, the social learning construct implies an underlying causal structure that exerts influenc e on the observed variables. Despite the different conceptualization, recall that researchers evaluate both approaches similarly. In the present research, analyses sugges t that differential associations,
216 definitions, rewards, and costs underlie one construct (eigenvalue = 1.85). The factor loadings for differential associati ons (.72), definitions (.84) and rewards (.70) each satisfy Hair and colleagueÂ’s ( 1998) criteria as being practically significant, whereas the costs loading (.37) falls in their minimally acceptable range. Researchers using SEM typically ignore factor loadings lower than .40 (Hatcher, 1994); however, recall that the costs measure was statistically significant in several of the OLS regr ession models (Tables 10, 13, 16, 17, 18, 21), including as a moderator to variables in the differential social organization (Tables, 11, 14) and differential location in the social structure (Table 15) dimensions. Dropping the costs measure ri sks altering the theoretical meaning of the construct, as well as the subst antive findings of the research. Figure 24 depicts the hypothesized so cial structure-social learning measurement model. A metric is established for each factor by fixing its variance at one, and each construct is allowed to covary. Table 22 presents the a priori goodness of fit measures, including the chisquare test statistic as a frame of reference.
217 definitions (X14), rewards (X15), and costs (X16).Social Structure-Social Learning Measurement Model Figure 24Note. Y = log10 delinquency. The "X" indicators correspond with the number s in correlation Tables 8 and 9: population density (X1), log10 race composition (X2), sex composition (X3), age composition (X4), near poverty (X5), individual sex (X6), individual race (X7) individual age (X8), SES (X9), log10 ethnic heterogeneity (X10), residential mobility (X11), fam ily disruption (X12), differential associations (X13), Location in Social Structure Social Learning X 10 X8 X 13 X 14 X 15 Y Location in Social Structure Theoretical Structural Causes Delinquency Differential Social Organization X1 X4 X2 X3 X5 Location in Social Structure X7 X6 X 12 X 11 X 9 X 16
218 The goodness of fit analysis implies that the initial meas urement model is a poor fit (RMSEA > .06; NFI, NNFI < .90; CFI < .95). The indexes suggest that the model is little differ ent from a null model. Although identifying the m easurement model is a c onfirmatory technique, one tool researchers have available in SEM is the ability to revise the model (Hatcher, 1994). Although that option is lim ited in this research as the model derives from AkersÂ’ (1998) theoretical assertions, examining each dimension individually may aid in the meas urement model identification. Figure 25 depicts a stand-alone measurem ent model for differential social organization. Table 23 reports its goodness of fit indexes. Indi vidually, the model for this dimension still fits the data poor ly. All measures fall outside of Bentler (1989) and Hu and BentlerÂ’s (1998) cutoff poi nts for suggesting a good model fit. Model 2df RMSEANFINNFICFI Social Structure-Social Learning3898.24*188.8.131.52.46* p < .05Table 22 Goodness of Fit Indices for the Social StructureSocial Learning Measurement Model (N = 1062 )Note. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index; CFI = comparative fit index. Values satisfying part of the a priori criteria are in bold.
219 An examination of the factor loadings revealed that sex composition is the Model 2df RMSEANFINNFICFI Social Structure-Social Learning724.60*184.108.40.206.76* p < .05Table 23 Goodness of Fit Indices for the Differential Soci al Organization Measurement Model (N = 1062 )Note. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index; CFI = comparative fit index. Values satisfying part of the a priori criteria are in bold. Figure 25Note. Y = log10 delinquency. The "X" indicators correspond with the numbers in correlation Tables 8 and 9: population density (X1), log10 race composition (X2), sex composition (X3), age composition (X4), poverty (X5), differential associations (X13), definitions (X14), rewards (X15), and costs (X16). Differential Social Organization Measurement Model Differential Social Organization X1 Social Learning X4 X2 X3 X 13 X 14 X 16 X 15 Delinquency X5 Y
220 only variable that is not statistically si gnificant. Akers (1998) asserts that this dimension represents social structural variables that empirically influence delinquency, and that social learning vari ables will mediate t heir effects. In addition to not being significant in t he measurement model, recall that sex composition was not significant in any of the OLS moderator models (Table 12). Table 24 reports the goodness of fit index es for a revised differential social organization measurement model in which the sex composition variable path is fixed at zero (removed from the equati on). Each of the index values in the revised model meet Bentler (1989) and Hu and BentlerÂ’s (1998) adopted a priori cutoffs for suggesting a good model fit. Figure 26 visually depicts t he differential location in the social structure measurement model, and Table 25 provides the values for its goodness of fit tests. The model results for this dimens ion are mixed. Although the index value Model 2df RMSEANFINNFICFI Social Structure-Social Learning98.72*24 .05.96.95.97* p < .05Model (N = 1062 ) Table 24 Goodness of Fit Indices for the Revised Di fferential Social Organization Measurement Note. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index; CFI = comparative fit index. Values satisfying part of the a priori criteria are in bold.
221 satisfies the Bentler (1989) and Hu and Bent ler (1998) criterion for the NFI, the values for the RMSEA, as well as the two measures that take the large sample size into account, the NNFI and CF I, suggest a poor model fit. Figure 26Note. Y = log10 delinquency. The "X" indicators correspond with the numbers in correlation Tables 8 and 9: individual sex (X6), individual race (X7), individual age (X8), differential associations (X13), definitions (X14), rewards (X15), and costs (X16). Differential Location in the Social Structure Measurement Model Location in Social Structure Social Learning X8 X7 X 13 X 14 X 15 X6 Y Location in Social Structure Location in Social Structure Delinquency X 16
222 Lastly, Figure 27 shows the theoretic ally defined structural causes individual measurement m odel, and Table 26 reports the results from the goodness of fit tests. The findings are agai n mixed. Three of the four indexes suggest a good fitting model according to t he a priori criteria, but the RMSEA value falls outside of Hu and Bent lerÂ’s (1998) specified range. Model 2df RMSEANFINNFICFI Social Structure-Social Learning145.95*10.11 .93 .82.94* p < .05Model (N = 1062 ) Table 25 Goodness of Fit Indices for the Differential Loc ation in the Social Structure Measurement Note. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index; CFI = comparative fit index. Values satisfying part of the a priori criteria are in bold.
223 Analyses of each dimension indivi dually suggest that the overall measurement model needs revision to a ccount for the differential social Model 2df RMSEANFINNFICFI Social Structure-Social Learning140.97*17.08 .96.93.96* p < .05Model (N = 1062 ) Table 26 Goodness of Fit Indices for the Theoretically Derived Structural Causes MeasurementNote. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index; CFI = comparative fit index. Values satisfying part of the a priori criteria are in bold. Figure 27Note. Y = log10 delinquency. The "X" indicators correspond with the numbers in correlation Tables 8 and 9: SES (X9), log10 ethnic heterogeneity (X10), residential mobility (X11), family disruption (X12), differential associations (X13), definitions (X14), rewards (X15), and costs (X16). Theoretically Derived Structural Causes Measurement Model Theoretical Structural Causes X 9 Social Learning X 12 X 10 X 11 X 13 X 14 X 15 Delinquency Y X 16
224 organization null path to sex compos ition. Further, although neither the differential location in the social structur e or the theoretically defined structural causes dimensions satisfied all four a pr iori criteria for indicating a good fitting model, each dimension had at least one indicator that suggested a good fit. Figure 28 presents a revised social stru cture-social learning measurement model with the sex composition path remo ved from the model. Table 27 presents the goodness of fit indexes.
225 definitions (X14), rewards (X15), and costs (X16).Revised Social Stru cture-Social Learning Measurement Model Figure 28Note. Y = log10 delinquency. The "X" indicators correspond with the numb ers in correlation Tables 8 and 9: population density (X1), log10 race composition (X2), age composition (X4), near poverty (X5), individual sex (X6), individual race (X7), individual age (X8), SES (X9), log10 ethnic heterogeneity (X10), residential mobility (X11), family disruption (X12), differential associations (X13), Location in Social Structure Social Learning X 10 X8 X 13 X 14 X 15 Y Location in Social Structure Theoretical Structural Causes Delinquency Differential Social Organization X1 X5 X2 X4 Location in Social Structure X7 X6 X 12 X 11 X 9 X 16
226 The indexes suggest that the revised model does not fit the data. Although the measurement models representing differ ential location in the social structure and theoretically defined struct ural causes did not satisfy the four criteria set a priori as suggesting a good model fit, t he indexes did suggest that the models require further examination. Table 28 describes the properties of the three measurement models. Model 2df RMSEANFINNFICFI Social Structure-Social Learning3533.24*220.127.116.11.49CFI = comparative fit index. Values satisfying part of the a priori criteria are in bold. p < .05Table 27 Goodness of Fit Indices for the Revised Soci al Structure-Social Learning Measurement Model (N = 1062 )Note. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index;
227 Structural models. The analyses now turn toward test ing its structural model. In SEM, standardized loadings represent the st andardized correlation coefficient for a latent constructÂ’s manifest variable i ndicator (Hatcher, 1994). The one-indicator variables suggest no measurement erro r because the measurement models did not estimate their variances. Those paths we re set at one. The indicator reliability Variance StandardizedExtracted Constructs and Indicators LoadingReliabilityEstimate Delinquency Construct1.00a1.00 Log10 Delinquency 1.00*1.00 Differential Social Organization Construct .96a.94 Population Density.43.18 Log10 Race Composition .76*.58 Age Composition.62*.38 Near Poverty.67*.45 Differential Location in the Social Structure Construct1.00a1.00 Individual Sex1.00*1.00 Individual Race1.00*1.00 Individual Age1.00*1.00 Theoretically Defined Structural Causes Construct.98a.97 SES.83*.69 Log10 Ethnic Heterogeneity -.52* .27 Residential Mobility-.44* .19 Family Disruption-.87* .76 Social Learning Construct.86a.82 Differential Associations.74*.55 Definitions.92*.85 Rewards.69*.48 Costs.37*.14and individual age variables assume no measurment error.Structure, and Theoretically Defined Structural Causes Measurement Models (N = 1062 )Note. p < .05 a Denotes composite reliability. The one-indicato r delinquency, individual sex, individual race, Table 28 Properties of the Final Differential Social Organ ization, Differential Location in the Social
228 represents the square of the standardized loadi ng (Hatcher, 1994). The composite reliability equates to the rationale of CronbachÂ’s (1951) alpha, reflecting internal consistency. Similarly, researchers seek a composite reliability coefficient greater than .70 (Hatc her, 1994). The index labeled Â“variance extractedÂ” estimates the amount of va riance that is not due to measurement error. Fornell and Larcker recommend that the value for a suitable model be greater than .50. Figures 29-31 depict the three tested structural models, and Table 29 presents their goodness of fit i ndexes. The criteria for sele cting which variable to set the path equal to one derive from Jo reskog & Sorbom (1989), who suggest picking the variable that best represents the factor.
229 e1e2e4e5e13e14e15e16 * * * * 1.00 ***1.00* e * ** 1.00 construct error (disturbance). Y = log10 delinquency. The "X" indicators correspond with the numbers in correlation Tables 8 and 9: population density (X1), log10 race composition (X2), age composition (X4), near poverty (X5), individual sex (X6), differential associations (X13), definitions (X14), rewards (X15), and costs (X16). Differential Social Organization Multifactor Structural Model (N= 1062) Figure 29 Note. An denotes an estimated path. A "1.00" represents a fixed path. An "e" denotes variable error, and "d" represents Social Learning Delinquency X4 X14 X13 X15 Differential Social Organization Y d d X2 X5 X1 X16
230 e6e13e14e15e16 * * 1.00*1.00* e * * e71.00 ** * 1.00 ** e8* * 1.00 Differential Location in the Social Structure Multifactor Structural Model (N= 1062) Figure 30 Note. An denotes an estimated path. A "1.00" represents a fixed path. An "e" denotes variable error, and "d" represents construct error (disturbance). Y = log10 delinquency. The "X" indicators correspond with the numbers in correlation Tables 8 and 9: individual sex (X6), individual race (X7), individual age (X8), differential associations (X13), definitions (X 14), rewards (X15), and costs (X16). Social Learning Delinquency X6 X14 X13 X15 Location in Social Structure Y d d X16 X7 Location in Social Structure X8 Location in Social Structure
231 In this research, the paths set equal to one are the indicator paths for the variables with the highest measurement model factor loading. Although sex composition was dropped from an earlier model because it contributed nothing to Model 2df RMSEANFINNFICFI Differential Social Organization2201.15*25.29.22-.13a.22 Differential Location in the Social Structure145.95*13.10 .93 .87.94 Theoretically Defined Structural Causes3292.47*27.340.00-.34a0.00* p < .05 part of the a priori criteria are in bold. Table 29 Goodness-of-Fit Indices for the Social Structure-Social Learning Structural Models (N = 1062)Note. RMSEA = root mean square error of approximation; NFI = normed fit index; NNFI = non-normed fit index; CFI = comparative fit index. aThe RMSEA sometimes produces values below 0 and above 1 (Hatcher, 1994). Values satisfying e9e10e11e12e13e14e15e16 * * * * * *1.00*1.00* e * ** 1.00 Theoretically Defined Structural Causes Multifactor Structural Model (N = 1062) Figure 31 Note. An denotes an estimated path. A "1.00" represents a fixed path. An "e" denotes variable error, and "d" represents construct error (disturbance). Y = log10 delinquency. The "X" indicators correspond with the numbers in correlation Tables 8 and 9: SES (X9), log10 ethnic heterogeneity (X10), residential mobility (X11), family disruption (X12), differential associations (X13), definitions (X14), rewards (X15), and costs (X16). Social Learning Delinquency X11 X14 X13 X15 Theoretically Defined Structural Causes Y d d X10 X12 X9 X16 d
232 the construct it was meant to measure, the circumstances for the non-significant differential social organiza tion population density loading are different. Beyond its non-significant factor loading, the race composition indicator was further nonsignificant in all OLS moderator model s. The population density variable, in contrast, was statistically significant as part of an interaction term with both differential associations and definitions (see Table 10). This research reasons that removing this variable from analysi s risks altering if not the theoretical meaning of the construct, the subs tantive empirical findings. In sum, the first overall social st ructure-social learning measurement model appeared to fit the data poorly Each dimension was examined individually, and a revised measurem ent model was tested with the sex composition path affixed at zero. The re vised model still fit the data poorly, but the individual dimension analyses suggested that the revised differential social organization measurement model was a goo d fit with the data. Further, the other two dimensions, although not satisfying the a priori criteria fo r a good model fit, had at least one indicator suggest a good fit. Structural models were estimated for each social structure-social learning dimension individually. None of the three dimensions satisfied the a priori criteria for a good model fit. Although the differential location in the social structureÂ’s NFI suggested that the model r easonably fit the data, the NNFI, the criterion that corrects for large sample sizes, suggests that the model fits the data poorly.
233 Chapter Seven Discussion Summary of the Problem The purpose of the pres ent study was to test a portion of AkersÂ’ (1998) cross-level social structure-social lear ning model. Elaborating on social learning theory, Akers suggest ed that the social learning process mediates social structural effect s on individual crime and deviancy. Although tests of the theory are s parse, and have limitations, they have provided a first glimpse of the effectiveness of the model. This research sought to improve on previous research by examining the model with more comple te measures of two of its social structural dimensions, and by more fu lly fleshing out how exactly social structure might impinge on the social learning pr ocess, areas suggested by Akers (1998, 1999) and colleagues (Lee et al., 2004) as needing more attention. The social structure-social l earning model is an elaboration of social learning theory (Akers, 1973, 1977, 1985, 1998; Burgess & Akers, 1966), which itself derived from SutherlandÂ’s (1947) differential association theory. Dissatisfied with the theoretical expl anations of his
234 time, Sutherland (1939) sought a general explanation of crime that would advance criminology as a science and pr ovide for the meaningful control of crime. Sutherland believed that the body of science was scattered, and he sought to organize the known correla tes of crime in a meaningful way (see Sutherland, 1924, 1934, 1939, 1947, 1970a, 1970b, 1970c). Sutherland first offered a tentative explanation fo r both crime and criminal behavior (Sutherland, 1939), before settli ng on his single-level theory of differential association (Sutherland, 1947). Social learning theory addresses a major criticism of differential association theory, that it does not explicitly specify the learning mechanisms inherent in the model (Akers, 1973, 1977, 1985, 1998; Burgess & Akers, 1966). Rather than a competing explanation for deviant, delinquent, and criminal behavior, social learning theory has subs umed differential association tenets (Akers, 1998). As a microsocial explanation for devi ant behavior, social learning theory has received much empirical attention. T he literature review revealed that social learning theoryÂ’s concepts and variables find moderate to strong support with survey, official, cross-sectional, and long itudinal data. Further, when researchers employ theory competition, social learning theory concepts and propositions generally find more support than those deriv ed from other simultaneously tested theories. When researchers apply social learning concepts and propositions to
235 integrated theory, social learning variabl es generally have the strongest effect. Although social learning theory offers a plausible explanation for deviant behavior, in its strictly processual form social learning theory cannot answer why some individuals and not other s encounter configurations of the social learning elements conducive to deviant behavior. Burgess and Akers (1966) originally argued that SutherlandÂ’s (1947) supposition that learning occurs thr ough interaction with others in social environments was compatible with the operant theory notion that environment shapes individual behavior. Burge ss and Akers expounded that because differential association theory was essent ially a learning theory, and that both criminal behavior and noncriminal behavior are learned through the same process, it was reasonabl e to incorporate modern l earning knowledge into the theory. AkersÂ’ (1998) social structure-so cial learning elaboration emphasizes the notion that social environments shape indi vidual behavior, and like SutherlandÂ’s (1939) original attempt to resolve perce ived failings in the criminological literature, Akers (1998) tackled the ta sk of simultaneously addressing both epidemiological and etiologica l explanations for crime. Starting from a social learning fram ework, Akers (1998) positioned social learning theory as the proxim ate cause mediator of distal social structural causes of crime. Although the model has received little empirical attention, its rationale has received strong theoretical oppositio n. Two main critics, Sampson (1999)
236 and Krohn (1999), collectively argue that t he social structure-social learning model does not adequately specify refutable propositions linking social structure to the social learning process. Samp son rejects the model outright, finding it Â“uninteresting,Â” and Krohn sees potential in the model but does not at present find it useful. Akers (1999) responded by noting that he his less concerned with understanding the macrosocial link ages than he is with understanding crime. However, AkersÂ’ (1998) seem ingly prescient remarks on the topic when explicating the model are mo re illuminating. Akers perhaps too subtly explained that although others we re welcome to view the model as a cross-level theoretical integration, that whic h requires the linking of propositions, he viewed the model differently. The social structure-social l earning model that Akers (1998) presented is a cross-level, concept ual integration that following the thinking of ThornberryÂ’s (1989) theoret ical elaboration, starts with the premise of social learning and expands it outward such that it becomes the process that explains macrosocial covariates of crime. The idea that drives theory elaboration is that re searchers add variables to an existing theory in order to improve it s adequacy (Bernard & Snipes, 1986). Whereas theory competition (Hir schi, 1979, 1989) attempts to refute opposing theoretical expositions, and theory integration (Bernard &
237 Snipes, 1996; Elliott et al., 1979; Liska et al., 1989) attempts to reconcile the differences, theory elaboration tr ies to advance science by working toward integration as if on a cont inuum, adding compatible concepts when applicable. Those that demand linking propositions from AkersÂ’ (1998) elaboration are not viewing it from the framework in which it was offered. They are starting from a different vi ewpoint than Akers, and although their position may be valid from their framew ork, the criteria they use to judge theory do not apply to AkersÂ’ elaboration by definition. Substantively, Akers (1998) is presumably less concerned with linking macrosocial explanations of cr ime to the social learning process through propositional integration, bec ause he views social structure generally as important to shaping the so cial learning process. He is not concerned with the source of that structure or any specific meaning attached to it by other t heorists (see Akers, 1998, 1999). Like Sutherland (1947), Akers (1998) views crime as rooted in societal social organization. He posit s differential social organization, as well as theoretically defined structural causes such as social disorganization theory, that which wa s measured in the present research, and only important to Akers because ot hers have already identified it as explaining the relationship between several correlates of crime, as cornerstones to the social structural dimensions of his social structure-
238 social learning model. Akers views soci al learning as the process by which social structure influences individ ual criminal and deviant behavior, and consequently crime rates. Akers (1999) believes the model is te stable as it is, and that rather than more theoretical specification, it needs better empirical testing, particularly through the incorporation of good empirical and theoretically derived social structural measures (see Akers, 1999; Lee et al., 2004). Responding to Sampson (1999) and Krohn (1999), Akers did acknowledge, however, t hat the lack of linking pr opositions was the least developed portion of the theory and he invited others to help with the specification. Akers (1998) concl uded his introduction to the social structure-social learning model with the comments, Â“I welcome othersÂ’ critiques, tests, and modifications.Â” Implications of the Present Research Nuances of the research question. The present research argued that Ak ers (1999) correctly characterizes social structure-social learning theory (Akers, 1998) as testable, but that his insistence on conceptual rather than propos itional integration is only adequate if the theory works as suggestedÂ—if social learning theory mediat es the effects of social structure on crime and criminal behavior. Although the lack of linking propositions may exacerbate the interp retation of less than clear empirical
239 findings, the present study reasoned that the theoretical adequacy of social structure-social learning t heory instead more likely hinges on AkersÂ’ standard for findings that empirically support the theory, substantial rather than full statistical mediation, and his description of the process. Akers (1998) suggests that expecting full statistical su pport of modeled sociological phenomena is unrea sonable. Because its main premise is that social structure has no effect on individual crimi nal behavior, if not fo r its effect on the social learning process, Akers argues t hat an observed statistical reduction in effects supports the theory in varying degrees: weakly to fully. Akers advances the notion of substantial m ediation as suitable for conc luding that the theory is plausible. He loosely defines the term s ubstantial mediation as that which is generally accepted by normal social science standards. Akers does not define the term more specifically, and the st udies in the literat ure that have found promise for the model have used the substantial mediation standard. The present research argued that the term substant ial mediation, as well as the notion of mediation generally, requi res more scrutiny than previously afforded. A review of the methodological literature suggested that although Akers may use the term mediation correctly when characterizing the process of statistically testing his model, account ing for mediational effects is more complicated than his (Lee and et al, 2004) and the other (Bellair et al., 2003; Lanza-Kaduce & Capece, 2003) two tests of the model have allowed. Because
240 social learning variables are expected to correlate with bo th social structural and outcome variables, the procedure of adding social learning variables to a model that includes social structural variabl es, and observing the new effects, cannot discern mediation from moderation. In such circumstances of expected correlation with the social learning variables, an incomplete mediation of effe cts may signal statis tical mediation or statistical moderation (see Baron & Kenny 1986). In order to conclude that mediation is plausible, researchers must first rule out moderation (see Friedrich, 1982). None of the three cited tests of so cial structure-social learning theory report testing the possibility of moderating effects. Adding to the complexity, some of AkersÂ’ (e.g., 1968, 1973, 1977, 1985, 1992, 1998) characterizations of the rela tionship between social learning and social structure cloud the theoretical distinction between mediation and moderation. Some of AkersÂ’ characteri zations seemingly describe a moderating relationship between social learning and soci al structure rather than a mediating relationship. The issue is important because the idea of moderation versus mediation is essentially what distinguishes the positions of Sampson (1999), and perhaps macrosocial researchers generally, from that of Akers (1998, 1999). Akers seems to view social learning theory as the pr ocess by which social structure impacts individual behavior. If not for the interv ening social learning process, social
241 structure would have no effect on crime. Akers (1998) makes this point more obvious in his illustration of his model (p. 331), his discussion of full versus substantial mediation, and in his test of the model (Lee et al., 2004). Sampson (1999) in contrast, which is parti cularly clear in his test of social disorganization theory (Sampson & Gr oves, 1989), views the relationship between social structure and individual behavior differently In that test, macrosocial variables measured a struct ure that was antecedent to a social disorganization construct that comprised measures of community control. Social disorganization was modeled as the mediator of the same types of variables that Akers (1998) views as the distal causes of crime, through their direct effect on the social learning process. However, AkersÂ’ (1998) model is not merely a one-for-one exchange of the social learning process with Sa mpson and GrovesÂ’ (1989) social disorganization measure. Sampson and Gr ovesÂ’ model serves as an explanation for crime rates, whereas AkersÂ’ model pr oposes that social structure influences social learning, which influences crimi nal behavior, which aggregate to crime rates. When discussing AkersÂ’ (1998) social structure-social learning model, Sampson (1999) is not viewing the problem from the same perspective as Akers. Whereas Akers sees a mediation relations hip between social structure and social learning, it seems more likely that Sampson sees moderation. To Sampson
242 (Sampson & Groves, 1989), social structur e serves as the antecedent cause of community control, the am ount of influence various lo cal networks are able to exert over its members, and the indivi dual level process is presumably only important through its interact ion with the predictor (soc ial disorganization) of crime rates. Overview of the Findings The present research tested a portion of AkersÂ’ (1998) social structuresocial learning model, emphasizing broad m easures of the differential social organization dimension (population density, race, sex, age, near poverty), known social structural correlates of crime, and four theoretically def ined measures of social disorganization theory (SES, et hnic heterogeneity, residential mobility, family disruption). The theoretical vari ables derived from Sampson and GrovesÂ’ (1989) test of social disorganizati on theory, Sun and co lleaguesÂ’ (2004) replication of Sampson and GrovesÂ’ test using U.S. census data, and from D. Gottfredson and colleagues (1991) who id entified additional important U.S. measures. In addition to modeling the theoretic al dimension more thoroughly than previous research, between the two dim ensions, the study included the three concentrated disadvantage variables (racial composition, fam ily disruption, and poverty) that Pratt and Cullen (2005) conc luded must be estimated or controlled in any test of crime causes to avoid t he risk of model misspecification. The study
243 also modeled the differential location in t he social structure as the mean survey sample respondent age, as well as the pr oportions of the respondents who were male and nonWhite. The study first examined the ques tion of moderation, using OLS regression to estimate 12 models that incl uded an interaction term for each social structure indicator and each social lear ning measure. At least one social structure and social learning indicato r interaction was found statistically significant in each dimension. In the differential social organiza tion dimension, population density statistically interacted separately wit h both differential associations and definitions, though in directions opposite than those hypothesized. The directions were, however, consistent with the opposite than predicted zero-order coefficient direction for population density and log10 delinquency. Researchers must interpret and assess in teractive models differently than standard OLS regression models becaus e the depicted relationships are conditional rather than general (Friedrich, 1982). An in teraction model measures joint impacts. The impact of one indepen dent variable on the dependent variable depends on the level of another independent va riable: The effect of the social structural variable on delinquency depends on the level of the social learning variable, and equally important, the effect s of the social learning variable on delinquency depend on the level of the social structural variable.
244 As to the combined effects negative c oefficient, the findings suggest that the impact of high populati on density levels on log10 delinquency is more substantial when the responde nt reports having fewer friends that engage in delinquent behavior, or having fewer defin itions favorable to self-reported delinquency (see Braumoeller, 2004). Said the other way, the results suggest that the negative impact of differe ntial associations and definitions on delinquency is more substantial as t he population density increases. Rather, having friends who skip school, steal items worth less than $50, hit to hurt, and use marijuana, or having neutralizing or lack of guilt definitions supportive of such behavior, only influences delinquency at t he lower ends of population density. The present research draws subs tantively similar conclusions and statements from the race composition and costs interaction term and from the age composition and differential associat ion term. Both interaction terms produced coefficients with negative values consistent with the zero-order correlation between the social structural variable and log10 delinquency. The results of the theor etically defined structural causes dimension suggest that ethnic heterogeneity (a stat istically non-significant zero-order correlate of log10 delinquency) and definitions likewise combine to produce opposite than expected results on t he delinquency measure. The SES and definitions interaction term moved in the direction anticipated, but the coefficient was trivial and SES was not a statistically si gnificant zero-order correlate of the
245 delinquency measure. In the differentia l location in the social structure dimension, sex composition statistically interacted separately with differential associations, definitions, and costs, pr oducing statements in the anticipated directions. Baron and Kenny (1986) remarked that results support moderation if an interactional term is statistically signifi cant, and they advised that the statistical significance of the other two paths (e .g., population density and differential associations in the described interactional model) is irrelevant to the moderation hypothesis. Following that standard, the present research concludes that differential associations moderate rather than mediate the effects of population density, age composition, and individual sex on log10 delinquency; definitions moderate rather than mediate the effects of populati on density, individual sex, SES, and log10 ethnic heterogeneity on the delinquency m easure; and costs moderate rather than medi ate the effects of log10 race composition and individual sex on log10 delinquency. However, Baron and Kenny (1986) also observe that when testing for moderation, a presumed moder ator should ideally not correlate with either the dependent or independent variable. Social le arning variables generally correlate with outcome measures, of course, and the social structure-social learning model predicts that the social learning variables will correlate with the social structure measures. Otherwise, t he model would be misspec ified because the theory
246 suggests that social structure is only im portant to crime thr ough its effect on the social learning process. Such interplay between the variables does not invalidate the test of moderation, but it does cloud interpretati on of significant findings (Judd & Kenny, 1981). Moreover, none of the interaction models received support for a dimension indicator across all social learning variables, nor did one social learning variable statistically intera ct with all macrosocial measures. The analyses proceeded to the tests for mediation. That decision was reasoned not only by the notion that so me variables had no statistically significant interactions, but further in consideration that a parsimonious SEM model would contain a social learning construct rather than the individual measures, thereby having broader m easurement than the OLS regression models and the possibility of not yet known results. Various measurement models were test ed, and none of the estimated, full social structure-social learning models fit the data well. The study rejected the original and two revised models. The st udy also examined measurement models separately for each dimension, however, and the a priori indexes for the revised differential social organization measurem ent model (sex composition path set = 0) suggested that the model was a good fi t with the data. Models for the other two dimensions seemed close eno ugh to warrant further scrutiny. The study tested three separate dim ension structural models. Following
247 the a priori goodness of fit measures st rictly, the study accepted none of the models as plausible fits with the data. The study did not support AkersÂ’ (1998) mediation assertions. Reconciliation of the result s with previous research. The results of the present study contradict the th ree reported tests of the social structure-social learning model (Bellair et al., 2003; Lanza-Kaduce & Capece, 2003; Lee et al, 2004). Each prev ious test found at least suggestive support for their mediation hypotheses. However, none of the previous tests reported testing for moderation. Moreov er, the tests used various methodologies (e.g., adding an additional intervening me asure into the model between social structure and social learning) and st atistical tests (e.g., standardized OLS regression) that may have affected the results. Lee and colleagues (2004) bot h examined the social structure-social learning model with fidelity to AkersÂ’ (1998) explication and a ssessed their model with a statistical technique (SEM) that the present research argued is most appropriate for examining AkersÂ’ medi ation assertions. Lee and colleagues presented the most rigorous published exam ination of the model to date, and it most closely compares (methodologically and statistically) to the present research. The contradictory findi ngs warrant close examination. Recall that Lee and colle agues (2004) estimated a full model that measured three of the four social structural dimensio ns and three of the four
248 social learning variables (excluding t heir separate test for imitation). They measured differential social organization as a one-indicator construct: community size (rural, urban, or suburban). They meas ured differential location in the social structure as two one-indicator constr ucts, the proportion of their survey respondents who were male and the mean age of their survey respondents, and one two-indicator construct, a composit e survey SES variable that measured the occupation and education of the repondentsÂ’ parents. T hey measured differential social location in primary, secondary, and reference groups as a one-indicator construct: a continuum of whether the respondent lived in a household with no parent present, with one biological parent present, or a household with both biological parents present. Lee and colleague s did not measure the theoretically defined structural causes dimension. Lee and colleagues (2004) measured di fferential peer association, definitions, and differential reinforcement consistent with the social learning literature, though they uncommonly modeled a social learning construct with the three concepts as indicators without ex plaining their rationale. They examined imitation separately because an SEM model would not converge with the measure in the equation. They drew simila r substantive conclusions from the full and partial models. Referring to the overall results, Lee and colleagues commented, The findings of the LISREL analy sis sustained the conclusion that variations in the behavioral and cognitive variables specified in the social learning process (1) account for subst antial portions of the variations in
249 adolescent use of drugs and alcohol and (2) mediate substantial, and in some instances virtually all, of the effects of gender, socio-economic status, age, family structure, and community size on these forms of adolescent deviance. (p. 29) The present research conc luded that rather than mediate the relationship between the effects of social structur e and delinquency, social learning more likely moderates the social structural e ffects. The present research measured social learning similarly to Lee and colle agues (2004) and althou gh incorporating SEM as a major part of the analytic strategy, the present study did not substantiate their conclusion. In contra st, the present study seemingly refutes their finding. The present study differed methodolog ically from Lee and colleaguesÂ’ (2004) test in three major ways. First, the present study model ed the theoretically defined structural causes dimensi on that Lee and colleagues were unable to incorporate, and it included much broader meas ures of the social structural crime correlates dimension. Secondly, the pr esent study estimated OLS regression interaction models, reasoning that a test of the social struct ure-social learning mediation statement was inappropriate unl ess moderation could at first be ruled out. Thirdly, the present study used diffe rent SEM model fit measures than those employed by Lee and colleagues. The rationale behind using more comple te measures of the differential social organization and theoretically def ined structural causes dimension was explained earlier. If these dimensions are indeed important to the social
250 structure-social learning model, then t he disparity between Lee and colleaguesÂ’ (2004) conclusions and those of the pr esent research may be the result of misspecification of Lee and colleaguesÂ’ tes t. They may have interpreted a model that does not adequately capture the full re lationship inherent in the theoretical explanation. The reasons why the present study tested for moderation were also explained earlier. Similar to the social structure di mensions explanation, if moderation is important to the true relationship between the social structural indicators and the social learning i ndicators, Lee and colleaguesÂ’ (2004) tested model is misspecified, which may in par t explain the discrepant results between their study and the present research. Lastly, the rationale for why the pres ent study used its selected a priori model fit measures, along with the reasons for the cutoff values, was also explained earlier. However, no attent ion was given to the goodness of fit measures used by Lee and colleagues (2004). Following convention, Lee and colleagues (2004) reported a chi-square test statistic that suggest ed the model did not fit the dat a, but they reasoned that the indicator was not reliable in th eir research (also common in the methodological literature). The two indicato rs they relied on to conclude that the model fit the data were the goodness of fit index (GFI ) and the adjusted goodness of fit index (AGFI). In the alcohol model, they reported that the GFI =
251 .93 and the AGFI = .95. Fo r the marijuana model, they reported that the GFI = .93 and the AGFI = .94. T he imitation model for alc ohol GFI was .84 and the AGFI was .53. For marijuana, the imitat ion GFI was .82 and the AGFI was .45. Lee and colleagues did not explain their ra tionale for their chosen fit measures, nor did they report their cutoff values for a good fitting model. They described the model fit in the body of the article by noting that the reported measures suggested a good fit. It is unclear if they meant that description to refer to the imitation models. As mentioned earlier, researchers have many SEM goodness of fit measures at their disposal, and there is little agreement on which indicator is the best measure of a modelÂ’s fit. One agreem ent in the literature tends to be the notion that using the chi-square test as t he indicator of model fit tends to produce biased results. If sample size is too sma ll, the chi-square test statistic is prone to Type I error and if sample size is too lar ge, the statistic may lead researchers to reject a good fitting model (see Hatcher, 1994; Mulaik, James, Van Alstine, Bennett, Lind & Stilwell, 1989; Tabachnick & Fidell, 2001). The GFI (Bentler, 1983; Joreskog & So rbom, 1984) measures model fit by examining a weighted propor tion of sample variance against an estimated covariance matrix. The idea is to produce a statistic that is analogous to the R2 (Tanaka & Huba, 1989). Because less rest ricted models (estimating many data points) produce better fitting models, t he AGFI adjusts the GFI based on the
252 number of parameters that the model is required to estimate. It penalizes the model for having many parameter esti mates (Mulaik et al., 1989; Tabachnick & Fidell, 2001), and thus is a conservative, presumably lower value than that of the GFI. Generally, researchers view .90 as the cutoff for the GFI and the AGFI (Joreskog & Sorbom, 1984), and some researchers suggest no fit measure should be accepted with a value below 90 (Hu & Bentler, 1999). Hu and Bentler (1999) noted that the GFI and AGFI are sensitive to sample size, with large samples increasing the opportunity for Type I error. Alth ough Tanaka (1987) and La Du and Tanaka (1989) found the GFI to be a good estimator in a wide range of examples, Shevlin & Miles (1998) c oncluded that based on a simulation study, Â“a cut-off value of 0.9 would result in an unacceptable number of misspecified models being acceptedÂ” (p. 85). Moreover, they concluded that any value below .95 in a model with low factor loadings will generally be unsatisfactory regardless of sample size. The suitability of the GFI and AGFI as SEM goodness of fit indicators appears mixed. McDonald and Ho (2002) re veal that although the GFI and AGFI appear often in the literatur e, they are not the most commonly used measures. Reviewing 41 studies in the psychological literature, they found that the two most commonly reported global fit indicators were the unbiased relative fit indicator (21 studies) and the CFI (21 studies), follo wed by the RMSEA (20 studies). Among
253 the other notables, the GFI was repor ted in 15 studies and the NNFI was reported in 13 studies. Though the effectiveness of the GFI and AG FI is mixed in the literature, researchers tend to agree that .90 is the minimum value that should be interpreted, and that the measure is sens itive to Type I error with large sample sizes. Lee and colleagues (2004) tested mode ls with sample sizes of 2,700 and larger, and they interpreted their imitati on models with a GFI as low as .82 and an AGFI as low as .45. They interpret ed their main models with a GFI as low as .93 and an AGFI as low as .94. Lee and colleagues (2004) did not explai n their reasons for interpreting the two models with fit index values below the generally ascribed .90 cutoff. They additionally did not address the issue of their reported full model AGFI values being higher than the GFI values, an illogical occurrence as the AGFI conservatively adjusts the GFI in order to penalize parameter estimation, nor did they discuss the implications of their lar ge sample sizes, or the implications of their low factor loadings. A third explan ation for the disparity between Lee and colleaguesÂ’ (2004) conclusions and those of the present study may be that the GFI and AGFI main model resu lts signify Type I error. Nuances of the findings. Although seemingly trying to have it both ways, hypothesizing about mediation and moderati on, the present study was prim arily interested in AkersÂ’
254 (1998) notion of mediation. The requisite to first test for moderation derived from a review of the literature. In doing so the study was unable to accept the mediation hypotheses, and instead, seve ral moderation hypotheses found statistical support. Before testing the social structur e-social learning model, the present research specified the hypothesized effe cts for the moderation and mediation models, and it also explicated a possible me chanism that links social structure to social learning: contingencies of reinforcement. Although the explicated functional relationships derived from a social structure-social l earning framework, which contrasts with the relationship dep icted by the moderation hypotheses, the unexpected results do not in validate the specification of this mechanism. It was earlier argued that social struct ure impinges on the social learning process through the notion of various reinforcement contingencies influencing individual reinforcement schedules. Alt hough it was anticipated that social structure set the contingency that would ot herwise not affect individual behavior if not through its impact on the so cial learning process, the mechanism itself is not inconsistent with a moderating relationship. Akers (1998) and Sutherland (1939, 1947) both view crime as an expression of social organization. Such te rms, as noted earlier, lend themselves to interpretation as a moderator rather than a mediator. At other times, Akers (1998) specifically describes the relati onship between social learning and social
255 structure as mediation. The idea that social structure sets various contingencies of reinforcement that are differentially reinforced individual ly, allows dual characterization. The notions of contingencies of reinforc ement and reinforcement schedules do not rely on the characterization of the st atistical relationship between the two variables. The described linking mechanism between social structure and social learning is invariant to the m ediation or moderation terminology. The point is important because this research suggests that social structural and social learning variables relate they do go together, just not in the precise way that Akers (1998) most often refers to the relationship. Although the depiction of a linking mechanism that ex plains the relationship between social structure and social learning at firs t seems incapable of being an a priori statement of the social st ructure-social learning model, or perhaps even not refutable as it fits both a moderating or mediating relationship, such is not the case. Recall that Akers has not fully s pecified his model, according to Sampson (1999), and Krohn (1999), and even Akers ( 1998, 1999) admits that he has made no linking propositions. Akers (1998) sometimes refers to his model in contradictory ways. Although it was reasoned that AkersÂ’ model must be te sted by SEM, in order to assess the mediational effects advanced by Akers, as opposed to HLM, which was the preferred macrosoc ial approach of Hoffmann (2002), for example, the
256 finding of moderating effects over mediat ing effects does not invalidate AkersÂ’ model. Social learning does relate to t he social structural variables and their impact on delinquency. If the social learning and social struct ure relationship generalizes beyond this research, Akers (1998) needs to change his verbiage. As was demonstrated earlier, the literature is alr eady full of studies that mi suse the terms moderate and mediate, some in the same study, and by it self, such causes little problem for the model. That AkersÂ’ (1998) model is not discr edited by the notion of a moderating relationship instead of a mediating relations hip, should that i ndeed be the reality, is demonstrated in part by elaboration of a point made earlier that refuted his mediation assertions. Recall the quotat ion that Lee and colleagues (2004) used to announce the findings of their test of t he social structure-social learning model. Lee and colleagues concluded that the tested model mediated the relationship between social structure and their devi ancy measures. The present research contradicted that assertion. However, in the next paragraph, Lee and colleagues (2004) commented, Â“We found, as proposed by the SSSL model, th at social learning theory offers a useful and empirically supported set of concepts and principles for understanding how social environmental factors have an impact on behavior (Burgess & Youngblade 1998)Â” (p. 29). The present re search supports t hat findingÂ—the well-
257 tested and empirically supported social l earning concepts moderate the impact of social structure on delinquency. The distinction between moderation and m ediation, as it turns out, does not speak to the validity of the model. Ho wever, if the present study generalizes, and if contingencies of reinforcement and reinforcement schedules adequately serve as the linking mechanism between so cial structure and the social learning process, the social structure-social learning statement r equires modification. Modification of the theoretical statement. Recall that the present research f ound that the combined effects of the social learning variables and indicators of the differential social organization and theoretically defined structural causes di mensions tended to impact delinquency in a direction opposite of that hypothesized. The present research suggests that the differential social organization and the theoretically defined structural causes dimension indicators combine with t he social learning process to reduce delinquency. The conclusion was that so cial learning measures moderate the relationship of social structural variables on delinquency in an unexpected direction. Recall the finding between differential associations and popul ation density, for example. The model was statistically significant ( R2 = .35, p < .05), and both differential associations and the populat ion density-differe ntial association interaction term contributed to the model The interaction term coefficient was
258 negative. Although the statistical significance of non-interaction terms is irrelevant to the moderation hypothesis (Baron & Kenny 1986), a statistically significant contributor does have meaning (Friedrich, 1982). As the relationship between an independent and dependent variable is condit ioned upon the level of another independent variable in an interaction OL S regression model, the coefficients of the non-multiplicative terms represent their independent effect on the dependent variable when the other variable is zero. In the population density and different ial associations OLS regression moderator model, the st atistically significant value of the differential associations coefficient was .77. The characterizati on for the whole model described earlier suggested that high levels of population density and high levels of delinquent peers result in a reduction of self-reported delinquency. The study further suggests that although having friends who engage in delinquent behavior generally results in an increase in delinquency, as reported in the literature, it condi tionally relates to self-r eported delinquency only at low levels of population dens ity. Differential associat ions affect delinquency equivalent to the .77 coefficient when the population density is equal to zero, thus leading to the statement that as populat ion density increases the effects of differential associations on delinquency reduces such that high levels of population density and high le vels of differential associations reduce
259 delinquency. The present study f ound similar opposite than expected characterizations for several combinatio ns of macrosocial and individual-level interaction terms. The findings of the present research suggest that the effects of social structure and social learning on de linquency are not cons tant. Moderation effects, regardless of the direction of im pact, are contrary to AkersÂ’ (1998) most prominent characterization of social stru cture-social learning model. Moreover, social learning concepts have not prev iously been characterized as having conditional effects. The moderation e ffects suggest that in addition to the misspecification of the social structure-so cial learning model, the social learning model is likewise misspecified. The effe cts of social stru cture on delinquency are conditioned by the level of social learning, and the effects of social learning on delinquency are likewise conditioned by the leve l of various social structures. Although such lack of constant effe cts is the outcom e of a moderation relationship by definition, interpretation of the contingent relationship between the social structural and social learni ng variables may be further complicated because the social structural dimensions advanced by Akers (1998) vary in their proximity to the mechanism that operates at the individual level. Recall that social learning variables have feedback effects gener ally, and that Akers suggests that there is some overlap between the soci al learning process and the meso-level variables advanced in the social structural elaboration.
260 In the differential location in the so cial structure individual sex and differential associations moderator model, fo r example, the statistically significant interaction term moved in the direction expected. Elaborate explanation is not needed. The interaction of maleness and di fferential associations combine to increase log10 delinquency. In this dimension, some other process appears to be going on than that of the diffe rential social organization or theoretically defined structural causes dimensions, which inte racted with social learning variables to reduce delinquency. To understand the differential location in social structure dimension, it is important to remember that its indi cators do not represent broad social structures, rather they represent an aggregate of the individual sample characteristics. Individual sex is the proportion of respondents in the sample who are male. The differential location in the soci al structure dimension described by Akers (1998) seems to represent a meso-lev el structure. It seems more in line with the differential social location in primary, secondary, and reference groups dimension, that which provides the imm ediate context for lar ger groupings, than the implied structures of the differential social or ganization or theoretically defined structural causes dimensions. Being around a small group of males, for example, may provide the opportunity for translating the messages of a larger grouping of males.
261 The present study concludes that the social learning process may moderate social structural variables t hat represent the differential social organization and theoretically defined struct ural causes dimensions in such a way that the combined effects reduce rather than increase delinquency. The study further concludes that these dimens ions represent more distal causes of crime than variables that repr esent the differential location in the social structure dimension, as well as the differential soci al location in primary, secondary, and reference groups, which was not modeled in the present study. Further, the present study finds that the social learning process might interact with differential location in the so cial structure indicators in such a way that the combined effects increase the propensity of delinquency. However, the study realizes that this di mension also closely resembled a mediator relationship in the SEM models, if not for the stringent a priori fit measures. Although its structural model was reje cted in the present resear ch, the model would have found support with the less stringent m easures utilized by Lee and colleagues (2004). Although only the NFI suggested suppor t for a mediational relationship in the present research, the GFI (.97) and AGFI (.91) met the standards used by Lee and colleagues. One possible explanation for this apparent discrepancy stems from the notion of moderated mediatio n (James & Brett, 1984). Re call that when testing interaction, it is ideal that the suspected moderators not correlate with
262 independent or dependent variables (Bar on & Kenny, 1986). In the present research, social learning variables co rrelate with both social structural and delinquency variables. The moderation interpretation was not clean. As moderated mediation is possibl e, the question becomes, how might social learning variables act both as a m oderator and as a mediator of social structural variables? If the present st udyÂ’s tested models ar e not misspecified, the alternative is that Aker sÂ’ (1998) social structuresocial learning theoretical model is misspecified. Social lear ning serves as bot h a moderator and a mediator of social structural variables because the model does not account for some unknown relationship. If variables do indeed operate as both a moderator and a mediator of social structure, t hen Akers is not describing the process correctly. Recall reinforcement contingencies and reinforcement schedules as the possible mechanism that links social struct ure to the social learning process. Also, recall Figure 4, or the bottom model in Figure 5, path diagrams that show social structure indirectly influenci ng delinquency through the social learning process. If the findings of Lee and coll eagues (2004) are correct, AkersÂ’ (1998) model finds support. If the m oderator models of the pres ent study are correct, the first reaction is to presume that the Lee and colleagues, and thus AkersÂ’, mediation model is incorrect. However, social structural reinforcement contingencies and individual reinforcement schedules may interact in such a way
263 that portions of both t he moderator and mediator hypotheses are correct. It was presented earlier that social structure may set reinforcement contingencies that are reinforced at the i ndividual level differentially. The process of reinforcement and extinction was descr ibed as an explanation for the aging out effect, for example. As described, rein forcement contingencies and reinforcement schedules are a dichotomy that equate to the structural and individual levels. AkersÂ’ (1998) social structure-social learning model, in contrast, does not present a dichotomy between social struct ure and individual behavior so much as it presents a continuum of social structure, wh ich was thought to impact individual behavior, and crim e rates, only through the social learning process. Differential social location in the prim ary, secondary, and reference groups, along with differential location in the soci al structure repres ent the proximate interpretation of more distal structures such as those empirically or theoretically derived. If AkersÂ’ (1998) social stru cture-social learning is conceptualized more as a dichotomy, the question becomes not how does social structure impinge on the social learning process, but rather how are reinforc ement contingencies, which are produced from the social structure, transmitted to reinforcement schedules, which occur at the individual level? One possible framewor k is that the transmittal process occurs through the small groups that actually reinforce or punish behavior. As such, social learning-socia l structure is not comprised of two
264 empirical and theoretical di mensions and two smaller-group dimensions, rather it more logically comprises one distal (macro-level) dimension and one more proximate (meso-level) dimension. Rather than social learning mediati ng the social structural effects on delinquency, distal macrosocial correla tes of crime may influence criminal behavior through their interaction with the so cial learning process, whereas more proximate meso-level crim e correlates may provide the messages social learning mediates. This explanation accounts for both the moderation effects observed in the present research and fo r the mediation effects noted in the literature (Bellair et al., 2003; Lanza-Kaduce & Capec e, 2003; Lee, et al., 2004). Relating the interpretation of the pr esent studyÂ’s results to the LanzaKaduce and Capece (2003) findings is straightforward. They, like Lee and colleagues (2004) did not measure strong macrosocial indicators, instead modeling measures that the present study views as meso-level. Their findings relate to the present study in sim ilar fashion to the findings of Lee and colleagues. As to Bellair and colleagues (2003) their findings require more interpretation to relate to the present re search. They used theoretically derived measures of concentrated disadvantage si milar to those used in the present research. They concluded that the conc entrated disadvantage measures had no relationship with social learning or delinq uency, but that other social structural
265 effects on the outcome measure were mediated upon introducing social learning variables to the equation, along with a family well-being construct. Bellair and colleagues ( 2003) added an additional cons truct to the model than that posited by Akers (1998), and it was this family well-being construct, combined with its direct effect on soci al learning variables, which mainly mediated the effects of occupational struct ure. They modified AkersÂ’ model using the rationale that the new construct comp rised of family income and family structure (single parent household) helped translate the contextual messages offered in the broader social structure. In essence, though not describing it as such, Bellair and colleagues (2003) measured AkersÂ’ (1998) differential soci al location in primary, secondary and reference groups dimension, as indexed by Lee and colleagues (2004), and placed it between social stru cture and the social learning process as a mediator. Consequently, their findi ng that the family wellbeing and social learning measures mediated the impact of their social structure measures on their outcome measure is consistent with the c onclusion of the present research. The present research characterizes the family well-being variables as the meso-level structure that affects de linquency through the mediation of social learning. Although Bellair and co lleagues (2003) modeled what the present research considers a meso-level variable as a mediator of social structureÂ’s effects on criminal behavior, rather than social learning as specified by Akers
266 (1998) and adopted by the present res earch, their model is nonetheless consistent with the present studyÂ’s descr iption of the functional relationships because the differential social location in primary, secondary, and reference groups dimension overlaps with the social learning process. In specifying the dimension, Akers qualified his statem ents by noting that the meso-level dimension may be difficult to distinguish fr om the individual level social learning process. Lastly, this studyÂ’s interpretation of ambiguous data (Sampson, 1999) is also consistent with the main conclusion s drawn by Sampson and Groves (1989) in their test of social disorganization theory. They found t hat local community control mediated the effects of their so cial structure measures (indexed in a similar way in the present resear ch) on their outcome measures. Sampson and Groves (1989) descr ibe and measure local community control in a manner that is similar to the social structure-social learning dimension of differential location in primary, secondary, and reference groups. When viewing AkersÂ’ (1998) social struct ure-social learning model as a macrolevel and meso-level dichotomy, Samp son and GrovesÂ’ intervening construct equates to the role of the meso-level di mension in the modified social structuresocial learning model. Moreover, reca ll that Veysey and Messner (1999), upon reexamining Sampson and GrovesÂ’ model with SEM, concluded that Sampson and GrovesÂ’ intervening mechanism comprised more than one dimension, one of
267 which, they concluded, was a social learning construct. One explanation for how social stru cture-social learning (Akers, 1998) might mediate crime at the me so-level, yet interact at th e more distal macrosocial level to reduce crime might stem from WirthÂ’s (1938) characterization of urbanism. Recall that he considered lar ge cities as a place of superficial relations. Using the present research findings that population density and differential associations interact to reduce delin quency as an example, large communities might represent a place wher e individuals not only have little in common, but may also tend to know lots of people in a superficial way. In the Largo sample, respondents in the areas with higher populations may know many people in a superficial way, may characterize the relationship as friendship, because such superficial interaction is normal, yet the individual may not be influenced by the individuals they have identified as fri ends that engage in delinquent behavior. Such a characterization holds less for the race composition, age composition, and ethnic het erogeneity interactions, particularly for those interactions that included social le arning concepts other than differential associations, such as the costs measure. However, the functional relationships between social structure and social lear ning may nonetheless be consistent with macrosocial lit erature. Whereas Wirth (1938) anticipates social stratification from urbanicity to be
268 represented by race and age, as well as high population density, and for such social structure to take on the characteri stics he describes as inherent in large, densely populated areas, Park and Bur gess (1925) characterize the innerworkings of the urban communities diffe rently than Wirth. Instead of being unconstrained by superficial urban relation s, as suggested by Wirth, Park and Burgess suggested that urban neighborhood s provide a sense of community. In the community depiction, high levels of stratification based on social structures such as race, age, sex, and poverty might create opportunities for stronger interpersonal relationships ra ther than weaker in teractions. This depiction follows the notion of community social control depicted earlier in the discussions of Shaw and McKay (1942, 1969), Sampson and Groves (1989), and the like. Rather than allowing greater anonymity, high levels of race and age composition and ethnic heter ogeneity, important in the present research, might combine with high levels of social learning variables to reduce delinquency because contrary small group social l earning processes may be overridden by strong community structures that provide ample opportu nity for reinforcement contingencies that reward conformity. This research argued that the functi onal relationship between macrosocial contingencies of reinforcement, micr osocial reinforcement schedules, and delinquency includes the notion that indivi duals seek opportunities for social reinforcement. The interplay between macr osocial structure and the meso-level
269 groups that actually reward or punish behav ior might be most noticeable in areas that are socially stratified. In such areas, the macrosocial c ontingencies of reinforcement, more normally distal, and bearing weaker messages than the more proximate structures that translate the messages into rewards or puni shment, may take on the same role as the meso-level struct ures. Areas of high stratification may have higher area cohesiveness that influences individual behavior similar to the ways otherwise shaped by small group networks. Such highly stratified areas may get the message to individual behavior direct ly, without the translation from smaller group networks. Individuals might still rece ive messages from smaller groups that are conducive to law violation, but as the larger community messages are cohesive, and amply rewarding, or punish ing, the messages of conformity are acted uponÂ—in this way, high levels of stru ctural stratification might interact with high levels of deviant social learning processes to reduce rather than increase delinquency. Limitations of the Present Research The present research has several limitations. The first pertains to generalizability. Although the micro-level data comprise a random sample of students in the select schools, the st udy does not purport to generalize beyond the schools. Particularly, the research may not generalize to youth less protected than those attending school (see discussion of street criminology versus school
270 criminology in Hagan and McCarthy, 1998). A second limitation has to do with scope. Like much of the social learning literature, the present study focu sed on minor forms of delinquency. The remaining limitations have to do with methodology. Skew and kurtosis were present in several variables, and t he study relied on several transformations to normalize the data. Study analyses assume normality, multivariate normality in the case of SEM, and the implications of nonnormality in these data mainly represent misinterpretation of the inferential procedures Although there is much literature to suggest that the analyses us ed in the present study are robust to assumptions of normality, the literat ure is mixed on some points. Further, the possibility of misinterpretation may have been exacerbated by the selection of strict model fit criteria in the SEM analyses, particularly in respect to the CFI. Many researchers use .90 as a cutoff, but the present research specified the CFI value according to t he more conservative views of Hu and Bentler (1998, 1999), who sugge st .95 or higher as an indicator of a good model fit. This decision made the difference bet ween the final SEM structural model having one out of four indexes suggest a good fit instead of two out of the four. However, because the study set four fi t measures a priori, the final model would have been rejected regardless. More over, the moderator analyses, also subject to the possibility of erro r stemming from nonnormal data (for an explanation of why concerns of multic ollinearity distorting coefficients in
271 interactive regression are warrantless s ee Friedrich, 1982), further suggested that mediation was not how the variables interrelated. Despite the possibility of biased coefficients in the SEM analyses, t he moderator results suggest that the substantive conclusions would not have differed. Lastly there is the issue of relative model fit. The literature offers a wide range of SEM measures by which to judge a modelÂ’s fit. The rationale for selecting the specific measures and t heir cutoff points was explained earlier. However, the variety of measures exist, in part, because of a lack of consensus over what type of support is actually needed to be assured of a reasonable fit, and because of the growi ng dissatisfaction with the chi-square statisticÂ’s stringency on requiring a perfe ct fit (see Hatcher, 1994). The various measures intentionally rela x certain criteria in order to find an approximate fit. Measures that start with an Â“rÂ” tend to model relative fit, like the RMSEA in the present research, and the NFI and NNFI are designed to be more in line with the purpose of the chi-square statistic, acc ounting for its tendency to underestimate in small samples and overestimate in large samples (see Hatcher, 1994). Researchers that use .90 as a cuto ff for the NFI, NNFI, and the CFI, as well as those who use .95 for the CFI, tend to qualify their lower limit by suggesting that the closer to 1.00 the better (e.g., Bent ler, 1989; Hu & Bentler, 1999). What is not addressed in the literat ure is whether a model that falls below the cutoff is Â“almost there,Â” such as might be suggested considering that there
272 seems to be a scale between .90-1.00, or whether the model should be rejected outright such as what was done in the pr esent research, following the rationale used in OLS regression that a non-signific ant model is not interpreted, no matter how close the p -value. Strict adherence to a priori indicators of hypothesis plausibility is what drives the scientific processes, and the present study argues that as the research was not exploratory, instead test ing a theory, such formal hypothesis testing proc edures were mandated. Conclusion The present research sought to test AkersÂ’ (1998) assertion that social learning theory mediates social structur al influences on delinquency. The study utilized the three measures (race poverty, and family disruption) that Pratt and Cullen (2005) identified in a macro-level predictors meta-analysis as Â“among the strongest and most stable predictors Â“ (p 373) of crime. Further, the study measured social disorganization theory vari ables in a manner similar to that used by Sampson (Sampson & Groves, 1989), on e of the social structure-social learning modelÂ’s more vocal skeptics (Sampson, 1999). Secondly, the study introduced possible linkages between social structure and the social learning process in an a ttempt to address the concerns of Krohn (1999), who suggested that the theory does not adequately do so, and Sampson (1999), who suggested that t he theory is incapable of producing a priori, refutable macrosocial propositions. Fu rther, the present research critically examined
273 AkersÂ’ (1998) notion that so cial learning mediates the relationship between social structure and crime, introduc ing the possibility that so cial learning may instead moderate social structureÂ’s effect on crime and criminal behavior. The study argued that clarifying this distinction may contribute to understanding how exactly social structure might influence the social learning process. Combined, the two aims of the study, utilizing more complete social structural measures and explaining how social structure might impinge on the social learning process, responded to AkersÂ’ (1999) plea to hel p specify the most underdeveloped portion of the model. Although finding a relationship between social structure and social learning, the study found no support for AkersÂ’ (1998) description of the relationship as mediation. The study in stead found support for several moderator hypotheses, concluding that AkersÂ’ model requires modification. Reconciling the discrepancies of the pr esent research with previous tests of AkersÂ’ (1998) model, the present res earch explored a t heoretical argument that links social structure to soci al learning through the mechanisms of macrosocial reinforcement contingenc ies. The study argued that such an explanation accounts for the findings in the present research (moderation) and the findings in the literatur e (mediation). The study offered a reconceptualization of the model such that social struct ure is viewed as influencing individual behavior by sets of reinforcement contingenc ies that are transmitted to the social
274 learning process through meso-level groups. The implications of the present study suggest that future research should focus on distinguishing macrosocial struct ures from meso-level groups most likely to have the most impact on the so cial learning process. Although the present study suggests that ma crosocial structure interact s with social learning to affect delinquency, and it argued that soci al learning mediates the effects of meso-level structure on individual delinq uency, the study further argued that the mechanisms by which these structur es impinge on individual behavior, macrosocial reinforcement contingencies influencing individual reinforcement schedules, might work dichotomously. The study suggests that the proximity of the social structural contingencies of rein forcement in relation to the translating macro-level structures is important, and t hat this distinction needs attention in future tests of the model.
275 References Akers, R.L. (1968). Problems in the sociol ogy of deviance: Social definitions and behavior. Social Forces, 46 (4), 455-465. Akers, R.L. (1973). Deviant behavior: A social learning approach. Belmont, CA: Wadsworth Publishing Company. Akers, R.L. (1977). Deviant behavior: A social learning approach (2nd ed) Belmont, CA: Wadsworth Publishing Company. Akers, R.L. (1985). Deviant behavior: A social learning approach (3rd ed) Belmont, CA: Wadsworth Publishing Company. Akers, R.L. (1989). A social behavioristÂ’s perspective on integration of theories of crime and deviance. In S.F. Messner, M.D. Krohn, & A.E. Liska (Eds.), Theoretical integration in t he study of deviance and crime (pp. 23-36). Albany, NY: State University of New York Press. Akers, R.L. (1990). Rational choice, dete rrence, and social learning theory in criminology: The path not taken. Journal of Criminal Law and Criminology, 81 (3), 653-676. Akers, R.L. (1992). Drugs, alcohol, and society: So cial structure, process, and policy. Belmont, CA: Wadsworth Publishing Company. Akers, R.L. (1998). Social learning and social stru cture: A general theory of crime and deviance Boston: Northeastern University Press. Akers, R.L. (1999). Social learning and social structure: Reply to Sampson, Morash, and Krohn. Theoretical Criminology, 3 (4), 477-493. Akers, R.L., & Cochran, J.K. (1985). Adol escent marijuana use: A test of three theories of deviant behavior. Deviant Behavior, 6 323-346. Akers, R.L., & La Greca, A.J. (1991). Alcohol use among the elderly: Social learning, community context, and life events. In D.J. Pittman, & H.R. White (Eds.), Society, culture, and drinking patterns re-examined (pp. 242-262).
276 New Brunswick, NJ: Rutgers Center of Alcohol Studies. Akers, R.L., & Lee, G. (1996) A longitudinal test of social learning theory: Adolescent smoking. Journal of Drug Issues, 26 (2), 317-343. Akers, R.L., & Lee, G. (1999) Age, social learning, and social bonding in adolescent substance use. Deviant Behavior, 20 (1), 1-25. Akers, R.L., & Sellers, C.S. (2004). Criminological theories: Introduction, evaluation, and application (4th ed.). Los Angeles: Roxbury. Akers, R.L., Krohn, M.D., Lanza-Kaduce, L., & Radosevich, M. (1979). Social learning and deviant behavior: A specific test of a general theory. American Sociological Review, 44 636-655. Akers, R.L., La Greca, A.J., Cochran, J.K., & Sellers, C. S. (1989). Social learning theory and alcohol behav ior among the elderly. Sociological Quarterly, 30 625-638. Alarid, L.F., Burton, V.S., Jr., & Culle n, F.T. (2000). Gender and crime among felony offenders: Assessing the generalit y of social control and differential association theories. Journal of Research in Crime and Delinquency, 37 (2), 171-199. Allen, M., Donohue, W.A., Griffin, A., Ryan, D., & Mitchell-Turner, M.M. (2003). Comparing the influence of parents and peers on the choice to use drug: A meta-analytic summary of the literature. Criminal Justice and Behavior, 30 (2), 163-186. Allison, J.P. (1972). Economic fa ctors and the rate of crime. Land Economics, 48 193-196. American Psychological Association [APA]. (1992). Ethical principles of psychologists Washington, DC: American Psychological Association. Anderman, C., Cheadle, A., Curry, S. Diehr P., Shultz, L., & Wagner, E. (1995). Selection bias related to parental consent in school-based survey research. Evaluation Review, 19 663-674. Anderson, J.C., & Gerbing, D.W. ( 1988). Structural equ ation modeling in practice: A review and recommended two-step approach. Psychological Bulletin, 103 (3), 411-423.
277 Archer, D., Gartner, R., Aker t, R., & Lockwood, T. (1978). Cities and homicide: A new look at an old paradox. Comparative studies in sociology, 1 73-95. Arnold, H.J. (1982). Moderat or variables: A clarificat ion of conceptual, analytic, and psychometric issues. Organizational Behavior and Human Performance, 29 143-174. Arnold, H.J. (1984). Testing moderator variable hypotheses: A reply to Stone and Hollenbeck. Organizational Behavior and Human Performance, 34 214224. Asher, H.B. (1988). Causal modeling (2nd ed.). Beverly Hills, CA: Sage. Bailey, W.C. (1984). Poverty, i nequality, and city homicide rates. Criminology, 22 531-550. Bailey, W.C. (1999). The socioeconomic st atus of women and patterns of forcible rape for major U.S. cities. Sociological Focus, 32 43-63. Ball, J.C. (1957). Delinquent and Non-delinquent attitudes toward the prevalence of stealing. Journal of Criminal Law, Cr iminology, and Police Science, 48 259-274. Bandura, A. (1977). Social learning theory Englewood Cliffs, NJ : Prentice-Hall, Inc. Baron, R.M., & Kenny, D.A. (1986). The moderator-mediator variable distinction in social psychological research: C onceptual, strategi c, and statistical considerations. Journal of Personality and Social Psychology, 51 (6), 11731182. Baum, S. (1999). An aggregate level analysi s of the socioeconomic correlates of drink driving offenders. Accident analysis and prevention, 31 213-220. Bauman, K.E., Foshee, V.A., Linzer, M.A., & Koch, G.G. (1990). Effect of parental smoking on the associati on between parental and adolescent smoking. Addictive Behaviors, 15 413-422. Bellair, P.E. (1997). Social interact ion and community crim e: Examining the importance of neighbor networks. Criminology, 35 677-703. Bellair, P.E., Roscigno, V.J., & Vlez, M. B. (2003). Occupational structure, social learning, and adolescent violence. In R.L. Akers, & G.F. Jensen (Eds.),
278 Social learning theory and the explanati on of crime: A guide for the new century, volume 11 advances in criminological theory (pp. 197-225). New Brunswick, NJ: Transaction Publishers. Benda, B.B. (1994). Testing competing theoretical conc epts: Adolescent alcohol consumption. Deviant Behavior, 15 375-396. Benda, B.B., & Corwyn, R.F. (2002). The effect of abus e in childhood and in adolescence on violence among adolescents. Youth and Society, 33 (3), 339-365. Bentler, P.M. (1983). Some contributions to efficient statistics in structural models: Specifications and esti mation of moment structures. Psychometrika, 48 493-517. Bentler, P.M. (1989). EQS structural equati ons program manual Los Angeles: BMDP Statistical Software. Bentler, P.M., & Bonett, D.G. (1980). Si gnificance tests and goodness-of-fit in the analysis of covariance structures. Psychological Bulletin, 88 588-606. Bentler, P.M., & Chou, C.P. (1987). Practical issues in structural modeling. Sociological Methods and Research, 16 (1), 78-117. Bergesen, A., & Herman, M. (1998). Immi gration, race, and riot: The 1992 Los Angeles uprising. American Sociological Review, 63 39-54. Bernard, T.J. (1990). Twenty years of testing theories. Journal of Research in Crime and Delinquency, 24 (4), 325-347. Bernard, T.J. (2001). Integrat ing theories in criminology. In R. Paternoster & R. Bachman (Eds.), Essays in contemporary criminol ogical theory: Explaining criminals and crime (pp. 335-346). Los Angeles: Roxbury Publishing. Bernard, T.J., & Ritti, R. (1990). The ro le of theory in scientific research. In K. Kempf (Ed.), Measurement issues in criminology (pp. 1-20). New York: Springer-Verlag. Bernard, T.J., & Snipes, J.B. (1996). Theoretical integrat ion in criminology. In M. Tonry (Ed.), Crime and justice: A review of the research, volume 20 (pp. 301-348). Chicago: University of Chicago Press. Blalock, H.M. (1984). Cont extual-effects models: Theor etical and methodological
279 issues. Annual Review of Sociology, 10 (2), 353-372. Blau, P.M. (1960). St ructural effects. American Sociological Review, 25 (2) 178193. Blau, P.M. (1977). Inequality and heterogene ity: A primitive theory of social structure New York: Free Press. Blau, J., & Blau, P.M. ( 1982). The cost of inequality: Metropolitan structure and violent crime. American Sociological Review, 47 114-129. Bollen, K.A. (1989). Structural equations with latent variables New York: Wiley. Bollen, K.A., & Paxton, P. ( 1998). Interactions of latent variables in structural equation models. Structural Equation Modeling, 5 267-293. Brady, D. (2003). Rethinking the soci ological measurement of poverty. Social Forces, 81 (3), 715-752. Braithwaite, J. (1989). Crime, shame, and reintegration. New York: Cambridge University Press. Braumoeller, B.F. (2004). Hy pothesis testing and multiplicative interaction terms. International Organization, 58 (4), 807-820. Brick, J.M., & Kalton, G. (1996). Handling missing dat a in survey research. Statistical Methods in Medical Research, 5 215-238. Browne, M.W. (1984). Asympt otically distribution-free methods for analysis of covariance structures. British Journal of Mathematical and Statistical Psychology, 37 62-83. Brownfield, D., & Thompson, K. (2002) Distinguishing the effects of peer delinquency and gang membership on self-reported delinquency. Journal of Gang Research, 9 (2), 1-10. Bryk, A., & Raudenbus h, S.W. (1992). Hierarchical linear models for social and behavioral research: Applicat ions and data analysis methods. Newbury Park, CA: Sage. Burgess, R.L., & Akers, R.L. (1966). A differential association-reinforcement theory of criminal behavior. Social Forces, 14 128-147.
280 Bursik, R.J., Jr., (1988). Social di sorganization and theories of crime and delinquency: Problems and prospects. Criminology, 26 (4), 519-551. Bursik, R.J., Jr., & Grasmick, H.G. ( 1992). Longitudinal neighb orhood profiles in delinquency: The decom position of change. Journal of Quantitative Criminology, 8 247-263. Bursik, R.J., Jr., & Grasmick, H.G. (1993). Neighborhoods and crime: The dimensions of the effe ctive community control. New York: Lexington Books. Bursik, R.J., Jr., & Grasmick, H.G. ( 1996). The use of contextual analysis in models of criminal behavior. In J.D. Hawkins (Ed.), Delinquency and crime: Current theories (pp. 236-267). New York: Cambridge University Press. Bursik, R.J., & Webb, J. (1982). Community change and patterns of delinquency. American Journal of Sociology, 88 24-42. Burton, V., Cullen, F., Evans, D., & Dunaw ay, R.G. (1994). Reconsidering strain theory: Operationalization, rival theories, and adult criminality. Journal of Quantitative Criminology, 10 213-239. Byrne, J.M. (1986). Cities, citizens, and crime: The ecological/nonecological debate reconsidered. In R.J Sa mpson & J.M. Byrne (Eds.), The social ecology of crime (pp.77-101). New York: Springer-Verlag. Byrne, J.M., & Sampson, R.J. (1986). Key issues in the social ecology of crime. In R.J Sampson & J.M. Byrne (Eds.), The social ecology of crime (pp.122). New York: Springer-Verlag. Cameron, A.C., & Trivedi, P.K. (1998). Regression analysis of count data New York: Cambridge University Press. Chamlin, M.B. ( 1989). Conflict theory and police killings. Deviant Behavior, 10 353-368. Chappell, A.T., & Piquero, A. R. (2004). Applying social learning theory to police misconduct. Deviant Behavior, 25 89-108. Chiricos, T.G. (1987). Rates of cr ime and unemployment: An analysis of aggregate research evidence. Social Problems, 34 187-212.
281 Chunrong, A., & Norton, E.C. (2003). Interaction terms in logit and probit models. Economic Letters, 80 123-129. Cleary, P.D., & Kessler, R.C. (1982). The estimation and interpretation of modified effects. Journal of Health and Social Behavior, 23 159-169. Cloward, R.A. (1959). Illegitimate means, anomie, and deviant behavior. American Sociological Review, 24 (2), 164-176. Cohen, A.K. (1962). Multiple factor approac hes. In M.E. Wolfgang, L. Savitz, & N. Johnston (Eds.), The sociology of crime and delinquency (pp.77-80). New York: John Wiley. Cohen, J., & Cohen, P. (1983). Applied multiple regre ssion/correlation analysis for the behavioral sciences (2nd ed.). Hillside, NJ: Erlbaum. Cohen, L.E., & Land, K.C. (1987). Age structure and crime: Symmetry versus asymmetry and the projection of crime rates through the 1990s. American Sociological Review, 52 170-183. Comrey, A.L. (1978). Common methodolog ical problems in factor analysis studies. Journal of Consulting and Clinical Psychology, 46 648-659. Conger, R.D. (1976). Social control and social learning models of delinquent behavior: A synthesis. Criminology, 14 (1), 17-40. Conway, K.P., & McCord, J. (2002). A longi tudinal examination of the relation between co-offending with violent a ccomplices and violent crime. Aggressive Behavior, 28 97-108. Conway, R.W., & McClain, J.O. (2003). The conduct of an effective simulation study. Transactions on Education, 3 (3), 13-2. Copes, H. (1999). Routine activities and motor vehicle theft: A crime specific approach. Journal of Crime and Justice, 22 125-146. Cortina, J.M. (1993). What is coeffici ent alpha? An examination of theory and applications. Journal of Applied Psychology, 78 (1), 98-104. Cressey, D.R. (1960). Ep idemiology and individual conduct: A case from criminology. The Pacific Sociological Review, 3 (2), 47-58. Cronbach, L.J. (1951). Coefficient alpha and the internal stru cture of tests.
282 Psychometrika, 16 (3), 297-334. Cronbach, L.J. (1976). Research on classrooms and schools: Formulation of questions, design and analysis Stanford, CA: Stanford University Evaluation Consortium. Crutchfield, R.D., Garken, M.R., & Grove, W.R. ( 1982). Crime rates and social integration: The impact of metropolitan mobility. Criminology, 20 467-478. Curran, P.J., West, S.G., & Fi nch, J.F. (1996). The robust ness of test statistics to nonnormality and specification error in confirmatory factor analysis. Psychological Methods, 1 (1), 16-29). Curry, G.D., & Spergel, I.A. ( 1988). Gang homicide, delinquency, and community. Criminology, 26 381-405. Davis, J.A., Spaeth, J.L., & Huson, C. (1961). A technique for analyzing the effects of group composition. American Sociological Review, 26 (2), 215225. DeCarlo, L.T. (1997). On the meaning of kurtosis. Psychological Methods, 2 (3), 292-307. DeMaris, A. (2004). Regression with social data: M odeling continuous and limited response variables Hoboken, NJ: Wiley & Sons. Dembo, R., Grandon, G., La Voie, L., Schmeidler, J., & Burgos, W. (1986). Parents and drugs revisited: Some fu rther evidence in support of social learning theory. Criminology, 24 (1), 85-104. Dewey, J. (1931). Social science and social control. New Republic, 67 276-277. Diez-Roux, A.V. (1998). Bringing context back into epidemiology: Variables and fallacies in multilevel analysis. American Journal of Public Health, 88 (2), 588-594. Diez-Roux, A.V. (2003). A glossa ry for multilevel analysis. Journal of Epidemiology and Community Health, 56 588-594. Durkheim. E. (2002). Suicide. In K. Thompson (Ed.), Readings from Emile Durkheim (pp. 91-115). New York: Routledge. (Original work published
283 1897) Elliott, D.S. (1985). The assumption t hat theories can be combined with increased explanatory power. In R.F. Meier (Ed.), Theoretical methods in criminology (pp. 123-149). Beverly Hills, CA: Sage. Elliott, D.S., Ageton, S.S., & Cantor, R. J. (1979). An integrated theoretical perspective on delinquent behavior. Journal of Research in Crime and Delinquency, 16 (1), 3-27. Elliott, D.S., Huizinga, D ., & Ageton, S.S. (1985). Explaining delinquency and drug use. Beverly Hills, CA: Sage Publications. Elliott, D.S., & Menard, S. (1996). De linquent friends and delinquent behavior: Temporal and developmental patterns. In J.D. Hawkins (Ed.), Delinquency and crime: Current theories (pp. 28-67). New York: Cambridge University Press. Erickson, M.L. & Jensen, G.F. (1997) Delinquency is still group behavior: Toward revitalizing the group premise in the sociology of deviance. Journal of Criminal Law and Criminology, 88 262-273. Fan, X., & Wang, L. (1998). Effects of potential confoundi ng factors on fit indices and parameter estimates for true and misspecified models. Structural Equation Modeling, 5 701-735. Farnworth, M. (1989). Theory integration versus model bui lding. In S.F. Messner, M.D. Krohn, & A.E. Liska (Eds.), Theoretical integratio n in the study of deviance and crime (pp. 93-100). Albany, NY : STATE UNIVERSITY OF NEW YORKPress. Farrell, A.D., & Danish, S.J. (1993). P eer drug associations and emotional restraint: Causes or consequences of adolescentsÂ’ drug use? Journal of Consulting and Clinical Psychology, 61 (2), 327-334. Farrington, D.P., Loeber, R., Stoutham er-Loeber, M., Van Kammen, W.R., & Schmidt, L. (1996). Self-repor ted delinquency and a combined delinquency seriousness scale based on boys, mothers, and teachers: Concurrent and predictive validity for African-Americans and Caucasians. Criminology, 34 (4), 493-517. Finch, J.F., West, S.G., & MacKinnon, D. P. (1997). Effects of sample size and
284 nonnormality on the estimation of medi ated effects in latent variable models. Structural Equation Modeling, 2 87-105. Findley, M.J., & Cooper, H.M. (1983) Locus of control and academic achievement: A literature review. Journal of Personality and Social Psychology, 44 419-427. Fleisher, B.M. (1966). The effe ct of income on delinquency. American Economic Review, 56 118-137. Flom, P.L., Friedman, S.R., Kottiri, B. J., Neaigus, A., & Curtis, R. (2001). Recalled adolescent peer norms towa rds drug use in young adulthood in a low-income, minority urban neighborhood. Journal of Drug Issues, 31 (2), 425-443. Florida Department of Education. (2003). Florida school indicators report [Data file]. Available at http://www.fldoe.org Florida Department of Law Enforcement. (1999). Crimes and crime rates statistics [Data file]. Available at http ://fl.rand.org/cgi-bin/annual.cgi Ford, J.K., MacCallum, R.C., & Tait, M. (1986). The applicat ion of exploratory factor analysis in applied psychology: A critical review and analysis. Personnel Psychology, 39 291-314. Franzblau, A.N. (1958). A primer of statistics for non-statisticians. New York: Harcourt, Brace, and World. Friedrich, R.J. (1982). The workshop: In defense of multiplicative terms in multiple regression equations. American Journal of Political Science, 26 (4), 797-833. Galtung, J. (1969). Theories and methods of social research. New York: Columbia University Press. Gardner, W., Mulvey, E.P., & Shaw, E.S. (1995). Regression analyses of counts and rates: Poisson, overdisper sed Poisson, and negative binomial models. Psychological Bulletin, 118 (3), 392-404. Gartner, R., Baker, K., & Pampel, F.C. (1990). Gender stra tification and the gender gap in homicid e victimization. Social Problems, 37 593-612. Gauthier, D.K., & Bankst on, W.B. (1997). Gender equa lity and the sex ratio of
285 intimate killing. Criminology, 35 577-600. Gibbs, J.P. (1972). Sociological theory construction. Hinsdale, IL: Dryden Press. Gibbs, J.P., & Erickson, M.L. (1976). Crime rates of Amer ican cities in ecological context. American Journal of Sociology, 82 605-620. Glaser, D. (1954). A rec onsideration of some par ole prediction factors. American Sociological Review, 19 335-351. Glaser, D. (1960). Differential associat ion and criminological prediction. Social Problems, 8 6-14. Glaser, D., & Rice, K. (1959). Crime, age, and employment. American Sociological Review, 24 (5), 679-686. Glueck, S. (1956). Theory and fact in criminology. British Journal of Delinquency, 7 92-109. Glueck, S., & Glueck, E. (1950). Unraveling juvenile delinquency. Cambridge, MA: Harvard University Press. Gold, M. (1970). Delinquent behavior in an American city. Belmont, CA: Wadsworth. Gordon, D.M. (1972). Theories of poverty and underemployment: Orthodox, radical, and dual market perspectives Lexington, MA: Lexington Books. Gottfredson, D., McNeil, R.J., & Gottfredson, G.D. (1991). Social area influences on delinquency: A mu ltilevel analysis. Journal of Research in Crime and Delinquency, 28 (2), 197-226. Gottfredson, M.R., & Hirschi, T. ( 1987). The methodological adequacy of longitudinal research on crime. Criminology, 25 (3), 581-614. Gottfredson, M.R., & Hirschi, T. (1990). A general theory of crime. Stanford, CA: Stanford University Press. Greenberg, D.E. (1985). Age, crim e, and social explanation. American Journal of Sociology, 91 1-12. Guadagnoli, S.W., & Velicer, W. (1988). Relation of sample size to the stability of component patterns. Psychological Bulletin, 103 265-275.
286 Hagan, J., & McCarthy, B. (1998). Mean streets: Youth cr ime and homelessness NY: Cambridge University Press. Hagenaars, A.J.M. (1991). The defin ition and measurement of poverty. In L. Osberg (Ed.), Economic inequality and poverty: International perspectives (pp. 134-156). Armonk, New York: M.E. Sharpe. Hair, J.F., Jr., Anderson, R.E., Ta tham, R.L., & Bla ck, W.C. (1998). Multivariate data analysis (5th ed.). Upper Saddle Riv er, NJ: Prentice Hall. Hamblin, R.L. (1979). Behavio ral choice and social rein forcement: Step function versus matching. Social Forces, 57 (4), 1141-1156. Hannan, M.T. (1971). Aggregation and disaggregations in sociology Lexington, MA: Heath-Lexington. Hannan, M.T. (1985). Problems of aggregation. In H.M. Blalock, Jr. (Ed.), Causal models in the social sciences (2nd ed.) (pp. 403-439). New York: Aldine. Harkins, S.G., Latane, B., & Williams, K. ( 1980). Social loafing: Allocating effort or taking it easy? Journal of Experimental Social Psychology, 16 457-465. Hartung, F.E. (1965). Crime, law, and society Detroit, MI: Wayne State University Press. Hatcher, L. (1994). A step-by-step approach to using the SAS system for factor analysis and structur al equation modeling. Cary, NC: SAS Institute. Hauser, R.M. (1970). Context and consex: A cautionary tale. American Journal of Sociology, 75 (4, Part 2), 645-664. Haynie, D.L. (2002). Friendship networks and delinquency: The relative nature of peer delinquency. Journal of Quantitative Criminology, 18 (2), 99-134. Heitgard, J.L., & Bursik, R.J. (1987). Ex tracommunity dynamics and the ecology of delinquency. American Journal of Sociology, 92 775-787. Herrnstein, R.J. (1974). Formal properties of the matching law. Journal of the Experimental Analysis of Behavior, 21 159-164.
287 Herrnstein, R.J., & Loveland, D.H. (1975). Maximizing and matching on concurrent ratio schedules. Journal of the Experimental Analysis of Behavior, 24 107-116. Hinkle, D.E., Wiersma, W., & Jurs, S.G. (1988). Applied statistics for the behavioral sciences (2nd. ed.). Boston: Houghton Mifflin. Hirschi, T. (1969). Causes of delinquency Berkeley, CA: University of California Press. Hirschi, T. (1979). Separat e but unequal is better. Journal of Research in Crime and Delinquency, 16 (1), 34-38. Hirschi, T. (1989). Exploring alternatives to integrated theory. In S.F. Messner, M.D. Krohn, & A.E. Liska (Eds.), Theoretical integratio n in the study of deviance and crime (pp. 37-49). Albany, NY: Stat e University of New York Press. Hirschi, T., & Gottfredson, M. (1983) Age and the explanation crime. American Journal of Sociology, 89 553-584. Hirschi, T., & Selvin, H.C. (1967). Delinquency research: An appraisal of analytic methods New York: Free Press. Hoffmann, J.P. (2002). A contex tual analysis of different ial association, social control, and strain theories of delinquency. Social Forces, 81 (3), 753-785. Holmbeck, G. N. (1997). Toward terminologica l, conceptual, and statistical clarity in the study of mediators and moderato rs: Examples from the child-clinical and pediatric psychology literatures. Journal of Consulting and Clinical Psychology, 65 (4), 599-610. Hox, J.J., & Kreft, I.G.G. ( 1994). Multilevel analysis methods. Sociological Methods and Research, 22 (3), 283-299. Hu, L., & Bentler, P.M. (1998). Fit indi ces in covariance structure modeling: Sensitivity to underparameteriz ed model misspecification. Psychological Methods, 3 (4), 424-453. Hu, L., & Bentler, P.M. (1999). Cutoff cr iteria for fit indexes in covariance
288 structure analysis: Conv entional criteria versus new alternatives. Structural Equat ion Modeling, 6 (1), 1-55. Huizinga, D., &Elliott, D.S. (1986). Reassessi ng the reliability and validity of selfreport delinquency measures. Journal of Quantitat ive Criminology, 2 293327. Huizinga, D., Esbensen, F.A., & Weiher, A. W, (1991). Are there multiple paths to delinquency? Journal of Criminal Law and Criminology, 82 83-118. Iannotti, R.J., & Bush, P.J. (1992). Percei ved vs. actual friendsÂ’ use of alcohol, cigarettes, marijuana, and cocaine: Which has the most influence? Journal of Youth and Adolescence, 21 375-389. Jaccard, J., & Wan, C.K. ( 1995). Measurement error in the analysis of interaction effects between continuous predictors us ing multiple regression: Multiple indicator and structural equation approaches. Psychological Bulletin, 117 (2), 348-357. Jaccard, J., & Wan, C.K. (1996). LISREL approaches to interaction effects in multiple regression Thousand Oaks, CA: Sage. Jackson, P.I. (1984). Opportunities and crime: A function of city size. Sociology and Social Research, 68, 173-193. James, L.R., & Brett, J.M. (1984). Mediators, moderat ors, and tests for mediation. Journal of Applied Psychology, 69 (2), 307-321. Jang, S.J. (1999). Age-varying effe cts of family, school, and peers on delinquency: A multilevel modeling te st of interactional theory. Criminology, 37 (3), 643-696 Jang, S.J. (2002). The effects of fa mily, school, peers, and attitudes on adolescentsÂ’ drug use: Do they vary with age? Justice Quarterly, 19 (1), 97-127. Jaquith, S.M. (1981). Adolescent marijuana and alcohol use: An empirical test of differential association theory. Criminology, 19 (2), 271-280. Jeffery, C.R. (1965). Criminal behavior and learning theory. The Journal of Criminal Law, Criminol ogy and Police Science, 56 (3), 294-300. Jessor, R., Jessor, S.L., & Finney, J. (1973). A social psychology of marijuana
289 use: Longitudinal studies of high school and college youth. Journal of Personality and Social Psychology, 26 1-15. Johnson, R.E., Marcos, A.C., & Bahr, S. J. (1987). The role of peers in the complex etiology of adolescent drug use. Criminology, 25 (2), 324-339. Johnson, V. (1988). Adolescent alc ohol and marijuana use: A longitudinal assessment of a social learning perspective. American Journal of Drug and Alcohol Abuse, 14 419-439. Johnstone, J.W.C. (1978). Social cl ass, social areas and delinquency. Sociology and Social Research, 63 49-77. Joreskog, K.G., & Sorbom, D. (1984). LISREL VI Mooresville, IN: Scientific Software, Inc. Joreskog, K.G., & Sorbom, D. (1989). LISREL 7 : A guide to the program and applications (2nd ed.). Chicago: SPSS Inc. Joreskog, K.G., & Yang, F. (1996). Non linear structural equation models: The Kenny-Judd model with interaction effe cts. In G.A. Marcoulides, & R.E. Schumacker (Eds.), Advanced structural equation modeling: Issues and techniques (p. 239-250). Mahwah, NJ: Erlbaum. Judd, C.M., & Kenny, D.A. (1981). Estimating the effects of social interventions New York: Cambridge University Press. Judd, C.M., Kenny, D.A., & McClella nd, G.H. (2001). Estimating and testing mediation and moderation in within-subject designs. Psychological Methods, 6 (2), 115-134. Julian, M.W. (2001). The cons equences of ignoring multile vel data structures in nonhierarchical covariance modeling. Structural Equation Modeling, 8 (3), 325-352. Kaiser, H.F. (1960). The appl ication of electronic comp uters to factor analysis. Educational and Psychological Measurement, 20 141-151. Kalton, G., & Kasprzyk, D. (1986). The treatment of missing survey data. Survey Methodology, 12 (1), 1-16. Kandel, D.B. (1996). The par ental and peer contexts of adolescent deviance: An algebra of interpersonal influences. Journal of Drug Issues, 26 289-315.
290 Kandel, D.B., & Andrews, K. (1987). Pr ocess of adolescent socialization by parents and peers. International Journal of the Addictions, 22 319-342. Kandel, D.B., & Davies, M. (1991). Friendship networks, in timacy, and illicit drug use in young adulthood: A compar ison of two competing theories. Criminology, 29 (3), 441-470. Kaplan, D. (2000). Structural equation modeling: Foundations and extensions. Thousand Oaks, CA: Sage. Kaplan, H.B., Martin, S.S., & Robbins, C. (1984). Pathways to adolescent drug use: Self-derogation, peer influence, weakening of social controls, and early substance use. Journal of Health and Social Behavior, 25 270-289. Kapuskinski, C.A., Braithwaite, J., & Chapman, B. (1998). Unemployment and crime: Toward resolving the paradox. Journal of Quantitative Criminology, 14 215-244. Kasarda, J.D., & Janowitz, M. (1974). Community attachment in mass society. American Sociological Review, 39 (3), 328-339. Kempf, K.L. (1993). The empirical st atus of HirschiÂ’ s control theory. In F. Adler and W.S. Laufer (Eds.), New directions in criminological theory, volume 4 advances in criminological theory (pp. 143-185). New Brunswick, NJ: Transaction Publishers. King, G. (1988). Statistical models for po litical science event counts: Bias in conventional procedures and evidenc e for the exponential Poisson regression model. American Journal of Political Science, 32 838-863. King, D.W., & King, L.A. (1997). A brief introduction to structural equation modeling. PTSD Research Quarterly, 8 (4), 1-4. Klein, A., & Moosbrugger, H. (2000). Maximum likelihood estimation of latent interaction effects with the LMS method. Psychometrika, 65 457-474. Kline, R.B. (1998). Principles and practices of structural equation modeling. New York: Guilford Press. Kline, R.B. (2005). Principles and practices of structural equation modeling (2nd ed.) New York: Guilford Press.
291 Kornhauser, R.R. (1978). Social sources of delinquency: An appraisal of analytic models Chicago: University of Chicago Press. Kraemer, H.C., Stice, E., Kazdin, A., O fford, D., & Kupfer, D. (2001). How do risk factors work together? Mediat ors, moderators, and independent, overlapping, and proxy risk factors. American Journal of Psychiatry, 158 (6), 848-856. Kreft, I.G.G. (1996). Are multilevel techniques necessary? An overview, including simulation studies [on-line]. Los Angeles, CA; California State University. Available at http://www.calstatela.edu /faculty/ikreft/quarterly/quarterly.html Kreft, I., & de Leeuw, J. (1998). Introducing multilevel modeling Thousand Oaks, CA: Sage. Krivo, L.J., & Peterson, R.D. (1996). Extremely disadvantaged neighborhoods and urban crime. Social Forces, 75 619-650. Krohn, M.D. (1986). The web of conf ormity: A network approach to the explanation of delinquent behavior. Social Problems, 33 (6), 81-93. Krohn, M.D. (1999). Social learning theory: The cont inuing development of a perspective. Theoretical Criminology, 3 (4), 462-476. Krohn, M.D., Lanza-Kaduce, L., & Akers, R.L. (1984). Comm unity context and theories of deviant behavior: An examination of soci al learning and social bonding theories. Sociological Quarterly, 25 353-372. Krohn, M.D., Lizotte, A.J., Thornberry, T.P. Smith, C., & McDo wall, D. (1996). Reciprocal causal relationships among drug use, peers, and beliefs: A five-wave panel model. Journal of Drug Issues, 26 405-428. Krohn, M.D., Skinner, W.F., Massey, J.L., & Akers, R.L. (1985) Social learning theory and adolescent cigarette sm oking: A longitudinal study. Social Problems, 32 455-473. Krull, J.L., & MacKinnon, D.P. (2001). Multilevel mode ling of individual and group level mediated effects. Multivariate Behavioral Research, 36 (2), 249-277. La Du, T.J., & Tanaka, J.S. (1989). The influence of sample size, estimation method, and model specification on goodness-of-fit assessments in structural equation models. Journal of Applied Psychology, 71 625-636.
292 Lakatos, I. (1978). The methodology of scientif ic research programmes London, UK: Cambridge University Press. Land, K.C., McCall, P.L., & C ohen, L.E. (1990). Structural covariates of homicide rates: Are there any invariances across time and social space? American Journal of Sociology, 95 (4), 922-963. Lanza-Kaduce, L., & Capece. M. (2003). Soci al structure-social learning (SSSL) and binge drinking: A specific test of an integrated theory. In R.L. Akers and G. F. Jensen (Eds.), Social learning theory and the explanation of crime: A guide for the new century, vo lume 11 advances in criminological theory (pp. 179-196). New Brunswick, NJ: Transaction Publishers. Lanza-Kaduce, L., & Klug, M. (1986). Learni ng to cheat: The interaction of moraldevelopment and social learning theories. Deviant Behavior, 7 243-259. Lanza-Kaduce, L., Akers, R.L., Krohn, M.D., & Radosevich, M. (1982). Conceptual and analytical models in test ing social learning theory: Reply to Stafford and Ekland -Olson and Strickland. American Sociological Deviant Behavior, 5 79-96. Lanza-Kaduce, L., Akers, R.L., Krohn, M. D., & Radosevich, M. (1984). Cessation of alcohol and drug use among adolesce nts: A social learning model. American Sociological Review, 47 (1), 169-173. Largo Chamber of Commerce. (1998). Largo demographics [Data file]. Available at http://www.largochamber.com Lazarsfeld, P.F., & Menzel, H. (1961). On the relation between individual and collective properties. In A. Etzioni (Ed.), Complex organizations (pp. 422440). New York: Rinehart & Winston. Lei, M., & Lomax, R.G. (2005). The effect of varying degrees of nonnormality in structural equation modeling. Structural Equation Modeling, 12 (1), 1-27. Lee, G., Akers, R.L., & Borg, M.J. (2004). Social learning and structural factors in adolescent substance use. Western Criminology Review, 5 (1), 17-34. Lewis, D.A., & Salem, G. (1981). Community crime prevention: An analysis of a developing strategy. Crime and Delinquency, 27 405-421. Lewis, C.J., Sims, L.S., & Shannon, B. (1989). Examination of specific nutrition/health behaviors using a social cognitive model. Journal of the
293 American Dietetic Association, 89 (2), 194-202. Lindesmith, A.R. (1938). A sociologi cal theory of drug addiction. American Journal of Sociology, 43 593-613. Liska, A.E. (1969). Uses and misuses of tautologies in social psychology. Sociometry, 32 (4), 444-457. Liska, A.E., Krohn, M.D., & Messner, S.F. (1989). Strategies and requisites for theoretical integration in the study of crime and deviance. In S.F. Messner, M.D. Krohn, & A.E. Liska (Eds.), Theoretical integratio n in the study of deviance and crime (pp. 1-20). Albany, NY: Stat e University of New York Press. Liska, A.E., Logan, J.R., & Bellair, P.E. (1998). Race and violent crime in the suburbs. American Sociological Review, 63 27-38. Loehlin, J.C. (1992). Latent variable models: An in troduction to factor, path, and structural analysis (2nd ed.). Hillsdale, NJ: Lawrence Erlbaum. Loftin, C., & Hill, R.H. ( 1974). Regional subculture and homicide: An examination of the Gastil-Hackney thesis. American Sociological Review, 39 714-724. Long, J.S. (1997). Regression models for categor ical and limited dependent variables Thousand Oaks, CA: Sage. Lowenkamp, C.T., Cullen, F.T., & Pratt, T.C. (2003). Replic ating Sampson and GrovesÂ’s test of social disorganizati on theory: Revisiting a criminological classic. Journal of Research in Crime and Delinquency, 40 (4), 351-373. MacKinnon, D.P., Lockwood, C.M., Hoffm an, J.M., West, S.G., & Sheets, V. (2002). A comparison of methods to te st mediation and other intervening variable effects. Psychological Methods, 7 (1), 83-104. Marcos, A.C., Bahr, S.J., & Johnson, R.E. (1986). Test of a bonding/association theory of adolescent drug use. Social Forces, 65 (1), 135-161. Mardia, K.V., Kent, J.T ., & Bibby, J.M. (1979). Multivariate analysis New York: Academic. Matsueda, R.L. (1988). The cu rrent state of different ial association theory. Crime & Delinquency, 34 (3), 277-306.
294 Matsueda, R.L., & Anderson, K. (1998) The dynamics of delinquent peers and delinquent behavior. Criminology, 36 (2), 269-308. Matsueda, R.L., & Heimer, K. (1987). Race family structure, and delinquency: A test of differential associati on and social control theories. American Sociological Review, 52 826-840. Maly, M.T. (2000). The nei ghborhood diversity index: A complementary measure of racial residential settlement. Journal of Urban Affairs, 22 (1), 37-47. McDonald, R.P., Ho, M.R. ( 2002). Principles and practice in reporting structural equation analyses. Psychological Methods, 7 (1), 64-82. McKay, H.D. (1960). Differential associ ation and crime prevent ion: Problems of utilization. Social Problems, 8 (1), 25-37). Mead, G.H. (1934). Mind, self, and society. Chicago: University of Chicago Press. Mencken, F.C., & Barnett, C. (1999) Murder, nonnegligent manslaughter, and spatial autocorrelation in mid-South counties. Journal of Quantitative Criminology, 15 407-422. Messner, S.F. (1982). Poverty, i nequality, and the urban homicide rate. Criminology, 20 1030115. Messner, S.F., & Sampson, R.J. (1991). T he sex ratio, family disruption, and rates of violent crime. Social Forces, 69 (3), 693-713. Michaels, J.W., & Miethe, T. D. (1989). Applying theories of deviance to academic cheating. Social Science Quarterly, 70 (4), 870-885. Miethe, T.D., Hughes, M., & McDowall, D. (1991). Social change and crime rates: An evaluation of alternative theoretical approaches. Social Forces, 70 165-185. Mladenka, K.R., & Hill, K.Q. (1976). A reex amination of the etiology of urban crime. Criminology, 13 491-506. Moffitt, T.E. (1993). Adole scence-limited and life-cour se-persistent antisocial behavior: A developmental taxonomy. Psychological Review, 100 (4). 674701.
295 Morenoff, J.D., & Sampson, R.J. (1997). Violent crime and t he spatial dynamics of neighborhood transition: Chicago, 1970-1990. Social Forces, 76 31-64. Mulaik, S.A. (1987). A brief history of the philosophical foundations of exploratory factor analysis. Multivariate Behavioral Research, 22 267-305. Mulaik, S.A., James, L.R., Van Alstine, J., Bennett, N., Lind, S., & Stilwell, C.D. (1989). Evaluation of goodne ss-of-fit indices for structural equation models. Psychological Bulletin, 105 (3), 430-445. Muthen, B.O. (1989). Latent variable m odeling in heterogeneous populations. Psychometrika, 54 557-85. Muthen, B., & Kaplan, D. (1992). A comparison of some methodologies for the factor analysis of non-normal Likert variables: A note on the size of the model. British Journal of Mathematical and Statistical Psychology, 45 1930. Myers, D. (1992). Analysis with local census data: Portraits of change San Diego, CA: Academic Press. Neapolitan, J.L. (1998). Cross-national vari ation in homicides: Is race a factor? Criminology, 36 139-156. Nunnally, J.C. (1978). Psychometric theory (2nd ed.). New York: McGraw-Hill. OÂ’Brien, R.M. (1991). Se x ratios and rape rates: A power control theory. Criminology, 29 99-114. Oberwittler, D. (2004). A multilevel analysis of neighbor hood contextual effects on serious juvenile offending. European Journal of Criminology, 1 (2), 201235. Oetting, E.R., & Beauvais, F. (1987) Common elements in youth drug abuse: Peer clusters and other psychosocial factors. Journal of Drug Issues, 2 133-151. Olsson, U.H., Foss, T., Troye, S.V., & Howell, R.D. (2000). The performance of ML, GLS, and WLS estimation in st ructural equation modeling under conditions of misspecification and nonnormality. Structural Equation Modeling, 7 557-595. Osborn, D.R., Trickett, A., & Elder, R. ( 1992). Area characteristics and regional
296 variates as determinants of area property crime levels. Journal of Quantitative Criminology, 8 265-285. Osgood, D.W. (2000). Poisson-based regr ession analysis of aggregate crime rates. Journal of Quantitative Criminology, 16 (1), 21-43. Osgood, D.W., & Anderson, A.L. (2004). Unstructured socializing and rates of delinquency. Criminology, 42 (3), 519-549. Park, R.E., & Burgess, W. (1925). The city Chicago: University of Chicago press. Patterson, E.B. (1991). Poverty, income inequality, and community crime rates. Criminology, 29 755-776. Patterson, G.R., & Dishion, T.J. (1985). Contributions of families and peers to delinquency. Criminology, 23 (1), 63-79. Pearson, F.S., & Weiner, N.A. (1985). To ward an integration of criminological theories. Journal of Criminal Law and Criminology, 76 (1), 116-150. Peeples, F., & Loeber, R. ( 1994). Do individual factor s and neighborhood context explain ethnic differences in juvenile delinquency? Journal of Quantitative Criminology, 10 (2),141-157 Peterson, R.A. (2000). A meta-analysis of variance accounted for and factor loadings in explorat ory factor analysis. Marketing Letters, 11 (3), 261-275. Peterson, R.D., & Bailey, W.C. (1988). Fo rcible rape, poverty, and economic inequality in U.S. metropolitan communities. Journal of Quantitative Criminology, 4 99-119. Philips, L., & Votey, H. L. (1972). Crime, youth, and the labor market. Journal of Political Economy, 80 491-504. Ping, R.A. (1996). Latent variable interact ion and quadratic effect estimation: A two-step technique using structur al equation analysis. Psychological Bulletin, 119, 166-175. Piquero, A.R., MacIntosh, R., & Hickm an, M. (2002). The va lidity of a selfreported delinquency scale: Com parisons across gender, age, race, and place of residence. Sociological Methods and Research, 30 (4), 492-529. Pokorny, S.B., Jason, L.A., Schoeny, M. E., Townsend, S.M. & Curie, C.J.
297 (2001). Do participation rates cha nge when active consent procedures replace passive consent? Evaluation Review, 25 (5), 567-580. Popper, K. (2002). The logic of scientific discovery (6th ed.). New York: Routledge. Pratt, T.C., & Cullen, F. T. (2005). Assessing macro-le vel predictors and theories of crime: A meta-analysis. Crime and Justice, 32 373-450. Pressman, I., & Carol, A. (1971). Crime as diseconomy of scale. Review of Social Economy, 19 227-236. Quetelet, A. (1984). Research on the propensity for crime at different ages (S. Sylvester, trans.). Cincinnati, OH: A nderson. (Original work published 1831) Raudenbush, S., Bryk, A., Cheong, Y.F., & Congdon, R. (2001). Hierarchical linear and nonlinear modeling Lincolnwood, IL: Scientific Software International. Raykov, T., & Marcoulides, G.A. (2000). A first course in structural equation modeling Mahwah, NJ: Lawrence Erlbaum. Rebellon, C.J. (2002). Rec onsidering the broken homes /delinquency relationship and exploring its mediating mechanism(s). Criminology, 40 (1), 103-136. Regnerus, M.D. (2002). FriendsÂ’ infl uence on adolescent theft and minor delinquency: A developmental test of peer-reported effects. Social Science Research, 31 681-705. Reiss, A.J., Jr. (1951). Deli nquency as the failure of per sonal and social controls. American Sociological Review, 16 196-208. Reiss, A.J., Jr., & Rhodes, A.L. (1961). The distribution of juvenile delinquency in the social class structure. American Sociological Review, 26 720-732. Reiss, A.J., Jr., & Rhodes A.L. (1964). An empirica l test of differential association theory. Journal of Research in Crime and Delinquency, 1 518. Riedel, M. (2000). Research strategies for sec ondary data: A perspective for criminology and criminal justice Thousand Oaks, CA: Sage.
298 Robbins, L.N. (1974). Deviant children grown up: A sociological and psychiatric study of sociopathic personality Huntington, NY: R.E. Krieger. Robinson, W.S. (1950). Ecol ogical correlations and the behavior of individuals. American Sociological Review, 15 (3), 351-357. Roncek, D.W., & Maier, P.A. (1991). Bars blocks, and crimes revisited: Linking the theory of routine activities to the empiricism of hot spots. Criminology, 29 725-755. Rountree, P.W., Land, K.C., & Miethe. T.D. (1994). Macr o-micro integration in the study of victimizati on: A hierarchical logistic model analysis across Seattle Neighborhoods. Criminology, 32 (3), 387-414. Rozeboom, W.W. (1956). Mediation variables in scientific theory. Psychological Review, 63 249-264. Sampson, R.J. (1985). Structural sources of variation in race-age-specific rates of offending across major U.S. cities. Criminology, 23 (4), 647-673. Sampson, R.J. (1986). Neighborhood family structure and the risk of personal victimization. In R.J. Sampson and J.M. Byrne (Eds.), The social ecology of crime (pp. 25-46). New York: Springer-Verlag. Sampson, R.J. (1987). Urban Black violence: The effect of male joblessness and family disruption. American Journal of Sociology, 93 348-382. Sampson, R.J. (1988). Local friendship ties and community a ttachment in mass society: A multilevel systemic model. American Sociological Review, 53 766-779. Sampson, R.J. (1999). Techniques of research neutralization. Theoretical Criminology, 3 (4), 438-450. Sampson, R.J., & Groves, W.B. (1989). Community structure and crime: Testing social-disorganization theory. American Journal of Sociology, 94 (4), 774802. Sampson, R.J., & Laub, J.H. (1993). Crime in the making: Pathways and turning points through life Cambridge, MA: Harvard University Press.
299 Sampson, R.J., & Raudenbush, S.W. (1999) Systematic observation of public spaces: A new look at disorder in urban neighborhoods. American Journal of Sociology, 105 603-651. Sampson, R.J., Raudenbush, S.W., & Earls, F. (1997). Neighborhoods and violent crime: A multilevel st udy of collective efficacy. Science, 277 (5328), 918-924. SAS Institute. (1999). SAS/STAT userÂ’s guide (8th ed.). Cary, NC: SAS Institute. Saunders, D.R. (1956) Moderator variables in prediction. Educational and Psychological Measurement, 16 209-222. Sellers, C.S., Cochran, J.K. & Winfree, T.L., Jr. (2003 ). Social learning theory and courtship violence: An empirical test. In R.L. Akers and G. F. Jensen (Eds.), Social learning theory and the explanation of cr ime: A guide for the new century, volume 11 advanc es in criminological theory (pp. 109-129). New Brunswick, NJ: Transaction Publishers. Sellers, C.S., & Winfree, T.L., Jr. (1990). Differential associations and definitions: A panel study of youthful drinking behavior. International Journal of the Addictions, 25 755-771. Sellin, T. (1938). Culture conflict and crime: A report of the subcommittee on delinquency of the committee on personality and culture New York: Social Science Research Council. Shaw, C.R., & McKay, H. (1942). Juvenile delinquency a nd urban areas: A study of rates of delinquency in relation to differential characteristics of local communities in American Cities. Chicago: University of Chicago Press. Shaw, C.R., & McKay, H. (1969). Juvenile delinquency a nd urban areas: A study of rates of delinquency in relation to differential characteristics of local communities in American Cities (2nd ed.) Chicago: University of Chicago Press. Shaw, C.R., Zorbaugh, F.M., McKay, H.D. & Cottrell, L.S. (1929). Delinquency areas: A study of the geographic distri bution of school truants, juvenile delinquents, and adult offenders in Chicago Chicago: The University of Chicago Press. Shevlin, M., & Miles, J.N.V. (1998). Effe cts of sample size, model specification and factor loadings on the GFI in confirmatory factor analysis. Personality
300 and Individual Differences, 25 85-90. Short, J.F., Jr. (1957). Different ial association and delinquency. Social Problems, 4 233-239. Short, J.F., Jr. (1958). Di fferential association with delinquent friends and delinquent behavior. Pacific Sociological Review, 1 20-24. Short, J.F., Jr. (1960). Diffe rential association as a hypothesis: Problems of empirical testing. Social Problems, 8 14-24. Shrout, P.E., & Bolger, N. (2002) Mediation in experimental and Nonexperimental studies: New procedures and recommendations. Psychological Methods, 7 (4), 422-445. Silver, E., & Miller, L.L. (2004). Sources of informal social control in Chicago neighborhoods. Criminology, 42 (3), 551-583. Simcha-Fagan, O., & Schwartz, J.E. (1986). Neighborhood and delinquency: An assessment of contextual effects. Criminology, 24 (4), 667-703. Skinner, B.F. (1953). Science and human behavior New York: Free Press. Skinner, W.F., & Fream, A.M. (1997). A social learning theory analysis of computer crime am ong college students. Journal of Research in Crime and Delinquency, 34 495-518. Skrondal, A., & Rabe-Hesketh, S. (2004). Generalized latent variable modeling: Multilevel, longitudinal and structural equation models Boca Raton, FL: Chapman & Hall. Smith, D. (2003). 10 ways practitioner s can avoid frequent ethical pitfalls. Monitor on Psychology, 34 (1), 50-55. Smith, M.D., & Bennett, N. (1985). Poverty, inequality, and theori es of forcible rape. Crime and Delinquency, 31 295-305. Smith, M.D., & Brewer, V.E. (1992). A sex-specific ana lysis of correlates of homicide victimization in United States cities. Violence and Victims, 7 279-286. Smith, D.A., & Parker, R.N. (1980). Type of homicide and variation in regional rates. Social Forces, 59 136-147.
301 Smith, D.A., & Jarjoura, G.R. (1988). Social structure an d criminal victimization. Journal of Research in Crime and Delinquency, 25 27-52. Snook, S.C., & Gorsuch, R.L. (1989). Component analysi s versus common factor analysis: A Monte Carlo study. Psychological Bulletin, 106 148-154. Specht, D.A. (1975). On the evaluation of causal models. Social Science Research, 4 113-133. Stafford, M.C., & Ekland-Olson, S. ( 1982). On social learning and deviant behavior: A reappraisal of the findi ngs: A comment on Akers et al. ASR August 1979. American Sociological Review, 47 (1), 167-169. Stafford, M.C., & Gibbs, J.P. (1980). Cr ime rates in an ecological context: Extension of a proposition. Social Science Quarterly, 61 63-65. Steffensmeier, D.J., Streifel, C., & Hare r, M.D. (1987). Relative cohort size and youth crime in the United States, 1953-1984. American Sociological Review, 52 702-710. Steffensmeier, D.J., Streifel, C., & Shi hadeh, E.H. (1992). Cohort-size and arrest rates over the life course: The Ea sterlin hypothesis reconsidered. American Sociological Review, 57 306-314. Steffensmeier, D.J., Allan, E.A., Harer, M.D., & Streifel, C. (1989). Age and the distribution of crime. American Journal of Sociology, 94 (4), 803-831. Steiger, J.H. (1990). Struct ural model evaluation modi fication: An interval estimation approach. Multivariate Research, 25 173-180. Stevens, J. (2002). Applied multivariate statis tics for the social sciences (4th ed.). Mahwah, NJ: Lawrence Erlbaum. Stewart, D.W. (1981). The application and misapplicati on of factor analysis in marketing research. Journal of Marketing Research, 18 51-62.
302 Stone, E.F., & Hollenbeck, J.R. (1984). So me issues associated with the use of moderated regression. Organizational Behavior and Human Performance, 34 195-213. Stone, E.F., & Hollenbeck, J.R. (1989). Clarifying some controversial issues surrounding statistical procedures fo r detecting moderator variables: Empirical evidence and related matters. Journal of Applied Psychology, 74 (1), 3-10. Strickland, D.E. (1982). Social learning and de viant behavior: A specific test of a general theory: A comment and critique. American Sociological Review, 47 (1), 162-167. Sun, I.Y., Triplett, R., & Gainey, R.R. (2004). Neighborhood characteristics and crime: A test of Sampson and Groves Â’ model of social disorganization. Western Criminology Review, 5 (1), 1-16. Sutherland, E.H. (1924). Criminology Philadelphia: J.B. Lippincott Company. Sutherland, E.H. (1932). Social process in behavioral problems. Publications of the American Sociological Society, 26 55-61. Sutherland, E.H. (1934). Principles of criminology (2nd ed.). Philadelphia: J.B. Lippincott Company. Sutherland, E.H. (1937). The professional thief Chicago: University of Chicago Press. Sutherland, E.H. (1939). Criminology (3rd ed.). Philadelphia: J.B. Lippincott Company. Sutherland, E.H. (1947). Criminology (4th ed.). Philadelphia: J.B. Lippincott Company. Sutherland, E.H. (1973a). De velopment of the theory. In K. Schuessler (Ed.), Edwin H. Sutherland on analyzing crime (pp. 13-29). Chicago: University of Chicago Press. Sutherland, E.H. (1973b). The person and the situation in the treatment of prisoners. In K. Schuessler (Ed.), Edwin H. Sutherland on analyzing crime (pp. 160-166). Chicago: Univer sity of Chicago Press. Sutherland, E.H. (1973c). Cr ime and the conflict process. In K. Schuessler (Ed.),
303 Edwin H. Sutherland on analyzing crime (pp. 99-111). Chicago: University of Chicago Press. Sutherland, E.H. (1973d). Critique of the theory. In K. Schuessler (Ed.), Edwin H. Sutherland on analyzing crime (pp. 30-41). Chicago: University of Chicago Press. Sutherland, E.H., & Cressey, D.R. (1969) A sociological t heory of criminal behavior. In D.R. Cressey & D.A. Ward (Eds.), Delinquency, crime, and social process (pp. 426-432). Albany, NY: State University of New York Press. Sutherland, E.H., & Cressey, D.R. (1970). Criminology (8th ed.). Philadelphia, PA: J.B. Lippincott Company. Sutherland, E.H., & Cressey, D.R. (1974). Criminology (9th ed.). Philadelphia, PA: J.B. Lippincott Company. Sykes, G., & Matza, D. (1957). Techni ques of neutralization: A theory of delinquency. American Journal of Sociology, 22 664-670. Tabachnick, B.G., & Fidell, L.S. (2001). Using multivariate statistics (4th ed.) Boston: Allyn & Bacon. Tanaka, J.S. (1987). How bi g is big enough? Sample si ze and goodness-of-fit in structural equation model s with latent variables. Child Development, 58 134-146. Tanaka, J.S., & Huba, G.J. (1989). A gener al coefficient of determination for covariance structure models under arbitrary GLS estimation. British Journal of Mathematical and Statistical Psychology, 42 233-239. Tarde, G. (1912). Penal philosophy (R. Howell, Trans.). Boston: Little Brown. Thompson, B., & Daniel, L.G. (1996). Fa ctor analytic evidence for the construct validity of scores: A historical overview and some guidelines. Educational and Psychological Measurement, 56 (2), 197-208. Thornberry, T.P. (1987). Toward an in teractional theory of delinquency. Criminology, 25 863-887. Thornberry, T.P. (1989). Reflections on the advantages and disadvantages of theoretical integration. In S.F. Messner, M.D. Kr ohn, & A.E. Liska (Eds.),
304 Theoretical integration in t he study of deviance and crime (pp. 51-60). Albany, NY: State University of New York Press. Thornberry, T.P., Lizotte, A.J., Krohn, M.D. Farnworth, M., & Jang, S.J. (1994). Delinquent peers, beliefs, and delinquent behavior: A longitudinal test of interactional theory. Criminology, 32 (1), 47-84. Tinsley, H.E.A., & Tinsley, D.J. (1987). Uses of factor analysis in counseling psychology research. Journal of Counseling Psychology, 34 (4), 414-424. Tittle, C.R. (1985). The assumption t hat general theories are not possible. In R.F. Meier (Ed.), Theoretical methods in criminology (pp. 93-121). Beverly Hills, CA: Sage. Tittle, C.R. (1989). Influences on urbani sm: A test of three perspectives. Social Problems, 36 270-288. Tittle, C.R. (1995) Control balance: Toward a general theory of deviance Boulder, CO: Westview Press. U.S. Census Bureau. (1990). United States census 1990 [Data file]. Washington, DC: Available at http://www.census.gov U.S. Census Bureau. (2000). United States census 2000 [Data file]. Washington, DC: Available at http://www.census.gov Urberg, K.A. (1992). Locus of peer infl uence: Social crowd and best friend. Journal of Youth and Adolescence, 21 439-450. Urberg, K.A. (1997). Close friend and group influence on adolescent cigarette smoking and alcohol use. Developmental Psychology, 33 834-844. Veysey, B.M., & Messner, S.F. (1999). Furt her testing of social disorganization theory: An elaboration of Sampson and GrovesÂ’s Â“community structure and crime.Â” Journal of Research in Crime and Delinquency, 36 (2), 156174. Vold, G.B. (1958). Theoretical criminology New York: Oxford University Press. Voss, H.L. (1964). Different ial association and report ed delinquent behavior: A replication. Social Problems, 12 78-85. Wareham, J., Cochran, J.K., Dembo, R ., & Sellers, C.S. (2005). Community,
305 strain, and delinquency: A test of a mu lti-level model of general strain theory. Western Criminology Review, 6 (1), 117-133. Warner, B.D., & Pierce, G.L. (1993). Reex amining social disorganization theory using calls to the police as a measure of crime. Criminology, 31 493-517. Warner, B.D., & Roundtree, P.W. (1997). Local ties in a community and crime model: Questioning the systemic natur e of informal social control. Social Problems, 44 520-536. Warr, M. (1993). Age, peers, and delinquency. Criminology, 31 (1), 17-40. Warr, M. (1996). Organi zation and instigation in delinquent groups. Criminology, 34 (1), 11-37. Warr, M. (1998). Life-c ourse transitions and desistance from crime. Criminology, 36 183-216. Warr, M. (2002). Companions in crime: The social aspects of criminal conduct. New York: Cambridge University Press. Webb, S.D. (1972). Crime and the divisi on of labor: Testing a Durkheimian model. American Journal of Sociology, 78 643-656. Weicher, J.C. (1970). The effect of income on delinquency: Comment. American Economic Review, 60 249-256. Wendorf, C.A. (2002). Comparisons of structural equation modeling and hierarchical linear modeling approaches to couplesÂ’ data. Structural Equation Modeling, 9 (1), 126-140. West, S.G., Finch, J.F., & Curran, P.J. (1995). Structural equation models with nonnormal variables: Problems and rem edies. In R.H. Hoyle (Ed.), Structural equation mo deling: Concepts, issues, and applications (pp. 5675). Thousand Oaks, CA: Sage. White, H.R., Johnson, V., & Horowitz, V. (1986). An application of three deviance theories for adolescent substance use. International Journal of the Addictions, 21 347-366. White, H.R., & LaGrange, R.L. (1987). An assessment of gender effects in self report delinquency. Sociological Focus, 20 (3), 195-213.
306 White, H.R., Pandina, R.J., & LaGrange, R.L. (1987). Longitudi nal predictors of serious substance abuse and delinquency. Criminology, 25 (3), 715-740. Wikstrom, P.H., Loeber, R. (2000). Do disadvantaged neighborhoods cause welladjusted children to become adolesc ent delinquents? A study of male juvenile serious offending, individu al risk and protective factors, and neighborhood context. Criminology, 38 (4), 1109-1142. Williams, K.R. (1984). Economic sources of homicide: Reestimating the effects of poverty and inequality. American Sociological Review, 49 283-289. Williams, K.R., & Flewelli ng, R.L. (1988). The social production of criminal homicide. American Sociological Review, 53 421-431. Winfree, L.T., Jr., & Bernat, F.P. ( 1998). Social learning, self-control, and substance abuse by eighth grade st udents: A tale of two cities. Journal of Drug Issues, 28 (2), 530-558. Winfree, L.T., Jr., & Griffiths, C.T. (1983). Social learning and adolescent marijuana use: A trend study of devi ant behavior in a rural middle school. Rural Sociology, 48 (3), 219-239. Winfree, L.T., Jr., Griffith s, C.T., & Sellers, C.S. (1989). Social learning theory, drug use, and American Indian yout hs: A cross-cultural test. Justice Quarterly, 6 395-417. Winfree, L.T., Jr., Mays, G.L., & Backstrom, T.V. (1994). Youth gangs and incarcerated delinquents: Exploring the ties between gang membership, delinquency, and social learning theory. Justice Quarterly, 11 229-256. Winfree, L.T., Jr., Sellers, C.S., & Clason, D.L. (1993). Social learning and adolescent deviance abstention: Towa rd understanding the reasons for initiating, quitting, and avoiding drugs. Journal of Quantitative Criminology, 9 (1), 101-125. Wirth, L. (1938). Urbanism as a way of life. American Journal of Sociology, 44 (1), 1-24. Wooldredge, J. (2002). Ex amining the (ir)relevance of aggregation bias for multilevel studies of neighborhoods and crime wit h an example comparing census tracts to official neighborhoods in Cincinnati. Criminology, 40 (3), 681-709.
307 Zedeck, S. (1971). Problems with t he use of Â“moderatorÂ” variables. Psychological Bulletin, 76 295-310. Zeger, S.L., & Liang, K.Y. (1986). Longi tudinal data analysis for discrete and continuous outcomes. Biometrics, 42 121-130. Zhang, L., & Messner, S.F. (2000). The effects of alternative measures of delinquent peers on self-r eported delinquency. Journal of Research in Crime and Delinquency, 37 (3), 324-337.
About the Author Stephen W. Verrill received an A ssociate Degree in Business Administration from the University of Southern Maine in 1987, a BachelorÂ’s Degree in Business Administ ration from the Universi ty of Southern Maine in 1989, a BachelorÂ’s Degree in Criminal Just ice from Florida Gulf Coast University in 2002, and a MasterÂ’s of Arts in Criminology from the University of South Florida in 2003. Prior to entering the Ph.D program at the University of South Florida, Mr. Verrill served in varying capac ities of either priv ate or public law enforcement for approxim ately 20 years. Mr. Verrill has authored or coaut hored three articles in the LAE Journal of the American Criminal Justice Association and he coauthored a publication in the Journal of Ethnicity in Criminal Justice Mr. Verrill also contributed to three presentations at annual meetings of the American Society of Criminology.