|USFDC Home | USF Electronic Theses and Dissertations||| RSS|
This item is only available as the following downloads:
Dose Time Response Modeling of Neurobehavioral Screening Data: Application of Physiologically Releva nt Parameters to Allow for Dose Dependent Time of Peak Effects by Michael Raymond Wessel A thesis submitted in partial fulfillment of the requirements for the degree of Master of Science Department of Epidemiology and Biostatistics College of Public Health University of South Florida Major Professor: Yiliang Zhu, Ph.D. Getachew Dagne, Ph.D. Jeff Gift, Ph.D. Date of Approval July 18, 2005 Keywords: dose-response, biphasic neurotoxicity, risk, toxicology Copyright 2005, Michael Raymond Wessel
DEDICATION To Joyce Ann Who allowed me to chase the snakes, and make my own mistakes, supporting me all the while. Thanks for the dance. Michael
Acknowledgments I am grateful to Dr. Yiliang Zhu, my advisor and major professor, who has given up countle ss hours of well deserved rest to provide me with his guidance, expertise and the support necessary to complete this thesis. I wish to express my sincere gr atitude to Dr. Jeff Gift of the USEPA for agreeing to serve on my committee and provide invaluable insights regarding this work. I also thank Dr. Gifts colleagues; Dr. Allan Marcus, Dr. Virginia Moser and Dr. Rob Dewoskin for reviewing the content of the thesis and provid ing their insight on the potential application of this effort to their very important work I further thank Dr. Getachew Dagne for his constructive comments and willingness to share his insights and joyful personality. In running out of room I finally thank my family; Mom, Maury, Mark, Craig, Linda, Alex, J.T., Jane, Mathew, Bob, Laura and, unendingly, my wife Megan who bore the brunt of my frustration and still encouraged me to fight on. With Love, Mike
i Table of Contents List of Tables..............................................................................iii List of Figures.............................................................................v Abstract...................................................................................vii Chapter One: In troduction...........................................................1 1.1 Health Risk Assessment.................................................2 1.2 Neurotoxicity................................................................3 1.3 Thesis Objectives..........................................................6 Chapter Two: Neurobehavioral Screening Experiments.....................8 2.1 Screening Tests............................................................8 2.2 Functional Obse rvational Battery.....................................9 2.3 Assessment Methodologies...........................................11 Chapter Three: Pharmacokinetics and Pharmacodynamics..............13 3.1 Compartmental Pharmacokinetics..................................13 3.2 Nonlinear Kinetics.......................................................17 3.3 Pharmaco dynamics......................................................19 3.4 Pharmacokinetic/Pharmacodynamic Models.....................20 3.5 Michaelis-Menten Kinetics.............................................23 Chapter Four: Nonlinear Mixed Effects Models...............................26 4.1 Toxico-Diffusion Model.................................................26 4.2 Parameter Estimation Methods......................................28 4.3 Model Selection and Diagnostics....................................29 Chapter Five: Expanding the Toxico-Diffusion Model......................31 5.1 Biphasic Toxico Diffusion Model.....................................31 5.2 Parathion Activity Counts.............................................34 5.3 Fitted Response..........................................................35 5.4 Dose Dependent TOPE.................................................38 Chapter Six: Simulation.............................................................43 6.1 Simulation Rationale....................................................43 6.2 Simulation Methods.....................................................44 6.3 Simulation Results.......................................................51
ii 6.3.1 Four Time Point Designs: Ten subjects...................51 6.3.2 Five Time Point Designs: Ten Subjects...................53 6.3.3 Designs with 6-9 Time Points...............................56 6.3.4 Designs with In creasing Number of Subjects..........57 6.3.5 Five Time Po int Designs with 20 Subjects...............58 6.3.6 Five Time Point Designs with 30 Subjects...............60 6.3.7 Bias and MS E in Time of Peak Effect......................63 Chapter 7: Conclusions..............................................................68 List of Re ferences.....................................................................76
iii List of Tables Table 1. Descriptive statistics for motor activity counts after acute parathion exposure includin g dose group average activity counts and standard deviation in parenthesis................................35 Table 2. Summary of biphasic model fit to parathion activity counts.....................................................................................36 Table 3. Results of log likelihood ratio test comparing the Toxicodiffusion and Biphasic toxico-diffusion models. .............................41 Table 4. Simulated experimental FOB designs using 10, 20, and 30 subjects per dose group and testing times between 0-168 hours with experimental testing times added between 4 and 20 hours......................................................................................50 Table 5. Summary of simulated design trials with 10 subjects per dose group and 4 testing times. .................................................51 Table 6. Bias and MSE for each parameter of interest for the 10 subject, 4 time point design.......................................................53 Table 7. Summary of simulated de sign trials on FOB parathion data with 10 subjects per dose group and 5 testing times. .............54 Table 8. Bias and MSE for each parameter of interest for the 10 subject, 5 time point design.......................................................55 Table 9. Summary of simulated de sign trials on FOB Parathion data with 10 subjects per dose group and between 6 and 9 experimental test ing times.........................................................56 Table 10. Bias and MSE for each pa rameter of interest for the 10 subject designs with between 6 and 9 experimental testing times......................................................................................57
iv Table 11. Summary of simulated de sign trials on FOB Parathion data with 20 subjects per dose group and 5 testing times...............58 Table 12. Bias and MSE for each pa rameter of interest for the 20 subject, 5 time point design.......................................................59 Table 13. Summary of simulated de sign trials on FOB Parathion data with 30 subjects per dose group and 5 testing times. .............60 Table 14. Bias and MSE for each pa rameter of interest for the 30 subject, 5 time point design.......................................................62 Table 15. Confidence intervals fo r dose group specific TOPE for each of the 5 time point, 20 subject designs.................................67 Table 16. Confidence intervals fo r dose group specific TOPE for each of the 5 time point, 30 subject designs.................................68
v List of Figures Figure 1. Illustration of simple one-compartment pharmacokinetic model..............................................................15 Figure 2. Two compartment PK/P D model describing the pathway of chemical introduction and e limination from an organism.............16 Figure 3. Bi-exponential decline of concentration after iv bolus injection when drug can be described using two-compartment model.....................................................................................18 Figure 4. Illustration of th e Effect-Compartment model..................21 Figure 5.The Michaelis-Menten equation with illustration of the relationship between su bstrate concentration (a) and velocity of the reaction..............................................................24 Figure 6. Observed response in ac tivity scores of rats subjected to acute Parathio n exposure.......................................................34 Figure 7. Fitted response of bi phasic numerator model for rats subjected to acute Pa rathion exposure.........................................36 Figure 8. Residual plots of bi phasic numerator model for rats subjected to acute Pa rathion exposure.........................................37 Figure 9. Dose specific trajecto ries for biphasic numerator model on rats subjected to ac ute Parathion exposure..............................38 Figure 10. Individual specific tr ajectories for biphasic numerator model on rats subjected to acute Parathion exposure.....................39 Figure 11. Dose group specific response to parathion predicted by the toxico-diffu sion model. ....................................................41 Figure 12. Theoretical true dose-dependent biphasic model response between 0 ho urs and 24 hours......................................47
vi Figure 13. Bias (Top) and Mean Square Error (Bottom) of the TOPE estimator under the 5-time point 20 subject designs.............64 Figure 14. Bias (Top) and Mean Square Error (Bottom) of the TOPE estimator under the 5-time point 30 subject designs.............65
vii Dose Time Response Modeling of Neurobehavioral Screening Data: Application of Physiologically Releva nt Parameters to Allow for Dose Dependent Time of Peak Effects Michael Raymond Wessel ABSTRACT In collaboration with the United States Environmental Protection Agency (USEPA), the University of South Florida Health Risk Methodology Group has developed dose-time-response models to characterize neurobehavioral resp onse to chemical exposure. The application of dose-time-response models to neurobehavioral screening tests on laboratory anim als allows for benchmark dose estimation to establish exposure limits in environmental risk assessment. This thesis has adva nced dose-time-response modeling by generalizing a published toxico diffusion model to allow for dose dependent time of peak effects. To accomplish this, a biphasic model was developed which adopted the ef fect compartment model paradigm used in pharmacokinetics/phar macodynamics to estimate a distributional rate constant to acco unt for dose related variation in the time of peak effect. The biphasic model was able to describe dosedependent time of peak effects as observed in the data on acute exposure to parathion and adeq uately predicted the observed response. However, the experimental design appeared insufficient in statistical power to confirm statisti cal significance for each parameter of interest. Motivated by the ques tion of what design requirement might be necessary to validate th e biphasic model, Monte Carlo simulation was adopted. Simulations were performed to assess the efficacy and efficiency of various experimental designs for detecting and evaluating some critical characteristics of the biphasic model, including the TOPE. The results of si mulation suggest that the location of measurement times around the TOPE have important implications for assessing the statistical sign ificance of the parameter that describes dose-dependent TOPE and that the mean squared error of the parameter estimator was improved most when testing times were
viii chosen to bracket the TOPE. While dose dependent time of peak effects has underlying physiological me chanisms such as synergistic or capacity limited kinetics, the biphasic model estimates these physiological properties through a mathematical function which may be physiologically relevant bu t does not necessarily define physiological mechanisms underlying the response. However, if verified through further testing, th e biphasic model may contribute to the USEPAs aim of developing phys iologically relevant dose-response models for assessing risk of neurot oxicity with repeated measurements of response.
1 Chapter One Introduction Chemical exposure has become a certainty of human life in the 21st century. In fact, it is difficult to imagine a day goes by in which one is not exposed to a chemical, natural or synthetic, about which there is some uncertainty of risk. Fo rmal attempts to characterize the risks associated with chemical exposure date back to Hippocrates (ancient 400 BC) who developed toxicological principles related to clinical observations on the bioava ilability and absorption of common therapies and poisons. Paracelsus (~ 1500 AD) is credited with the idea that all substances are poisons and it is the dose that determines its potential risk and benefit. In modern times, exponential growth in the production of chemicals occurred as a consequence of the industrial revolution and World War II and in the United States led to the establishment of regulatory agenci es such as the Food and Drug Administration and later the Envi ronmental Protection Agency (US EPA). Because of the many natural and synthetic chemicals introduced in todays environment, governme ntal agencies throughout the industrialized world have become k eenly interested in assessing the potential risks to humans from toxic agents (US EPA, 1998). In the
2 US, the EPA has registered more than 65,000 chemical substances manufactured, imported, or processe d in the United States under the Toxic Substance Control Act (TSCA) (US EPA, 1996). Chemical exposure has become one of the ten leading causes of workplace disorder (Anger, 1984) and the pote ntial for adverse effects on the nervous system is becoming common in the workplace as approximately 70 chemicals of known neurotoxic potential have potential exposure to more than 1 million workers (Anger,1990). 1.1 Health Risk Assessment Environmental risk assessment is an emerging field that relies on three basic assessment principles : exposure assessment, hazard characterization, and risk quantifica tion (McCarty and Mackay, 1993). The EPAs National Center for Environmental Assessment (NCEA) conducts risk assessment for an arra y of health effects that may result from exposure to environmental agents. This process includes a thorough evaluation of all the ava ilable data as well as conducting scientific experiments to underst and the relationship between exposure and risk. Historically, th ese analyses have been done very differently for cancer and non-cancer health effects because of perceived differences in the mechan istic underpinnings of cancer and other toxic effects. As our understan ding of the underlying biology of toxic effects has grown, however, the apparent differences between
3 cancer and non-cancer effects have lessened to the point where it seems reasonable to develop quanti tative methods based on similar considerations for all ty pes of health effects, and to make approaches to risk assessment as consistent across health endpoints as our current mechanistic understanding allows (US EPA, 2000). Neurotoxicity risk assessment is one area in particular where the EPA has expressed the need for consiste nt guidance on how to evaluate data on neurotoxic substances an d assess their potential to cause transient or persistent and direct an d indirect effects on human health. 1.2 Neurotoxicity Neurotoxicity is defined as an adverse change in the structure or function of the central and/or pe ripheral nervous system following exposure to a chemical, physical or biological agent (Tilson, 1990). The central nervous system is particul arly vulnerable to chemical insult and has limited ability to regenera te. Functional neurotoxic effects include adverse changes in somatic/ autonomic, sensory, motor and/or cognitive function (US EPA, 1998). The effects can be transient (the organism returns to pre-exposure condition) or persistent (the organism is permanently and adversely changed by exposure). However, even transient effects can signify underlying resultant damage to the organism (US EPA, 1998). Animal studies make up the largest portion of controlled exposu re assessments and allow for the
4 use of high concentrations of chemic als to be administered to achieve responses that may define the mechanism of action as well as the magnitude of response to chemic al exposure. Neurotoxic screening tests such as the Functional Observational Battery (FOB) test (Moser et al., 1995, 1997a), along with neuro-physiological, biochemical, neuro-pathological and neuro-endocr inological studies are now being used by the EPA as an overall stra tegy to detect the full range of chemical induced alterations in th e structure and function of the nervous system. The advancement of neurotoxicity testing methods and experimental design has coincided with advances in statistical methodologies and computer applicat ions to allow for more effective methods of performing risk assessmen ts on neurotoxins. A significant advance in neurotoxicity risk assessment is the use of benchmark dose (BMD) methodologies for establishi ng safety levels of chemical exposure (US EPA, 2000). While tradit ional analysis of FOB data has used Analysis of Variance (ANO VA) to set a No Observed Adverse Effects Level (NOAEL), dose-time-response models provide continuous estimation of response over the ti me course of the study providing beneficial information for BMD estimation. The University of South Flor idas Health Risk Assessment Methodology Group (HRAMG) has been working to develop new
5 statistical methods for explicit dose-time-response models (DTR) of the FOB data, which may serve as the foundation for benchmark dose estimation. The development of ex plicit DTR models to describe neurotoxic potential of chemical expo sure has progressed from strictly mathematical models such as polynomial models to those incorporating simple toxicokineti cs. The potential for physiologic interpretation enhances comparability with other available data and increases confidence in the interspe cies extrapolation of results of these screening tests to characterize potential risk to human health. Zhu (2005a) and Zhu et al. (2005b,c ) have developed a family of dose-response models and illustrated their application through several published datasets generated from the EPA Superfund study (Moser et al, 1995) and a study conducted in co llaboration with the International Program on Chemical Safety (IPC S) (Moser et al. 1997a). These models incorporate basic toxicokine tic principles into a family of mathematical models in consideration of the physical properties underlying responses to chemical ex posure observed in the FOB data. Zhu (2005 b,c) found that the toxico diffusion model often satisfactorily described the observed dose-response relationship in FOB data and is useful in application of benchmark dose estimation methods. While the toxico diffusion model has proven robust in describing FOB data, it is limited in describing the full possible range of
6 dose related response to acute expo sure. Most importantly this model is limited by imposing a dose indepe ndent Time of Peak Effect (TOPE) across every exposure level. 1.3 Thesis Objectives The first objective of this thesis was to expand on a published toxico-diffusion function used to model neurobehavioral screening data by considering a situation wher e the TOPE may take on dosedependent characteristics. Throug h the incorporation of a dosedependent distributional parameter, this thesis proposes a biphasic diffusion model and illustrates the utilit y of this model on the analysis of a FOB dataset on the motor activi ty of laboratory rats exposed to the organophosphate pesticide parathion. A second objective of the thesis was to investigate the study design requirements necessary to recover the key characteristics of the biphasic model, especially dose-dependent time of peak effect. Monte Carlo simulation was adopte d to explore the potential FOB testing times as well as sample size with respect to the efficiency and effectiveness of recovering key co mponents of the biphasic model. Various designs were compared using statistical measures of bias, power, mean squared error, and confidence interval. Utility of alternative study designs may contri bute to the USEPAs aims toward
7 using dose-response modeling, part icularly with physiologically relevant models, to assess potentia l risks of chemical exposure to human health.
8 Chapter 2 Neurobehavioral Screening Tests 2.1 Screening Tests Screening tests for neurotox icity represent the most fundamental level of investigat ion for forecasting potential neurotoxicity in animal studie s (WHO 1986). Screening tests are widely used because they are simple, rapid and economical. A battery of measurements is acquired with these tests that include measurements of behavioral endpoints representing neurophysiological, neuro-muscular, auto nomic and sensorimotor functions. Behavioral endpoints reflect the in tegration of various functional components of the nervous system and are often used as surrogates for mechanistic processes involved in a subjects response to exposure of a chemical agent. The EPAs te sting guidelines developed for the Toxic Substances Control Act (TSCA) and the Federal Insecticide, Fungicide, and Rodenticide Act descr ibed the use of neurobehavioral screening tests and established protoc ols and procedures used in these experimental designs (US EPA, 1991). Since that time the protocols and procedures have been refine d and are evolving to become a
9 standardized first tier screening method for neurotoxicity (US EPA, 1998). These toxicological studies, kn own as Functional Observational Battery (FOB) tests, are simple to im plement, relatively non-invasive and generate behavioral change rapidly. 2.2 IPCS Functional Observational Battery Tests This thesis utilizes data from a Functional Observational Battery (FOB) as described in Moser et al (1997a) and implemented in the IPCS collaborative study. For acute exposure, the study protocol uses 5 dose levels including a control gr oup and 4 testing times (including a baseline measurement) to assess ti me-related response to chemical exposure. The dose levels and the second testing time were determined via a dose range find ing study. Given the cooperative nature of the IPCS sponsored studie s several laboratories participated in these studies yielding severa l independent estimates of doseresponse for many of the chemicals tested. Prior to initiation of the FOB te sting a dose range finding study was performed to determine the dosing regime and testing times for a hypothesized time of peak chemical effect (TOPE) (Moser et al., 1997b). The starting dose was chos en based loosely on published estimates of the LD50 (dose which would be lethal to 50% of the subjects). Three doses at consta nt intervals above and below the starting dose were used for a se ven day survival study to determine
10 the top nonlethal dose to use in the FOB studies. The dose levels used for the FOB designs were then 100% 50% 25% and 12.5% of the top nonlethal dose. A separate pilot stud y was then conducted to estimate the TOPE using gait and arousal score s. Thus, the TOPE estimated for gait and arousal scores were taken to represent a hy pothesized TOPE for all neurobehavioral endpoints. Ten animals were randomly assign ed to each dose group for the FOB studies. Testing times were established such that a FOB was conducted prior to exposu re and at 2-3 subsequent time points after exposure. The first post-exposure testing time was conducted to correspond with the TOPE of the chemical being tested. Subsequent tests were performed at one day an d then one week after exposure. Post exposure test times genera lly did not exceed one week. The FOB response variables cons isted of 2530 non-invasive measures designed to assess behavi oral alterations with respect to a wide range of neurobiological functi ons, including sensory, motor and autonomic functions, excitability, ne uromuscular strength, and activity level. The entire battery of tests required approximately 6-8 minutes per rat. Assessment of motor activities was used in conjunction with the FOB because of its long history of use for evaluating behavioral effects of chemicals (MacPhail et al., 1989). Motor activity counts represent a broad class of be haviors involving coordinated
11 participation of sensory, motor, an d integrative processes. Neurotoxic agents can lead to either increase s or decreases in motor activity counts and organophosphate pesticides such as parathion, the chemical studied in this thesis, ha ve been shown to decrease motor activity counts in neurobehaviora l screening studies (USEPA, 1998). The apparatus for motor activity test was left to the discretion of participating laboratories, provided se veral criteria were met (Moser et al., 1997a). This thesis used para thion motor activity counts to illustrate fitting the biphasic toxico -diffusion model to neurobehavioral screening data. 2.3 Assessment Methodologies The USF Health Risk Assessm ent Methodology Group (HRAMG) has used the FOB to develop and test several classes of mathematical models to predict neurobehavioral re sponse to chemical exposure. Liu (2000) developed polynomial models for continuous FOB outcomes; Woodruff (2001) developed a diffusion model that is flexible in describing both transient and pers istent nonlinear dose-response relationships seen in the FOB data; Zhu (2001) tested these models on a large number of FOB datasets from both the EPA Superfund and the IPCS studies. More recently Zhu (2005a) developed a class of mathematical models that allow for incorporation of toxicokinetics, and in a series of reports (Zhu et al. 2005a,b) applied these models to the
12 FOB datasets and illustrated their us e in benchmark dose estimation. Common in these statistical analyses was the use of random effects to adjust for biological variation in responses among animals. These models performed well for describ ing the dose-response patterns observed in the FOB studies. However, physiological relevance of these models relies on simplifying assump tions such as linear or singlecompartment kinetics. It would be beneficial to derive models capable of describing exposure related response as well as incorporate parameters that estimate well kn own physiological phenomena that regulate the time course of a chem icals presence in the body and its affect on the observed respon se when the system displays nonlinearities such as nonlinear upta ke or saturation kinetic processes. Understanding of this process begins with knowledge of how the body processes the chemical (pharmacokinetics/toxicokinetics) and subsequent characterization of how the concentration of the chemical affects the organism of study (pharmacodynamics/toxicodynamics).
13 Chapter 3 Pharmacokinetics an d Pharmacodynamics In this thesis, references ar e made to pharmacokinetic and pharmacodynamic (PK/PD) modelin g techniques to develop a background on which the generalized biphasic toxico diffusion model was based. It should be noted that while the terms pharmacokinetics and toxicokinetics have essentially pa rallel meaning in their respective fields, in the strictest sense the former term should be restricted to the field of pharmacology. Since much of the literature devoted to kinetic and dynamic properties of chemical exposure has arisen from the field of pharmacology, we define the modeling approach using the PK/PD modeling paradigm and apply the bi phasic model to a toxicological study. 3.1 Compartmental Pharmacokinetics Orthodox pharmacokinetic and to xicokinetic studies deal with changes of drug concentrations in th e plasma or an organ over time in an attempt to describe how the body absorbs, distributes and eliminates a drug (Torda et al. 1994). The compartmental approach to pharmacokinetic estimation views th e body as being composed of a
14 number of pharmacokinetically distinct compartments. Each compartment can be thought of as an imaginary space in the body representing a combination of va rious tissues and organs, among which the drug interacts. Anatomic al composition of the compartment is unknown and in most cases its anal ysis is often of little value (Kwon 2002). Compartmental models are design ed to provide a conceptual understanding of distributional behaviors of a drug between the plasma and other tissues or organs in the body and estimate various pharmacokinetic parameters incl uding plasma concentrations, apparent volumes of distribution and rates governing elimination and clearance. It is recognized that these compartment models, while estimating physiological properties, still represent empirical fits to the data. However, the physiological interpretations are important for extrapolating information from anim al studies to regulatory data on human health (US EPA 1998). Consider for example, the simplest of pharmacokinetic models describing IV bolus injection of a substance into the circulatory system and subsequent elimination (Ke) from the body (Figure 1).
15 Concentration of the chemical in the compartment is highest immediately after injection and concentration at any time t is governed by a single exponential rate of elimination: Cp( t ) = C0*e(-Ke* t ) Where; Cp( t ) = Concentration in the ci rculatory system at time t C0 = Concentration of chemical immedi ately after IV bolus injection e(-Ke* t ) = an exponential elimination rate constant. This first order, mono-exponential fu nction is typically used to model plasma concentration versus time curves with direct injection into a central compartment yielding estimates of compartment volume and chemical elimination. One comp artment behavior of plasma concentration does not ne cessarily imply that the chemical is at the same concentration in all the tissues and organs in the body. Rather, Ke Cp Chemical Figure 1. Illustration of simple one compartment pharmacokinetic model
16 this implies that the concentrations are in instantaneous equilibrium with those in the plasma upon drug administration. An expansion of the one-compartmental approach to pharmacokinetic modeling is to a llow for the distribution of the chemical from the plasma compartm ent into a peripheral compartment (brain, muscle tissue, etc). When dist ribution of a chemical from the plasma into certain organs or tissues is substantially different from the central compartment, multi-compartment models allow for the incorporation of one, or several, pe ripheral compartments (Figure 2). The two compartment model is typically defined where the administered chemical is delivered and eliminated from the central compartment and distribution betw een the central and peripheral compartments is controlled via two distribution micro-constants ( K12 Chemical C1 ( t ) C2 ( t ) K12 K21 Ke Ka Figure 2. Two compartment PK model describing the pathway of chemical introduc tion and elimination from an organism
17 and K21) characterized by the distribution constant Ka (Gabrielsson and Weiner 2000). When these rate cons tants reach a steady state, then the distribution of the chemic al reaches equilibrium and the mechanism governing concentration in the central compartment is controlled through the elimination rate Ke Methods of estimating the concentration in the central co mpartment via a two compartment paradigm include the bi-e xponential function Cp(t)= 1e-Ka*t + 2e-Ke*t where Ka represents the distribution phase from the plasma that includes absorption into the second compartment and Ke represents the post equilibrium or terminal elimination phase governed by elimination of the chemical from the central compartment. Chemical concentration in the plasma is high est immediately after injection in the two-compartment model and behaves similarly to the one compartment model except initially when absorption into the second compartment causes an accelera ted depletion of the chemical concentration from the centra l compartment (Figure 3).
18 3.2 Nonlinear Kinetics Any pharmacokinetic process of a chemical (i.e. absorption, distribution, metabolism or excretion) that cannot be described with first order (linear) kinetics can be considered nonlinear kinetics. Nonlinear kinetics implies deviations in the rate of change in chemical concentration from first order kinetics in a manner that is dose and /or time dependent. Dose-dependent nonlinearity can be due to any carrier mediated process such as me tabolism or active transport that displays transient saturation or c ooperativity at high concentrations (Kwon 2002). Dose-dependent ki netics may be observed as a Figure 3. Bi-exponential de cline of concentration in the central compartment after iv bo lus injection when chemical distribution can be described using two-compartment model (Taken from Kwon 2002). Distribution phase Terminal elimination phase Time Log Cp(t) Distribution phase Terminal elimination phase Time Log Cp(t)
19 stimulated or suppressed response function where increases in chemical concentration result in non additive increases in concentrations in peripheral compar tments in either a synergistic or capacity limited manner. 3.3 Pharmacodynamics To this point we have considered the relationship between administration of chemical into an animal and some of the basic pharmacokinetic parameters describin g the time course of chemical concentration in biological fluids. Obviously, there are many other factors, known and unknown that go vern the specific effects of chemical exposure on the ex posed subject. The study of Pharmacodynamics (PD) is designed to characterize the effect of the chemical on the body. If concentration in the plasma and effect site is in rapid equilibrium, PD models ma y adequately serve to estimate the pharmacological effects of chemical in the body. These models are valuable in capturing the nonlinear aspects of response patterns often seen in pharmacological experiments and in estimating various pharmacodynamic parameters used for establishing dosing regimens in clinical studies (Gabrielsson and Weiner 2000). However, advancements in the field of ph armacology and toxicology have included the realization that more often in in vivo experiments,
20 pharmacological effects take time to develop. Pharmacokinetic/pharmacodynamic (PK/PD) models attempt to link concentration in the plasma and re sponse to chemical exposure by linking the pharmacokinetic properties of a chemical to the biological response observed at the effect si te. While direct and simultaneous measurement of chemical concentrat ion at the effect site and its pharmacological effect is the most desirable approach to reveal the true pharmacodynamic profiles of a su bstance, it is seldom feasible to measure chemical concentration at th e effect site because of limited accessibility and availability to the site (Kwon, 2002). 3.4 Pharmacokinetic/Phar macodynamic Modeling Pharmacokinetic/Pharmacodyna mic modeling has become a common mechanism for elucidating th e chemical/effect relationship when a distributional delay exists or when the system is subject to time or dose-dependent kinetic and/or dynamic changes. Effect compartment models are a class of PK/PD models used to account for differences in concentration be tween the plasma and effect compartments by introducing an equilibrium rate constant Keo (Holford and Sheiner, 1981) (Figure 4).
21 This rate constant Keo eliminates the discrepan cies observed between the plasma and effect compartments resulting in an estimate of chemical concentration ( Ce) in the effect compar tment. For example, an equation used to link a sing le compartment pharmacokinetic equation assuming intravenous inje ction of drug into the central compartment to a pharmacodynamic model is the following biphasic kinetic equation. ()()eeoKtKt eo e eoeK D Ctee VKK Where: Ce(t)= Concentration in the effect site at time t D V= dose/volume equivalent to Ce(t=0) Ke = Elimination rate for the central compartment Chemical Cp(t) K eo Effect Ke Ce(t) Fi g ure 4. Illustration of the Effect-Compartment model.
22 Keo = Distribution rate for the effect compartment This model is used to account for differences between the central and effect compartments caused by di stributional delay (Kwon 2002). This function becomes central to the biphasic toxico-diffusion model discussed in Chapter 5. Once Ce(t) is derived, effect compartment models can then be used to estimate pharmacodynamic response Since observed response can be represented by either an elevatio n of the measured effect or an inhibition of the effect, these effect compartment models are classified as either Emax (elevated response) or Imax (inhibited response) models. ) ( ) ( ) (50 maxt C EC t C E t Ee e A baseline response E0 can be added to reflect change from baseline. The term Emax in the equation correspond s to the theoretical maximum effect while EC50 represents the half-life ti me at which the effect is half of the maximum response. A power term ( n ) can be added to the Emax or Imax models making them sigmoid models (sometimes referred to as the Hill equation) to repr esent sigmoid concentration versus effect curves related to carrier me diated processes such as enzyme cooperativity (negative or positive) affecting the concentration effect curve.
23 ) ( ) ( ) (50 maxt C EC t C E t En ne n e Exponents less than 1 increase the in itial slope of the curve while an exponent greater than 1 makes the curve more shallow initially. These curves converge around the EC50 value and then cross with larger exponents reaching Emax more quickly while smaller exponents yield asymptotic values less than Emax. An exponent of 1 reduces the equation to the standard hyperbolic Emax model. 3.5 Michaelis-Menten Equation The effect compartment models are based on a well defined equation developed in the early tw entieth century to describe the fermentation of cane sugar (i.e. su crose) via hydrolysis into glucose and fructose. Michaelis and Menten derived an equation built upon earlier work by Henri and others by carrying out definitive experiments using the enzyme invertase (Corni sh-Bowden 1995). They found that the rate of the reaction v was dependent on the substrate (sucrose) concentration (a) and limited by the concen tration of enzyme in the reaction. The reaction is described by the equation:
24 a K a V vm max Figure 5. The Michaelis-Menten equation with illustration of the relationship between substrate conc entration (a) and velocity of the reaction. Km is the concentration of substrate that equaled half the maximum velocity of the reaction. Where: Vmax is the limiting rate for the velocity of the reaction and Km is defined as the Michaelis co nstant describing the substrate concentration at which the velocity of the reaction is 1/2 that of Vmax. Among the many important contributi ons to the description of this reaction was the measurement of th e initial rate of the reaction ( Vmax/Km) at different sucrose concen trations thereby avoiding complicating factor such as revers e reaction, product inhibition and inactivation of the enzyme (Corni sh-Bowden, 1995). This equation is recognized as the fundamental equa tion of enzyme kinetics and has been widely used to describe biolog ical processes in areas outside of its original intent. In pharmacologica l studies, a relationship described a Vmax0.5V Km Initial slope Vmax/Km 2Km v0 a Vmax0.5V Km Initial slope Vmax/Km 2Km a Vmax0.5V Km Initial slope Vmax/Km 2Km v0
25 by this type of equation is refe rred to as having Michaelis-Menten kinetics.
26 Chapter 4 Nonlinear Mixed Effects Models The HRAMG has advanced the estima tion of dose-time-response in FOB studies by using non-linear mixe d effects methods for estimating response to chemical exposure. These models capture the dosedependent response to chemical expo sure while allowing for individual subject variation in their natural responses to the measurement instrument in the absence of chem ical exposure. Increasingly, these nonlinear models are evolving to incorporate parameters thought to describe well known physiological me chanisms governing response. 4.1 Toxico-Diffusion Model Zhu (2005) used a re-parameterized version of the Michaelis-Menten equation for the analysis of the FOB data. t t d conc C t t d conc B t d f ) ( 1 ) ( ) ( In this equation ) exp( ) ( t K V d t d conce
27 is dependent on the parameter Ke which can be interpreted as the elimination rate under intravascular administration (Zhu 2005a). The parameter V represents the volume of the circulation system or central compartment. In the toxi co diffusion function, V is absorbed into the parameters B and C in the function. This function resembles the Michaelis-Menten equation, but with concentration-dependent coefficients B*conc(d,t) and C*conc(d,t). The coefficient C can be negative as long as the denominator is positive within the experimental range (Zhu 2005a). By incorporating a baseline response ( A ), the equation descr ibes the change in response from the initial condit ion measured prior to exposure throughout the time course of th e study. Assuming that chemical concentration is directly linked to response, rapid elimination of the chemical corresponds the function f(d,t) quickly reaching a peak value at the time of peak effect (TOPE) and returning tobaseline. However, if the compound remains in the subjects system, f(d,t) may not return to the baseline line level, characterizing persistent dose effects. This function, relying on a one compartment kinetic paradigm, results in a dose independent TOPE. As t varies from 0 toward infinity, f(t,d) varies from baseline to a maximum at eK t / 1 then back to baseline, irrespective of dose level. Statistical evidence of neurotoxic
28 effects is present when either the coefficient B or C is statistically nonzero. 4.2 Parameter Estimation Methods Implementation of the toxico-di ffusion function represents a situation where the observed resp onse is predicted based on a nonlinear function of the estimated parameters in the model. Often in application of these nonlinear models it is advantageous to incorporate a random effects parameter in addition to the fixed effects coefficients to account for natural variation in biological response to the testing instrument. Thus, random effects represent deviations from the population average and can enha nce parameter estimation and hypothesis testing procedures by ac counting for a source of variation in the data otherwise subjected to the error term. Given the short temporal sequences of FOB data, it is rare that the data can accommodate more than one random effe ct even if biologically feasible (Zhu 2005a). Models that incorp orate both fixed and random components are often termed mix ed-effects models and a detailed description of their development an d implementation can be found in Pinhiero and Bates (2000). These models generally use maximum likelihood (ML) methods for paramete r estimation. Likelihood functions for mixed-effects models are gene rally complex, and closed form solutions are generally not available. Implementation of ML requires
29 iterative numerical procedures such as Newton-Raphson algorithm, the EM-algorithm, and more often the co mbinations of them (Lindstrom and Bates 1988). The model fitting techniques used in this thesis rely on the Newton-Raphson algorithm. 4.3 Model Selection and Diagnostics There are several important assumptions associated with mixed effects models that require validat ion. Namely, random effects are assumed to be normally distributed around a mean of zero; the random effects and the error term are assumed to be uncorrelated, and the variance is assumed consta nt across different experimental conditions. Choosing the most appr opriate model to represent the response observed in the data requ ires both objective criteria and sound investigative principle. Th e Likelihood Ratio Test (LRT) and Information Criteria (AIC and BIC: Pinhiero and Bates, 2000) are useful tools for guiding appropriate model selection. Significance testing using the LRT relies on the Chi-squared distribution (i.e. 2(log(L1)-log(L2)) ~ 2 p) where L1 is likelihood function of the expanded model and L2 is the existing model, and p is the number of additional parameters in L1 While this statistic is readily available in most computer software packages, the one-sided alternative test may require weight adjustments in some cases (Zhu, 2005). Another
30 restrictive requirement of LRT is th at one model must be a sub-model of the other comparison model. In formation Criteria (Akaike, 1973) are a generalization to the Likelihood Ratio Test. They add to the likelihood ratio a term to penalize the inclusion of excessive terms in the model. They do not require on e model being a sub-model of the other, but do require the mode ls follow the same family of distributions. Akaikes Informat ion Criteria (AIC) and Bayesian Information Criteria (BIC) are two popular methods that yield values for comparing the goodness of fit of two models. The model with the smaller value of AIC and/or BIC is favored. Investigative principles include graphical comparisons of th e fitted model with the raw data and residual plots to check for ra ndomness of error terms. These are helpful tools for selecting the appr opriate model in conjunction with statistical criteria such as the Log-Likelihood Ratio Test (LRT) and Information Criteria (AIC and BIC). Graphical tools are useful for visualizing not only the shape of the dose-response but also the agreement of the predicted and obse rved responses. Residual plots check the assumptions of independent, normally distributed errors and identify potential outlying observations.
31 Chapter Five Expanding the Toxico-Diffusion Function 5.1 Biphasic Toxico Diffusion Function This thesis generalizes the toxico-diffusion function developed by Zhu (2005) to estimate the dose-timeresponse relationship in which a dose-dependent TOPE is observed. The biphasic toxico-diffusion function is a modification of the bi phasic kinetic equation discussed in chapter 3 that accounts for discrep ancies between the central and effect compartments in pharmacokine tics. Its foundation is the toxicodiffusion function with the addition of a second exponential term to achieve a dose-dependent TOPE. The rate of the second exponential term is dependent on dose as well as time. This second exponential term forms a complex exponent with the elimination rate either in the numerator (Equation A) to describ e accelerated TOPE or in the denominator (Equation B) to de scribe the delayed TOPE often observed in capacity limited kinetics. (A) ) exp( 1 )) * exp( ) (exp( ) (0t K Ctd t dose K t K Btd d t fe e e
32 (B) Therefore, these models do require some a priori knowledge of the dose-response relationship to select the appropriate model to describe dose dependent TOPE. At some point in the time course of chemical exposure, assuming Keo >>Ke the second exponential term of the biphasic model reduces to zero and the model reverts to the standard toxico-diffusion function. As with the pharmacodynamic models, the biphasic toxicodiffusion models are not intended to be mechanistic in the literal sense of describing actual pharmacologic or kinetic processes. Rather, here the purpose of these models is to incorporate pharmacokinetic concepts into a mathematical model th at allows for a dose-related shift in TOPE that may be related to well known pharmacokinetic processes such as nonlinear disposition, tissu e or protein binding, metabolism or clearance kinetics observed in ph armacological and toxicological studies (Gabrielsson and Weiner 200 0). Irrespective of the actual mechanism governing the observed shift in TOPE, these proposed )) * exp( ) (exp( 1 ) exp( ) (0t dose K t K Ctd t K Btd d t fe e e ) exp( 1 ) exp( ) ( t K Ctd t K Btd d t fe e
33 models may be useful for describi ng the FOB data and elucidating dose-response relationships which exhibit dose-dependent TOPE. To illustrate the use of the biph asic toxico-diffusion models, parathion motor activity counts were chosen as the response variable of interest due to an observed shift in the time of maximal response at higher dose levels. In the IPCS st udies parathion was introduced by oral gavage in corn oil solution at doses of 0.85, 1.69, 3.38 and 6.75 mg/kg. Parathion is readily absorb ed through the digestive tract and detection in the bloodstream has been reported immediately after oral administration (INCHEM 2004). No nlinear binding of the toxic parathion metabolite, paraoxon, has been reported for red blood cells (with supra-linear dose-response) and brain tissue (with a sub-linear dose-response) (Vogel et al. 2002). Kramer et al. (2002) used a three compartment model to describe the pharmacokinetics of Methyl Parathion but failed to elucidate the pharmacodynamic properties associated with this hazardous or ganophosphorus pesticide. For the model described in this thesis to be related to physio logical processes, bioavailability is assumed to be 100% and absorption from the gastrointestinal tract is assumed to be nearly instantaneous.
34 5.2 Parathion Activity Counts Examination of the raw data plots reveal the relationship between dose, time and activity counts obse rved after a single exposure to parathion (Figure 6). These groupe d data plots are constructed such that the control group is the first gr aph (starting on the left), followed by panels of increasing dose to the right. The x axis is in log scale for display purposes only. It is apparent that dose increases resulted in changes of the response trajectories (Table 1). The two hi ghest dose groups (3.38mg/kg and 6.75 mg/kg ) appeared to have vast ly lower activity counts than the lower dose groups. The observed TOPE associated with the higher dose groups was 2H compared with a TOPE at 24H for the controls and the lowest dose group. Figure 6. Observed response in ac tivity counts of rats subjected to acute parathion exposure. Note : Panels are arranged in an increasing order of dose (0, 0.85, 1.68, 3.38, 6.75) from left to right.
35 Table 1. Dose group average activity counts for motor activity counts after acute parathion exposure with standard deviation in parenthesis Time Dose 0 2 24 168 0 323.6 (51.3) 225.6 (72.1) 180.2 (71.4) 222.3 (70.0) 0.85 316.0 (53.2) 217.8 (83.9) 178.9 (76.0) 257.5 (60.5) 1.69 291.9 (37.5) 207.6 (44.5) 185.5 (59.9) 232.5 (95.8) 3.38 336.2 (57.8) 139.6 (99.2) 144.0 (42.4) 224.9 (77.9) 6.75 296.2 (64.1) 15.9 (14.7) 93.6 (49.9) 201.7 (67.7) These aspects of the observed da ta suggested that the parathion activity count data may allow for an application of the biphasic numerator model. Therefore, the bi phasic numerator model (equation 1) was applied the parathion motor activity counts to describe the apparent acceleration of TOPE observed in the data. 5.3 Fitted Response The fitted biphasic model predicted a decline in activity counts after exposure to parathion, and the predicted trend resembles the observed pattern. However, only the intercept A and the rate parameter Ke were statistically significant, the slope factors B and C and Keo are insignificant (Table 2). A la rge amount of variation in the data was attributable to the between rat variation at baseline and was accounted for by a random effect for baseline activity count. The random effects had a standard deviation 35.9, about 54% of residual standard deviation (Table 2).
36 Figure 7. Fitted response of bi phasic numerator model for rats sub j ected to acute p arathion ex p osure. Table 2. Summary of biphasic mode l fit to parathion activity counts Parameter Estimate Standard Error DF t-value p-value A 258.0129 7.91744 146 32.588 <.0001 B -72.3701 83.48372 146 -0.867 0.387 C 0.2265 0.35207 146 0.643 0.521 Ke 0.1400 0.03561 146 3.931 0.0001 Keo 0.3198 0.21072 146 1.518 0.131 Random Effects Intercept Residual 35.93553 65.08784 The fitted biphasic model appeared to be a reasonable fit the observed data. The fitted dose-response pattern resembles the observed data in aspects of TOPE and magnit ude of response changes. Evidence of a good fi t of the model to the data comes through examination of the residual plots. Residual plots express the difference between the observed and predicte d responses, termed error. An assumption of nonlinear mixed effect s models is that the errors are
37 independent and normally distributed. Standardized residual plots for the parathion activity counts (Figur e 8) suggest that the errors are centered on a mean of zero and fairly randomly distributed. The NLME models further assume that the errors are constant across dose group while often in biological settings the variance will increase in proportion to the mean response When there is evidence of nonconstant standard deviation, NM LE can incorporate a separate standard deviation to each dose gr oup and verify the resulting model improvement using the Likelihood Rati o Test (LRT). The addition of dose group specific standard deviations to compensate for heteroscedascity did not significantl y improve the model fit according to the likelihood ratio test (log(L 2)-log(L1)) value of 7.28 and a corresponding p-value of 0.122. This suggests that a constant standard deviation across dose groups is acceptable. Figure 8. Residual plots of biphasic model for rats subjected to acute parathion exposure. No te: Panels are arranged in an increasing order of dose (0, 0.85, 1.68, 3.38, 6.75) from left to right. -2 -1 0 1 2 3 050100150200250300 dose 050100150200250300 dose 050100150200250300 dose 050100150200250300 dose 050100150200250300 dose Fitted values Standardized residuals
38 5.4 Dose-Dependent TOPE The fitted model also reveals different times of peak effect determined by dose level (Figure 9 ): the higher dose groups clearly had an accelerated TOPE. The TOPE varied from 11.5 hours for the lowest dose group (0.85mg/kg) to 7.3 and 7.1 hours for the 3.38mg/kg and 6.75mg/kg dose groups, respectively. TimePredicted Response 050100150 050100150200250 0.85mg 1.68mg 3.38mg 6.75mg The fitted model can also be ex amined for each subject after accounting for the random effect at baseline (Figure 10). Indeed, the predicted model appears to capture individual (red) deviation when compared to the population average (blue). Figure 9. Dose specific trajecto ries for biphasic model on rats subjected to acute parathion exposure.
39 0 100 200 300 400 050100150 1755 1759 050100150 1723 1728 050100150 1737 1763 050100150 1751 1742 050100150 1732 1767 1743 1747 1764 1760 1729 1756 1733 0 100 200 300 400 1738 0 100 200 300 400 1719 1768 1758 1746 1750 1754 1727 1762 1736 1722 1725 1731 1749 1745 1740 1726 1766 0 100 200 300 400 1753 0 100 200 300 400 1757 1721 1741 1730 1769 1744 1765 1739 1761 1752 050100150 1720 1734 050100150 1748 0 100 200 300 400 1724 Test Time (Hour)Activity Counts fixedrat. The biphasic model can be viewed as a generalization of the toxicodiffusion model to allow for a dose -dependent TOPE. Given that these models have this hierarchical stru cture, Information Criteria and the Log Likelihood Ratio test were used to test for improvement in the fit of the biphasic model relative to th e fit of the toxico diffusion model (Figure 11). Both Information Criteria and the Log Likelihood Ratio test suggest a significant improvement of the biphasic model overall (Table 3). Despite the fact that the dose-d ependent coefficients (B and C) Figure 10. Individual specific tr ajectories for biphasic model on rats subjected to acute parathion exposure.
40 and the second kinetic coefficient Keo in the biphasic model are statistically insignificant, due perhap s to limited number of data points (in time or subjects) in the experiment, the biphasic model is attractive because it permits dose-v arying TOPE as indicated by the motor activity count data. Further, the magnitude of response predicted by the biphasic model wa s dramatically different between the two models with the toxico-diffusion model predicting peak response (in this case lowest activi ty count) of 100 at the highest dose level while the biphasic model pr edicted a peak response (lowest activity count) of three; the true response observed in the data (see figure 6). Correctly predicting th e magnitude of peak effect is important as the peak effect could have dramatic implications on benchmark dose estimation. On th e grounds of toxicology, it is generally accepted that the TOPE is less likely to be a constant. With only a limited number of time points (4 in the case of motor activity counts) spread over a wide range, however, variance in TOPE is less likely to be detectable from the FOB data. Therefore, verification of a true dose-related TOPE is difficul t because the statistical power in detecting the variation in observe d TOPE from standard FOB assay data is low.
41 TimePredicted Response 050100150 120140160180 0.85mg 1.68mg 3.38mg 6.75mg Figure 11. Predicted response to parathion by the toxico-diffusion model. Table 3. Results of log likelihood ( LL) ratio test comp aring the toxicodiffusion (TD) and Biphasic Numerator (BNM) models. Model DF AIC BIC LL Chi Sq. Pr>Chi.Sq TD 6 2297.751 2317.541 -1142.876 BNM 7 2291.730 2314.818 -1138.865 8.021 0.0046 The analysis thus far has demonstrated that FOB assays using scarcely spaced experimental time points will likely fail to generate sufficient data for dose-response modeling especially when dose varying TOPEs are considered. The next chapter focuses on considering FOB designs that may provide enough information about the true dose-response necessary to capture dose-dependent response
42 patterns of neuro-toxicological effe cts, particularly dose-dependent time of peak effects, in a valid and reliable manner. In the next chapter, computer simulation is used to explore the most effective FOB design protocols to predicted dose-d ependent TOPE using the biphasic toxico diffusion model.
43 Chapter 6 Simulation 6.1 Simulation Rationale While the biphasic toxico-diffusion model appeared advantageous relative to the toxico diffusion model in permitting dosedependent TOPE, the para thion motor activity count data leaves some uncertainty about the model as the parameter estimates for the B, C and Keo were statistically insignificant. This lack of statistical power seemed to result from insufficient data generated under the current FOB design. Specifically, the sparse spacing of experimental testing times is believed to provide insufficient data for model parameterization. Therefore, Mont e Carlo simulation was adopted to investigate the design protocol s necessary to recover the key characteristics of dose-dependent re sponse along the time course as predicted by the biphasic model. The primary objective of the simulation was to investigate the be nefits of considering alternative time spacing and/or adding additional experimental testing times and/or experimental subjects to the FOB design to recover the key characteristics of the biphasic model with regard to parameter
44 estimation. Assuming that the biph asic model represents the true underlying response of rat motor acti vity to acute parathion exposure, we could construct a da taset through simulation that contained the underlying response at a sequence of possible testing times in addition to the experimental times in the FOB study. We could also control the number of subjects in each dose group. The simulation experiments then allowed us to evaluate, empirica lly, the efficiency of designs with various locations and frequencies of testing times as well as with additional subjects needed for the biphasic toxico-diffusion model to capture the dose and time depend ent characteristics of the true response. 6.2 Simulation Methods We obtained the parameter estimates from fitting the biphasic model to the motor activi ty count data of rats exposed to parathion. These estimates were then used as the population (true) parameters in the biphasic function to defi ne the true dose-response model. Dose levels remained the same as th ose used in the original parathion experiment. Within the content of simulation, the fitted model represents the true underlying mean response to parathion, and the random effects and random errors govern variation among rats in a given population. Generating simulation datasets required several steps:
45 Step 1. Generate the mean underly ing dose-time-response curve using the biphasic model (with the parameter estimates in Table 1) at selected dose and time levels. This step results in average response specific to the given dose and time involved in the designed experiment (Figure 12). Step 2. Add random effects. To account for each subjects deviation from the average score at ba seline, random effects were generated for each subject us ing a normal distribution, N(0, D), where the standard deviation ( D ) was taken from the estimate (35.94) of the fitted biphasic response model. The random effects were added to the mean response at every dose-time point in the form of the random intercepts. In the present simulation there was one random effect for each rat. Step 3. Add random errors. Random errors were generated for each individual observation from a normal distribution N(0, ) and added to the mean response to represent measurement variation ( = 65.09) around the mean response. Step 4. Repeat Steps 1-3 n times to generate data of n rats per dose group. This process generates one simulated experiment. Step 5. Repeat Step 4 one th ousand times to generate 1000 replications of the experiment.
46 Based on this simulation protocol, we generated motor activity scores for n rats in each dose group at a se quence of desired testing times. We simulated responses at each of th e time points used in the original study (i.e. 0,2,24,168) as well as addi tional points in 4 hour intervals between times 4 and 24 (i.e. hours 4, 8,12, 16,20). In this way we had information at a multitude of testing times around the time of peak effect, which allowed us to identify effective designs for recovering key characteristic s of the biphasic model. The purpose of analyzing these simulated experiments was to assess the efficiency of these expe riments with respect to recovering the true underlying dose -response information, specifically dosedependent TOPE, with a high level of statistical certainty (power). Thus, we first considered two groups of simulation experiments, 4time point designs and 5-time point designs with ten subjects per dose group. While a larger number of time points are statistically desirable, designs with 4-5 time points are more practical and more closely follow the EPA neurotoxicity risk gu ideline (USEPA, 1998). Intuitively, a time point in the neighborhood of the true TOPE would provide the most relevant information on the TOPE. Therefore, in the 4-point design, we retained three time points, 0, 24, and 168 hours from the original design protocol and let the 2nd time point vary between 2 h and 24 h. These design variations ai m at assessing the efficiency of
47 the original FOB experimental de sign. In the 5-point design, an additional time point between 4 and 24 hour was added to the original FOB experiment. Testing TimeActivity Counts 024812162024 0 50 100 150 200 250 300 350 0.85mg 1.69mg 3.38mg 6.75mg In addition to the 4-point an d 5-point designs, we also considered designs with 6, 7, 8, or even 9 time points for the purpose of statistical dose-response mode ling. These simulated experiments helped illustrate the relative gain of using additional points in Figure 12. Theoretical true dose-dependent biphasic model response between 0 hours and 24 hours. Candidate experimental testing times used for simulation are indicated by the broken vertical lines
48 increasing statistical power or gaining information on the doseresponse. In order to see the effect s of sample size on statistical power, we also considered using more subjects (i.e. 20, 30) at each dose level under the 4 and 5time point designs. Note there are three 5-point designs that did not use 2 ho urs as an experimental point but rather contain 2 testing times betw een 4 and 12 hours as well as at hours 0, 24 and 168. One thousand data replicates of each of the 35 design trials listed in Table 4 were generated and analyzed to evaluate each design. Efficiency of the design was ev aluated in the following ways: 1) The convergence rate the num ber of fits on the biphasic numerator model that reached su ccessful model convergence out of the 1000 replicate trials. 2) Statistical power: The percentage of times the t statistic for each parameter exceeded its 95th-percentile (p-value less than 0.05, or type I error at 0.05). 3) Bias = ( (i )/n): The difference between the true model parameter and the average of 1000 replications of the parameter estimate.
49 4) Mean Squared Error = ( (i )/n)2+ 2: The Bias squared + the sample variance of the 1000 re plications of the parameter estimate. 5) The Bias and MSE of the Ti me of Peak Effect (TOPE). These statistical criteria were used to compare the designs and identify the more efficient designs with re spect to dose-response modeling of the biphasic toxico-diffusion mode l. It should be noted that the assessment of power for any parameter estimate precludes the situation that the parameter is not meaningful both biologically and mathematically. This is because any parameter can be imposed and signified statistically with a suffici ently large dataset. However, the case of biphasic model indeed prec ludes the possibility of statistical manifestation. While found to be statistically insignificant in the original model fit, the model parameters have inherent importance in the model function to describe the resp onse (e.g. initial rate of change in response, half-peak effect dose and dose dependence in TOPE). Comparison with different models suggested the magnitude of the parameter estimates were non-trivial, but the design was underpowered to ascertain the value. Within the context of simulation, the scenario of such a statistical manifestation can be plainly excluded
50 because data were generated from th e biphasic model and all involved parameters are mathematical quantities in existence. Table 4. Simulated experimental FOB designs using 10, 20, and 30 subjects per dose group. Experi mental testing times were added between 4 and 20 hours. Testing Times (Hours) Number of Subjects Number of Time Points 4 time-point designs 0,2,24,168 10 4 0,4,24,168 10 4 0,8,24,168 10 4 0,12,24,168 10 4 0,16,24,168 10 4 5 time-point designs 0,2,4,24,168 10,20,30 5 0,2,8,24,168 10,20,30 5 0,2,12,24,168 10,20,30 5 0,2,16,24,168 10,20,30 5 0,2,20,24,168 10,20,30 5 Alt. 5 time-point designs 0,4,12,24,168 10,20,30 5 0,8,12,24,168 10,20,30 5 0,4,8,24,168 10,20,30 5 Extended time-point designs 0,2,4,12,24,168 10 6 0,2,8,12,24,168 10 6 0,2,4,8,12,24,168 10 7 0,4,8,12,16,24,168 10 7 0,2,4,8,12,16,24,168 10 8 0,2,4,8,12,16,20,24,168 10 9
51 6.3 Simulation Results 6.3.1 Four Time Point Designs: Ten Subjects In the 4-time point design, only the 2nd point differs from the original FOB experiment protocol. The 2nd time point varied from 2 to 16 hours, the results showed pron ounced effect on the power of detecting the parameter Keo (Table 5). When the second time point was chosen to be hours 8 or 12, th e statistical power for the non-zero Keo improved from 45 % to approx imately 80%. However, for the parameters B and C the power of detecting a non-zero value remained low, ranging from 0.1% to 15.5%. Th e convergence rate in fitting the biphasic model decreased markedly as the second testing time moved past 12 hours. For example, the co nvergence was only about 72.9% at 16 hours, suggesting the data did no t provide adequate information to even fit the model. Note that the sh aded row indicates the design that appeared to be most beneficial for re covering the true model response. Table 5. Summary of simulated design trials with 10 subjects per dose group and 4 testing times. Contd next page 4 Point Designs PARAMETER Power (% p<0.05) (average p value of t-statistic) Time Points B C Keo Convergence (%) 0,2,24,168 15.5(0.329) 0. 0(0.466) 45.5(0.118) 94.1 0,4,24,168 11.5(0.309) 0. 0(0.392) 69.1(0.068) 93.1 0,8,24,168 5.7(0.259) 0.0(0.329) 81.6(0.057) 88.5 0,12,24,168 1.2(0.271) 0. 0(0.341) 80.5(0.063) 82.6 0,16,24,168 0.1(0.331) 0. 0(0.398) 75.7(0.091) 72.9
52 Table 5 Contd 1) 2nd testing time was selected at various times from 2-16 hours and effects on power and statistical significance examined. 2) Numbers in columns 2, 3, and 4 indica te the percent of time a p value <0.05 was observed while numbers in parent heses represent the average p_value based on 1000 replications. 3) The first row (design 0,2,24,168) is the true FOB design on which this simulation was based Bias and MSE were also used to evaluate an efficient design for parameter estimation using the biphasic model. Bias represents the average difference in the location of the parameter estimate while MSE represents the imprecision of that estimate across the 1000 replications of each simulated experi mental design. Consistent with the results of the power analys is, bias and MSE for the Keo parameter were the smallest at 8 hours and therefor e represented the best choice of testing times within this group of de signs (Table 6) for this endpoint. However, the bias and MSE in parame ters B and C appeared to be the smallest when the second testing ti me was at 16 hours. Convergence decreased markedly for designs with the hour 16 being a testing time, which may have affected the MSE for this design by eliminating some of the larger variations in the pa rameter estimates. The MSE for the hour 8 testing time was not marke dly different from the hour 16 design for parameters B, C and Ke and had a higher convergence rate. This lends further support for the 0, 8,24,168 design as the best choice among the 4 time point designs. In general, the MSE was so large for the B parameter that it should be no surprise that statistical
53 significance was not achieved for this parameter in any of these designs. In view of the results in Table 6, one can conclude that 4point designs did not have the statis tical power to reliably detect the appropriate biphasic model for this endpoint. Table 6. Bias and MSE for each pa rameter of interest for the 4 time point designs with 10 subjects. Design Bias_B Bias_C Bias_Ke Bias_Keo 0,2,24,168 -1.59E+03 8. 91E+00 6.05E-03 7.47E-02 0,4,24,168 -1.76E+04 9. 81E+01 1.12E-02 2.88E-02 0,8,24,168 -1.33E+02 6.59E-01 1.12E-03 -1.89E-02 0,12,24,168 -1.29E+02 6.48E -01 -1.09E-02 -6.43E-02 0,16,24,168 -9.87E+01 5.29E-01 -2.80E-02 -1.10E-01 Design MSE_B MSE_C MSE_Ke MSE_Keo 0,2,24,168 6.55E+08 2.25E+04 3.03E-03 8.81E-02 0,4,24,168 7.94E+10 2.50E +06 4.34E-03 3.39E-02 0,8,24,168 3.53E+05 8.69E+00 3.16E-03 2.10E-02 0,12,24,168 1.63E+06 3.76E +01 4.52E-03 2.35E-02 01624168 3.08E+05 8.52E+00 6.59E-03 3.39E-02 6.3.2 Five Time Point Designs: Ten Subjects Due to the lack of statistical po wer in parameter estimation in the 4-time point designs, we exam ined the benefits of adding an additional testing time. We retained the 4-time points of the FOB design and allocated an addition al testing time between 4 and 20 hours. In a second situation of 5point designs we also removed the hour 2 time point and added 2 time points between 4 and 12 hours to
54 see if time points closer to the true TOPE would increase the statistical power. The 5-point designs generally improved the power for all parameters of interest (Table 7). The average p value for Keo reached a level below 0.05, and the power im proved to above 80% for designs with the additional testing time s between 4 and 12 hours. The improvement in power was less prono unced when the additional point was beyond 16 hours. For parameter C, the power remained at a low level of less than 10%. For the B parameter, the power reached a highest level of 28% when the added time point was at 16 hours, and stayed at a comparable level within the time range of 8 to 20 hours. Table 7. Summary of simulated desi gn trials on FOB parathion data with 10 subjects per dose group and 5 testing times. 1) The experimental testin g time (3rd testing time) was adjusted by 4 hour intervals and effects on power examined. 5 Point Designs PARAMETER % p<=0.05 (mean) Time Points B C Keo Convergence (%) 0, 2, 4, 24, 168 21.2(0. 220) 2.0(0.303) 80.9(0.030) 97.3 0, 2, 8, 24, 168 24.5(0.147) 9.8(0.213) 82.4(0.030) 99.0 0, 2, 12, 24, 168 24.5(0.147) 9.8(0.213) 82.0(0.033) 98.4 0 ,2, 16, 24, 168 27.7(0.158) 7.4(0.233) 73.5(0.047) 97.2 0, 2, 20, 24, 168 25.8(0.198) 2.6(0.292) 58.3(0.072) 96.4 Alternate 5 point designs 0, 4,12, 24,168 22.0(0.160) 7.8(0.222) 87.4(0.033) 95.2 0, 8,12, 24,168 17.6(0.144) 1.6(0.204) 89.6(0.032) 91.3 0, 4, 8, 24,168 21.2(0.169) 5.4(0.232) 90.4(0.024) 96.4
55 2) Numbers in columns 2, 3, and 4 indica te the percent of time a p value <0.05 was observed while numbers in parent hesis represent the average p_value based on 1000 replications. 3) The true design on which this simula tion was based includ ed actual testing times of 0,2,24, and 168 hours. Interestingly, dropping the hour 2 testing time and adding two testing times between 4-12 hours had little effect on the statistical power though in one case (0,4,8,24,168) a large bias was introduced (Table 8). The bias for B, C, and Ke was smallest with the design of 0,8,12,24,168 hours while the bias of Keo was the smallest with the 0,4,8,24,168 design. The MSE was smallest with the 0,8,12,24,168 design for three of the four parameters of interest. Table 8. Bias and MSE for each parameter of interest for the 10 subject, 5 time point design. Design Bias_B Bias_C Bias_Ke Bias_Keo 0,2,4,24,168 -3.50E+02 1. 86E+00 1.00E-02 4.11E-02 0,2,8,24,168 -1.33E+02 6. 95E-01 9.53E-03 4.80E-02 0,2,12,24,168 -1.33E+02 6. 93E-01 8.68E-03 4.56E-02 0,2,16,24,168 -1.23E+02 6. 41E-01 5.04E-03 5.52E-02 0,2,20,24,168 -1.17E+02 6. 05E-01 3.89E-03 7.47E-02 0,4,12,24,168 -5.71E+02 3. 11E+00 5.77E-03 1.96E-02 0,8,12,24,168 -6.99E+01 3.58E-01 -4.18E-04 -2.22E-02 0,4,8,24,168 -1.47E+ 08 8.49E+05 8.97E-03 1.56E-02 Design MSE_B MSE_C MSE_Ke MSE_Keo 0,2,4,24,168 3.39E+07 1. 06E+03 1.81E-03 3.43E-02 0,2,8,24,168 1.28E+06 3.70E+01 9.67E-04 5.17E-02 0,2,12,24,168 1.28E+06 3. 70E+01 1.08E-03 5.22E-02 0,2,16,24,168 2.94E+06 8. 05E+01 1.31E-03 5.20E-02 0,2,20,24,168 3.55E+05 9. 73E+00 1.67E-03 7.02E-02 0,4,12,24,168 2.21E+08 6. 78E+03 1.91E-03 3.16E-02 0,8,12,24,168 1.68E+05 4.05E+00 2.40E-03 1.91E-02 0,4,8,24,168 2.16E+19 7. 22E+14 2.36E-03 2.11E-02
56 6.3.3 Designs with Six Nine Time Points The results of the 5-time point designs begs the question How many testing times will be enough to achieve required statistical power for the B and C parameters and acceptable bias and MSE?. In an attempt to answer this question, we increased the number of testing times from 5, to between 6 to 9 po ints, by adding additional points between 0 and 24 hours. As the num ber of testing times increased, the statistical power improved but even with 9 time points, the statistical power was still below 70% for B and 37% for C (Table 9). It is clear that as the number of ti me points increases, the power increases generally. Furthermore, fi xing the number of time points yielded some data more informative than others. Table 9. Summary of simulated expe riments with six to nine points and 10 subjects per dose group. PARAMETER % p<=0.05 (mean) Time Points (Extended time point designs) B C Keo Convergence (%) 0,2,4,12,24,168 33.5(0.113) 13.2(0.171) 95.4(0.010) 98.8 0,2,8,12,24,168 43.7(0.084) 16.1(0.133) 93.2(0.017) 98.8 0,4,8,12,24,168 34.8(0.101) 12.6(0.152) 93.5(0.019) 96.9 0,2,4,8,12,24,168 47.4(0.073) 19.7(0.117) 97.3(0.006) 99.2 0,4,8,12,16,24,168 49.0(0.070) 18.8(0.114) 93.9(0.014) 96.8 0,2,4,8,12,16,24,168 62.6(0.049) 31.0(0.086) 98.0(0.004) 99.2 0,2,4,8,12,16,20,24,168 68.3(0. 043) 36.6(0.077) 98.0(0.005) 99.0 1) Two experimental test ing times were added (3rd and 4th testing times) and adjusted by 4 hour intervals an d effects on power examined. 2) Numbers in columns 2, 3, and 4 indica te the percent of time a p value <0.05 was observed while numbers in parent hesis represent the average p_value based on 1000 replications. 3) The true design on which this simulation was based included actual testing times of 0,2,24, and 168 hours.
57 The bias and MSE were generally reduced as the number of testing times increased for parameters B, C, and Ke, although the smallest bias in Keo was attained with the 7 time point design of 0,4,8,12,24,168 (Table 10) and smallest MSE for Keo was associated with the 8 time point de sign 0,2,4,8,12,16,24,168. Even with the 9 time point design, the coefficient of variation for parameter B was still 45% of the average. Table 10. Bias and MSE for Simulate d Experiments with 6 and 9 time points and 10 subjects Design Bias.B Bias.C Bias.Ke Bias.Keo 0,4,8,12,24,168 -6.00E+01 3.06E-01 6.42E-03 1.37E-02 0,4,8,12,16,24,168 -7.32E +01 3.96E-01 4.63E-03 1.15E-02 0,2,4,8,12,24,168 -8.85E+01 4.62E-01 9.12E-03 1.76E-02 0,2,4,8,12,16,24,168 -4.94E+ 01 2.58E-01 7.59E-03 1.77E-02 0,2,4,8,12,16,20,24, 168 -3.99E+01 2.10E-01 6.37E-03 2.14E-02 Design MSE.B MSE.C MSE.Ke MSE.Keo 0,4,8,12,24,168 3.46E+04 8. 02E-01 1.22E-03 2.24E-02 0,4,8,12,16,24,168 8.57E+05 2.74E+01 1.18E-03 1.71E-02 0,2,4,8,12,24,168 1.29E+06 3.77E+01 7.05E-04 1.69E-02 0,2,4,8,12,16,24,168 7.05E +04 1.97E+00 5.94E-04 1.44E-02 0,2,4,8,12,16,20,24, 168 1.09E+04 2.71E-01 5.82E-04 1.84E-02 6.3.4 Designs with Increasi ng Number of Subjects While the simulations have clearly demonstrated that increasing the number of time points and select ing informative time points can generally improve the efficiency of an experiment, there remains a considerable amount of MSE, particul arly associated with parameter B.
58 This suggests that while selection of time points is essential in soliciting information on the shap e of dose-response, to reduce variation increasing the number of subjects may also be important. Therefore, a second simulation was performed using the 5-point experiments with 20 or 30 subjects per dose group. 6.3.5 Five time point designs with twenty subjects Adding 10 additional subjects to each dose group resulted in pronounced improvement in statis tical power for estimating the parameters. It is seen from Table 11 that the power for Keo was consistently above 95% among all experiments, and the power for parameters B and C reached 77% and 44% respectively under the design of 0,8,12,24,168 hours. Table 11. Summary of simulated design trials on FOB Parathion data with 20 subjects per dose group and 5 testing times. PARAMETER % p<0.05 (average p value) Time Points 5 Point Design (20 subjects per dose group) B C Keo Convergence (%) 0,2,4,24,168 33.0 (0.104) 10. 1(0.168) 98.4 (0.004) 99.4 0,2,8,24,168 63.8 (0.049) 24.2 (0.089) 98.6 (0.003) 99.8 0,2,12,24,168 70.3 (0.041) 29.2 (0.081) 96.9 (0.006) 99.5 0,2,16,24,168 61.5 (0.053) 25.3 (0.100) 96.8 (0.009) 99.6 0,2,20,24,168 47.7(0.079) 18. 2(0.142) 93.7(0.016) 99.0 0,4,12,24,168 64.3(0.048) 28.3 (0.086) 98.2 (0.004) 98.6 0,8,12,24,168 76.9 (0.036) 43.7 (0.067) 96.4 (0.011) 97.0 0,4,8,24, 168 57.0(0.056) 25. 5 (0.097) 99.4 (0.002) 98.7 Contd next page
59 Table 10 Contd 1) The experimental testing time (3rd testing time) was adjusted and effects on power examined. 2) Numbers in columns 2,3, and 4 indicate the percent of time a p value <0.05 was observed while numbers in parent hesis represent the average p_value based on 1000 replications. 3) The true design on which this simulati on was based included actual testing times of 0,2,24, and 168 hours. Under these 20 subject experiment s, bias associated with the estimates of B, and C was rath er invariant while bias in Ke was minimized with the 0, 2, 20, 24, 168 design and Keo was minimized with the 0, 2, 8, 24, 168 design. The MSE for the 20 subject designs was generally reduced by at least an order of magnitude compared with the same designs of half th e number of subjects. The smallest MSE was achieved with designs wher e the third time was located at either 8 or 12 hours (Table 12). Th e best statistical properties of the Keo parameter again was achieved under a slightly different design with time points of 0, 8, 12, 24, 168. While increasing the number of subjects per dose group did not redu ce bias, it dramatically reduced the uncertainty of the statistical quantities as measured by MSE. Because of the disparity of stat istical power in estimating the parameters, we further investigated the case of 30 subjects per dose group. Table 12. Bias and MSE for each pa rameter of interest for the 20 subject, 5 time point design. Design Bias.B Bias.C Bias.Ke Bias.Keo 0,2,4,24,168 -5.39E+01 2. 74E-01 8.34E-03 2.03E-02 0,2,8,24,168 -3.96E+ 01 2.08E-01 7.67E-03 1.34E-02 Contd next page
60 Table 12 Contd 0,2,12,24,168 -3.20E+01 1. 73E-01 5.46E-03 2.61E-02 0,2,16,24,168 -3.17E+01 1. 70E-01 4.83E-03 3.66E-02 0,2,20,24,168 -3.17E+01 1.68E-01 2.96E-03 4.94E-02 0,4,12,24,168 -3.53E+01 1. 90E-01 5.41E-03 9.84E-03 0,8,12,24,168 -3.35E+ 01 1.83E-01 3.29E-03 -2.88E-03 0,4,8,24,168 -4.29E+01 2. 26E-01 7.11E-03 3.03E-03 Design MSE.B MSE.C MSE.Ke MSE.Keo 0,2,4,24,168 1.61E+04 3.65E -01 6.46E-04 1.30E-02 0,2,8,24,168 6.43E +03 1.46E-01 3.81E-04 1.29E-02 0,2,12,24,168 5.18E+03 1.21E-01 3.82E-04 2.05E-02 0,2,16,24,168 6.44E+03 1. 47E-01 4.22E-04 2.79E-02 0,2,20,24,168 9.26E+03 2. 08E-01 5.88E-04 3.21E-02 0,4,12,24,168 5.68E+03 1. 29E-01 6.07E-04 1.00E-02 0,8,12,24,168 8.01E+03 1. 80E-01 9.13E-04 8.36E-03 0,4,8,24,168 8.10E+03 1.86E-01 6.52E-04 6.69E-03 6.3.6 Five-Time Point Designs: Thirty Subjects A final simulation was conducted on the parathion data using 30 subjects per dose group. As expected adding an additional 20 subjects to the original design resulted in greater convergence rates and nearly satisfactory statistical power for all parameters. Table 13 clearly shows that the design of 0, 8,12,24,168 yielded the largest power for parameters B and C. We further note that the power of C is particularly sensitive to design whereas B, Ke, and Keo are less so. Table 13. Summary of simulated design trials on FOB Parathion data with 30 subjects per dose group and 5 testing times. PARAMETER % p<0.05 (mean) Time Points 5 Point Design (30 subjects per dose group) B C Keo Convergence (%) 0,2,4,24,168 50.5(0.057) 21.1(0.103) 100(<0.001) 100 0,2,8,24,168 93.0(0.016) 73. 4(0.038) 99.8(<0.001) 99.9 0,2,12,24,168 94.6(0.014) 78.2(0.035) 99.9(0.001) 100 Contd next page
61 Table 13 Contd 0,2,16,24,168 89.4(0.021) 56. 7(0.049) 99.8(0.002) 99.9 0,2,20,24,168 70.4(0.039) 35. 3(0.081) 99.5(0.003) 99.9 0,4,12,24,168 92.7(0.016) 73. 0(0.037) 97.8(0.002) 98.0 0,8,12,24,168 97.8(0.009) 92.0(0.023) 97.0(0.004) 98.5 0,4,8,24,168 88.5 (0.021) 67.3 (0.043) 99.6 (<0.001) 99.9 1) The experimental testing time (3rd testing time) was adjusted by 4 hour intervals and effects on power examined. 2) Numbers in columns 2, 3, and 4 indicate the percent of time a p value <0.05 was observed while numbers in parent hesis represent the average p_value based on 1000 replications. 3) The true design on which this simulati on was based included actual testing times of 0,2,24, and 168 hours We expect that adding additional subjects to the experimental design will reduce the MSE but that the relative performance of the designs will remain consistent with the results under 20 subjects. Indeed, as with the twenty-subject experiments, the 0,2,20,24,168 design with 30 subjects retained the smallest bias for parameters B, C, and Ke while the 0,8,12,24,168 design has the smallest bias for Keo (Table 14). It is intriguing that, except for Ke, bias was reduced under the 30-subject designs, although th e magnitude was small. A plausible explanation is the replication size (1 000) of the simulation may not be sufficiently large to stabilize the variation of the parameter estimators. The design which attained the smallest MSE for Ke was consistent with the 20-subject design (i.e. 0, 2, 12, 24, 168). However, the MSE for parameters B and C and Keo was slightly different (though the same order of magnitude) from the results using 20 subjects. Specifically, the smallest MSE for th ese parameters occurred with the
62 0, 8, 12, 24, 168 design under 30 subjects. This discrepancy may be attributed to (1) reduction in bias and (2) non-convergence in simulations. Reduction in bias may change the expected relationship between MSEs of the 20and 30-subject designs. When a simulated experiment resulted in non-co nvergence in model fitting, the underlying extreme value was excluded in the computation of bias and MSE. Since the non-convergence rate was slightly higher under the 20-subject design, it is possible that the designs with the smallest observed MSE would no longer be the one if the convergence rate improved. Finally, there is a po ssibility that chance alone was responsible for the inconsistent MSE. If this is the reason, increasing the size of simulation replication woul d be helpful. Despite this artifact of inconsistency, it is clear that in cluding testing times around the true TOPE was most beneficial in recoveri ng the true dose-r esponse profile. Table 14. Bias and MSE for each pa rameter of interest for the 30 subject, 5 time point design. Design Bias.B Bias.C Bias.Ke Bias.Keo 0,2,4,24,168 -4.70E+01 2. 41E-01 8.87E-03 6.82E-03 0,2,8,24,168 -3.21E+01 1. 72E-01 6.84E-03 2.15E-03 0,2,12,24,168 -2.71E+01 1. 48E-01 5.63E-03 8.42E-03 0,2,16,24,168 -2.55E+01 1. 39E-01 4.77E-03 1.75E-02 0,2,20,24,168 -2.53E+01 1.37E-01 3.69E-03 2.99E-02 0,4,12,24,168 -2.80E+01 1. 55E-01 3.50E-03 5.85E-03 0,8,12,24,168 -2.54E+ 01 1.43E-01 3.74E-03 -1.21E-03 0,4,8,24,168 -4.29E+01 2. 26E-01 7.11E-03 3.03E-03 Design MSE.B MSE.C MSE.Ke MSE.Keo 0,2,4,24,168 1.01E+04 2.30E -01 3.99E-04 4.37E-03 0,2,8,24,168 3.44E+03 7.98E -02 2.46E-04 4.03E-03 Contd next page
63 Table 14 Contd 0,2,12,24,168 3. 33E+03 8.01E-02 2.00E-04 5.67E-03 0,2,16,24,168 3.44E+03 7. 92E-02 2.57E-04 1.12E-02 0,2,20,24,168 5.43E+03 1. 26E-01 3.01E-04 1.32E-02 0,4,12,24,168 3.52E+03 8. 02E-02 6.18E-04 2.18E-02 0,8,12,24,168 2.21E+03 5.03E-02 4.97E-04 4.97E-03 0,4,8,24,168 4.73E+03 1.09E-01 2.97E-04 3.55E-03 6.3.7 Bias and MSE in Estimating Time of Peak Effect In addition to examining the Bi as and MSE associated with the parameter estimates it is especially useful for the purposes of benchmark dose estimation to in vestigate efficient designs for estimating the TOPE. Since the TOPE is critical to benchmark dose estimation, Bias and MSE of the TOPE estimator are useful criteria for comparing designs. The 5 time point designs with 20 and 30 subjects per dose group were assessed with respect to identifying the design that recorded the smallest Bias an d MSE for each dose group specific TOPE. For the both the 20 and 30 subject designs, the Bias was within one half hour of the true TOPE fo r each of the dose groups. The MSE appeared relatively invariant to desi gn for all dose gr oups except for the lowest dose group (0.85mg/kg) which had the highest magnitude in MSE overall (Figures 13 and 14) and seemed to benefit most from designs where the testing times were closest to the true TOPE. This lowest dose group is especially im portant to consider because of its
64 relevance in establishing a Refere nce Dose used to define safe exposure limits. 12345678 -1.0-0.50.00.51.0Dose=0.85 12345678 -1.0-0.50.00.51.0Dose=1.68 12345678 -1.0-0.50.00.51.0Dose=3.38 12345678 -1.0-0.50.00.51.0Dose=6.75 12345678 0.01.02.03.0Dose=0.85 12345678 0.01.02.03.0Dose=1.68 12345678 0.01.02.03.0Dose=3.38 12345678 0.01.02.03.0Dose=6.75 Design Key: 1=0,2,4,24,168; 2= 0,2,8,24,168; 3=0,2,12,24,168 4=0,2,16,24,168; 5=0,2,20,24,168; 6=0,4,8,24,168; 7=0,4,12,24,168; 8=0,8,12,24,168 Figure 13. Bias (Top) and Mean Sq uare Error (Bottom) of the TOPE estimator under the 5-time point 20 subject designs.
65 12345678 -1.0-0.50.00.51.0Dose=0.85 12345678 -1.0-0.50.00.51.0Dose=1.68 12345678 -1.0-0.50.00.51.0Dose=3.38 12345678 -1.0-0.50.00.51.0Dose=6.75 12345678 0.01.02.03.0Dose=0.85 12345678 0.01.02.03.0Dose=1.68 12345678 0.01.02.03.0Dose=3.38 12345678 0.01.02.03.0Dose=6.75 Figure 14. Bias (Top) and Mean Square Error (Bottom) of the TOPE estimator under the 5-time point 30 subject designs. Design Key: 1=0,2,4,24,168; 2= 0,2,8,24,168; 3=0,2,12,24,168 4=0,2,16,24,168; 5=0,2,20,24,168; 6=0,4,8,24,168; 7=0,4,12,24,168; 8=0,8,12,24,168
66 Confidence intervals around the TO PE estimates were generated for each of the 5 time point, 20 and 30 subject replicate design trials to compare the ability of each design to recover the true TOPE (Table 15). As shown by the bias and MS E figures the variability in TOPE estimation was largest for the lowest dose group. As dose levels increased, the confidence intervals were within +/one hour of the true TOPE (Table 15). The results of the TOPE analysis indicate that the biphasic model predicts the TO PE with validity (small bias) and that the reliability of the TOPE esti mate (MSE) seems to be dependent on the number of subjects as well as the location of testing times, especially for the lowest dose group). Testing times located in proximity to the true TOPE for each dose group appeared to increase the reliability of the estimate for that dose group. Table 15. Confidence intervals for dose group specific TOPE for each of the 5 time point, 20 subject desi gns. Confidence intervals were generated by calculating the 2.5% and 97.5% of the distribution of 1000 TOPE estimates for each design. Dose Group 20 Subject Designs 0.85 1.69 3.38 6.75 0,2,4,24,168 12 (9-15) 9 (8-11) 7 (7-8) 7 (5-8) 0,2,8,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (5-8) 0,2,12,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (5-8) 0,2,16,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (5-8) 0,2,20,24,168 12 (9-15) 9 (8-11) 7 (7-8) 7 (5-8) 0,4,8,24,168 12 (10-14) 9 (8-10) 7 (6-8) 7 (5-8) 0,4,12,24,168 12 (9-14) 9 (8-10) 7 (6-8) 7 (5-8) 0,8,12,24,168 12 (9-14) 9 (8-10) 7 (6-8) 7 (6-8)
67 Table 16. Confidence intervals for dose group specific TOPE for each of the 5 time point, 30 subject desi gns. Confidence intervals were generated by calculating the 2.5% and 97.5% of the distribution of 1000 TOPE estimates for each design. Dose Group 30 Subject Designs 0.85 1.69 3.38 6.75 0,2,4,24,168 12 (10-15) 9 (8-11) 7 (7-8) 7 (5-8) 0,2,8,24,168 12 (10-14) 9 (8-10) 7 (7-8) 7 (6-8) 0,2,12,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (6-8) 0,2,16,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (6-8) 0,2,20,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (6-8) 0,4,8,24,168 12 (9-14) 9 (8-10) 7 (7-8) 7 (6-8) 0,4,12,24,168 12 (10-14) 9 (8-10) 7 (7-8) 7 (6-8) 0,8,12,24,168 12 (10-14) 9 (8-10) 7 (6-8) 7 (6-8)
68 Chapter 7 Conclusions In this thesis we have developed a biphasic toxico-diffusion model to describe dose-dependent ti me of peak effect (TOPE) in neurobehavioral toxicity screenin g studies. The model has been applied to a dataset from the IPCS collaborative study testing the neurotoxic potential of the chemic al parathion using motor activity counts. The biphasic model predicted a decrease in motor activity counts as dose increased with a dose-accelerated TOPE: 11.5 hrs in the lowest dose group to 7.1 hrs in the highest dose group. The biphasic model was a significant impr ovement relative to the fit of the toxico-diffusion model according to the LRT as well as AIC and BIC. Due to the limited amount of data, however, estimates for parameters B and C were not statistically signif icant. Estimates of TOPE varied greatly between the two models wi th the toxico-diffusion model predicting a TOPE irrespective of dose at 26 hours while the biphasic model predicted a TOPE between 7-12 hours, decreasing with higher dose levels. Further, the magnitude of peak response (activity count at the TOPE) was dramatically differe nt between the two models. Under
69 the toxico-diffusion model the minimu m activity count was predicted to be 100 at the highest dose level, whereas under the biphasic model the peak response was predicted to be an activity count of 3 (the observed response in the data). Th is difference in magnitude could have dramatic impacts on benchmark dose estimation. The simulation experiments identified the limitations of the existing FOB design protocol for elic iting physiologically relevant doseresponse information and explored possible design improvements in order to generate adequate data for dose-response modeling. A specific focus of the simulation was to investigate the sensitivity of the experiments with respect to do se-dependent TOPE. Our results suggest that under the present ex perimental protocol of the FOB study, there is a substantial lack of statistical power when fitting a model such as the biphasic model. Only when the time spacing is adequate will information about the true shape in time be available to predict the dose-response relationsh ip with validity. Does-dependent TOPEs cannot be verified if there ar e only limited time points in the experiment. In this regard, the existi ng protocol of FOB is unlikely to generate data that support dose -dependent TOPE. The simulation results further demonstrated that the design protocol can be altered to gain sensitivity and statistical power by adding additional time points and/or additional subjects to each dose group in the experimental
70 design. By adding a fifth experimental testing time point between hour 2 and 24, the sensitivity of the biphas ic model improved with regard to estimating the dose-dependent ti me trajectory (e.g. TOPE) as witnessed by the increased model fi tting convergence rate. By adding an additional time point, there are also more data points leading to an increase in statistical power for detecting the kinetic parameter Keo, which determines the dose-depende nt TOPE. Further increases in power can be achieved by adding addi tional subjects. It is debatable, however, whether it is practical to have at least 30 subjects per dose group in these screening tests. In suggesting design protocol improvements, consideration must be given to the practicality of any suggested design for implementation within the fram ework of the EPA testing guidelines for acute neurotoxicity. An objective of the IPSC stud ies was to validate the study protocol across participating la boratories. As a result, several laboratories conducted experiment s on the same chemical using identical methodologies. However, ea ch individual laboratory identified the TOPE of the chemical through a pilot study and used that TOPE estimate as the second testing ti me for the FOB experiments. Since the definition of motor activity counts and the placement of the 2nd testing time were left to the discret ion of the individual laboratories, there are concerns about pooling data across labo ratories for statistical
71 inference. Studies such as th e IPCS would benefit from an experimental design which standa rdized the approach to TOPE estimation and motor activity count s in such a way that these data could be pooled across laborato ries. The example below considers possible alternatives to the IPCS de sign that may effectively allow for more physiologically relevant info rmation to be obtained without substantial investment of additional resources. The IPCS collaborative studies rely on TOPE estimates from range finding studies using gate and arousal scores; however, there is evidence that different functional domains may exhibit different time courses following acute chemical ex posure (Lammers and Kulig, 1997. Presuming that the TOPE for a gi ven endpoint is measured without error, testing endpoints in more functional domains would yield a range of possible TOPEs for chemical exposure. This range of TOPEs for the given chemical could be used to guide time point selection. For example, even using only gait and arousal scores it is quite possible that the TOPE was different for these scores yet only one time point was chosen to represent the TO PE for the IPCS studies. As an alternative to the design above, a testing time associated with each TOPE (in this case 2) could be randomly assigned to participating laboratories. The data across laboratories could then be pooled resulting in a dataset where testing times for each TOPE estimate were
72 captured while each participating laboratory would only perform the experiments using 4 testing times. Mixed models such as those developed in this thesis allow for missing information from the individual time trajectories that wo uld result from pooling these data. This flexibility a critical advantage of using mixed effects models for dose-response assessment compared to traditional methods such as analysis of variance. A scenario su ch as the example above could be expanded by considering a situation in which 5 metrics were used in the pilot studies to represent a neur obehavioral endpoint from each of 5 functional domains. A range of po tential TOPEs could be established from these experiments. Each labora tory could be randomly assigned a testing time within the establishe d range of TOPEs. Once the data were pooled, the resulting dataset would have numerous time points for analysis; a scenario more consiste nt with kinetic studies used to estimate physiological properties of th e time course of a chemical in an organism after exposure. Toyi nbo (2004) has shown that for functional domain composite score s, the precision of the TOPE estimate is somewhat insensitive to the actual time of TOPE but that the MSE appeared to be minimized when the time points were chosen prior to the actual time of peak effect. Our findings suggest that testing times bracketing the dose -dependent TOPEs minimized the imprecision of the TOPE estimato r and increased the power of the
73 model. By adding even two time poin ts bracketing the range of TOPEs from range finding studies, the expe rimental design would be capable of capturing more information for more neurobehavioral endpoints. Another possible design protocol would be to choose distinct time points for different dose group in anticipation that the TOPE may dose-dependent. Since the range finding studies used a single dose for the TOPE estimate, if the chemic al possesses dose-dependent TOPE, obviously a single testing time woul d capture the TOPE only for that dose. If information were available on potential dose-dependent TOPE for a given chemical under study, defi ning dose specific time points for the experiment would yield relevant dose-dependent information related to TOPE while maintainin g the current sampling effort. Adding additional subjects in each dose group would also improve estimates of the dose grou p specific average response which should increase the statistic power to detect significant response to chemical exposure, but only if the testing times are chosen appropriately would the benchmark dose estimation procedures be improved. This thesis makes several simplif ying assumptions in attempting to link mathematical models to pharmacologically relevant parameter estimates. Bioavailability of oral induced exposure to parathion is
74 assumed to be 100%. Further, abso rption from the gastrointestinal tract is assumed to be in rapid eq uilibrium with concentration in the plasma compartment. From there it is the distributional delay associated with the hypothesized effects compartment that governs the dose-dependent TOPE observe d in the data. There is some evidence that parathion and its toxic metabolite paraoxon is a chemical that induces non addi tive biochemical reactions in experimental subjects (INCHEM 2004). The biphasic toxico-diffusion imposed this dose-dependent TOPE but further study is necessary to validate this as a mechanistic prod uct of parathion exposure. The case of delayed TOPE was not illustrated in this thesis but a model was presented which may be useful in assessing dosedependent TOPE arisin g from capacity limited kinetics. These situations have been observed in th e literature (Dayneka et al. 1993, Jusko et al.,1995; Krzyzanski and Ju sko 1997) though no examples of delayed effects with increasing dose were observed in the IPCS data. The Nonlinear mixed effects models described in this thesis do not stand in isolation for asse ssment the neurotoxic potential associated with chemical exposure. Rather, the models serve to support a framework of differ ent tools including hazard characterization, exposure assessment, risk charac terization as well as
75 dose-response assessment utilized to build consensus on the potential risks associated with toxicity of the tested chemical. These tools are intended to support and not to replace the expert judgments of toxicologists who remain the key to understanding the implications of any inference of chemic al exposure. This thesis has provided an additional tool that can be used in this regard and has illustrated its potential benefits when assessing neurotoxic potential of chemical exposure. It has further expanded the capability of risk assessment researchers to capture the diverse and sometimes complex aspects of dose-response relationships and contributed to the EPAs aims to use more physiologically relevant mo dels to predict dose-response relationships over the time cour se of experimental studies.
76 List of References Akaike, H (1973) Information th eory and an extension of the maximum likelihood principle. In Proceedings of the Second International Symposium on Info rmation Theory, B. N. Petrov and F. Csaki, eds. Akademia i Kiado, Budapest, pp. 267-281. Anger, W.K. 1984. Neurobehavioral testing of chemicals: impact on recommended standards. Neurobehavioral Toxicology and Teratology. 6 147-153. Anger, W.K. 1990 Worksite Behavioral Research: results, sensitive methods, test batteries and the transition from laboratory data to human health. Neurotoxicology 11:627-718 Dayneka, N.L., Garg,V., Jusko, W. J. 1993. Comparison of four basic models of indirect pharmacody namic responses. Journal of Pharmacokinetics and Biopharmaceutics. 21(4):457-479 Cornish-Bowden, A. 1995. Fundamentals of Enzyme Kinetics Portland Press, London. Gabriellson, J. Weiner, D. 2000. Pharmacokinetic and Pharmacodynamic Data Analysis: Concepts and Applications. Swedish Pharmaceutical Press. Stockholm Sweden. Holford, N.H., Sheiner, L.B. 1981. Understanding the dose effect relationship: clinical applic ations of pharmacokineticpharmacodynamic models. Clinical Pharmacokinetics. 6(6):429453 INCHEM 1994. Guidance values for human exposure limits (EHC 170, 1994) : http://www.inchem.org/
77 Jusko, W. J. Hui, C. K., Ebling, W. F. 1995. Convergence of direct and indirect pharmocodynamic response models. Journal of Pharmacokinetics and Biopharmaceutics. V 23 (1) 5-8 Kramer, R.E., Wellman,S.E. Rock hold,R.W. and Baker,R.C. 2002. Phamacokinetics of Methyl Para thion: A Comparison following Single Intravenous, Oral or De rmal Administration. Journal of Biomedical Science 9:311-320. Krzyzanski, W. and Jusko, W.J. 1997. Mathematical formalism for the properties of four basic models of indirect pharmacodynamic responses. Journal of Pharmacoki netics and Biopharmaceuticals. 25(1):107-123 Kwon, Y. 2002 Handbook of Essential Pharmacokinetics and Drug Metabolism for Industrial Scientists Kluwer Academic Publishers. New York. Lammers, J.H. and Kulig, B.M. 1997. Mu ltivariate time of peak effects assessment for use in selecting time of testing in acute neurotoxicity studies. Neurotoxicology 18(4):1079-1083. Lindstrom, M., Bates, D. 1998 Newton Ralphson and EM algorithms for linear mixed effects models for re peated measures data. Journal of the American Statistical Association 4:1014-1022 Liu, T. (2000) Analyses of neurob ehavioral screening data and risk assessment of neurotoxicity. Master s thesis, University of South Florida. MacPhail, R.C., Peele, D.B., Crofto n, KM. (1989) Motor activity and screening for neurotoxicity. Journa l of the American College of Toxicology. 8:117-125. McCarty, L.S. and Mackay, D. 1993. Enhancing ecotoxicological modeling and assessment: Body residues and modes of toxic action. Environmental Science and Technology 27: 1719-28.
78 Moser, V.C. Cheek, B.M., MacPhail, R.C. 1995. A multidisciplinary approach to toxicological screenin g III: Neurobehavioral toxicity. Journal of Toxicology and En vironmental Health 45:173-210 Moser, V.C., Tilson, H.A., MacPhail, R.C., Becking, G.C., Cuomo, V., Frantik, E., Kulig, B., and Winneke, G. 1997a. The IPCS collaborative study on neurobeh avioral screening methods: II. Protocol design and testing pr ocedures. Neurotoxicology 18: 929-938 Moser, V.C., Becking, G.C., and Cuomo, V., Frantik, E., Kulig, B., MacPhail, R. C., Tilson, H.A., Winneke, G. Brightwell,W.S., DeSalvia,M.A., Gill, M,W, Haggerty, G.C., Hornychova, M. Lammers, J. Larsson,J., McDaniel, K.L., Nelson, B.K. and Ostergaard,G. 1997b. The IPCS study on neurobehavioral screening methods: results of chem ical testing. Neurotoxicology 18: 969-1056 Pinhiero,J.C. and Bates,D.M. 2000. Mixed Effects Models in S and Splus. Springer-Verlag: New York. Tilson, HA. (1990) Neurotoxicology in the 1990s. Neurotoxicology and Teratology 12:293-300. Torda, T.A. Graham,G., Warwick,N.R. 1994. The Kinetics of Effect of Neuromuscular Blocking Drugs. The Keio Journal of Medicine. 43(1):27-30. Toyinbo, P.A. On effective and e fficient experimental design for neurobehavioral screening tests: Th e choice of a testing time for estimating the time of peak effect s. Masters Thesis. University of South Florida. U.S. Environmental Protection Agency (EPA) (1991) Pesticide assessment guidelines, subdivisio n F. Hazard evaluation: human and domestic animals. Addendum 10: Neurotoxicity, series 81, 82,and 83. Office of Prevention, Pesticides and Toxic Substances, Washington DC. EPA 540/09-91123. Available from: NTIS, Springfield VA. PB91-154617
79 U.S. Environmental Protection Ag ency (1995) Reportable quantity adjustments; final rule. Fede ral Register 60(112):30926. U.S. Environmental Protection Agen cy (1996) Proposed guidelines for carcinogen risk assessment. Federal Register 61(79):1796018011. U.S. Environmental Protection Agency (1998) Guidelines for Neurotoxicity Risk Assessment Federal Register: 26926-26954 U.S. Environmental Protection Agency (2000) Benchmark dose technical guidance document, external review draft. Vogel, J.S., Keating, G.A. Buchho lz, B.A. 2002. Protein binding of isoflourophate in vivo after coex posure to multiple chemicals. Environmental Health Perspe ctives, 110 Suppl. 6: 1031-6. Abstract Woodruff, S. 2001. The use of mixed effects dose-response models for neurotoxicity risk assessment. Masters Thesis. University of South Florida. World Health Organization (WHO) Environmental Health Criteria 60: Principles and Methods for the assessment of Neurotoxicity Associated with Exposure to Ch emicals. Geneva: World Health Organization, 1986. Zhu, Y. (2001) Neurobehavioral Toxi city Risk Assessment: Dose-TimeResponse Modeling and Benchma rk Dose Estimation, US EPA Report Zhu, Y. (2005a) Dose-time-respon se modeling of longitudinal measurements for neurotoxicity risk assessment. Environmetrics 16.(In press) Zhu, Y., Wessel, M.R., Liu, T., and Moser, V.C. (2005b) Analyses of Neurobehavioral Screening Data I: Dose-Time-Response Modeling of Continuous Outc omes. Regulatory and Applied Toxicology. 41:240-255
80 Zhu, Y., Jia, Z., Wang, W., Gift, J., Moser, V.C., and B.J. Pierre-Louis (2005c), Data Analysis of Neur obehavioral Screening Data II: Benchmark Dose Estimation. Regu latory and Applied Toxicology (In press )
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 Ka
controlfield tag 001 001670398
007 cr mnu|||uuuuu
008 051128s2005 flu sbm s000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0001273
Wessel, Michael Raymond.
Dose time response modeling of neurobehavioral screening data
h [electronic resource] :
b application of physiologically relevant parameters to describe dose dependent time of peak effects /
by Michael Raymond Wessel.
[Tampa, Fla.] :
University of South Florida,
Thesis (M.S.P.H.)--University of South Florida, 2005.
Includes bibliographical references.
Text (Electronic thesis) 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 91 pages.
ABSTRACT: In collaboration with the United States Environmental Protection Agency (USEPA), the University of South Florida Health Risk Methodology Group has developed dose-time-response models to characterize neurobehavioral response to chemical exposure. The application of dose-time-response models to neurobehavioral screening tests on laboratory animals allows for benchmark dose estimation to establish exposure limits in environmental risk assessment. This thesis has advanced dose-time-response modeling by generalizing a published toxico diffusion model to allow for dose dependent time of peak effects. To accomplish this, a biphasic model was developed which adopted the effect compartment model paradigm used in pharmacokinetics/pharmacodynamics to estimate a distributional rate constant to account for dose related variation in the time of peak effect.The biphasic model was able to describe dose-dependent time of peak effects as observed in the data on acute exposure to parathion and adequately predicted the observed response. However, the experimental design appeared insufficient in statistical power to confirm statistical significance for each parameter of interest. Motivated by the question of what design requirement might be necessary to validate the biphasic model, Monte Carlo simulation was adopted. Simulations were performed to assess the efficacy and efficiency of various experimental designs for detecting and evaluating some critical characteristics of the biphasic model, including the TOPE.
Adviser: yiliang zhu.
x Public Health
t USF Electronic Theses and Dissertations.