analysis of data on spontaneous reports of adverse events associated with drugs

Glasgow Theses Service http://theses.gla.ac.uk/ theses@gla.ac.uk Baah, Emmanuel Mensah (2014) Analysis of data on spontaneous reports of adverse events associated with drugs. PhD thesis. http://theses.gla.ac.uk/4990/ Copyright and moral rights for this thesis are retained by the author A copy can be downloaded for personal non-commercial research or study, without prior permission or charge This thesis cannot be reproduced or quoted extensively from without first obtaining permission in writing from the Author The content must not be changed in any way or sold commercially in any format or medium without the formal permission of the Author When referring to this work, full bibliographic details including the author, title, awarding institution and date of the thesis must be given Analysis of Data on Spontaneous Reports of Adverse Events Associated with Drugs by Emmanuel Mensah Baah A thesis submitted to the College of Science and Engineering at the University of Glasgow for the degree of Doctor of Philosophy February 2014 i Abstract Some adverse drug reactions (ADRs) are not detected before marketing approval is given because clinical trials are not suited for their detection, for various reasons [5, 23]. Drug regulatory bodies therefore weigh the potential benefits of a drug against the harms and allow drugs to be marketed if felt that the potential benefits far outweigh the harms [ 26,48]. Associated adverse events are subsequently monitored through various means including reports submitted by health professionals and the general public in what is commonly referred to as spontaneous reporting system (SRS) [19, 23, 69]. The resulting database contains thousands of adverse event reports which must be assessed by expert panels to see if they are bona fide adverse drug reactions, but which are not easy to manage by virtue of the volume [6]. This thesis documents work aimed at developing a statistical model for assisting in the identification of bona fide drug side-effects using data from the United States of America’s Food and Drugs Administration’s (FDA) Sp ontaneous Reporting System (otherwise known as the Adverse Event Reporting System (AERS)) [28]. Four hierarchical models based on the Conway-Maxwell-Poisson (CMP) distribution [43,78] were explored and one of them was identified as the most suitable for modeling the data. It compares favourably with the Gamma Poisson Shrinker (GPS) of DuMouchel [19] but takes a dimmer view of drug and adverse event pairs with very small observed and expected count than the GPS. Two results are presented in this thesis; the first one, from a preliminary analysis, presented in Chapter 2, shows that problems such as missing values for age and sex that militate against the optimal use of SRS data, enumerated in the literature, remain. The second results, presented in Chapter 5, concern the main focus of the research mentioned in the previous paragraph. ii Acknowledgement I am indebted to my supervisors: Prof. Stephen J. Senn, Prof. Adrian W. Bowman and Dr. Agostino Nobile for their guidance, criticisms and suggestions; I could not have come this far without your tutelage. My appreciation to the staff and students of the School of Mathematics and Statistics, College of Science and Engineering and the University of Glasgow at large who in diverse ways have contributed to my studies in the University. Takoradi Polytechnic deserve plaudits for funding the research work which is recorded in this thesis. I am obliged to Thearch Daniel Arthur for the lessons in iots. I found your departure painful. My deepest gratitude to my father, Egya Baah, whose abiding faith in God is a source of inspiration and to my mother, Maame Akosua, who unfortunately, did not live to see the fruits of her sacrifice and hard work. To my siblings Alfred, Isaac, Grace and Daniel whose companionship, along with other things, have shaped my understanding of humanity, I say: I could not have had a better company! Contents 1 Introduction 1 1.1 Drug Safety and Related Issues . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1.1 Adverse Drug Reactions . . . . . . . . . . . . . . . . . . . . . . . . . 2 1.1.2 Nature and Types of ADRs . . . . . . . . . . . . . . . . . . . . . . . 2 1.1.3 Prevalence of ADRs . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.1.4 Detecting ADRs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Pharmacovigilance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2.1 Spontaneous Reporting System (SRS) . . . . . . . . . . . . . . . . . 6 1.2.2 Problems of the Spontaneous Reporting System . . . . . . . . . . . . 6 1.2.3 Effects of the Problems of Spontaneous Reporting System . . . . . . 7 1.2.4 Contribution of Spontaneous Reporting System to Pharmacovigilance 8 1.3 Motivation for this Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.4 Objective(s) of the Research . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.5 Outline of the Rest of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . 13 2 Preliminary Analysis 14 2.1 Data: Nature and Treatment . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.2 Results of Preliminary Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.2.1 Overall Number of Rep orts and Trend Over Time . . . . . . . . . . . 18 2.2.2 Patient Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.2.3 Occupation of Reporters . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.2.4 Types of Report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.2.5 Mode of Submission of Reports . . . . . . . . . . . . . . . . . . . . . 22 2.2.6 Sex of Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.2.7 Age of Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 iii CONTENTS iv 2.2.8 Age and Sex Load of Adverse Events . . . . . . . . . . . . . . . . . . 24 2.3 Discussion and Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3 Review of Background Theory 29 3.1 Bayesian Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.1.1 Bayes’ Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.1.2 Prior Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.1.3 Prior Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 3.1.4 Hierarchical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.1.5 Posterior Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2 Stochastic Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2.1 Markov Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 3.2.2 Metropolis-Hastings (MH) Algorithm . . . . . . . . . . . . . . . . . 32 3.2.3 Convergence and Related Issues . . . . . . . . . . . . . . . . . . . . . 34 4 Data Models 37 4.1 Simplified SRS Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4.2 Some Existing Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 4.2.1 Relative Rep ort Rate (RR) . . . . . . . . . . . . . . . . . . . . . . . 39 4.2.2 Proportional Reporting Ratio (PRR) . . . . . . . . . . . . . . . . . 39 4.2.3 Reporting Odds Ratio (ROR) . . . . . . . . . . . . . . . . . . . . . 40 4.2.4 Gamma Poisson Shrinker (GPS) . . . . . . . . . . . . . . . . . . . . 40 4.2.5 Bayesian Confidence Propagation Neural Network (BCPNN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 4.2.6 Simple shrinkage Method . . . . . . . . . . . . . . . . . . . . . . . . 43 4.2.7 Confounding and Other Methods . . . . . . . . . . . . . . . . . . . 44 4.3 Proposed Model(s) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 4.3.1 Background of Model(s) . . . . . . . . . . . . . . . . . . . . . . . . . 46 4.3.2 Models C-G and P-G . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 4.3.3 Models C-IG and P-IG . . . . . . . . . . . . . . . . . . . . . . . . . . 51 5 Application of Proposed Model(s) to FDA SRS Data 53 5.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 5.2 Results of Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 CONTENTS v 5.2.1 Performance of algorithm . . . . . . . . . . . . . . . . . . . . . . . . 55 5.2.2 Parameter Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 58 5.3 Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 5.3.1 Validation of the distribution of φ . . . . . . . . . . . . . . . . . . . 60 5.3.2 Posterior Predictive Check . . . . . . . . . . . . . . . . . . . . . . . 60 5.4 Other Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 5.4.1 Comparison of φ Values Generated from the Three Data Sets . . . . 63 5.4.2 Comparison of Mean Replicate Count with Observed Count (N) . . 65 5.4.3 Credible Intervals of φ . . . . . . . . . . . . . . . . . . . . . . . . . . 65 5.5 Selection of Drug and Adverse Event Pairs . . . . . . . . . . . . . . . . . . . 70 5.6 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 5.6.1 DIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 5.6.2 RJMCMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 6 Discussion of Results and Comments 78 6.1 Suitable Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 6.2 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 6.3 Model of Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80 6.3.1 Drugs Common and Uniquely Chosen by RR, C-G and GPS . . . . 81 6.3.2 Other Characteristics of C-G . . . . . . . . . . . . . . . . . . . . . . 86 6.3.3 Genuine Drug Problems Within the Top Fifty Drug and Adverse Event Combinations Selected by C-G, GPS and RR . . . . . . . . . 88 6.3.4 C-G values compared with that of GPS . . . . . . . . . . . . . . . . 88 7 Conclusion 91 7.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 7.2 Highlights of the Research . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 7.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 A Selected Variables and their Description 97 B Some Selected Plots and Tables 100 C C-G Model Results 105 C.1 Data 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 CONTENTS vi C.2 Data 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 C.3 Data 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 D P-G Model Results 124 D.1 Data 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 E C-IG Model Results 131 E.1 Data 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 E.2 Data 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 E.3 Data 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 F P-IG Model Results 150 F.1 Data 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 G Model C-G Compared with other Methods 157 G.1 Formula for Computing LogP . . . . . . . . . . . . . . . . . . . . . . . . . . 166 References 178 List of Figures 2.1 Number of reports per 10,000 people against time, 2004 – 2010. . . . . . . . 19 2.2a Chart showing the trends in the number of deaths, other outcomes and all non-missing cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.2b Chart showing the trends in the number of deaths, other outcomes and all events. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.3 Chart showing the trends in the percentage of death in all the reports and in the non-missing cases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 4.1 Density of the proposal distribution f or α and ν. . . . . . . . . . . . . . . . 49 5.1 Acf plots of α, β, ν and the logarithm of the target distribution using C-G and Data 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 5.2 Trace plots of α, β and the logarithm of the target distribution f or three chains using C-G and Data 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 57 5.3 Histogram of φs for C-G and Data 1. . . . . . . . . . . . . . . . . . . . . . . 61 5.4 Bayesian p-value scatter plots for C-G and Data 1. . . . . . . . . . . . . . . 62 5.5 Logarithm of posterior means of φ for Data 1 plotted against those of Data 2 and Data 3, for C-G. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 5.6 Scatter plot of logarithm of posterior means of φ for Data 1 against corre- sponding values for Data 2 using results from C-G. . . . . . . . . . . . . . . 64 5.7 Scatter plot of logarithm of posterior means of φ for Data 1 against corre- sponding values for Data 2 using results from C-IG. . . . . . . . . . . . . . . 64 5.8 Logarithm of posterior medians and 95% posterior intervals of φ plotted against the logarithm of the reporting rate RR, for C-G and Data 1. . . . . 66 5.9 Logarithm of posterior medians of φ plotted against the logarithm of the reporting rate RR for C-G. . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 vii LIST OF FIGURES viii 5.10 Scatter plot of logarithm of φ 025 against the logarithm of RR 025 . . . . . . . 68 5.11 Trace plots of α, β, ν, log of the target distribution and model for the RJMCMC based on P-G and C-G when they are thought to be equiprobable a priori. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 5.12 Trace plots of α, β, ν, log of the target distribution and model for the RJMCMC based on P-G and C-G when prior probability of P-G is s et at 0.999999. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 6.1 Plots of φ against λ (EBGM) for various combinations of the observed and expected counts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 6.2 Plots of φ 025 against λ 025 for various combinations of the observed and expected counts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 B.1 Percentage of reports from health professionals and consumers and lawyers for 2004-2010. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 B.2 Percentage of report types from 2004-2010. . . . . . . . . . . . . . . . . . . 101 B.3 Percentage of reports on male and female subjects from 2004-2010. . . . . . 101 B.4 Percentage of reports for the various age groups for the period 2004-2010. . 102 B.5 Age and gender load of reported adverse events associated with drug use. . 102 B.6 ‘Proportion’ of the various age groups reported on for the period 2004 – 2010.103 C.1 Histogram of φs for C-G and Data 1. . . . . . . . . . . . . . . . . . . . . . . 105 C.2 Acf plots of α, β, ν and the logarithm of the target distribution for C-G and Data 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 C.3 Trace plots of α, β, ν and the logarithm of the target distribution for three chains, for C-G and Data 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 C.4 Bayesian p-value scatter plots for C-G and Data 1. . . . . . . . . . . . . . . 107 C.5 Logarithm of posterior medians and 95% posterior intervals of φ plotted against the logarithm of the reporting rate RR, for C-G and Data 1. . . . . 107 C.6 Logarithm of posterior means of φ for Data 1 plotted against those of Data 2 and Data 3, for C-G. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 C.7 Acf plots of α, β, ν and the logarithm of the target distribution for C-G and Data 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 C.8 Trace plots of α, β, ν and the logarithm of the target distribution for three chains, for C-G and Data 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . 112 [...]... unknown adverse drug reactions associated with marketed drugs as soon as practicable Some adverse drug reactions are not easy to detect after marketing approval because of inaccurate reporting of adverse events and lack of information regarding the population of users of drugs Timely detection of new and unknown serious adverse drug reactions would ensure the aforementioned benefits via: ◦ “reduced morbidity,... Multiple reports of adverse event episodes arising from use of multiple channels or inappropriate tracking of events leading to misrepresentation of old cases as new is reported to be common [6, 29, 30] The number of people actually using a given medication at any point in time is unknown and the current information situation does not permit accurate estimation of it 1.2.3 Effects of the Problems of Spontaneous. .. medications being used at the same time [23,29], or drug-drug interaction [23,29,64], such as happen when isoniazid is administered concurrently with rifampicin [64] CHAPTER 1 INTRODUCTION 1.2.4 8 Contribution of Spontaneous Reporting System to Pharmacovigilance The spontaneous reporting system, nonetheless, has played and continues to play an important role in the identification of adverse drug reactions... prescription-onlymedications are more or less treated like over-the-counter drugs because of weak regulatory systems and virtually non-existent systems of reporting ADRs The sixty-fifth edition of the British National Formulary [74] presents a classification of ADR on the basis of prevalence as shown in Table 1.1 Table 1.1: A classification of ADRs based on prevalence Prevalence Description 1 in 10 Very Common... irregular periods of increased reports of adverse events some of which are not real Additionally, regulatory policy could also tilt the reporting rate in a given direction; regulatory bodies request reporting institutions to be particular about serious and uncommon events, which could bias reporting in favour of these events [4, 6, 70, 81] Reporting partial and erroneous information on adverse events are... medical product, once the problems have been identified They range from improving precautionary and warning messages on packages and information leaflets, labeling modification; limiting indications, mandatory monitoring of patients, dose modification; and limiting distribution and prescription of product, seeking informed consent of patients; to suspension of distribution and marketing, drug withdrawal from... other remedies Additionally, it is not difficult to perceive the existence of the huge market for drugs given the proliferation of diseases, in spite of the impressive advances in the science of medicine 1 CHAPTER 1 INTRODUCTION 1.1.1 2 Adverse Drug Reactions The harm(s) a drug can cause are discussed in terms of the adverse reaction(s) associated with it Put simply, an adverse drug reaction (ADR), otherwise... prevalence of adverse reactions [4, 70, 81] It is not easy to establish whether or not the relationship between a drug and an adverse event which occurred during the administration of the drug is causal on the basis of spontaneous reporting system data alone; as the event may have occurred accidentally or have been associated with the disease under treatment Other factors that may be responsible for the adverse. .. agents of adverse events They thus concluded that the problem of under-reporting is significant but not uniform across events and medications; it is more likely to involve common and non-serious events, and underscored that this phenomenon, in some way, augurs well for pharmacovigilance as rare or novel but serious adverse reactions are likely to show up in spontaneous reporting system as events with. .. studying the adverse reaction characteristics of a medication relative to those of the same therapeutic class [70] and “providing information to healthcare professionals and patients to optimize safe and effective use of medicines” [50] The action a regulatory body, in collaboration with a sponsor, may effect takes several forms depending on the enormity and urgency of the problems associated with a medical . institution and date of the thesis must be given Analysis of Data on Spontaneous Reports of Adverse Events Associated with Drugs by Emmanuel Mensah Baah A thesis submitted to the College of Science. theses@gla.ac.uk Baah, Emmanuel Mensah (2014) Analysis of data on spontaneous reports of adverse events associated with drugs. PhD thesis. http://theses.gla.ac.uk/4990/ . 26,48]. Associated adverse events are subsequently monitored through various means including reports submitted by health professionals and the general public in what is commonly referred to as spontaneous