Statistical Methods For The Analysis of Adverse Experience Data From Clinical Trials of New Drugs by Suzanne Edwards Department of Biostatistics University of North Carolina at Chapel Hill Institute of Statistics Mimeo Series No. 1497T December 1985 STATISTICAL METHODS FOR THE ANALYSIS OF ADVERSE EXPERIENCE OATA FROM CLINICAL TRIALS OF NEW DRUGS bv ~ Suzanne Edwards A Dissertation submitted to the faculty of The University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Doctor of Public Health in the Department of Biostatistics in the School of Public Health. Chapel Hill 1985 To My parents, Clint and Joame Edwards Abstract SUZANNE EDWARDS Statistical Methods for the Analysis of Adverse Experience Data from Clinical Trials of New Drugs (Under the direction of GARY G. KOCH) In premarketing clinical trials of new drugs, data are collected to evaluate both the efficacy and the safety of the new drug relative to the control(s). One aspect of the safety evaluation is the ascertainment of adverse experiences (e.g., headaches, nausea), which occur during the treatment period. The usual analysis of these adverse experience data has been simply to tabulate the percent of patients in each treatment group who had each of the adverse experiences of interest. In this presentation, alternative methods for summarization and analysis of this type of data are discussed. Four attributes of adverse experiences are delineated: occurrence/ non-occurrence, severity, frequency/duration, and timing. Summary measures characteriZing each of these attributes are constructed, and design- and model-based methods for analysis of these summary measures are discussed. The design-based methods include randomi zation statistics with covariate-adjustment and/or stratification, exact tests, and nonparametric tests for group comparisons of time to first adverse experience. Model-based methods include logistic regression for dichotomous response variables, extensions of logistic regression for ordinal variables, weighted least squares methods for analysis of counts and rates, and Cox regression and piecewise exponential modeling for time to first adverse experience. The analyses are illustrated using data from two randomized olacebo-controlled trials: a short-term trial of a new combination L antihistamine/decongestant drug, and a medium length trial of a new antidepressant. A third example involves the analysis of a series of trials for a new analgesic. ·e Ackrowledgements It is with great appreciation that I acknowledge the assistance, gUidance, and support of my very patient committee chairman, Gary G. Koch. I also wish to thank each of my committee members for their helpful suggestions: Michael Symons (my first advisor at UNC and the one who suggested I go on for a doctorate), Anthony Segreti, Craig Turnbull, and Caroline Becker. Thanks also are due to Walt Cummings, who suggested the topic and helped coordinate my assistantship at Burroughs Wellcome along with Dr. Koch; Dan Heatherington, who lent his expertise in the field of adverse experiences in a very thorough review of my Chapter I; Joseph Janis and Woody Setzer for their help in providing me a printer for my final copy; and my fellow students at UNC and coworkers at Burroughs Wellcome, of whom I have asked many questions (in particular, Sue Allen, Ingrid Amara, Gail Knowlton, Jerry Schindler, John Schoenfelder, Robert Starbuck, and Maura Stokes). Finally, I wish to thank my husband, Joseph Newberry, for his patience and understanding throughout the time my dissertation was in progress. Table of Contents I. Introdtaion to Adverse Experieree Data in Clinical Trials. 1 1.1 Introduction, 1 1.2 Terminology, 3 1.3 Overview of Adverse Experience Evaluations in Clinical Trials, 5 1.4 Adverse Drug Reaction Mechanisms and Predisposing Factors, 7 1.5 Adverse Experiences Not Due to Drugs, 10 1.6 Collection of Adverse Experience Data in Clinical Trials, 12 1.6.1 Methods of Obtaining Adverse Experience Data, 14 1.6.2 Length of Observation Period, 15 1.6.3 Collection of Concomitant Information, 17 1.7 Potential Sources of Bias, 19 II. Statistical Considerations 22 2. 1 Sampling Process Issues, 22 2.2 Measurement Process Issues, 25 2.2.1 Occurrence/Non-occurrence and Severity, 26 2.2.2 Frequency, Duration, and Timing, 29 2.3 Effect of Noncompliance to Protocol, 31 2.4 Statistical Analysis Issues, 33 2.4.1 Implications of Sample Size and Study Length, 34 2.4.2 Implications of Choice of Control Group, 37 2.4.3 Generalizability Considerations, 38 III. Statistical Methxfology . . . . . . . . . . . 41 3.1 The Mantel-Haenszel Method and Extensions, 41 3.1. 1 Unadjusted Randomization Statistics for Case Record Data, 43 3.1.2 Covariate-Adjusted Randomization Statistics, 46 3.1.3 Stratification-Adjusted Randomization Statistics, 48 3.2 Exact Tests, 51 3.3 Nonparametric Tests for Censored Survival Data, 53 3.4 Logistic Regression and Extensions to Ordinal Data, 56 3.5 Weighted Least Squares Methods, 61 3.6 Cox Regression, 64 3.7 Poisson RegreSSion and Piecewise Exponential Models, 67 IV. An Antihistamine/Decongestant Drug Trial Example . . . . . . . 71 4.1 Description of the Trial, 71 4.2 Overview of Statistical Analyses, 74 4.3 Analysis of Occurrence/Non-occurrence, 77 4.3.1 Contingency Table Analyses, 78 4.3.2 Covariate-Adjusted Randomization Statistics Analysis, 90 4.3.3 Logistic Regression, 93 4.4 Correlation Analysis, 98 4.5 Analysis of Severity and Frequency of Drowsiness, 99 V. An Antidepressant Orog Trial Example . .109 5.1 Description of the Trial, 109 5.2 Overview of Statistical Analyses, 114 5.3 Comparison of Adverse Experience Patterns for Various Criteria, 115 5.4 Stratified Covariate-Adjusted Randomization Statistics Analysis, 124 5.5 Logistic Regression, 130 VI. An Example Usi~a Series of Analgesic Orog Trials 137 6.1 Description of the Data, 137 6.2 Overview of Statistical Analyses, 140 6.3 Treatment Group Comparisons of Proportion Having Digestive System Adverse Experiences, 142 6.4 Survival Data Analysis Issues, 144 6.5 Analysis of First Order Associations, 148 6.6 Piecewise Exponential Model Fitting Via Poisson Regression, 155 6.7 Cox Regression, 165 6.8 Proportional Hazards Regression Compared to Logistic Regression for Days 1 to 14, 172 . 6.9 Weighted Least Squares Analysis of Frequency, 175 6.10 The Package Insert, 176 VII. St.nmary 180 7.1 Summary, 180 7.2 Topics Not Covered, 185 RefeJ"'el'lC:lE!S. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .188 Olapter I Introduction to Adwrse Experience Data in Clinical Trials 1.1 Introduction In clinical trials of new drugs, the analysis of safety data, particu larly the analysis of adverse experiences, is often not as extensive as the analysis of efficacy data. There are many reasons for this. Adverse experience data are difficult to quantify and standardize, since many of the commonly occurring reactions are subjective, e.g., head ache, nausea. Furthermore, the determination of a causal link between an adverse experience and a particular drug can be exceedingly diffi cult. During the course of a clinical trial, the adverse experiences occurring in the group treated with the experimental drug are not necessarily due to the pharmacological action of the drug. They may be due to the illness being treated or an unrelated illness, due to some aspect of the drug-taking experience (viZ. the placebo effect), or due to the action of or interaction with other medications. From a statistical point of view, the analysis of safety data 1s often awkward because one wants to confirm, rather than disprove, a null hypothesis, Le., that the drug is safe. Failure to show the drug is unsafe does not imply, of course, that the hypothesis of safety is true. Furthermore, clinical trials are generally not designed to proVide statistically powerful tests of safety. The early phase clinical trials enroll very few subjects and the later trials are designed primarily to address questions of efficacy rather than safety. The types of adverse reactions most amenable to statistical analy sis are those which occur with some frequency. The investigators may be able to anticipate some of these reactions based on their knowledge of the pharmacological activity of the drug and the results of animal experiments. Unexpected and unusual drug reactions may also occur, especially if a sufficiently large population is studied for a sufficiently long period of time. However, in the entire series of premarketing clinical trials, the new drug is tested in at most a few thousand subjects, most of whom are treated for less than a year. Therefore it is not possible before marketing to discover adverse drug reactions which have either a low incidence rate or a long latency period. Although post-marketing surveillance is undeniably important in the detection of adverse reactions, careful collection and analysis of the data from premarketing trials are important for several reasons (Skegg and Doll 1977): (1) The commonest reactions are often the most important, (2) Common minor reactions may be clues to rarer, more serious ones, and (3) Suspicions raised in premarketing trials may lead to more effective post-marketing monitoring. The drug sponsor needs to determine the extent and intensity of the adverse reactions the drug may be expected to produce in the potential user population in order to decide whether use of the drug poses an unacceptable level of discomfort or risk. If the drug is marketed, this information is also necessary for labeling, so that the drug can be 2