摘要
In the clinical development of new drugs, research pharmaceutical industry and biotech companies (known as sponsors) design and execute clinical trials in order to evaluate the efficacy and safety of the study drug. Clinical data are observed and collected from trial participants. These data are analyzed and organized to support clinical study reports. Sponsors that develop the new drugs will submit these study reports to U.S. Food and Drug Administration (FDA) for review. If FDA considers this new drug is efficacious, and, relatively safe, then this new drug would be approved for general patient population. Discrete data are very common in clinical trials. One example is known as “responder analysis.” After the completion of a clinical trial, the physician evaluates each patient and classifies the patient as a “responder” or a “nonresponder.” This is a binary classification of treatment results for each patient. The responder analysis is to study the proportion of responders among all trial participants. If this proportion is high, then it is likely that the study drug can be approved by FDA. In most of clinical trials, two treatment groups (one group of patients receives the test drug treatment, and the other receives control treatment) are compared for efficacy. If it turns out that more proportion of responders is observed from the test drug group than the control group, then the test drug can be approved for general patient population. In this book, Chapters 4 and 5 provide extensive materials to cover treatment comparisons of two independent binomial proportions. These methods can be very useful to practitioners. In some situations, when developing drugs for the treatment of rare diseases, because of difficulty in patient recruitment, “single arm trial” can be found in practice. A single arm trial means the clinical trial treats every participant only with the study medication, without any control group. When this is the case, methods for analysis of one binomial population can be very useful. Another application can be found in crossover clinical trials, or matched case-control epidemiology studies—these research work can be based on the paired binomial data. Chapter 3 of this book covers the one binomial proportion situation, and Chapter 6 focuses on paired binomial proportions. In clinical trial applications, sometimes the interest is in recurring events. For example, in treatment of diabetic patients, hypoglycemia can be a very severe adverse event. In the duration of the study treatment, most patients are not expected to develop hypoglycemia. Some patients could develop one episode. Other patients may develop repeated events. In this application, confidence intervals on count data can be useful. Another example could be chronic obstructive pulmonary disease (COPD) exacerbation. In treating COPD patients, the number of exacerbations experienced by patients in a treatment group may serve as one of the primary efficacy endpoints. This is another application of count data. Chapter 7 of this book covers count data and confidence intervals on rate. In development of new drugs, typically many molecules are discovered as drug candidates. Only the best candidates would have the opportunity entering the clinical development stage. Clinical trials are divided into four phases—Phases I, II, III, and IV, in a sequential manner. Every phase of clinical development can be viewed as a process of weeding out unsuccessful drug candidates. In other words, only the very best drug candidate gets the chance to enter the later phase clinical development, and only the ones that can objectively and scientifically demonstrate both drug efficacy and drug safety can be approved by FDA. Hence for every drug candidate at every stage of development process, a “Go/NoGo” decision has to be made about the candidate as to whether it deserves to be progressed into the next step of development. This “Go/NoGo” decision can be thought of as a confirmatory question in drug development. However, every stage of new drug development can also be viewed as a learning process. Development team members are always asking many questions about the drug candidate. All of these questions are exploratory in nature—people are curious, and it is always important to learn the efficacy and safety associated with each drug candidate. Based on this understanding, every clinical question can be divided into a confirmatory question or an exploratory question. The confirmatory question is about the “Go/NoGo” decision, and all other questions are exploratory. In statistical applications, the method to deal with a confirmatory question is hypothesis testing, and the method to deal with an exploratory question is estimation—either point estimation or interval estimation. Another consideration is about the noninferiority decisions. This happens in active-controlled clinical trials. In other words, the Phase III clinical trial is designed to compare a test drug against another drug for the same indication, and this other drug has already been approved by regulatory agencies. Such an approved drug is known as an active control agent. Noninferiority trials are designed with a prespecified margin, which must have been agreed between the sponsor and the regulatory agency. The Phase III comparison is to demonstrate that the lower confidence limit of treatment difference (test drug mean subtract active control mean) should be above the negative value of this margin (when a positive value means better efficacy). Although the comparison is based on the one-sided confidence limit, the decision itself is of a nature as the one-sided hypothesis testing. In other words, the efficacy of the test drug has to be above a predefined quantity, in order for the test drug to be approvable. Again, this approval decision is fundamentally a “Go/NoGo” decision, the only difference is that the noninferiority decision is expressed in terms of a “one-sided confidence interval.” Because the only confirmatory question in drug development is about a “Go/NoGo” decision, most of other clinical questions are exploratory. Estimation is more useful in most of clinical applications. Also, confidence intervals are much more informative than a simple p-value. Chapter 2 of this book covers the p-value discussion from the American Statistical Association (ASA) statement. This is a very valid and relevant point, indicating that estimation, either point estimation or interval estimation, is more useful in applications than statistical hypothesis testing. This book is about confidence intervals. In fact, the emphasis on the confidence interval makes this book very useful for practitioners. There are seven chapters in this book. The first chapter is a brief overview, which covers both the frequentist approach and the Bayesian approach of constructing confidence intervals on discrete data. This chapter also mentions statistical inference, and how statistical thinking established the scientific foundation of clinical trials and modern drug development. Chapter 2 discusses various viewpoints on the “p-value,” specifically, it brings up the ASA statement on the p-value. As mentioned earlier, in clinical trial applications, most of the scientific questions are exploratory in nature. Hence confidence intervals are much more useful than p-values. It is very sensible that this book focuses on confidence interval approaches to handle clinical research questions. In statistical applications to solve real-world problems, confidence interval approaches are much more useful than p-values. The main body of this book includes Chapters 3–7. In these chapters, the authors present a wide range of statistical methods for calculating confidence intervals. For each method, this book introduces the theoretical derivation, the background, and the implementation using either SAS code or R. Many methods covered in this book have existing SAS procedures available. On the other hand, a lot of methods have to be programmed using SAS or R. For most of the applications, the book brings a numerical example (mostly based on clinical trial data) and uses SAS or R to achieve the computation. The results are illustrated to help readers in understanding how each method can be implemented in practice. Furthermore, performance of these methods is also displayed and compared using coverage probabilities obtained from the calculated confidence intervals. Chapter 3 provides methods for calculating confidence intervals on one sample binomial proportion. Broadly speaking, there are three major categories in establishing confidence intervals on one binomial proportion—asymptotic approaches, exact methods, and Bayesian intervals. Within each category, the authors introduce many methods and offer details of each method. Some of these methods are well known and available for decades, and some are relatively new. The authors provide a rather comprehensive list of methods. As mentioned above, in the development of drugs treating rare diseases, because of difficulty in patient recruitment, “single arm trial” can be found in practice. When this is the case, methods introduced in this chapter can be useful to practitioners. Chapter 4 discusses confidence intervals on differences between two independent binomial proportions. Differences between two independent proportions is frequently observed in clinical trials. As mentioned earlier, “responder analysis” is a common practice in many clinical settings. When comparing the drug efficacy of two groups (e.g., test treatment vs. control), development teams often study the differences in the proportion of responders between these two treatments. This chapter focuses on the difference p1– p2 and proposes ways to construct confidence intervals on this difference. In this chapter, the authors introduce many methods to perform the calculation. It starts with asymptotic methods and then moves into exact methods. Also introduced in Chapter 4 of this book is a Bayesian approach to compute credible intervals. Relative risk or risk ratio (RR) is another commonly used metric for comparing two independent binomial proportions in clinical trial applications. Chapter 5 covers confidence interval methods for RR. Similar to Chapter 4, the book starts with asymptotic methods and then moves into exact methods. Furthermore, Chapter 5 also introduces approaches to establish Bayesian credible intervals. Chapter 6 presents methods to calculate confidence intervals for paired binomial proportions. The book mentioned that in crossover trials there may be applications of paired proportions. However, in my personal experience, I have never seen such an application. Most of crossover clinical trials are designed for Phase I bioequivalence studies (or other Phase I trials including bioavailability, food effect, and/or drug–drug-interaction trials). In these clinical trials, the primary variables of interest are pharmacokinetic (PK) parameters. PK parameters are generally believed to follow log-normal distribution, and they tend to be continuous variables. I can imagine that in certain situations, crossover trials are applied to study drug efficacy like in Phase II or III applications and that the primary efficacy endpoint is a binary variable (such as responder status). But those applications are not common in drug development practice. On the other hand, in epidemiology research, case control is an important study type. For case-control studies, first a number of cases (patients with the disease under study, e.g., a certain type of cancer) are identified and recruited into the study. Next, for each case, find one or a few matching controls. A control subject is a person without this disease, but with similar characteristics (e.g., same gender, similar age, occupation, …) as the case. After the cases and controls are identified, data are collected from these participants and, finally, these data are analyzed. Binomial data are very common in case-control studies. Therefore, the application of this book can also include epidemiology studies, or, more generally, the book could be useful for public health applications. In Chapter 6, the authors provide a number of asymptotic approaches including both frequentist and Bayesian methods. It also talks about exact confidence intervals for paired binomial proportions. One more section of Chapter 6 discusses confidence intervals with missing data. Finally, Chapter 7 presents confidence interval methods for rate. This chapter offers a rather comprehensive treatment of count data. Again, count data are an important data type in clinical research. A reference of confidence intervals on rate parameter observed from count data is useful. The chapter starts with Poisson distribution and focuses on the confidence interval of the rate parameter λ. Then the chapter covers confidence intervals for mean and other count data models. Before wrapping up this chapter, the authors cover Bayesian credible intervals. Overall speaking, this book delivered what the authors hoped to achieve—"Confidence Intervals for Discrete Data in Clinical Research." This book provides a comprehensive review of existing methods in constructing confidence intervals for binary and count data because these are the discrete data most frequently used in clinical research. Other types of discrete data such as nominal data, multinominal data, or general categorical data do not usually appear in clinical trials. In fact, the scope of application for this book can go beyond clinical trials. This book can also be a very useful reference tool for public health applications. Generally speaking, public health covers a much wider scope and includes many more practitioners. Most of the public health studies do not use randomization. Examples of these studies are epidemiology studies (including most of observational studies), electronic medical records, registries, and medical data collected from insurance companies like claim databases, and so on. In other words, this book not only serves the readers they it intended to serve but can also help potentially a much broader readership. One strength of this book is that it covers a wide range of methods with very good reference articles (published between 1927 up to 2020, pp 215–222). This feature makes this book to be one of the very useful references on this topic. For practitioners engage in clinical research, epidemiology, or public health, this book can be a very helpful tool. *(abstract via OpenAlex)*