Clinical Trial Design: Bayesian and Frequentist Adaptive Methods

PREFACE

Drug development is a long-term, complex, and expensive process. Each new drug originates from basic biochemical research, moves on to laboratory experiments and animal studies, and eventually reaches clinical trials. Clinical trials are prospective studies of new interventions—such as experimental treatments, combination therapies, or medical devices with human subjects. The entire procedures of clinical trials are rigorously specified and controlled to reduce bias and errors. Clinical trials can generally be classified into four sequential phases: I, II, III, and IV. Phase I trials mainly focus on the safety and toxicity profile of the investigational compound. Once the new agent is considered tolerable, a phase II trial will be undertaken to examine the efficacious activities based on a short-term efficacy endpoint. If the test drug shows promising anti-disease effects, the study will then move forward to a large-scale phase III trial for confirmative evaluation of the drug’s efficacy. If the new drug has successfully undergone extensive testing through phase I, II, and III trials, it will be filed to the regulatory authority (e.g., the United States Food and Drug Administration or the European Commission), for approval of widespread use in the general patient population. After the drug becomes available on the market, phase IV trials may be initiated to keep drugs’ efficacy, toxicity, and rare side effects under long-term surveillance. New warning labels may be added to the prescription of the drug and, even more seriously, some drugs exhibiting unforeseen excessive toxicities could be withdrawn from the market.

Every clinical trial starts from the design and planning stage, moves to trial conduct and monitoring, and finally to the data analysis and conclusions; each step along the way calls for statistical methods. Without a good design and proper implementation, the trial could be a mess (e.g., leading to inconclusive or false findings), or even a disaster (e.g., causing an undesirably large number of patients to suffer from toxicity or death). Clinical trials should be efficient and ethical; for example, saving resources, benefiting more patients, drawing correct conclusions quicker, and resulting in less unnecessary toxicities. Well-designed and carefully carried-out clinical trials are the most powerful tools for new drugs’ discovery. With a focus on the practicality, this book covers a wide range of statistical designs that are commonly used for each phase of clinical trials from both the Bayesian and frequentist perspectives. There has been great interest and extensive development in Bayesian adaptive designs, especially for early-phase clinical trials (i.e., phase I and phase II trials). Nevertheless, frequentist methods still dominate phase III trials by explicitly controlling the type I and type II errors in the hypothesis testing framework. Instead of biasing toward either Bayesian or frequentist methods, this book takes a pragmatic approach and introduces all clinical trial designs that are routinely used.

For beginners in this field, Chapters 1 and 2 provide an overview of the fundamentals of clinical trials and the related terminologies and concepts. For readers without a statistical background, Chapter 3 gives a brief introduction of basic knowledge of statistics including both Bayesian and frequentist estimation and inference procedures, with highlights on the key differences between the two. Chapters 4 through 6 discuss various Bayesian and frequentist designs and their statistical properties and operating characteristics for phase I, II, and III clinical trials, respectively. In particular, phase I and phase II trial designs are mainly based on Bayesian methods, due to small sample sizes in these early-phase studies. Chapters 4 and 5 also cover more advanced methodological development of early-phase trial designs, including Bayesian predictive probability trial monitoring, seamless phase I/II trial designs, and time-to-event toxicity and efficacy trade-offs. Chapter 6, which is dominated by frequentist approaches in the hypothesis testing framework, concentrates on power and sample size calculation for phase III clinical trials with continuous, dichotomous, and survival endpoints, respectively. Sample sizes may be calculated using fixed-sample designs, group sequential methods, or adaptively re-estimated in light of interim data. Noncompliance issues and intent-to-treat analysis are also discussed. In subsequent Chapters 7–10, more specific topics and more up-to-date developments in clinical trials are presented, such as Bayesian adaptive randomization, late-onset toxicity, dose finding in drug-combination studies, and targeted therapy designs.

The impetus of writing this book is to provide comprehensive and systematic coverage of statistical methodologies in clinical trial designs from practical perspectives. It may serve as a textbook for a graduate-level course and also as a reference for statisticians, medical doctors, research nurses, and other clinical trial practitioners, who are interested or involved in designing, conducting, and monitoring clinical trials. My goal is that readers would be able to design clinical trials for each phase on their own and would also understand and evaluate those designed by others. Many clinical trial designs and statistical methods discussed in this book are routinely used at the University of Texas M. D. Anderson Cancer Center and pharmaceutical industries. Most of the software used in this book can be freely downloaded from the website of the Department of Biostatistics at M. D. Anderson Cancer Center:

http://biostatistics.mdanderson.org/SoftwareDownload/

Approximately one-third of the book was written when I was Associate Professor in the Department of Biostatistics at M. D. Anderson Cancer Center, and the rest was finished after I joined the faculty of the Department of Statistics and Actuarial Science at the University of Hong Kong. I would like to express my sincere thanks and gratitude to my colleagues at both institutes for their enormous encouragement and support. In particular, I would like to thank Ying Yuan, J. Jack Lee, Donald Berry, Peter Müller, Peter Thall, Valen Johnson, Yu Shen, Jeffrey Morris, Bradley Carlin, Nan Chen, and Shurong Zheng for many insightful discussions on various issues arising in clinical trials and Bayesian adaptive designs, as well as Jianwen Cai, K. W. Ng, and W. K. Li for their consistent encouragement. Special thanks go to Lee Ann Chastain, Vicki Geall, Robert Golden, Guo-Liang Tian, Yuanshan Wu, and Jiajing Xu for proofreading, Susanne Steitz-Filler, Kristen Parrish, and Amy Hendrickson for editorial help, and all the students and colleagues who took my courses and workshops in this exciting area, which helped me to structure the book from the teaching materials. Finally, I would like to thank my mother, who has always encouraged me but eventually said to me Son, don’t write another book, this is too much work! and also dedicate this book to the memory of my father; without their guidance, encouragement, and love in my life, this dream would never have come true.

GUOSHENG YIN

Hong Kong, China; and Houston, USA

November, 2011

CHAPTER 1

INTRODUCTION

1.1 WHAT ARE CLINICAL TRIALS?

Clinical trials are prospective intervention studies with human subjects to investigate experimental drugs, new treatments, medical devices, or clinical procedures, under rigorously specified conditions. Clinical trials play a critical role in drug development and pharmaceutical research. Conventionally, clinical trials are classified into four sequential phases: I, II, III, and IV. The trial design for each phase is a complicated process, which often requires close collaborations and joint efforts from many stakeholders, such as academic institutions, medical centers/hospitals, pharmaceutical companies, contract research organizations, government organizations (e.g., the National Institute of Health), and regulatory agencies (e.g, the Food and Drug Administration—FDA). From phase I to phase IV trials and from fixed to adaptive designs, all study procedures need to ensure consistency and validity of the findings. Every aspect of the trial design, every stage of the trial conduct, and every interim monitoring and data analysis call upon statistical methods. Therefore, the importance of statistics in the applications of clinical trials can never be overemphasized.

Before stepping into statistical methods for clinical trial designs, we first provide an overview of clinical trials. If a clinical trial does not involve a comparison treatment or if the patient enrollment and administration of comparison treatments are not concurrent such as use of historical controls, the trial is said to be uncontrolled. A controlled clinical trial may include an active control (the standard treatment) or a placebo (an inert that mimics the look and the route of administration of the real treatment) for direct comparison so that the difference in the clinical outcome attributable to the experimental therapy can be evaluated objectively.

A clinical trial is said to have internal validity if the observed difference between treatment groups is real, not confounded nor due to any bias or chance. Generally speaking, a randomized, double-blind (masking of the identity of treatment to both patients and clinicians), placebo-controlled trial possesses a high level of internal validity. External validity of a clinical trial refers to whether the study conclusions can be generalized to a broader population. The external validity of a trial would not be relevant if its internal validity is questionable. External validity may be enhanced by relaxing patient eligibility criteria.

Clinical trials are the most effective approach to examining and comparing treatment effects of experimental drugs, medical therapies, or any clinical intervention in human beings. A carefully thought-out, well-designed, and appropriately conducted and analyzed clinical trial is a powerful tool for new drug discovery and pharmaceutical development. Most importantly, the findings in clinical trials have a direct and enormous impact on clinical practice.

In a clinical trial, patients are accrued over time and followed prospectively. While participants may not necessarily enter the trial on the same calendar date due to staggered entry, they all progress from a well-defined baseline point by meeting the eligibility criteria of the study. Investigators must take full responsibility to inform the participants of all aspects of the trial—in particular, of the potential benefits and adverse effects of the new intervention.

1.2 BRIEF HISTORY AND ADAPTIVE DESIGNS

The first controlled clinical trial may be traced back to a study of investigating treatments for scurvy conducted by Lind (1753). In that study, twelve patients aboard the Salisbury at sea were divided into six groups, with two in each group. Patients were in similar conditions and had the same diet. Two of the patients were given a quart of cider a day; two took elixir of vitriol three times a day; two took two spoons of vinegar three times a day; the worst two patients were put under a course of seawater; two others had two oranges and one lemon a day; and the remaining two patients took nutmeg three times a day. The most sudden and visible good effects were perceived from the use of oranges and lemons; one of those who had taken them was fit for duty at the end of six days. Obviously, this scurvy study lacks some essential characteristics of modern clinical trials. For one, patients were not properly randomized; for example, the worst two patients were treated with seawater. Second, the study was not blinded or masked; that is, both patients and the investigator knew what treatment was used. As a result, there could be selection bias and other confounding effects in the scurvy study.

Early applications of randomization were in agriculture to study which fertilizers affected the great crop yields (Fisher, 1926). The field was divided into plots, and each plot was randomly assigned a specific fertilizer. The goal of randomization here is to obtain a valid test of significance through independent replications, whereas randomization used in clinical trials—for example, the streptomycin trial in pulmonary tuberculosis by Hill (Medical Research Council, 1948)—is to produce comparable groups so that patients in different groups are alike in all aspects except for the treatment. Randomization is essential in clinical trials to control known and unknown biases during patient selection, treatment allocation, outcome evaluation, and so on. Blinding provides another way of reducing the treatment-related bias by intentionally concealing the identity of treatments.

Traditionally, clinical trials are designed with fixed sample sizes and equal randomization (patients are allocated to each treatment with the same probability). This can be illustrated with the following phase III clinical trial of human immunodeficiency virus (HIV) type 1. It is known that maternal-to-infant transmission is the primary means for newborns infected with HIV. To evaluate whether the antiviral therapy zidovudine reduces the risk of maternal-to-infant HIV transmission, the Pediatric AIDS (acquired immune deficiency syndrome) Clinical Trials Group conducted a randomized, double-blind, placebo-controlled, multi-center trial to evaluate the efficacy and safety of the zidovudine regimen (Connor et al., 1994). The primary binary endpoint was whether the newborn infants were HIV-positive (with at least one positive HIV culture of peripheral-blood mononuclear cells). At the first interim analysis, 239 pregnant women received zidovudine and 238 received placebo through equal randomization between zidovudine and placebo, while 12 women withdrew from the study before delivery. Among 363 births with known HIV-infection status, there were 180 newborns in the zidovudine group with 13 infants HIV-positive and 183 in the placebo group with 40 HIV-positive. The interim result was very compelling: Zidovudine reduced the risk of maternal-to-infant HIV transmission by approximately two-thirds. This finding led to early termination of the trial, and the data and safety monitoring board recommended that the patient enrollment be discontinued and that all patients in the trial be offered zidovudine treatment. In a later updated analysis, 20 newborns were HIV-positive in the zidovudine group and 60 were in the placebo group (Zelen and Wei, 1995; Rosenberger, 1996). The results were indeed overwhelming. Had those 60 women with HIV-positive newborns in the placebo group been given zidovudine, many infants would have been saved. This trial reveals a limitation of equal randomization; that is, regardless of the accumulating evidence in the trial, patients are always equally allocated to the experimental treatment and control. By contrast, adaptive randomization, which tends to assign more patients to better treatments based on the accumulating data, may appear to be a more ethical approach.

In fact, adaptive randomization is just one aspect of adaptive designs; adaptations in clinical trials have many other features and meanings (Berry, 2006; Chow and Chang, 2006; Chang, 2008; Berry et al., 2010). In general, adaptive designs may allow trial early stopping for superiority, noninferiority, or futility; adaptive dose escalation/de-escalation or dose insertion in dose-finding studies; dropping or adding treatment arms; adaptive randomization, seamless phase I/II or phase II/III transition; extending accrual or sample size re-estimation; enriching a subpopulation; and so on. No matter how adaptations are undertaken, they all should be completely specified in advance of the trial, so that the type I error rate can be properly controlled. Adaptive clinical trials are much more challenging and demanding than traditional fixed-sample trials. This is true not only in the design stage, but also during the trial conduct. Adaptive designs often require an integrated multidisciplinary research team and the infrastructure to allow for more frequent interim data monitoring. In particular, patients must be examined and treated on the regular basis along with biomarker analysis. We also need to design and oversee the entire trial conduct, implement real-time adaptive randomization, and carry out timely interim analyses.

As an example, adaptation in a phase I dose-finding trial means to escalate or de-escalate the dose based on the accumulating toxicity data. Patients are enrolled sequentially over time and are treated in cohorts. At any time of the trial, a new cohort may be treated at a lower, a higher, or the same dose, depending on whether the current dose is considered overly toxic, safe, or appropriate. Decision making on dose assignment is frequent and spontaneous upon each new cohort’s arrival. However, toxicity may be of late onset, such that the outcomes of previous patients are still not available when that information is needed for the next dose assignment. For example, in a dose-finding trial with the combination of oxaliplatin and gemcitabine along with concurrent radiation therapy, toxicity assessment required a nine-week follow-up, while the accrual was one patient every two weeks (Desai et al., 2007). Hence as a new patient entered the trial, some of the patients who had already been treated might have only been partially followed and their toxicity outcomes were missing. Such delayed outcomes inevitably pose great challenges to dose finding. More interestingly, that trial also raises a commonly encountered situation in which multiple therapies are combined for enhancing treatment synergistic effects. Here is another example of a drug-combination study: A seamlessly connected phase I/II trial evaluated both the safety and efficacy of the combination of decitabine and Ara-C in the treatment of acute myelogenous leukemia and myelodysplastic syndrome. Two doses of decitabine, two doses of Ara-C, and two treatment schedules were studied, which led to a total of eight different drug combinations.

1.3 MODERN CLINICAL TRIALS

Traditional cancer treatments, such as chemotherapies, take effect by impairing mitosis and act effectively on fast-dividing cancer cells. Unfortunately, these drugs (often known as cytotoxic agents) cannot discriminate fast-dividing normal cells and cancer cells and thus kill both blindly, which often results in substantial toxicity. Nowadays, with enormous expansion of our knowledge on the complex cancer pathways and networks, personalized medicine holds the most promise for the next generation of drug development. Targeted therapies are more specific toward certain disease pathways or inhibiting certain protein profiles. This type of agents utilize pathogenesis at a molecular level to differentiate patients who are more likely to respond from those who are not. Consequently, each patient is treated with individually tailored treatments. For example, imatinib (also called Gleevec) is highly effective in chronic myelogenous leukemia by inhibiting the BCR-ABL fusion protein that promotes cancer cell growth. This wonder drug works by seeking out and destroying cancer cells only, while leaving healthy cells virtually untouched. Another example is a monoclonal antibody, called trastuzumab with a trade name of Herceptin, which interferes with the human epidermal growth factor receptor 2 (HER2). Trastuzumab only works effectively in a subset of breast cancer patients with HER2 positive status. As the trend of personalized medicine grows, it is desirable to identify each patient’s biomarker profile in order to provide the best available treatment accordingly (Lee, Gu, and Liu, 2010).

Although many new agents are waiting in the pipeline to be tested and a large number of biomarkers (e.g., molecular profiles or protein pathways) have shown promising evidence to be therapeutically useful, efficient diagnosis and treatment as well as biomarker validation have proven to be extremely difficult. To overcome the biomarker barrier, Bayesian adaptive designs appear to be well-suited because they ideally adapt to information that accrues during the trial. In the modern era of clinical trials, the study design and trial conduct become more sophisticated than ever, which, in turn, demands more advanced and adaptive statistical methods. To appreciate the importance and complexity of the process, we present three recent high-profile clinical trials in the following.

The first trial is known as BATTLE (Biomarkers-Integrated Approaches of Targeted Therapy for Lung Cancer Elimination) at the University of Texas M. D. Anderson Cancer Center (Zhou et al., 2008). This umbrella study consisted of four parallel phase II trials for patients with advanced non-small-cell lung cancer (NSCLC). The trial assessed four targeted agents and four biomarkers simultaneously. Through timely tissue collection and biomarker analysis, BATTLE provided biomarker-based targeted therapies for NSCLC patients.

The statistical design of BATTLE is adaptive, flexible, and ethical. Based on Bayesian hierarchical modeling, the design enhanced borrowing information across different subtypes of biomarker groups. In addition, Bayesian adaptive randomization was used to favor treatments that were more likely to be effective during patient allocation. The trial continued to learn about treatment effects aligning with patients’ biomarker profiles. The BATTLE design possesses many desirable operating characteristics, such as

selecting effective drugs with high probabilities and ineffective drugs with low probabilities;

treating more patients with more effective drugs according to their tumor biomarker profiles; and

dropping inefficacious arms with high probabilities based on an early stopping rule.

In conjunction with an early stopping rule, Bayesian adaptive randomization appears to be a rational and smart choice for treating patients and underpinning effective treatments. As a follow-up study, BATTLE 2 is under the way.

Second, we introduce the highly anticipated multi-agent trial, called I-SPY 2 (Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis 2). This is an adaptive neoadjuvant phase II trial for women with newly diagnosed locally advanced breast cancer (Barker et al., 2009). The goal is to examine whether combinations of investigational drugs targeting molecular pathways with standard chemotherapy are better than standard chemotherapy alone. I-SPY 2 evolves from I-SPY 1, which has built an infrastructure to integrate enormous amounts of complex and disparate data from many resources and to facilitate real-time adaptive learning.

The standard biomarkers in I-SPY 2 are hormone receptor, HER2, and MammaPrint, while many other exploratory biomarkers are also involved. Based on practical and clinical relevance, the number of biomarker groups is narrowed down to ten for identifying molecularly tailored treatments. The standard neoadjuvant chemotherapy regimen include weekly paclitaxel (plus trastuzumab for HER2+ patients) followed by doxorubicin and cyclophosphamide. At any time of the trial, up to five novel targeted agents are investigated simultaneously, with the standard therapy added to each.

The primary endpoint in I-SPY 2 is pathologic complete response (pCR) at the six-month follow-up. Patients in each subgroup are adaptively assigned to the treatments that are believed to benefit them the most. However, potentially delayed outcomes may hamper the real-time implementation of adaptive randomization. To overcome this difficulty, the statistical design provides joint modeling of some surrogate endpoints and pCR. For each biomarker signature, the trial continuously updates drugs’ predictive probabilities of success in a phase III trial and, consequently, decisions are made on whether an experimental treatment should

graduate along with the corresponding biomarker signature and move forward to a more informed phase III trial,

be dropped for futility, or

continue for further evaluation after accruing more information.

During the trial, new drugs may be added to replace those that have either graduated or been dropped. Unlike the BATTLE trial which considered four fixed therapies, I-SPY 2, being more adaptive, allows treatments to come and go as the trial progresses.

At last, IPASS (Iressa Pan Asia Study) is a phase III trial to compare oral gefitinib (commonly known as Iressa) monotherapy with intravenous carboplatin and paclitaxel chemotherapy as first-line treatment in chemotherapy-naive Asian patients with advanced NSCLC (Mok et al., 2009). Prior to the IPASS study, there have been several randomized, controlled phase III trials of the epidermal growth factor receptor (EGFR) tyrosine kinase inhibitors for NSCLC treatment, but the results were confusing with some positive and some negative findings (Saijo, Takeuchi, and Kunitoh, 2009).

From a more selective patient population, IPASS enrolled a total of 1,217 patients from Asian countries and equally randomized them to gefitinib or chemotherapy (the combination of carboplatin and paclitaxel). The primary endpoint of IPASS was progression-free survival, and the primary objective of the study was the noninferiority of gefitinib to chemotherapy. Not only the trial concluded the noninferiority of gefitinib, but also demonstrated its superiority over chemotherapy. An interestingly finding was that the survival curves of the two treatment groups crossed at month six, favoring chemotherapy during the first six months and gefitinib thereafter. This suggested that patients could be a mixture of two possible subpopulations that were differentially responsive to the molecular-targeted therapy and cytotoxic agents. Further biomarker analyses showed that patients with EGFR mutations had longer progression-free survival in the gefitinib arm, while patients with wild-type EGFR had longer survival in the chemotherapy arm.

Following the findings of the IPASS study, European Commission granted the marketing authorization for Iressa as treatment of adults with locally advanced or metastatic NSCLC with an EGFR mutation. The endorsement of Iressa’s use in a subset of NSCLC patients reflects the growing importance of personalized treatment.

1.4 NEW DRUG DEVELOPMENT

Before any further discussion on new drug development, it is important to make a distinction between different types of agents. First of all, most of the oncology drugs are cytotoxic agents, which damage or destroy rapidly growing cancer cells. For example, carboplatin and paclitaxel typically shrink the tumor in a dose-dependent manner: A higher dose would result in more shrinkage of the tumor. The second type of agents are cytostatic; many targeted therapies belong to this family, and they are often directed at molecular targets to inhibit tumor growth or prevent the proliferation of cancer cells (Korn, 2004). Patients may benefit from cytostatic agents even without explicitly shrinking the tumor. For cytostatic agents, lower doses may be as effective as higher doses. For example, a tyrosine kinase inhibitor, lapatinib, specifically targets HER2+ in breast cancer patients. In the treatment of lung cancer, gefitinib prevents cancer cells from growing and multiplying by targeting EGFR through the disruption of EGFR signal transduction for cell division, apoptosis, and angiogenesis. Finally as the third type, biologic agents are substances from a living organism, such as interleukins and vaccines, which are often used in the prevention, diagnosis, or treatment of cancer and other diseases.

For illustration, we describe the intuition behind the four successive phases of clinical trials using cancer drug development as an example. Before initiating a clinical trial for a new chemical compound, extensive preclinical studies must have been carried out. In preclinical settings, in vitro (within a controlled environment—e.g., on glass slides or in test tubes) and in vivo studies (within a living organism, such as rodents) are performed to test a wide range of doses of the experimental agent. These cell-line and animal experiments mainly provide the preliminary toxicity and efficacy data, along with pharmacokinetics (PK) and pharmacodynamics (PD) information. PK refers to how the body processes the drug, characterizing the relationship between the dosage regimen and the drug concentration in the blood over time; PD studies how the drug works in the body by modeling the relationship between the drug concentration-time profile and therapeutic and adverse effects.

Suppose that laboratory scientists conducted extensive basic research in biochemistry and identified a new chemical compound that appears to be promising to eradicate cancer cells. Every drug comes with certain amount of risk. This chemical compound has never been tested in human subjects; thus the first task is to examine whether the new drug can be tolerated by human beings. We may be willing to accept a certain level of toxicity if the drug’s therapeutic benefits outweigh its adverse effects. This is particularly sensible with cancer drugs, because they often induce various levels of toxicity and adverse events.

As the first human study, a phase I clinical trial is launched to investigate the toxicity and side effects of the new agent on a small number of cancer patients. Often these patients are disease-relapsed or refractory to standard treatments, and sometimes there is no other better treatment option for them. In oncology, the goal of a typical phase I trial is to identify the maximum tolerated dose (MTD) and evaluate the drug’s dose-limiting toxicities. The MTD is defined as the dose that has a toxicity probability closest to the maximally tolerable level predetermined by the investigator. It is common to assume that both toxicity and efficacy effects of the drug increase as the dose increases. Thus a set of doses of the new drug is explored to find the most toxic dose (presumably also the most therapeutically effective) that can be reasonably tolerated by patients. In a dose-finding study, there is a trade-off between toxicity and efficacy. If the trial design is too conservative, the amount of the dosage may not be sufficient to fully impart the drug’s therapeutic effects; however, if the design is too aggressive, the administered dose may be too much to tolerate, and thus the study may result in excessive toxicity, or even death.

After the MTD of the new drug is determined, the next step is to assess whether the drug has sufficient biologic activity in opposition to the disease. For this purpose, a phase II clinical trial is undertaken, in which the drug is often administered at the MTD or the dose immediately lower than the MTD (sometimes called the recommended phase II dose—RP2D). The MTD is the highest dose that can still be tolerated, while use of the RP2D is a more conservative approach. Phase II is a proof-of-concept stage, which examines the drug’s short-term therapeutic effects and also continues monitoring severe adverse events. Nonworking or unsafe drugs should be killed as early as possible. Phase II trials often use a quick endpoint to guard the door so that drugs of little therapeutic effects will be blocked out. Once a phase II trial is completed, a decision is made on whether the drug is promising to warrant further investigation. At this stage, compounds found to be ineffective or unsafe should be dropped to avoid wasting more resources.

If a new drug successfully passes through the phase II testing, it will be moved forward to a phase III trial for definitive comparison with the current standard treatment or placebo. Phase III trials are large-scale and long-term randomized studies that may involve hundreds or even thousands of patients. If the drug is proven to be truly effective in such confirmative trials (typically two separate positive phase III trials are required for FDA approval), it will be filed to the regulatory agency for authorization of marketing. If granted approval, the drug prescription will be available to the general patient population in public.

Due to the restrictive eligibility criteria and rigorously specified conditions in phases I–III trials, some rare but serious adverse effects of the drug might not have been surfaced in the previous studies. Hence after the approval, a phase IV trial may be launched with more relaxed eligibility criteria, which will follow a larger number of patients over a much longer period of time. It provides an opportunity to learn more about rare side effects of the approved agent and its interaction with other treatments. Sometimes the findings in a phase IV trial may add a warning label to the drug, or even result in the removal of a drug from the market due to severe adverse events that were unforeseen when the drug was approved.

Conventionally, these four phases of trials are conducted sequentially and separately without any kind of formal borrowing information or strength across them. Each trial, regardless of the phase, requires an independent study design and a completely separate protocol for its own. There is typically a gap between two consecutive phases because it takes time and effort to complete and analyze the previous trial and also to initiate a new trial. Nevertheless, there is an increasing trend of combining phase I and phase II trials—seamless phase I/II trials, and combining phase II and phase III trials—seamless phase II/III trials, in order to expedite the drug development process.