The Assessment Of Response-adaptive Randomization

Randomized clinical trials (RCTs) can present a scientific/ethical dilemma for clinical investigators. From statistical point of view, statisticians prefer RCTs because of the statistical and scientific advantages. While from the ethical consideration, we should make sure that every patient can get the best benefit since human beings are subject to research. The widely used RCT is presently regarded as the most scientific and sound method, but it still doesn't fully consider the ethical related problems. So is there an approach that can balance the dilemma between ethical and statistical arguments? The adaptive designs appeared for this purpose.In this research, we focused on the response-adaptive randomization, one type of the adaptive designs. We first described the techniques of response-adaptive random allocation. Two urn models, the randomized play the winner (RPW) and the drop the loser (DL), were investigated, and had been compared with the complete random allocation that each treatment's probability were fixed through the trial. Simulation study and datasets from real clinical trials were used to examine the statistical properties of the methods. The power of the tests, the proportion of patients assigned to treatment A, the total number of treatment failures, and the expected reduced number of failures were considered.1. Updated the urn after each patient responseThe first method investigated was using response-adaptive randomization to update the urn after each patient response, assuming immediate patient response. We considered two situations: (1) Fixed the sample size in advance to compare the power, the proportion of patients assigned to treatment A, the total number of treatment failures and the average reduced failures when used RPW rule, DL rule and equal allocation respectively in clinical trials. The sample size was calculated that yielded power of 90 percent under equal allocation. (2) Fixed the sample size to be 240, 480, 960 respectively to compare the properties mentioned above under three randomization rules.2. Updated the urn after fixed number of patient responsesThe second method investigated was using response-adaptive randomization to update the urn after fixed number of patient responses, assuming immediate patient response. We also considered the two situations which had been used in our first method. Meanwhile, we fixed the number of patients by setting the number of stage to be 2, 3,4, 5. Except comparing the RPW rule and DL rule with equal allocation, we also compared the differences among them when we chose different number of stages.3. Cases studyFor illustrate purpose, two real datasets were analyzed. One was the clinical trial which published on the New England Journal of Medicine in 1994 by Corner et al, the other was the clinical trial which we attended to analyze the dataset in 2006. The differences between the four properties were explored when the patients were re-randomized by RPW rule and DL rule, assuming immediate patient response. The two ways of updating the urn, updating after each patient response and after fixed number of patient responses were also considered in this part.The main results of this research are as following:1. When the urn is updated after each patient response, the DL rule, compared with the RPW rule, is more effective for reducing the total number of treatment failures, increasing the proportion of patients assigned to treatment A and the average reduced failures whilst still maintaining the power.2. When the urn is updated after fixed number of patient responses, the RPW rule, compared with the DL rule, tends to be the more effective rule when the success probabilities are small, while the DL rule performs better when the success probabilities are relatively large.3. When compare the results between the two different ways of updating the urn, the adaptiveness is reduced if we updating the urn after fixed number of patient responses for the same randomization rule.Based on these conclusions, we propose the situations where the response-adaptive randomization could be used: (1) superior trial; (2) the differences between groups are large enough; (3) the outcome of the investigated diseases are serious; (4) the patients responses are immediate.