| One important feature of any randomized phase III clinical trial is heterogeneity in the patient population. There has been a long debate in both statistical and regulatory communities as to what are appropriate ways to deal with covariate information in the design and analysis of clinical trials. Today, a general standard is to pre-stratify a randomization procedure on important known covariates and perform a covariate-adjusted analysis in the end of the trial.; We have conducted a systematic study of the methods of handling covariates in the design of clinical trials. In clinical trials where outcomes are modeled using a homoscedastic linear regression model, the most efficient design is the one which balances treatment assignments across covariates. For these trials, various covariate-adaptive randomization procedures (also known as "dynamic allocation" procedures) have been proposed in the literature. These methods achieve well-balanced treatment groups, but their theoretical properties have been poorly researched.; We have explored the merits and validity of several covariate-adaptive randomization procedures using the well-defined concepts of balance and bias. In particular, we have developed a probability model for Pocock and Simon's procedure and found the stationary distribution of covariate imbalances in the case of a single polytomous covariate. We have also extended Efron's idea of accidental bias and provided measures which can be used to facilitate comparison of different covariate-adaptive randomization procedures.; In clinical trials with binary or survival responses the most efficient designs will be unbalanced, as they are based on heterogeneous models. One way to handle this is through the use of covariate-adjusted response adaptive (CARA) randomization procedures. We have proposed several CARA designs under the assumption of the logistic regression model and the exponential model. These designs are fully randomized, account for patient heterogeneity, and target specific allocations which satisfy compound optimality criteria (maximizing efficiency and minimizing ethical costs, such as expected number of treatment failures). In addition, some of our proposed procedures have small variability and good balancing properties, which can be psychologically reassuring for physicians. We evaluate the proposed CARA procedures both theoretically and by simulation in the situations when covariates are discrete and continuous. Our developed methods have been shown to be superior over some of the existing procedures which use covariate information in that they result in additional ethical savings without losing much in statistical efficiency. |
