Clinical Trials and Statistical Methods for Cancer Research
Our major research activities involve the design and analysis of clinical trials and the study of related statistical issues. The Statistical Center for the Eastern Cooperative Oncology Group, headed by Dr. Gray, is involved in dozens of studies which often include biological correlative endpoints and quality-of-life evaluations, as well as evaluations of primary efficacy. Current research interests include methods of adjustment for treatment noncompliance, design and analysis of studies with competing-risks endpoints, methods for analyzing correlated failure-time data, dependent censoring (related to incomplete follow-up of disease status), issues of interim analysis of clinical data, and analysis of high-dimensional laboratory data.Correlated failure-time data can arise in a variety of settings, such as multicenter clinical trials, in which patient outcomes may vary by center. Correlated data methods for estimating the effects of baseline characteristics on the prognosis of cancer patients have been developed. These methods generally involve solving estimating equations, which depend on families of weight functions that can be used to incorporate information on the degree and structure of the dependence in the data. In theory, incorporating such information should improve prognostic estimates of effects, but attempts to do so have not been very successful. We have investigated the magnitude of the gains possible in both linear failure-time models and proportional hazards models. From this work, we have developed new approaches for estimation using linear-models that closely approximate the weight functions that minimize the sampling variation of the estimated effects. The presence of competing causes of failure for the primary end point in a clinical trial is another problem that makes interpretation and analysis difficult. For example, in studies of treatment for elderly patients with breast cancer, the primary end point might be prolonging time to disease recurrence or time to breast cancer mortality. But some patients may develop other diseases and die without experiencing failure of the primary end point for the study, thus creating complications for the statistical analysis. Related problems can occur in examining the effect of radiotherapy on controlling disease at the site where it is applied, when there are competing risks of disease at other sites. We have developed models that directly estimate the effect of factors on the probability of observable events in these competing-risks settings, and have successfully applied these models to local failure rates in breast cancer patients. Progression-based failure-time endpoints in clinical trials require adherence to a regular schedule of disease evaluations. When protocol treatment is terminated without documented progression, the disease evaluation schedule may not be followed, or results of evaluations may not be reported. Patients withdrawn from study without documented progression often have different risk for progression than those continuing on study, and the lack of further information on disease status for these patients potentially biases statistical analyses. In current work, we are investigating the potential impact of these biases and developing methods for exploring the sensitivity of analyses to incomplete follow-up on disease status.