STATISTICAL METHODS FOR COMMUNITY INTERVENTION TRIALS

点击翻译按钮获取中文摘要

DESCRIPTION (Applicant's abstract): Group Randomized Trials (GRTs), in which groups rather than individuals are randomized into treatment conditions, are of central importance to community-based cancer prevention research, evidenced by the large number of GRTs conducted in the past decade. The overall goal of this proposal is to develop improved statistical evaluation methods for GRTS. The research carried out under this proposal consists of development and evaluation of analytical methods in GRTs including: 1) methods to make randomization-based inference more powerful for GRTS; 2) methods to evaluate trial results in the matched pair GRTs regardless of whether matching is effective or not; and 3) methods to make Generalized Linear Mixed Models more suitable for GRTS. The project will involve both theoretical and empirical work, drawing on data sources and collaborative opportunities provided by a large number of completed, and ongoing studies. 1). Theoretical work on randomzation-based inference will use a weighted permutation test and examine its properties, in particular the statistical power to detect an intervention effect. A weighted permutation-based confidence interval will be developed allowing for individual covariate adjustment. 2). Theoretical work on matched pair analysis will develop a test conditional on the observed correlation between matched conununities. The properties of this new method will be compared with traditional unconditional methods both analytically and via simulation. Such a test will allow investigators carrying out GRTs to use matching or blocking to control for factors potentially related to outcomes and yet recapture the power at the time of analysis if the matching or blocking is not effective. 3). Theoretical work on Generalized Linear Mixed Models will develop a new analysis method which is properly suited for GRTs where the number of survey subjects per community is usually large while the number of communities is usually small. The bias and efficiency of this method compared to competing procedures will be examined analytically and via simulation.