American Statistical Association New York City Metropolitan Area Chapter Memorial Sloan-Kettering Cancer Center Biostatistics Seminar

Cancer epidemiology studies routinely examine multiple putative risk factors to identify those associated with disease or disease-related traits. A standard analytic approach is to fit a regression model using all the risk factors. But, since any individual risk factor may confer only a modest association with the outcome, there is now an increasing trend in evaluating user-defined linear combinations of the risk factors. For example, several studies now examine the effect of the cumulative number of variant alleles from multiple genes in candidate biochemical pathways instead of including the individual variants in the regression model. There is now an increasing number of publications reporting significant associations based on such analyses. Some papers even conclude that using the aggregate number of genetic variants can be a useful strategy for improving disease risk prediction to identify high-risk individuals in clinical settings. These analyses, however, rely on the presumption that the user-defined linear combinations are accurately specified. These methods may lead to biased effect estimates and substantially inflated type I errors and, hence, incorrect inference about the individual risk factors if the linear combinations are incorrectly specified. This talk will examine how this issue may be potentially remedied using preliminary test estimators and some simple empirical Bayes-type shrinkage estimators, and illustrate the benefits and limitations of these intuitive methods using simulations and a questionnaire data.