Subgroup Analysis: Guidelines for Department of Veterans Affairs Clinical Trials

• Activity Number: 559
• Type: Topic Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Government Statistics
• Author(s): Robert Lew*+ and Ying Lu and Mei-Chiung Shih and Grant Huang
• Companies: VA Boston Cooperative Studies Program Coordinating Center and Palo Alto VA CSP Coordinating Center/Stanford and Veterans Administration and Veterans Administration
• Keywords: subgroup analysis ; type I error ; policy guidelines

Abstract

The VA has done over 500 multicenter clinical trials drawing most subjects from the millions of US veterans cared for at VA facilities. The typical trial primary hypothesis has 90% power and a 5% Type I error. Subgroup analyses are exploratory because the sample size adequate to power all of them would make the trial infeasible. Thus, what is a fair level of Type I error for subgroup analyses? We propose guidelines that would: 1) pre-specify a list of subgroup analyses, 2) require analytic reports to reference the primary and other pertinent results based on the primary data, and 3) set the level below 5% but above the Bonferroni level. The pre-specified list would include hypotheses in the proposal and those recommended by the DMC during the trial and by disinterested experts. We discuss approaches to Type I error that account for correlation among the test statistics associated with a set of hypotheses; Efron's method to control for the false discovery rate, Simon's empirical Bayes method for clinical trials, and Bayesian subgroup analysis methods. Adherence to such guidelines would refine secondary analyses, inter-relate conclusions, and clarify the import of major results.