Direct Inference Requiring Only a Touch of Bayes: An Alternative to the Nissen-Wolski Meta-Analysis Questioning the Safety of Avandia (Rosiglitazone)

• Activity Number: 412
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Ralph G. O'Brien*+
• Companies: Case Western Reserve University

Abstract

In raising concerns about the safety of Avandia, the meta analyses by Nissen and Wolski (2007, 2010) helped precipitate the 9/2010 FDA decision to markedly restrict the drug's use. However, the N&W methodology drew sharp criticism, and analyses by others reached different conclusions. Here, the NW2010 data is used to illustrate a proposed inference method, and its associated point estimate and interval. Let theta be the true difference in event proportions, p.Avandia - p.comparator. Consider the three hypothesis intervals, {-} = {theta < -0.0005}; a null interval, {0} = {-0.0005 <= theta <= 0.0005}; and {+} = {theta > 0.0005}. Let the a priori priors be p{-} = 0.025, p{0} = 0.95, p{+} = 0.025. Using NW2010, the posteriors are p{+ | data} = 0.014 for death and 0.809 for MI. Consider the point null, {.} = {theta = theta0}. The theta0 that maximizes the cumulative Bayes factor for {.}, provides a point estimate; here, 0.0000 for death and 0.0001 for MI. 9:1 support intervals come from finding delta such that p{0} = p{theta0 - delta <= theta <= theta0 + delta} = 0.10 goes to p{0 | data} = 0.90. For NW2010, these intervals are [-0.003, 0.003] for death and [-0.004, 0.005] for MI.