Likelihood-Based Meta-Analysis: Conditional vs. Unconditional Approaches

• Activity Number: 692
• Type: Contributed
• Date/Time: Thursday, August 8, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Jingjing Yan*+ and Eloise E. Kaizar and Steven MacEachern
• Companies: The Ohio State University and The Ohio State University and Ohio State University
• Keywords: Meta-Analysis ; Likelihood-Based ; Baseline risks ; Unconditional ; Conditional

Abstract

Likelihood-based approaches to combining multiple similar randomized controlled trials in a single analysis(i.e.,meta-analysis)are gaining popularity in practice.In these models, we can either specify a single fixed effect of treatment or a distribution for random effects of treatment across all the studies.However,there is no agreement for how to treat the baseline risks, or the risks for patients in the control groups.It is generally recognized that specifying a single fixed baseline risk may produce unexpected results such as Simpson's paradox.Thus, alternatives to this model should be explored.Two possibilities are to 1) explicitly model the baseline risks or 2)work with a conditional likelihood that does not involve the baseline risks.While the second approach is appealing because no model assumptions for the baseline risks are apparent,we show that in fact the model assumptions are hidden in the conditional likelihood.In particular, we show that when the baseline risk is not independent of the treatment effect(log odds ratio),the unconditional model can produce estimates of treatment effect that are quite biased, whereas estimates based on explicit models perform much better.