Caution Should Be Used in Applying Propensity Scores Estimated in a Full Cohort to Adjust for Confounding in Subgroup Analyses

• Activity Number: 272
• Type: Invited
• Date/Time: Tuesday, July 31, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Sue M Marcus*+ and Robert D Gibbons
• Companies: Columbia University and The University of Chicago
• Keywords: propensity score ; confounding ; bias ; observational studies ; matching

Abstract

Propensity score matching (Rosenbaum and Rubin, 1983) is a popular and useful tool for removing bias in observational studies. In theory (Rosenbaum and Rubin, 1983), a correctly-specified propensity score should remain valid in a subgroup population. However, it is generally more difficult to correctly specify propensity scores for a full cohort when many subgroups have different processes related to confounding by indication. For example, Patorno and colleagues (2010) used a 'full' cohort who had a variety of indications for using AEDs, including epilepsy, convulsions, neuropathic pain, migraine, depressive disorder, manic depressive disorder, psychoses, alcohol and drug abuse or dependence, delirium and dementia; Gibbons et al used a subgroup, i.e. those with bipolar disorder only. Their conclusions were quite different, perhaps due to the use of propensity scores estimated in a full cohort versus a subgroup. The propensity to use a particular antiepileptic drug may depend upon the disorder, and each disorder has different underlying suicide rates and may have different patterns of confounding.