Matching with Propensity Scores, Prognostic Scores, and Prognostic Propensity Scores

• Activity Number: 22
• Type: Topic Contributed
• Date/Time: Sunday, July 29, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Ben Hansen*+
• Companies: University of Michigan
• Keywords: matching ; confounding ; risk ; observational study

Abstract

Suppose a comparison of exposed and unexposed groups with sufficient baseline information to overcome any confounding of exposure effects with preexisting differences. It suffices in principle to match exposed and unexposed subjects on a propensity score constructed from all of the baseline variables, then construct effect estimates from matched differences in the outcome; but including _all_ baseline variables, even those with little relevance to the outcome, can backfire in various ways. A natural alternative is to match on risk or "prognostic" scores instead (Hansen,2009, Biometrika), but this can add subtle biases to later effect estimation that are difficult to compensate for. On the other hand certain compromise procedures, to be explained in this talk, appear to shed the two approaches' liabilities while retaining most all of their benefits.