Other Lecture or Seminar

Adjustment for Treatment Selection Bias in Retrospective Studies

Dr. Bob Obenchain
Eli Lilly & Company

Start Date and Time: Mon, 19 Nov 2001, 4:30 PM

End Date and Time: Mon, 19 Nov 2001, 6:00 PM

Venue: LAEB B222

Abstract:

My clients in Lilly US Medical Outcomes Research perform studies to establish that marketed Lilly products are cost-effective. While "efficacy" refers to the potential effects of treatments under ideal conditions, "effectiveness" refers to their actual effects under real-word conditions (e.g. without artificial compliance incentives.) Thus my clients end up using a "mosaic" of research designs that includes retrospective study of claims databases. In these "observational" studies, one cannot rely on randomization of patients to treatments to provide unbiased estimates. 

We first review the conventional Propensity Score (PS) binning approach of Rosenbaum and Rubin [1984: J Amer Stat Assoc 79, 516-524.] We use trellis graphics in Splus to illustrate how within-bin distributions of covariates become independent of treatment choice when an appropriate logistic regression predictor of treatment selection has been found. Bagged classification tree PS estimates are useful in sensitivity analyses, and clusters of PS estimates for 3 or 4 treatments can be combined to form bins (sub-classes, strata.) Finally, I illustrate a new PS methodology based upon spline or lowess smoothing. 

My experience has been that propensity (patient matching) methods are not only fairly easy to understand and apply but also are fairly robust. By way of contrast, simultaneous equation (econometric) methods, like those of Heckman [1979: Econometrica 47,153-161] and Murnane, Newstead and Olsen [1985: J Bus & Econ Stat 3, 23-35], can be frustratingly sensitive to model mis-specifications and departures from normality.