Abstract:
|
One post-stratifies on propensity, risk or prognostic scores in order to tell causal stories with observational data; estimation uncertainty attaching to the index scoring model ordinarily plays no role. This talk proposes a mechanism through which such uncertainty can support and improve the telling of the primary causal story. It begins with a novel appraisal of sampling variability inherent in fitting of parametric index models, offers a new diagnostic for matching on the basis of the index, and finally gives advice on the setting of matching radii or subclass widths. The method is applied to evaluation of the 2006 Massachusetts health care reform. It is seen with the method's help that propensity score subclasses should be drawn more narrowly than those used in Sommers, Long and Baicker's (2014) widely reported, but contested (Kaestner 2016), analysis of mortality changes following the reform. At the same time, from the conceptual perspective with which propensity scores usually are interpreted, the method's recommended subclass widths are surprisingly wide.
|