Primary Submission Category: Matching
Asymptotically exact propensity and prognostic score matching, with application to recovery from pandemic-related learning loss in K-12 education
Authors: Ben B. Hansen, Mark Fredrickson, Josh Wasserman,
Presenting Author: Ben Hansen*
Matches made using estimated propensity or prognostic scores are inherently inexact, but with suitable combinations of caliper restrictions make them exact in the asymptotic limit. Such matches lead to directly adjusted causal effect estimates that are consistent, assuming a parametric scoring model and the absence of hidden bias, under the weakest possible overlap condition. Asymptotic exactness is often simultaneously attainable for a propensity and a prognostic score, if not also for multiple specifications of either or both scores. This engenders a multiple robustness for matched, as opposed to inverse-probability weighted, estimation.
Following a sketch of their supporting theoretical arguments, these points will be illustrated in an observational study of an innovative Texas K-8 public education program. Research on effectivness of this program was selected for federal funding in an Institute for Education Sciences competition aimed at identifying policies to promote recovery from pandemic-related learning loss. Texas happens to have begun program rollout on the eve of the pandemic, and happens to make public sufficient school-level information to make a matched study feasible. While end analysis of the federally funded project will involve protected student data and a number of additional design and analysis elements, the simpler preliminary analysis to be described in this talk centers publicly available data, and asymptotically exact matching.