[Ml-stat-talks] Wed: Marc Ratkovic on covariate balance

David Mimno mimno at CS.Princeton.EDU
Mon Mar 12 12:04:26 EDT 2012


For this week's ML lunchtime talk, we welcome Marc Ratkovic from the
Politics department. Marc's method is an exciting new approach to
controlling the effect of covariates in a classification setting.

CS402, Wed 3/14, 12:30

Title: Achieving Optimal Covariate Balance Under General Treatment Regimes

Abstract:
Balancing covariates across treatment levels provides an effective and
increasingly popular strategy for reducing bias in observational
studies. Matching procedures, as a means of achieving balance,
pre-process the data through identifying a subset of control
observations whose background characteristics are similar to the
treated observations. The proposed method adapts the support vector
machine classifier to the matching problem. The method provides a
fully automated, nonparametric procedure for identifying a balanced
subset that accommodates both binary and continuous treatment regimes.
I show that, for the identified subset, the expected treatment level
is jointly independent of the pre-treatment covariates. Unlike
existing methods, the proposed method maximizes balance across all
covariates simultaneously, rather than along a summary measure of
balance. Two applications, a benchmark dataset measuring the impact of
a job training program and survey data with information on medical
expenditures and smoking habits, illustrate the method's use and
efficacy.


More information about the Ml-stat-talks mailing list