Unsupervised Propensity Scoring: NN and IV Plots

• Activity Number: 177
• Type: Topic Contributed
• Date/Time: Tuesday, August 10, 2004 : 8:30 AM to 10:20 AM
• Sponsor: Section on Health Policy Statistics
• Author(s): Robert L. Obenchain*+
• Companies: Eli Lilly and Company
• Keywords: causal inference ; instrumental variables ; nonrandomized study ; clustering ; nearest neighbors ; treatment effects

Abstract

The traditional propensity scoring (PS) approach to adjustment for treatment selection bias in nonrandomized studies is to fit a logistic regression model predictive of treatment choice and to then match or subclassify patients using either the resulting linear functional or actual (nonlinear) PS. While patients with identical PS predictions must lie on the same hyperplane (linear subspace) of covariate X-space, they are not necessarily close to each other within that hyperplane. In contrast, clustering of patients in X-space can assure this sort of closeness, at least when clusters are both compact (due to complete patient linkage) and numerous. Because clustering is an unsupervised method, any number of treatments and health outcomes can then be compared using a single, hierarchical set of clusters. Here, we stress use of nearest neighbor (NN) and instrumental variable (IV) plotting approaches because they not only reveal underlying assumptions about covariates but also depict the sensitivity of estimated treatment effects to choice of tuning parameters ...such as the number of clusters, choice of clustering algorithm, and patient dissimilarity metric.