Use of propensity scores in non-linear response models
*Anirban Basu, University of Chicago
Willard G. Manning, University of Chicago
Daniel Polsky, University of Pennsylvania
Keywords: propensity scores, non-linear regression, doubly robust estimators, costs
The theory of propensity scores and their robustness issues have been extensively tested and supported in the context of a linear model and under the normality assumption. However, these issues have remained largely unexplored in the context of non-linear outcomes, especially for continuous and count data, where these methods have been recently applied. In this work, using extensive simulations, we explore the advantages and disadvantages to conditioning on propensity scores compared to non-linear covariate adjustment and combinations of both in estimating unbiased effects of alternative treatments or diseases on costs and other outcomes, marginal to other covariates in the model. We expect that our results and discussions will provide necessary evidence and guidance for applied researchers, who plan to use these methods on observational studies for cost-effectiveness analyses.