Semi-supervised Propensity Score Methods

semi-supervised-propensity is a repo containing simulation studies underlying the paper “Semi-supervised learning and the question of true versus estimated propensity scores.” Specifically, the code generates data from a complicated nonlinear model in which the probability of being assigned treatment is a function of the expected outcome. We then compare four methods that make use of the unlabeled treatment-covariate data:

1. IPW, with a misspecified parametric logistic propensity model
2. IPW, with a flexible, nonparametric BART propensity
3. TMLE
4. BCF

We also run each of the four methods above on simulated data with randomized treatment assignment.