Propensity score modeling strategies for the causal analysis of observational data

Propensity score methods are used to estimate a treatment effect with observational data. This paper considers the formation of propensity score subclasses by investigating different methods for determining subclass boundaries and the number of subclasses used. We compare several methods: balancing a summary of the observed information matrix and equal-frequency subclasses. Subclasses that balance the inverse variance of the treatment effect reduce the mean squared error of the estimates and maximize the number of usable subclasses.