Iterative proportional fitting as a balancing method in observational studies

This study compares Iterative Proportion Fitting (IPF) as a direct balancing method with five traditional propensity score modeling methods using 10-years of administrative and claims data from a regional health plan. Each method is assessed for internal and external covariate balancing between treated and controls, bias impact of control exclusions, and the design effect of the models. A scaled effect summary score (lower is better) shows that IPF performs better overall with a score of 0.09 while the propensity models have scores ranging from 0.12 to 0.31. All models show internal and external validity with average standardized covariate differences ranging between 0.0 and 0.2, while most models have design effects ranging between 1.0 and 5.6 suggesting modest variance inflation due to balancing. Four of the five propensity models exclude some control observations, and a Wilcoxon signed-rank test verifies that these exclusions were not random suggesting the existence of bias. This study demonstrates that IPF performs better than propensity score balancing methods primarily because IPF utilizes all observations which eliminates any control exclusion bias and reduces the external balancing effect due to perfect external treatment alignment among covariates.