Doubly Robust Criterion for Causal Inference

TL;DR

This paper develops a new AIC-type criterion for propensity score analysis in causal inference that is doubly robust, allowing for accurate model selection even if either the outcome or assignment model is misspecified.

Contribution

It introduces a novel doubly robust information criterion based on Kullback-Leibler divergence for causal inference models, extending AIC to non-linear and complex models.

Findings

The new criterion outperforms existing criteria in simulations.

It more accurately identifies the true model structure.

Real data analysis shows significant differences in variable selection results.

Abstract

The semiparametric estimation approach, which includes inverse-probability-weighted and doubly robust estimation using propensity scores, is a standard tool in causal inference, and it is rapidly being extended in various directions. On the other hand, although model selection is indispensable in statistical analysis, an information criterion for selecting an appropriate regression structure has just started to be developed. In this paper, based on the original definition of Akaike information criterion (AIC; \citealt{Aka73}), we derive an AIC-type criterion for propensity score analysis. Here, we define a risk function based on the Kullback-Leibler divergence as the cornerstone of the information criterion and treat a general causal inference model that is not necessarily a linear one. The causal effects to be estimated are those in the general population, such as the average treatment effect on the treated or the average treatment effect on the untreated. In light of the fact that this field attaches importance to doubly robust estimation, which allows either the model of the assignment variable or the model of the outcome variable to be wrong, we make the information criterion itself doubly robust so that either one can be wrong and it will still be an asymptotically unbiased estimator of the risk function. In simulation studies, we compare the derived criterion with an existing criterion obtained from a formal argument and confirm that the former outperforms the latter. Specifically, we check that the divergence between the estimated structure from the derived criterion and the true structure is clearly small in all simulation settings and that the probability of selecting the true or nearly true model is clearly higher. Real data analyses confirm that the results of variable selection using the two criteria differ significantly.