🤖 AI Summary
This study addresses policy evaluation when the outcome variable is non-Gaussian—exhibiting skewness or heavy tails—and the error distribution is unknown. Under a structural model assuming a low-dimensional parametric form for the mean function and independence between errors and both treatment assignment and covariates, the authors derive the semiparametric efficiency bound and the corresponding efficient influence function. They propose a targeted maximum likelihood estimator based on cross-fitting and efficient regression scores. The method substantially outperforms Gaussian working models, Bayesian additive regression trees, and augmented inverse probability weighting in settings with correctly specified mean structures and imbalanced treatment allocation, yielding lower root mean squared error and shorter confidence intervals. Its practical advantages are further demonstrated through an application to earnings data from the National Supported Work program.
📝 Abstract
We study semiparametric estimation of average treatment effects in a structured outcome model whose mean function is indexed by a finite-dimensional parameter, while the additive error distribution is left otherwise unspecified apart from mild regularity conditions and independence from treatment and baseline covariates. The framework is motivated by policy-evaluation settings in which the main economic structure is plausibly low dimensional but outcome distributions are distinctly non-Gaussian, for example because earnings are skewed or heavy tailed. We derive the efficient influence function and semiparametric efficiency bound for the average treatment effect under this model, and we show how the resulting estimator can be implemented through a cross-fitted targeted updating step driven by the efficient regression score. Simulation evidence indicates that when the mean structure is correctly specified and the main difficulty lies in the error distribution, the proposed estimator can deliver smaller root mean squared error and shorter confidence intervals than Gaussian working-model inference, Bayesian additive regression trees, and augmented inverse-probability weighting under more imbalanced treatment assignment. An application to the National Supported Work program illustrates the empirical relevance of the approach for transformed earnings outcomes.