Averaging causal estimators in high dimensions

摘要: There has been increasing interest in recent years the development of approaches to estimate causal effects when number potential confounders is prohibitively large. This growth led a estimators one could use this setting. Each these different operating characteristics, and it unlikely that estimator will outperform all others across possible scenarios. Coupling with fact an analyst can never know which approach best for their particular data, we propose synthetic averages over set candidate estimators. Averaging widely used statistics problems such as prediction, where there are many models, averaging improve performance increase robustness using incorrect models. We show ideas carry into estimation high-dimensional theoretically provides against choosing bad model, empirically via simulation performs quite well, most cases nearly well among Finally, illustrate environmental wide association study see largest benefit more difficult scenarios have large numbers confounders.