A0231
Title: Efficient federated learning of the average treatment effect
Authors: Sijia Li - Harvard University (United States) [presenting]
Abstract: A new data fusion method is introduced that utilizes multiple data sources to estimate the average treatment effect. Most existing methods only make use of fully aligned data sources that share common conditional distributions of the outcome. However, in many settings, the scarcity of fully aligned sources can make existing methods require unduly large sample sizes to be useful. The approach enables the incorporation of weakly aligned data sources that are not perfectly aligned, provided their degree of misalignment is known up to finite-dimensional parameters. The canonical gradient is derived to estimate the average treatment effect in such a data fusion setting, and the semiparametric efficiency bound is characterized. Furthermore, the proposed approach is decentralized and only requires individual-level data from the user-specified target data and summary-level statistics from other sources.
