A1503
Title: Causal inference with multiple versions of treatment via mixture-of-experts
Authors: Kohei Yoshikawa - Kyushu University (Japan) [presenting]
Shuichi Kawano - Kyushu University (Japan)
Abstract: SUTVA (Stable Unit treatment Value Assumption) includes the requirement that there are no multiple versions of treatment. In observational studies, although the implementation of treatment is not always controlled, multiple versions of treatment may exist in the treatment. Ignoring such latent versions violates the no-multiple-versions assumption and may obscure causal interpretation. Existing approaches clarify comparisons between treatments, however, they do not directly identify causal contrasts between latent versions within the treatment. A causal inference framework for multiple versions of treatment based on a mixture-of-experts model is proposed. This framework allows identification and estimation of pairwise average causal effects for treatment-version pairs even when the versions are unobserved. A maximum likelihood estimator is developed by using the EM algorithm and combined with inverse probability weighting to estimate the expectation of potential outcome for each version. Identification and large-sample properties under standard regularity conditions are also studied. Simulation studies show that the proposed method recovers the latent version structure and provides reliable estimates of causal contrasts.
