An Application of Matching After Learning To Stretch (MALTS) to the ACIC 2018 Causal Inference Challenge Data
In the learning-to-match framework for causal inference, a parameterized distance metric is trained on a holdout train set so that the matching yields accurate estimated conditional average treatment effects. This way, the matching can be as accurate as other black box machine learning techniques for causal inference. We use a new learning-to-match algorithm called Matching-After-Learning-To-Stretch (MALTS) (Parikh et al., 2018) to study an observational dataset from the Atlantic Causal Inference Challenge. Other than pro-viding estimates for (conditional) average treatment effects, the MALTS procedure allows practitioners to evaluate matched groups directly, understand where more data might need to be collected and gain an understanding of when estimates can be trusted.1.
