Interpretable Almost Matching Exactly

In observational studies, confounding biases estimation of treatment effects. To remedy this, similar units can be matched together, with each matched group emulating a randomized experiment. Treatment effects can then be estimated by aggregating across matched groups. There are many ways to match units; I focus on interpretable methods, which are crucial in high-stakes decision contexts. Such methods allow a practitioner to go back to the raw data and determine precisely why units were matched together and therefore why the treatment effect estimate is what it is. I’ve worked on scalable matching methods for categorical data that match units exactly on as many covariates as possible, prioritizing matches on covariates learned to be more important. In the context of continuous data, this translates to learning regions of the space where the treatment effect is roughly constant and units are therefore “close enough” to be accurately matched. I’ve also applied this work in non-standard contexts, such as when units experience network interference. Lastly, I’ve created an R package implementing some of these methods to make them accesible to practitioners; more are forthcoming. Currently, I’m working on how best to generate variance estimates for resulting treatment effects. For more information, see my work with the Almost Matching Exactly Lab.