Commentary: Matched Instrumental Variables: A Possible Solution to Severe Confounding in Matched Observational Studies?

Matching in ObservatiOnal studies and shi et al. matching is a popular technique to deduce causal effects of a treatment on an outcome in observational data. In brief, matching individuals in groups with different values of the treatment (for cohort designs), but similar values of the observed covariates, so that within each group, the only difference between the individuals is their treatment values; for case–control designs, matching is done on the outcome, instead of on the treatment. then, under the usual set of causal identifying assumptions (conditional ignorability, consistency, and positivity), one can estimate the average causal effect of a treatment on an outcome. While there are many other techniques to estimate causal effects in observational data, including standardization, g-formula, g-estimation, inverse probability weighting, stratification, and targeted maximum likelihood estimation (see references 5–10 for textbook discussions), matching has some advantages over these methods. First, matching is transparent in assessing covariate balance. that is, if there are values of covariates for which almost all individuals have a high (or low) value of the treatment, then matching and its associated diagnostics will tell us that matched sets cannot be formed. Second, matching is blind to the outcome data; a matching algorithm only requires the measured covariates and the treatment values. more importantly, matching diagnostics and covariate balance checks can be done all without looking at the outcome data. Finally, for estimation, matching is nonparametric; it does not use any parametric modeling assumptions, such as linearity. For more discussions and recent overviews, see. In reference 1, the authors used matching to analyze the causal effect of trauma care from different trauma centers (treatment) on emergency department mortality (outcome). Because there were three treatment arms/trauma centers, level I and II trauma centers and nontrauma centers, the authors employed triplet matching where each patient from nontrauma centers were matched exactly with one patient from level I and II trauma centers, forming a triplet. Ideally, the patients in a matched triplet were similar with respect to the eight covariates about patient demographics and health.