Title: Contrasting identification criteria of average causal effects: Asymptotic variances and semiparametric estimators
Authors: Tetiana Gorbach - Umea University (Sweden) [presenting]
Xavier de Luna - Umea University (Sweden)
Juha Karvanen - University of Jyvaskyla (Finland)
Ingeborg Waernbaum - Uppsala University (Sweden)
Abstract: The back-door criterion (based on pre-treatment covariates) and the front-door criterion (based on mediators) are commonly used to identify an average causal effect from observational data given that the data generating mechanism is compatible with a directed acyclic graph (DAG). Even if the back- and the front-door criteria are not fulfilled, the causal effect might also be identified using mediators and pre-treatment covariates together when a condition that we call the two-door criterion holds. When several criteria hold for the DAG at hand, one may want to choose the criterion that provides the most efficient estimator. We give theoretical and numerical comparisons of asymptotic variances of semiparametric estimators based on the back-, the front-, and the two-door identification assumptions when any of the criteria hold simultaneously. The theoretical and simulation results obtained show that none of the criteria systematically yields the lowest asymptotic variance, or in other words, no estimation strategy is going to be most efficient in all situations. We, however, give conditions under which the two-door criterion is known to outperform the back-door criterion.