Discussion of "Causal and Counterfactual Views of Missing Data Models'' by Razieh Nabi, Rohit Bhattacharya, Ilya Shpitser, James Robins

Viewing Missingness As InterventionViewing missingness as longitudinal intervention enabled closed-form efficient estimation for challenging bivariate censoring problems.Opposite to the common view of causal inference as missing data problems, viewing missingness as (longitudinal) intervention has also been a fruitful strategy, leveraging identification and estimation techniques from causal inference literature.For example, in the challenging bivariate censoring problem, viewing the complex censoring mechanism as longitudinal interventions under sequential randomization assumptions (SRA; it can be shown that SRA MAR for multivariate right-censored data and SRA = MAR for univariate right-censored data) leads to closed-form efficient influence curves which would have not been possible with only the MAR assumptions (Rubin, 1976; van der Laan and Robins, 2003).For example, if the full data is Xj(t) = I(Tj t), j = 1, 2 defined by survival times T1, T2 that are subject to censoring times C1, C2, then only assuming MAR is significantly weaker than assuming (C1, C2) (X1, X2), resulting in an efficient influence curve that does not exist in closed form.Even though such