We focus on causal inference for longitudinal treatments, where units areassigned to treatments at multiple time points, aiming to assess the effect ofdifferent treatment sequences on an outcome observed at a final point. A commonassumption in similar studies is Sequential Ignorability (SI): treatmentassignment at each time point is assumed independent of unobserved past andfuture potential outcomes given past observed outcomes and covariates. SI isquestionable when treatment participation depends on individual choices, andtreatment assignment may depend on unobservable quantities associated withfuture outcomes. We rely on Principal Stratification to formulate a relaxedversion of SI: Latent Sequential Ignorability (LSI) assumes that treatmentassignment is conditionally independent on future potential outcomes given pasttreatments, covariates and principal stratum membership, a latent variabledefined by the joint value of observed and missing intermediate outcomes. Weevaluate SI and LSI, using theoretical arguments and simulation studies toinvestigate the performance of the two assumptions when one holds and inferenceis conducted under both. Simulations show that when SI does not hold, inferenceperformed under SI leads to misleading conclusions. Conversely, LSI generallyleads to correct posterior distributions, irrespective of which assumptionholds.
展开▼