Sensitivity analysis for non-ignorable dropout of marginal treatment effect in longitudinal trials for G-computation based estimators

We specify identifying assumptions under which linear increments (LI) estimator can be used to estimate unconditional expectation for longitudinal data from a clinical trial in the presence of dropout. We show that these are analog conditions under which extended linear SWEEP estimator achieves unbiased estimation of the identical parameter in the same setting. Within a class of linear autoregressive models we specify how strategies implemented in LI and extended SWEEP relate to each other w.r.t. the conditional expectation of increments and outcomes respectively. We utilize conceptual overlap of these two methods to define a sensitivity analysis for both of them in presence of non-ignorable dropout. Interdependency of these two approaches offers a natural solution to a prominent problem of asynchronous association between outcome and dropout inevitably encountered in sensitivity analysis for dropout in longitudinal data. Validation of our approach is done on the data coming from a randomized, longitudinal trial of behavioral economic interventions to reduce CVD risk. We subsequently show that our approach to sensitivity analysis can be perceived as extension of the pattern mixture method defined by Daniels and Hogan in 2007. to longer sequences of observations. For T=3 we give the explicit expression for bias of our approach w.r.t. mentioned pattern mixture approach. We further show on a subset of the data from the same study that this bias does not invalidate our sensitivity analysis for LI when it comes to evaluating the robustness of findings under increasingly less ignorable dropout