Primary Submission Category: Dynamic Treatment Regimes
Robust Estimation of Stochastic Intervention Effects under Multistage Missingness
Authors: Shengfang Song, Hongxiang Qiu, Honglei Chen, Christine Parks, Zhehui Luo,
Presenting Author: Shengfang Song*
We develop estimation and inference methods for stochastic intervention effects in longitudinal studies with outcome-dependent sampling and multistage missing data. We consider estimation of the mean counterfactual outcome under a proportional modified treatment policy for a semi-continuous exposure, with outcomes measured decades after exposure and data collected through three stages.
Under a nonparametric model, we derive the efficient influence function and propose a one-step estimator that integrates stochastic treatment reweighting with stage-specific missingness mechanisms. The estimator accommodates semi-continuous exposures and monotone nonresponse, extending existing stochastic intervention methods that typically assume fully observed data or simpler missingness structures. Assuming correct specification of the treatment mechanism, the estimator is sequentially multiply robust, remaining consistent if at each stage either the outcome model or the missingness mechanism is correctly specified. When all nuisance functions are consistently estimated, the estimator is asymptotically linear and achieves the semiparametric efficiency bound.
This work fills a key gap by providing a single estimator that jointly handles stochastic interventions, outcome-dependent sampling, and multistage missingness. Simulation studies are underway, and an application to data from the Agricultural Health Study and its Pesticide and the Sense of Smell sub-study is ongoing.