Pushing Boundaries on Solutions for Non-ignorable Missingness

In clinical settings, accurately weighing the probability of treatment benefits against the risks of side effects or adverse events is critical. However, inferring causal effects of treatments on multiple outcomes is often complicated by missing data on outcomes, which typically includes primary efficacy measures and secondary safety assessments. Building on foundational concepts that utilize graphical models to align missing data models with their causal counterparts, this talk focuses on an MNAR mechanism known as the block-conditional MNAR model. This model accounts for the influence of measured, unmeasured, or partially measured variables on the missingness indicators. We introduce novel nonparametric estimation methods designed to evaluate the trade-offs between population-averaged benefits and risks in clinical treatment decisions. These methods can accommodate mixed-type responses, such as binary and continuous data. The presentation will illustrate the methodological advancements and practical implications for robust inference on treatment effects in clinical settings. 

The instrumental variable model of Imbens and Angrist (1994) and Angrist et al. (1996) allow for the identification of the local average treatment effect, also known as the complier average causal effect. However, many empirical studies are challenged by the missingness in the treatment and outcome. When the treatment and outcome are missing not at random (MNAR), the CACE is in general not identifiable because the underlying data distribution can not be identified without further assumptions. We study the identifiability of the CACE even when the treatment and outcome are MNAR. Through an exhaustive search, we identify all MNAR mechanisms that enable the identification of the CACE without the need for auxiliary information. This is achieved under the following two scenarios: (1) when missing data are present exclusively in either the treatment or the outcome, and (2) when missing data occur in both the treatment and outcome in the context of prospectively collected data. We review the existing results and obtain many new results to complete the discussion.
 

Missing data is a common problem that challenges the study of treatment effects. In the context of mediation analysis, this paper addresses the missingness in the following two key variables, mediator and outcome, focusing on identification. We first consider self-separated missingness models, where identification is achieved by conditional independence assumptions only. This class of models is somewhat limited because it is constrained by the need to remove a certain number of connections from the model. We then turn to self-connected missingness models, where identification relies on information from so-called shadow variables. This class of models
turns out to contain substantial variation, allowing models with built-in shadow variables (mediator, outcome or covariates) and models with auxiliary shadow variables at different positions in the causal structure (relative to the mediator and outcome). In order to improve the practical plausibility of the missingness mechanisms, when constructing the models, we allow for dependencies due to unobserved causes of the missingness wherever possible. In this exploration, we develop theory where needed. This results in templates for identification in the mediation setting, generally useful identification techniques, and perhaps most importantly a synthesis and substantial extension of shadow variable theory. This is joint work with Razieh Nabi, Fan Yang and Elizabeth Stuart. 

Scharfstein et al. (2021) developed a sensitivity analysis model for analyzing randomized trials with repeatedly measured binary outcomes that are subject to nonmonotone missingness. Their approach becomes computationally intractable when the number of measurements is large (e.g., greater than 15). In this paper, we repair this problem by introducing mth-order Markovian restrictions. We establish identification results by representing the model as a directed acyclic graph (DAG). We illustrate our methodology in the context of a randomized trial designed to evaluate a web-delivered psychosocial intervention to reduce substance use, assessed by testing urine samples twice weekly for 12 weeks, among patients entering outpatient addiction treatment. We evaluate the finite sample properties of our method in a realistic simulation. This work is joint with Jaron Lee, Ilya Shpitser, Agatha Mallett, Aimee Campbell and Edward Nunes.