A Two-Stage Least Squares-Based Sensitivity Analysis Approach for Assessing the Role of Engagement with Text Message-Delivered Interventions
Jamie Joseph, PhD Candidate Department of Biostatistics Vanderbilt University
There has been a recent proliferation of clinical trials studying the use of interactive text messages to support medication adherence. In general, it is understood that a mobile intervention’s benefit should be derived at least in part by engagement with the intervention (quantifiable via, e.g., subject-specific response rate). Isolating the role of engagement is complex in this setting due to unmeasured confounding, and traditional instrumental variable analyses fail due to potential violations of the exclusion restriction assumption (i.e., that a treatment’s benefit must be solely derived through engagement). In this talk, we formalize the causal definitions and assumptions necessary to examine the role of engagement in the intervention’s effect using a modified instrumental variable analysis. We expand on a prior approach that allows for the bounding of local average treatment effects in people with theoretical levels of compliance by extending it to a two-stage least squares framework. We also formulate a closed-form sandwich variance estimator and demonstrate that this method can accommodate causal pathways disallowed in traditional instrumental variable methods. Through simulation and application to a recent study, we demonstrate that this approach can be successfully implemented in mobile health intervention settings.