| dc.description.abstract | Recent technological advancement has resulted in the proliferation of interactive text message-based interventions to support medication adherence in patients managing chronic illnesses. Several recent clinical trials have identified these interventions as a strategy to improve outcomes, particularly when used in combination with other interventions. In these settings, patient engagement with these text messages may drive a portion of the intervention’s effects on key outcomes. Such trials typically include a control arm with no opportunity to engage with text messages. Nevertheless, the relationship between engagement and outcomes may be subject to unmeasured confounding. Quantifying treatment effects using engagement as a post-randomization variable is therefore challenging. In this dissertation, we develop approaches to handle these challenges and provide researchers with principled tools to understand the role of engagement with mobile health interventions. Our first focus involves methods to estimate and bound functional local average treatment effects (i.e. an effect of treatment at theoretical levels of engagement under the intervention), when the exclusion restriction cannot reasonably be assumed. We investigate these methods cross-sectionally and longitudinally in a regression-based framework, and derive closed-form sandwich variance estimators for key contrasts of interest. We further show that this method accommodates multiple pathways from treatment to outcome, and consider how operationalizing engagement over time can affect these approaches. Our next focus involves direct investigation of engagement as a mediator, suitable for the setting in which we believe key common causes of engagement and the outcome have been measured. The first fundamental goal of this aim is to delineate (and interpret) the mediation effects that are applicable to studies of mobile health interventions under strong access monotonicity, and the second is to formalize the assumptions under which they can be identified. We propose using a parametric g-computation based approach to estimating key effects, and evaluate finite-sample properties through simulation studies. We illustrate the utility of our proposed methods through application to a recent clinical trial of patients with type 2 diabetes that showed significant overall effects on key psychosocial outcome measures. | |
