Modeling Causal Effects of Components of Bundled Interventions with Application to a Multilevel Dental Caries Clinical Trial

This proposal seeks to develop new statistical methods applicable to studies of bundled interventions. Randomized clinical trials, including in dental research, often focus on multi-component and other complex interventions. Using such ‘bundled’ interventions is appealing as a way to increase power – as well as simplicity - of the study. An apparent disadvantage is that it is not clear how to assess the effects of individual components of bundled interventions, which of also of frequent interest. While the measurement of treatment compliance, and use of causal mediation analysis is commonly recognized as a possible approach, rigorous methods to identify and estimate causal effects of components are not available. The present research seeks to fill this important gap. We will first elucidate the assumptions under which causal mediation/path analysis can be used to determine causal effects of individual intervention components. We propose, as a novel and relevant estimand, what we refer to as a cluster-specific interventional effect. We will develop an extended mediation formula/simulation approach to estimating these causal effects. In our second aim, we will extend methods to handle repeatedly measured mediators and outcomes. As a novel aspect of this aim, new methods will be developed to analyze summary (or cumulative) measures in a way that respects the causal order of model variables. We will perform simulation studies to evaluate the properties of the new methods, and compare them to possible alternative approaches. In addition, we will develop sensitivity analysis methods to examine the impact of violations of model assumptions, including extended sequential ignorability as well as structural (e.g., no direct effect) assumptions. In particular, we extend a copula model approach, previously developed for the single mediator case, to perform sensitivity analyses in the context of more complex path models. We will develop an R package to allow user-friendly implementation of the new methods. The new methods will be applied to data from a recently completed cluster-randomized clinical trial of a multi-component intervention (including multiple provider-level components) to improve dental care utilization among 3 to 6 year old Medicaid-enrolled children attending well-child visits in primary care settings. Our analysis will assess the causal effects of individual components of this bundled intervention.