A Model Averaging Approach to Causal Inference in Substance Abuse Prevention Research Many evaluations of school-based preventions for alcohol and other drugs (AOD) use are either observational by design or by implementation given noncompliance and dropouts. The observational nature of prevention studies is a major challenge to researchers trying to understand an intervention's effectiveness because of the serious threats of selection and confounding biases (i.e., individuals who receive more of the intervention are often very different from those who receive less). This application proposes a three-year R01 study to develop a novel causal inference approach using model averaging that will provide a more robust solution than current approaches to this major methodological problem in prevention research. The Rubin Causal Model (RCM) is a general framework for causal inference with studies in which randomization is not possible or is compromised by implementation difficulties. While classic statistical techniques can be severely biased when the distribution of confounding variables differ between treated and control individuals, the RCM can reduce such biases from effectiveness estimates. NIDA has made the continued development of methods under the RCM framework a high research priority. Currently, a major difficulty for practitioners is to choose among the numerous RCM available approaches. A preliminary review suggests that more than 40 distinct RCM approaches have been proposed. Further, numerical and empirical studies show that the conclusions across methods can be highly variable and that many distinct approaches have been recommended by different authors. Thus, the most recommendable RCM approach for a specific application is often uncertain. To address this challenge, we propose to develop a novel model averaging approach to causal inference. When there are many candidate estimators, the optimal model averaging estimator has been shown to offer the best statistical efficiency among all candidate estimators and eliminate sensitivity from model choice. Despite their key advantages, model averaging methods for causal effects have not been thoroughly investigated in the literature to tackle the issue of choosing an RCM approach. We propose to develop causal inference model averaging methodology and develop a software tool to implement the new method. We will evaluate practical advantages of the method in numerical studies and in an application study evaluating the effectiveness of CHOICE, a prominent school-based prevention for AOD use.
A Model Averaging Approach to Causal Inference in Substance Abuse Prevention Research Lack of randomization, non-compliance, and dropout are major challenges common to school-based preventions of alcohol and other drug (AOD) use in adolescents. While causal inference methods are viable in reducing the biases from these issues, there is notable sensitivity in estimates among numerous available causal inference methods. This study will develop a novel approach to causal inference using model averaging that will provide a more robust approach to address these methodological challenges when estimating intervention effectiveness. We will apply the new method to evaluate a school-based prevention (CHOICE).
| Antonelli, Joseph; Cefalu, Matthew; Palmer, Nathan et al. (2018) Doubly robust matching estimators for high dimensional confounding adjustment. Biometrics : |
| Antonelli, Joseph; Han, Bing; Cefalu, Matthew (2017) A synthetic estimator for the efficacy of clinical trials with all-or-nothing compliance. Stat Med 36:4604-4615 |
