Modern Longitudinal Mediation Methods for Prevention Studies

点击翻译按钮获取中文摘要

Project Summary Critical to the health of the American public are interventions to encourage health promoting behaviors and to discourage health diminishing behaviors. Designing interventions and evaluating their impact over time requires an understanding of the processes by which interventions bring about behavior change. Statistical mediation analysis is used to reveal these processes and to identify the impact of individual components of the intervention that lead to health promoting behaviors and to detect possible counterproductive components that lead to health diminishing behaviors. Statistical mediation analysis of outcomes that emerge over time following an intervention (e.g., sustained exercise and healthy diet, reduced alcohol consumption) pose special challenges from confounding variables that may produce changes in behavior that are not due to the intervention. Separating the effects of these confounding variables on behavior over time is important for assessing true intervention effects. This proposal addresses a class of confounding variables—time-varying confounders (variables over time that are affected by the intervention and affect both the mediator and the outcome). The proposed research addresses two lines of development in statistical mediation; traditional approaches used in prevention science and potential outcomes approaches used in epidemiology and biostatistics, and evaluates their ability to accurately separate confounding influences from longitudinal effects of treatment interventions. The work will improve statistical methods for evaluating long-term treatment effects of health protective interventions. The proposed research training plan will bridge the gap between methods traditional in prevention science and new methods in epidemiology and biostatistics for estimating longitudinal mediated effects in the following aims: 1) Delineate and compare the statistical and causal inference assumptions across traditional methods and new potential outcomes methods for estimating longitudinal mediated effects with time-varying confounding. 2) Conduct a Monte Carlo simulation study to investigate the statistical properties of longitudinal mediated effect estimators of all methods. 3) Apply these methods to extract more information from three NIH funded datasets to compare longitudinal mediated effect estimates on substance use prevention datasets.