RESEARCH METHODS CORE

RMC提出的工作将在对干预至关重要的领域推进统计方法

The work proposed by the RMC will advance statistical methodology in areas crucial to intervention studies. It is important to carefully design intervention studies to facilitate learning about the effectiveness of the interventions under study. With appropriate and careful design, more robust conclusions can be made. In the RMC, we propose to bring together researchers who are addressing a number of methodological issues critical to the evaluation of the Center's proposed pilot intervention and assessment initiatives and the RO1 supported effectiveness trials that will evolve out of these pilot initiatives. Study attrition plagues many studies as it becomes more and more difficult to follow up all study subjects; new study designs are needed to reduce the effects of attrition on study results. Long-term followup is often necessary to determine the long-term effects of preventive interventions, such as the Good Behavior Game and PATHS+GBG interventions studied by the JHU PIRC. But this long-term follow-up leads to challenges in dealing with study attrition. Most statistical work developing methods to deal with attrition have focused on statistical analyses, for example weighting or imputation methods to adjust for the missing data (Little & Rubin, 2002; Groves et al., 2004). However, in some cases better study design and careful selection of subjects to follow-up can reduce the need for complex modeling assumptions at the analysis stage (Brown et al., 2000; Graham et al., 2001). However, to fully understand the benefits of these designs, further methodological work is needed. The importance of economic impacts as outcomes of early prevention programs, and the length of time required to observe them, also poses special challenges for economic assessments of these programs (Aos et al., 2004; Kellam and Langevin, 2003). While some longitudinal studies have tracked treatment and control subjects from early interventions overextended periods (Barnett, 1996; Maase & Barnett, 2003), doing such longterm follow-up often is difficult because of the costs involved in obtaining high response rates over an extended period of time. Long-term follow-ups also present challenges to analysis because of factors such as non-random sample attrition One potential solution is to use multiple-stage predictive models to infer impacts of early preventive interventions on distal economic outcomes, using information on more proximal outcomes, and the relationship between the proximal and distal outcomes. This has the potential to allow lower cost predictions of long-term effects of early preventive interventions. However, more work is needed to fully develop the methods and determine when the distal predictions would be appropriate. It is important to detect variation in intervention response that is mediated by post-randomization variables. Intervention research at JHU (e.g., lalongo et al., 1999) and elsewhere (e.g., Reid et al., 1999) suggests that variation in impact is found almost as frequently as significant main effects (Brown & Liao, 1999). An improved understanding of sub-group variation in intervention response and the factors contributing to it would facilitate the design of preventive and early interventions that more precisely target those youth who fail to benefit from existing interventions. The failure of intervention researchers to address issues related to variations in outcomes stems in part from limitations in our statistical procedures for examining subgroup variation. Improved analytic strategies and wider dissemination of these strategies are needed if we are to understand sub-group variation and the factors contributing to it. Previous work by members of the RMC has investigated in detail how to detect subgroup variation in intervention response that is governed by post-randomization variables (Jo, 2002a-c). This work builds on the framework of principal stratification set out by Frangakis & Rubin (2002). Further work is needed to consider settings where the post-treatment mediators are themselves measured longitudinally, such as compliance behavior over time. The work by members of the RMC will extend their previous work in this area in this important direction. Policymakers need ways of determining whether the results seen in randomized trial samples are likely to generalize to target populations, which may be somewhat different from the trial sample. Even effectiveness trials rarely are done using subjects that are fully representative of the target populations in which the interventions being evaluated may eventually be implemented (Rothwell, 2005). Statistical methods to assess the generalizability of results from effectiveness trials to those target populations are needed, as highlighted in recent government reports (National Institute of Mental Health 1999; Institute of Medicine 2006). Work proposed in this RMC will build on research being done by members of the RMC (Frangakis & Rubin, 2002; Stuart 2007b) to develop such methods, bridging internal and external validity. Complementary work will extend the "target efficiency" methods developed by members of the RMC (Salkever et al., 2008), which consider the optimal targeting of preventive interventions so that they reach those individuals whom they will most benefit. These efforts will guide the design and implementation of research conducted by the Center's investigators.