There are over two million admissions to substance abuse treatment programs in the United States each year, and there is a need to ensure that the services offered to these clients are producing positive and meaningful impacts on their lives. To address this need, the government has begun to require that treatment providers receiving funding collect data on their clients at intake, during treatment, and after treatment. A a result of these efforts, researchers now has available an increasing amount of observational data that track individuals from multiple treatment programs and that contain a wealth of information regarding program effectiveness. These data can be used to test the relative effectiveness of many different treatment programs and treatment services, but to fully capitalize on these data and assess the true causal impact of these programs requires appropriate and often cutting-edge statistical modeling tools. One particularly promising tool for estimating causal effects of treatment, developed by our team at RAND, is the Toolkit for Weighting and Analysis of Non- Equivalent Groups (TWANG) package developed in the R statistical computing environment. The TWANG package utilizes a sophisticated weighting technique based on an individual's probability of receiving a treatment given his or her pretreatment variables (i.e., th propensity score) to adjust for imbalances between treatment programs on the observed pretreatment characteristics of their clients, thereby enabling analysts and researchers to draw more robust inferences about the relative effectiveness of two treatment programs on outcomes than traditional methods. TWANG is an exceptional tool for estimating the relative effectiveness of two treatments, but many pressing research questions involve more complex settings, such as comparing multinomial (more than two) treatments and studying the relative effectiveness of time-varying sequences of treatments. Moreover, the R environment in which TWANG is currently available may be unfamiliar to some researchers and analysts, creating a barrier to use of TWANG. This proposal aims to extend the TWANG package to be more versatile and better able to meet the current and future needs of addiction researchers. It also aims to improve dissemination of the package. Specifically, this proposal aims to (1) extend TWANG to estimate propensity scores and assess balance for multinomial and time-varying treatments, (2) develop software to provide access to TWANG via environments other than R (e.g., SAS and Stata), and (3) develop and implement a dissemination strategy to make the updated software available to the addiction research community and to promote its uptake in that community. The contributions from this grant in the short and long term will be to improve both the TWANG package as a health services research tool and the statistical practices of addiction health service researchers. This grant will encourage broader use of modern causal modeling methods through our dissemination efforts. Thus, this grant will not only improve a promising new causal inference tool but more effectively place it directly into the hands of addiction researchers.
The proposed research aims to improve a promising health services research tool (the TWANG package) for estimating causal effects of treatment using data on clients receiving different treatment services or from different treatment providers. In addition, the proposal will enhance the statistical practices of addiction health service researchers through a series of meaningful dissemination efforts which include conducting workshops and webinars on the tool and the causal inference methods it utilizes and developing a website for the tool. In doing so, the proposed research aims to greatly improve the scientific information upon which clinicians, public health officials, and policymakers make decisions about which substance abuse treatment services are most effective and should be recommended and supported for treating people addicted to or abusing alcohol or other drugs.
|Setodji, Claude M; McCaffrey, Daniel F; Burgette, Lane F et al. (2017) The Right Tool for the Job: Choosing Between Covariate-balancing and Generalized Boosted Model Propensity Scores. Epidemiology 28:802-811|
|Griffin, B A; McCaffrey, D; Almirall, D et al. (2017) Chasing balance and other recommendations for improving nonparametric propensity score models. J Causal Inference 5:|
|Parast, Layla; McCaffrey, Daniel F; Burgette, Lane F et al. (2017) Optimizing Variance-Bias Trade-off in the TWANG Package for Estimation of Propensity Scores. Health Serv Outcomes Res Methodol 17:175-197|
|Parast, Layla; Griffin, Beth Ann (2017) Landmark estimation of survival and treatment effects in observational studies. Lifetime Data Anal 23:161-182|
|Ridgeway, Greg; Kovalchik, Stephanie Ann; Griffin, Beth Ann et al. (2015) Propensity Score Analysis with Survey Weighted Data. J Causal Inference 3:237-249|