Propensity Score Estimation With Boosted Regression for Evaluating Causal Effects in Observational Studies.

Abstract

Causal effect modeling with naturalistic rather than experimental data is challenging. In observational studies participants in different treatment conditions may also differ on pretreatment characteristics that influence outcomes. Propensity score methods can theoretically eliminate these confounds for all observed covariates, but accurate estimation of propensity scores is impeded by large numbers of covariates, uncertain functional forms for their associations with treatment selection, and other problems. This article demonstrates that boosting, a modern statistical technique, can overcome many of these obstacles. The authors illustrate this approach with a study of adolescent probationers in substance abuse treatment programs. Propensity score weights estimated using boosting eliminate most pretreatment group differences and substantially alter the apparent relative effects of adolescent substance abuse treatment.

Evaluating the Effectiveness of Correctional Education: A Meta-Analysis of Programs That Provide Education to Incarcerated Adults

“…Some studies achieve this through matching or weighting the comparison group so that it is similar to the treatment group on a number of possibly confounding characteristics. When the number of characteristics to be used in the weighting or matching is large, balance can sometimes be achieved by using these characteristics to estimate the probability of receiving the treatment and matching treated to comparison cases based on these fitted probabilities, or propensity scores (McCaffrey, Ridgeway, and Morral, 2004;Rosenbaum and Rubin, 1983;Rubin, 1997). Matching or weighting on observed characteristics helps ensure that the observed characteristics are not responsible for any apparent treatment effects, but it leaves open the possibility that unmeasured differences may be driving such effects.…”

The Air Force Deployment Transition Center: Assessment of Program Structure, Process, and Outcomes

“…The control sample is then weighted by (p/1 -p), where p is the estimated propensity score (i.e., the predicted probability of being in the treatment group conditioned on the available information). This weighting method results in substantially better balance across groups than can be achieved with logistic regression or other common methods (Lee, Lessler, and Stuart, 2010;McCaffrey, Ridgeway, and Morral, 2004;Ridgeway and McCaffrey, 2007). The TWANG algorithm continues to add iterations that modify the weights (i.e., that increase the complexity of the model producing the propensity score) until the resulting weights show no further improvements in the balance across groups on the full set of background variables relative to the previous iteration.…”

“…The use of the TWANG algorithm to produce the weights allows researchers to balance across more variables (including collinear variables) and generally produces better balance than logistic regression (Lee, Lessler and Stuart, 2010;McCaffrey, Ridgeway, and Morral, 2004;Ridgeway and McCaffrey, 2007). This is largely attributable to the fact that the TWANG algorithm uses an iterative procedure that directly optimizes the weighted covariate balance between groups.…”

Secular Changes in Late-Life Cognition and Well-Being: Towards a Long Bright Future With a Short Brisk Ending?

“…To minimize possible confounds and equate the cohort samples as closely as possible on age, gender, and education, we used propensity score-matching procedures (Coffman, 2011;Foster, 2010;McCaffrey, Ridgeway, & Morral, 2004;Thoemmes & Kim, 2011). Calculating a logistic regression, we used 1:1 matching methods to select for each participant from the BASE cohort (n ϭ 447) a "twin" participant from the BASE-II cohort (n ϭ 708) who was the same age (or as similar as possible) at baseline, same gender, and same cohort-normed education.…”