Impact of the treatment assignment model on propensity score-based methodsSrivastava, Avantika (2019) Impact of the treatment assignment model on propensity score-based methods. Master's Thesis, University of Pittsburgh. (Unpublished)
AbstractStandard statistical methods assess the association between treatments or exposures and the outcome. To better estimate the causal effect, defined as the expected difference in potential outcomes, several methods have been proposed, including propensity-score (PS) based methods. PS-based methods use the predicted probability of being assigned to a given treatment to produce a pseudo population which better resembles that of a randomized trial. Standard statistical methods can then be applied to that pseudo population to better estimate the causal effect. The goal of this thesis is to describe how choice of treatment assignment model can impact the results of PS-based methods, using an empirical example of a binary outcome with a binary treatment variable and both continuous and categorical confounders. We hypothesize that using different treatment assignment mechanisms will produce differing degrees of overlap (between treatment groups), which will then lead to different pseudo populations and different estimates of treatment effect. The treatment assignment mechanism was modeled using three approaches: logistic regression, classification trees, and a random forest model. Next, three pseudo populations were created from each of the resulting PS distributions: one using 1:1 propensity score matching, one using stratification into quintiles, and one using inverse probability of treatment weighting (IPTW). Covariate balance was assessed by calculating the standardized mean differences of covariates in the entire sample and in each pseudo population. A Cox proportional hazard model was then fit for each pseudo population to estimate the treatment effect. Results varied for different outcomes models. The forest model gave a significant estimate in matched pseudo populations, no model was significant in stratified pseudo populations, and both the unpruned tree and forest models gave significant estimates in IPTW pseudo populations. In conclusion, these mixed results indicated that the assignment mechanism model, the approach for forming the pseudo population, and the choice of outcomes model, can all significantly influence results. This thesis is significant to public health because it illustrates a comparative effectiveness research analysis of the causal effect of two treatments using only observational data. The methods used are a frequent approach in public health studies involving nonrandomized data. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}