Date of Award

January 2012

Document Type

Open Access Thesis

Degree Name

Master of Public Health (MPH)


School of Public Health

First Advisor

Peter N. Peduzzi

Second Advisor

Haiqun Lin


We always strive to minimize the impact of bias in observational studies due to possible nonrandom treatment assignment. Propensity score and inverse weighting methods both attempt to achieve this goal. Inverse probability weighting is the method based on Horvitz and Thompson (1952) while propensity score is based on Rosenbaum and Rubin (1983). Because they are the most prevalent methods in longitudinal studies, these methods should be evaluated to find out which is better in reducing bias and producing accurate estimates. However, there are few studies comparing the two approaches. In a study of theory and simulated data, Ertefaie and Stephens (2010) demonstrated that, in simple cases, multivariate generalized propensity score (MGPS) routinely produced estimators with lower Mean-Square Error (MSE) when compared to inverse probability weighting (IPW). In the same paper, however, they were unable to show the same result in a longitudinal dataset. In this paper, I will perform similar comparisons in the treatment effect hazard ratio estimates as well as the efficiency of the estimates, specifically the variance of the two methods in an observational longitudinal public health study. I will only compare the direct effect of treatment, or the unconfounded and unmediated effect on expected response, since this is the only place where Propensity score and Inverse Weighting methods are comparable, and demonstrate that PS may not be the best method of analysis for reducing bias in longitudinal time-to-event studies, despite theoretical studies to the contrary. The results show that the treatment effect hazard ratio estimates with the two approaches are indistinguishable, although PS is consistently efficient while IPW varies based on whether stabilization occurs and on covariates.


This is an Open Access Thesis.