Article Title

Application of Propensity Score Models in Observational Studies Using VDW Data

Publication Date



propensity score, observational study methods


Background/Aims: Treatment effects from observational studies may be biased since the patients were not randomly allocated to a treated or untreated group. Propensity score methods are increasingly being used to address this bias. The propensity score is the probability of treatment assignment conditional on observed baseline characteristics. After appropriately adjusting for the propensity score, the distribution of observed baseline covariates will be similar between treated and untreated patients.

Methods: The estimation of propensity scores and how they can help in adjusting the treatment effect for differences in baseline imbalances are reviewed. Variable selection for the propensity score model, balancing the propensity score, and suggestions on what to do when the propensity score does not balance also are reviewed. The strengths and limitations of each of the different methods of propensity score adjustment are presented (e.g. matching, covariate adjustment, stratification, inverse probability of treatment weighted [IPTW], stabilized IPTW) as well as recommendations for different sensitivity analyses to look at the magnitude of hidden biases. Finally, suggestions are provided on what to present in the report of the final analysis.

Results: Propensity score methods are illustrated by estimating the effect of angiotensin-converting enzyme inhibitor (ACEI) and angiotensin receptor blocker (ARB) use in stage IIIb/IV non-small cell lung cancer (a project funded by Kaiser Permanente’s Center for Effectiveness & Safety Research). All steps of the propensity score methods used to analyze the effect of ACEI and/or ARB antihypertension treatment on survival in advanced stage non-small cell lung cancer patients are described.

Conclusion: Although propensity scores cannot control for unobserved or unmeasured confounding, propensity score analyses can address residual confounding by simulating randomized control trials and are a tool one can use when comparing two treatment groups.