Propensity score weighting with multilevel data

Abstract

Propensity score methods are being increasingly used as a less parametric alternative to traditional regression to balance observed differences across groups in both descriptive and causal comparisons. Data collected in many disciplines often have analytically relevant multilevel or clustered structure. The propensity score, however, was developed and has been used primarily with unstructured data. We present and compare several propensity-score-weighted estimators for clustered data, including marginal, cluster-weighted and doubly-robust estimators. Using both analytical derivations and Monte Carlo simulations, we illustrate bias arising when the usual assumptions of propensity score analysis do not hold for multilevel data. We show that exploiting the multilevel structure, either parametrically or nonparametrically, in at least one stage of the propensity score analysis can greatly reduce these biases. These methods are applied to a study of racial disparities in breast cancer screening among beneficiaries in Medicare health plans.

Frequency and Outcomes of Reduced Dose Non–Vitamin K Antagonist Anticoagulants: Results From ORBIT‐AF II (The Outcomes Registry for Better Informed Treatment of Atrial Fibrillation II)

“…Only these 2 drugs were considered because of limited power in patients receiving other agents. Overlap propensity weighting was performed for adjustment of these outcomes 14. Overlap weights produce exact covariate balance and greater precision by more heavily weighting those with a reasonable chance of receiving either treatment compared with those with extreme propensities who are very likely to receive a particular treatment.…”

Long-term outcomes of mitral valve annuloplasty versus subvalvular sparing replacement for severe ischemic mitral regurgitation

“…To reduce the impact of treatment selection bias and potential confounding in the observational study, rigorous adjustment for baseline differences by use of propensity score matching was performed [10]. A PS representing the probability of having subvalvular sparing MVR as opposed to MVA was calculated for each patient by using a non-parsimonious multivariable logistic regression model.…”

A comparison of marginal odds ratio estimators

“…Due to the lack of potential outcomes, these estimates can no longer be considered causal, but are still useful as they allow comparisons among entire subpopulations while controlling for imbalance in relevant covariates. For a more detailed explanation and worked example, the reader is referred to Li et al 12 …”