|
Abstract:
|
n recent years there has been an increase in the use of methodologies involving propensity scores to minimize selection bias in observational studies aimed at identifying causal relationships. Much of the research using the propensity score approach assumes that the observations are independent. This underlying assumption, however, can be rather unreasonable when repeated measures of the outcome are collected in longitudinal studies. More recently, propensity score methods have been applied considering multilevel settings, particularly when clustered designs were applied. Two aspects of using propensity scores within a multilevel modeling framework are the quality of the balance achieved and the accuracy and precision of the treatment effect estimates. In this paper, we review methodologies that propose the use of random effects models for the definition of propensity scores and/or for modeling the outcome, as well as methods for dealing with time-varying interventions or confounders, such as the marginal structural models. We analyze two datasets to illustrate the common challenges we face when adopting these approaches for analysis of longitudinal data.
|
