Introduction to propensity scores

Abstract

Although randomization provides a gold-standard method of assessing causal relationships, it is not always possible to randomly allocate exposures. Where exposures are not randomized, estimating exposure effects is complicated by confounding. The traditional approach to dealing with confounding is to adjust for measured confounding variables within a regression model for the outcome variable. An alternative approach--propensity scoring--instead fits a regression model to the exposure variable. For a binary exposure, the propensity score is the probability of being exposed, given the measured confounders. These scores can be estimated from the data, for example by fitting a logistic regression model for the exposure including the confounders as explanatory variables and obtaining the estimated propensity scores from the predicted exposure probabilities from this model. These estimated propensity scores can then be used in various ways-matching, stratification, covariate-adjustment or inverse-probability weighting-to obtain estimates of the exposure effect. In this paper, we provide an introduction to propensity score methodology and review its use within respiratory health research. We illustrate propensity score methods by investigating the research question: 'Does personal smoking affect the risk of subsequent asthma?' using data taken from the Tasmanian Longitudinal Health Study.

Keywords: causal inference; confounding; environmental and occupational health and epidemiology; observational studies; statistics.