Advanced statistics: The propensity score - A method for estimating treatment effect in observational research

Abstract

Observational studies assessing the effect of a particular treatment or exposure may be subject to bias, which can be difficult to eliminate using standard analytic techniques. Multivariable models are commonly used in observational research to assess the relationship between a certain exposure or treatment and an outcome, while adjusting for important variables necessary to ensure comparability between the groups. Large differences in the observed covariates between two study groups may exist in observational studies in which the investigator has no control over who was allocated to each treatment group, and these differences may lead to biased estimates of treatment effect. When there are large differences in important prognostic characteristics between the treatment groups, adjusting for these differences with conventional multivariable techniques may not adequately balance the groups, and the remaining bias may limit valid causal inference. Use of a propensity score, described as a conditional probability that a subject will be "treated" based on an observed group of covariates, may better adjust covariates between the groups and reduce bias. The purpose of this article is to describe the use of propensity scores to adjust for bias when estimating treatment effects in observational research and to compare use of this technique with conventional multivariable regression. The authors present three methods for integrating propensity scores into observational analyses using a database collected on head-injured trauma patients. The article details the methods for creating a propensity score, analyzing data with the score, and explores differences between propensity score methods and conventional multivariable methods, including potential benefits and limitations. Graphical representations of the analyses are provided as well.