Article type
Year
Abstract
Background: Randomized controlled trials (RCTs) are regarded as the gold standard for the comparison of treatments in clinical trials. Systematic reviews performed by the Cochrane Collaboration focus mainly on results of RCTs. Nevertheless, evidence from non-randomized studies and their inclusion in systematic reviews is a current topic, since in some situations a randomization is not feasible. In the analysis of non-randomized studies, a simple overall comparison of the treatment arms may lead to a biased estimate of the treatment effect due to possible confounding factors (covariates). To cope with this problem, different methods for adjustment of the treatment comparison have been proposed.
Objectives: So-called propensity score based methods have gained popularity in recent years, which may be seen from the growing number of articles containing propensity score based methods. The propensity score, the conditional probability of being assigned to a treatment group given the covariates, is used to stratify patients into homogenous groups with respect to their probability to receive one of two treatments under comparison. This procedure yields pseudo randomization within strata. Thus, a stratified analysis theoretically leads to unbiased estimation of the treatment effect, and such methods seem to be a reasonable alternative to classical methods for covariate adjustment, as e.g. multiple regression analysis.
In practice, different procedures for estimation of the propensity score and for adjustment for the propensity score are used. Most papers deal with rather unclear and uncritical applications for the analysis of non-randomized studies leading to a difficult assessment of results.This calls for a discussion of the fundamental properties and the usefulness of propensity score based methods and for a comparison to classical approaches for covariate adjustment.
Results: Adjustment using the propensity score and classical adjustment methods show some fundamental differences. In the propensity score approach, the relationship of covariates to treatment assignment is essential, whereas in a multiple regression model, the relationship to outcome is crucial. The consequences of these underlying concepts will be discussed and the impact on properties of treatment effect estimates illustrated by examples.
Objectives: So-called propensity score based methods have gained popularity in recent years, which may be seen from the growing number of articles containing propensity score based methods. The propensity score, the conditional probability of being assigned to a treatment group given the covariates, is used to stratify patients into homogenous groups with respect to their probability to receive one of two treatments under comparison. This procedure yields pseudo randomization within strata. Thus, a stratified analysis theoretically leads to unbiased estimation of the treatment effect, and such methods seem to be a reasonable alternative to classical methods for covariate adjustment, as e.g. multiple regression analysis.
In practice, different procedures for estimation of the propensity score and for adjustment for the propensity score are used. Most papers deal with rather unclear and uncritical applications for the analysis of non-randomized studies leading to a difficult assessment of results.This calls for a discussion of the fundamental properties and the usefulness of propensity score based methods and for a comparison to classical approaches for covariate adjustment.
Results: Adjustment using the propensity score and classical adjustment methods show some fundamental differences. In the propensity score approach, the relationship of covariates to treatment assignment is essential, whereas in a multiple regression model, the relationship to outcome is crucial. The consequences of these underlying concepts will be discussed and the impact on properties of treatment effect estimates illustrated by examples.