2013 Volume 24 Issue 2 Pages 73-85
Propensity score (PS) is increasingly used to establish causal relationships in observational data. PS is a valuable tool because health economics research such as economic evaluation of health guidance for patients with metabolic syndrome must rely on observational data such as health insurance claims. PS is the probability that a case will be assigned to treatment. If cases from treated and untreated groups of the same PS are compared, then the comparison becomes a quasi-randomized trial and hence establishes a causal relationship. Covariates must be those preceding treatment and outcome. Health check data are appropriate as covariates for PS calculation because health checks precede health guidance and health insurance claims. However, PS requires a certain condition of "strongly ignorable" treatment assignment for it to be valid. Selecting covariates that best fulfill the strongly ignorable assignment constitutes the most difficult task.
The Hosmer-Lemeshow goodness-of-fit test result, as well as the discriminatory power measured by the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, serve as the two main indicators for evaluating the appropriateness of PS calculation. Both indicators should be close to 1 but if AUC is too close to 1, there may not be enough overlapping of PS to secure a good sample size. After PS is properly calculated, analyses will be conducted by 1) matching, 2) subclassification and 3) covariance analysis. One-to-one matching is the first choice but it may reduce the sample size because not all cases can be matched. Subclassification can use all the cases but interpretation may be difficult when the results are not consistent across subclasses. Covariance analysis simply uses PS as an additional explanatory variable in the regression model. A Tobit regression model should be employed when the outcome is health care cost, which will be a zero-truncated value. Although PS programs are readily available, the most important task, selecting covariates, is left to the researcher's personal skills and knowledge. Therefore, researchers must develop such skills and knowledge to use PS appropriately. In observational studies, selection of control groups is crucial. Results may be manipulated when control groups are arbitrarily selected. Hence, PS requires high ethical standards and morals among researchers. To ensure research integrity, it would be better to have different researchers perform the PS calculation and outcome measurement tasks. It is also recommended that multiple researchers analyze the same datasets and combine them using systematic review.
