Which covariates should be used to match patients in Propensity Score Matching?

When using propensity score matching (PSM), it's important to choose covariates that are both predictive of treatment assignment and related to the outcome of interest. This will help to ensure that the individuals in the treatment and control groups are as similar as possible and that any differences in outcomes can be attributed more confidently to the treatment.

Here are some general guidelines for choosing covariates for PSM:

1. Include all covariates that are related to both the treatment assignment and the outcome of interest: These covariates are likely to be important confounders and should be included in the propensity score model. Examples might include age, sex, pre-existing medical conditions, and socioeconomic status.
   
   
2. Include all covariates that are strongly associated with the treatment assignment: Even if these covariates are not directly related to the outcome of interest, they should still be included in the propensity score model, because they can help to ensure that the treatment and control groups are as similar as possible.
   
   
3. Include any covariates that are important for the interpretation of the results: Sometimes, including certain covariates in the propensity score model can be important for interpreting the results. For example, if the study is comparing a treatment that is only available to a certain subgroup of the population, such as a certain age or sex, then it may be important to include this subgroup as a covariate in the propensity score model.
   
   
4. It's also important to choose the right level of measurement for the covariates to use in PSM. For example, using a continuous variable when a categorical variable would be more appropriate can lead to poor estimates of the propensity score.

It's important to note that these are general guidelines and the appropriate covariates for PSM will depend on the specific research question and data. Additionally, researchers should consider the possibility of unmeasured confounding variables and should interpret the results with caution.