지난 6월 4일 좋은 강의해주셔서 감사드립니다.
몇 가지 관련된 질문을 드릴 것이 있어 질문게시판에 글을 올립니다.
저는 지금까지 propensity score matching을 SPSS 또는 R로 1:1~1:N 매칭을 해왔었는데, 보통 age, sex, BMI를 PS matching에 넣은 뒤 다른 변수들은 회귀 분석 모형에서 보정하며 모델 1, 2, 3 식으로 제시했었는데요, 제가 고려하는 모든 변수를 PS matching에 넣고 회귀분석에서 추가적으로 모델을 보여주지 않는 것이 보편적인지요? 그리고 web-r에서 PS matching을 한 뒤에 매칭된 결과가 담긴 엑셀 파일을 다운 받을 수 있을지요?
너무나 좋은 강의에 다시 한 번 감사드립니다.
또한 매칭된 결과는 매칭 후 download file 버튼을 누르시면 됩니다.
이 글을 참조하세요. https://cran.r-project.org/web/packages/MatchIt/vignettes/estimating-effects.html
With covariate adjustment. Including covariates in the outcome model is straightforward with a continuous outcome. We can include main effects for each variable and use the coefficient on the treatment as the treatment effect estimate. It can also be helpful to include interactions between the covariates and treatment if effect modification is suspected, though it is important to center the covariates at their means in the focal (i.e., treated) group if the coefficient on the treatment is to be interpreted as an average treatment effect estimate (which we demonstrate with full matching). A marginal effects procedure can also be used, which we demonstrate with binary outcomes. Below we simply include main effects of each covariate, which is the most typical way to include covariates.
As previously mentioned, it is important to avoid interpreting coefficients in the outcome model on predictors other than the treatment, as the coefficients do no correspond to causal effects and are likely confounded. In the code above, we restricted the output to just the treatment effect. Note that sometimes the intercept of the outcome model can be useful as an estimate of the average potential outcome under control, but the covariates in the model (if any) must be centered at their means in the target population to allow for this interpretation. Without centering the covariates, the intercept is uninterpretable.