Account for two levels of robust standard errors in school data

Question

I am relatively new to Statistics and R, so please forgive my naive question. I have clustered school data and would like to employ matching followed by a regression. In the paper I read ("Matching Methods for Clustered Observationsl Studies in Education" by Keele et. al) it was recommended to account for both with-in school correlations and pair-wise correlations (from matching) with robust standard errors or random effects.

As I am not familiar with random effects and only know how to use robust standard errors on one level, how would I use robust standard errors for both levels (with-in school correlations and pair-wise correlations) in R?

Answer 1

In R this is simple. Using the sandwich package, which is the package used for robust standard errors, you can simply run the following:

library(sandwich); library(lmtest)
fit <- lm(Y ~ treat, data = matched_data)
coeftest(fit, vcov. = vcovCL, cluster = ~subclass + school)

The assumes matched_data contains the results of your matching and has a column called subclass containing pair membership (e.g., as the output of a call to match.data() after MatchIt::matchit()) and school the name of the school variable. This performs multi-way clustering, accounting for both types of clustering, which is explained in the vignette for the sandwich package.