We all know that, at least, we should be clustering at the level of treatment, but we also know that many times during seminars, we are asked to cluster at different levels. One potential take-away people might get from this is that we should cluster in all of the ways. The paper by THE econometricians said "not so fast." (See my previous post about this issue).
And now for more on clustering, the Declare Design folks explain when we should cluster above the level of treatment. I learned of this on David Mackenzie's blog, and there you can see his thoughts on clustering as well.
To my applied micro students (who are not doing RCTs), here is a reminder: These posts are all written using RCT language. You may think none of this applies to you. It does! It is true that we as researchers don't randomly put people into treatment and control groups. But, by using our tricks (RD, differences in differences, IV, etc.), we are assuming that the world is putting people into treatment and control groups for us. This is so important to remember as we think about our own work and evaluate the work of others. Conclusion: (almost?) everything the "randomistas" say about clustering applies to us, too.
No comments:
Post a Comment