More Sandwiches, Anyone?

Consider this my Good Deed for the Day!

A re-tweet from a colleague whom I follow on Twitter brought an important paper to my attention. I thought I'd share it more widely.

The paper is titled, "Small-sample methods for cluster-robust variance estimation and hypothesis testing in fixed effect models", by James Pustejovski (@jepusto) and Beth Tipton (@stats-tipton). It appears in The Journal of Business and Economic Statistics.  

You can tell right away, from its title, that this paper is going to be a must-read for empirical economists. And note the words, "Small-sample" in the title - that sounds interesting.

 Here's a compilation of Beth's six tweets:

"Econ friends, @jepusto and I have a new paper out that we would love to share. It’s about clustering your standard errors (more below).

‏Any suggestions for how to get these methods out to economists given that we aren’t NBER?

Summary: Our paper provides small-sample adjustments to cluster robust variance estimation (CRVE). It can be used with panel data, experimental data, and regression. You can implement the method in a Stata macro called REG_SANDWICH and an R package called clubSandwich.  

Why do you need this? Regular CRVE doesn’t do so well, even with as many as 100 clusters (!). In fact, CRVE only gives you appropriate Type I error when your covariates are balanced. 

What did we do? We extended the bias-robust linearization method (BRL) by Bell & McCaffrey in three ways: (1) in addition to a t-test, there is now an F-test; (2) We can handle the inclusion of fixed effects; (3) You get the same results whether you use FE or absorption. 

How does it work? The adjustment inflates the standard errors a small bit. But more importantly, it provides Satterthwaite-type degrees of freedom that are more appropriate. The result is a test we call the ‘Approximate Hotelling’s T-squared’ (AHT) test. 

We’d love to share the work broadly, so if you have ideas, please let us know. Thanks!"

I've added the links to the R and Stata software in the quote above. 

There are also some nice R vignettes available:

• Cluster-robust variance estimation with clubSandwich
• Meta-analysis with cluster-robust variance estimation
• Cluster-robust standard errors and hypothesis tests in panel data models

Well, Beth, who can resist a club sandwich? I don't know if this post will help you and James, but I do hope so. 

These new results, and the associated software certainly deserve to be publicized, and used, widely.

Check them out!

(Disclaimer: I do not know either James of Beth, personally or professionally.)

© 2018, David E. Giles