Thread by @mattdwebb

Hey #EconTwitter, do you use clustered standard errors? Do you worry about reliable inference? Our new guide explains when you should worry, when you shouldn't, and how to make better inferences. A joint 🧵 with @jgmQED and @MortenEcon. 1/

We are very excited that our new Journal of Econometrics overview paper: "Cluster-robust inference: A guide to empirical practice" is now #openaccess at:
doi.org/10.1016/j.jeco… 2/

Our guide connects theory and practice to show what works and when. The methods we recommend in most cases are available in the @Stata packages boottest and summclust. These are easy to use and (usually) not very computationally demanding. 3/

Our paper concludes with a summary guide, which we condense here. For empirical analysis with possibly clustered error terms we recommend the following: 4/

Make an informed decision about how to cluster by considering all the plausible dimensions and levels for the data at hand. Formal tests and placebo regressions can help. A conservative approach is to choose the level with the largest standard errors. 5/

For each plausible level of clustering, report the number of clusters and summary measures of the variability of their sizes. The min, median, mean, and max are particularly informative. 6/

For key regressions, report information about leverage, partial leverage, influence, and the effective number of clusters for the coefficients of interest. When some of these measures are extreme, then conventional cluster robust inference may be unreliable. 7/

In addition to the usual CV1 standard errors, employ CV3 and/or wild cluster bootstrap inference for both tests and confidence intervals. When they differ, do not rely on conventional CV1 inference, and consider using alternative procedures as well. 8/

For models with treatment at the cluster level and few treated clusters, reliable inference is challenging even when based on CV3 or the wild cluster bootstrap. Randomization Inference can offer improved inference in this setting. End🧵. 9/

We're overwhelmed by the number of likes and retweets, thank you very much everyone.