1: \begin{abstract}

2: For linear regression models with cross\tkk-section or panel data, it

3: is natural to assume that the disturbances are clustered in two

4: dimensions. However, the finite\tkk-sample properties of two\tkk-way

5: cluster-robust tests and confidence intervals are often poor. We

6: discuss several ways to improve inference with two\tkk-way clustering.

7: Two of these are existing methods for avoiding, or at least

8: ameliorating, the problem of undefined standard errors when a

9: cluster-robust variance matrix estimator (CRVE) is not positive

10: definite. One is a new method that always avoids the problem. More

11: importantly, we propose a family of new two\tkk-way CRVEs based on the

12: cluster jackknife. Simulations for models with two\tkk-way fixed

13: effects suggest that, in many cases, the cluster-jackknife CRVE

14: combined with our new method yields surprisingly accurate inferences.

15: We provide a simple software package, \texttt{twowayjack} for 

16: \texttt{Stata}, that implements our recommended variance estimator.

17: 

18: \vskip 12pt

19: 

20: \medskip \noindent \textbf{Keywords:} cluster jackknife, cluster sizes,

21: clustered data, cluster-robust variance estimator, CRVE, grouped data,

22: two\tkk-way fixed effects.

23: 

24: \medskip \noindent \textbf{JEL Codes:} C10, C12, C21, C23.

25: 

26: \end{abstract}

27: