Robust Inference with Multi-way Clustering
read more
Citations
A Practitioner’s Guide to Cluster-Robust Inference
Bootstrap-Based Improvements for Inference with Clustered Errors
Why Do U.S. Firms Hold So Much More Cash than They Used To
Credit Spreads and Business Cycle Fluctuations
On making causal claims: A review and recommendations
References
Econometric Analysis of Cross Section and Panel Data
A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity
Longitudinal data analysis using generalized linear models
How Much Should We Trust Differences-In-Differences Estimates?
Microeconometrics: Methods and Applications
Related Papers (5)
Frequently Asked Questions (11)
Q2. What are some examples of nonlinear estimators to which this method can be applied?
Commonly-used examples of nonlinear estimators to which this method can be applied are nonlinear-least squares, just-identified instrumental variables estimation, logit, probit and Poisson.
Q3. What is the Wald test for iid errors?
The Wald test based on assuming iid errors is exactly T distributed with (GH − 3) degrees of freedom under the current dgp, so that even in the smallest design with G = H = 10 the theoretical rejection rate is 5.3% (since Pr [|t| > 1.96|t ∼ T (97)] = 0.053), still quite close to 5%.
Q4. What is the way to reduce the rejection rate of the random effects model?
One possibility is to adapt the random effects model to allow dampening serial correlation in the error, similar to the dgp used by Kezdi (2004) and Hansen (2005) in studying one-way clustering, with addition of a common shock.
Q5. What is the reason for the rejection rate of the Wald test statistic?
Then rejection rates may exceed 5%, as even with a Gaussian dgp, the Wald test statistic has a distribution fatter than the standard normal, due to the need to estimate the unknown error variance (even if the standard error estimate is unbiased).
Q6. What is the 95% confidence interval for the Wald test?
For methods 1-3 with larger designs, specifically G ×H > 1600, the authors use only 1,000 simulations due to computational cost; the 95% confidence interval is (3.6%, 6.4%).
Q7. What is the effect of multi-way clustering?
7In a variety of Monte Carlo experiments and replications, the authors find that accounting for multi-way clustering can have important quantitative impacts on the estimated standard errors.
Q8. What is the problem with the two-way robust estimatoris?
A practical matter that can arise when implementing the two-way robust estimatoris that the resulting variance estimate bV[bβ] may have negative elements on the diagonal.
Q9. What is the general approach of Bertrand et al. (2004) in investigating a?
The second follows the general approach of Bertrand et al. (2004) in investigating a placebo law in an earnings regression, except that in their example the induced error dependence is two-way (over both states and years) rather than one-way.
Q10. How much is the maximum possible increase in standard errors due to error correlation at the household level?
The maximum possible increase in standard errors due to error correlation at the household level is about forty percent (corresponding to a doubling of the variance estimate: √ 2 = 1.41).
Q11. What is the simplest way to calculate the cluster-robust standard errors?
The N × N selection matrix SGH may be large in some problems, however, and even if N is manageable many users will prefer to use readily available software that calculates cluster-robust standard errors for one-way clustering.