Bootstrap-Based Improvements for Inference with Clustered Errors
read more
Citations
A Practitioner’s Guide to Cluster-Robust Inference
Robust Inference with Multi-way Clustering
Mostly harmless econometrics
Teacher training, teacher quality and student achievement
Does Management Matter? Evidence from India
References
An introduction to the bootstrap
A Heteroskedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity
Longitudinal data analysis using generalized linear models
Bootstrap Methods: Another Look at the Jackknife
How Much Should We Trust Differences-In-Differences Estimates?
Related Papers (5)
Frequently Asked Questions (19)
Q2. What is the primary contribution of this paper?
The primary contribution of this paper is to use bootstrap procedures to obtain more accurate cluster-robust inference when there are few clusters.
Q3. What is the variation that the authors use to impose the null hypothesis?
The variation the authors use is one that uses equal weights and probability, and uses residuals from OLS estimation that imposes the null hypothesis.
Q4. What is the standard method for resampling that preserves the within-cluster?
The standard method for resampling that preserves the within-cluster features of the error is a pairscluster bootstrap that resamples at the cluster level, so that if the gth cluster is selected then all data (dependent and regressor variables) in that cluster appear in the resample.
Q5. What is the method for resampling?
The data are clustered into G independent groups, so the resampling method should be one that assumes independence across clusters but preserves correlation within clusters.
Q6. What is the obvious method for a regression model with additive error?
The obvious method is a residual cluster bootstrap that resamples with re-placement from the original sample residual vectors to give residuals {bu∗1, ..., bu∗G} and hence pseudo-sample {(by∗1,X1), ..., (by∗G,Xg)} where by∗g = X0gbβ + bu∗g.
Q7. What is the practical limitation of cluster-robust standard errors?
A practical limitation of inference with cluster-robust standard errors is that the asymptotic justification assumes that the number of clusters goes to infinity.
Q8. What is the significance of the BDM (2004) study?
One important conclusion of BDM (2004) is that for few (six) clusters the cluster-robust estimator performs poorly, and for moderate (ten and twenty) number of clusters theirbootstrap based method also does poorly.
Q9. What is the way to resample the residuals?
In particular, one can hold regressors X constant throughout the pseudo-samples, while resampling the residuals which can be then used to construct new values of the dependent variable y.
Q10. What is the common correction for clustering?
A common correction is to compute cluster-robust standard errors that generalize the White (1980) heteroskedastic-consistent estimate of OLS standard errors to the clustered setting.
Q11. How does the BCA method perform for low numbers of clusters?
They find that (1) default standard errors do poorly; (2) cluster-robust standard errors do well for all but G = 6; and (3) their bootstrap, which the authors discuss in their section 3.1, does poorly for low numbers of clusters, with actual rejection rates 0.44, 0.23 and 0.13 for G = 6, 10 and 20, respectively.
Q12. What are the different methods used in the BCA bootstrap resampling?
The authors use three different cluster bootstrap resampling methods, respectively, the pairs cluster bootstrap, the residual clusters bootstrap with H0 imposed, and the wild bootstrap with H0 imposed.
Q13. What are the other methods that yield the lowest rejection rates?
The remaining bootstrap-t methods all yield rejection rates less than 0.08, with the residual cluster bootstrap-t and wild cluster bootstrap-t doing best.
Q14. What is the p-value for the pairs cluster bootstrap-t?
If the authors instead bootstrap this Wald statistic with B = 999 replications, the pairs cluster bootstrap-t yields p = 0.209, the residual cluster bootstrap-tgives p = 0.112, and the wild cluster bootstrap-t gives a p-value of 0.070.12
Q15. What is the p-value for the pairs cluster bootstrap?
The authors believe that the p-value for the pairs cluster bootstrap is implausibly large, for reasons discussed in the BDM replication, while the other two bootstraps lead to plausible p-values that, as expected, are larger than those obtained by using asymptotic normal critical values.
Q16. What does the paper propose for DID studies with policy invariant within state?
Donald and Lang (2007) also demonstrate this and propose, for DID studies with policy invariant within state, an alternative two-step GLS estimator that leads to Tdistributed Wald tests in some special circumstances.
Q17. What is the alternative method with asymptotic refinement?
An alternative method with asymptotic refinement is the bias-corrected accelerated (BCA) procedure, defined in Efron (1987), Hall (1992, pp. 128- 141), and in (Cameron, Gelbach, and Miller, 2006).
Q18. What is the procedure that bootstraps w?
The bootstrap-t procedure directly bootstraps w which is asymptotically pivotal since the standard normal has no unknown parameters.
Q19. What are the other options for a bootstrap?
in contrast to the bootstrap-t procedure, it does not offer asymptotic refinement, and so may perform worse 6Alternative names used in the literature include cluster bootstrap, case bootstrap, nonparametric bootstrap, and nonoverlapping block bootstrap.