Refined Cluster Robust Inference

Abstract: It has become standard for empirical studies to conduct inference robust to cluster dependence and heterogeneity. With a small number of clusters, the normal approximation for the $t$-statistics of regression coefficients may be poor. This paper tackles this problem using a critical value based on the conditional Cramér-Edgeworth expansion for the $t$-statistics. Our approach guarantees third-order refinement, regardless of whether a regressor is discrete or not, and, unlike the cluster pairs bootstrap, avoids resampling data. Simulations show that our proposal can make a difference in size control with as few as 10 clusters.

• The paper introduces an analytic correction to cluster-robust t-statistics by inverting a Cramér–Edgeworth expansion, ensuring third-order asymptotic refinement.
• It presents a closed-form, resampling-free methodology that accommodates both continuous and discrete regressors without relying on strict Gaussian assumptions.
• Simulation evidence confirms that the proposed method outperforms conventional bootstrap techniques, offering improved size control in small-sample applications.

Refined Cluster-Robust Inference: Asymptotic Third-Order Refinement via Cramér–Edgeworth Expansion

Introduction

Cluster-robust inference has become an indispensable component of empirical work in applied microeconometrics, where within-cluster dependence can result in severe over-rejection if not properly addressed, particularly when the number of clusters is limited. One longstanding challenge has been the poor fidelity of normal approximations to the $t$-statistic distribution in this context, leading to procedures that either rely on resampling (bootstrap) or various degrees of parametric assumption. This paper introduces an analytical approach to cluster-robust inference, leveraging the conditional Cramér–Edgeworth expansion to achieve uniformly third-order asymptotic refinement for the null distribution of $t$-statistics under clustered dependence, without recourse to data resampling or parametric restrictions on regressors.

Methodological Contributions

The authors propose an analytic correction to the critical values used for cluster-robust $t$-statistics by inverting the Cramér–Edgeworth expansion of the null distribution. This expansion is computed up to the second order, ensuring that the size distortion of the resulting test is $o(G^{-1})$ (with $G$ the number of clusters), i.e., it achieves third-order refinement as defined by the Edgeworth expansion literature.

Notably, the method accommodates both continuous and discrete regressors, sidestepping both the Gaussianity assumption and the classical Cramér’s condition. This contrasts with typical bootstrap methods such as the cluster-pairs bootstrap, which hinge critically on Cramér’s condition and thus may fail or behave erratically with discrete or highly collinear covariates (see [bhattacharya2010normal]).

The construction proceeds as follows:

• Begin with the cluster-robust $t$-statistic based on the OLS estimator and a cluster-robust variance estimator, treating the regressors as fixed.
• Apply the Cramér–Edgeworth expansion to the null distribution, analytically characterizing the expansion’s coefficients (skewness, kurtosis, and higher cumulants) in terms of observable sample quantities.
• Derive an explicit correction term for the critical value, using the estimated cumulants.

Resulting confidence intervals take the form $$\lambda'\hat\beta \pm \hat{cv}\sqrt{\hat\sigma^2 / G}$$, where $\hat{cv}$ is the refined critical value.

The principal theoretical result is that, under mild moment and invertibility conditions, the corrected critical value yields a test with actual size diverging from nominal size by $o(G^{-1})$, regardless of the presence of discrete regressors or non-Gaussian clustered errors. This covers settings where bootstrap refinements do not generally apply.

Comparison to Existing Methods

The analytic approach exhibits several advantages over bootstrap methods:

• Improved small-sample behavior: The method avoids the singular $X'X$ issues present in the pairs bootstrap [cameron2008bootstrap] and behaves robustly to small numbers of clusters (as few as 10 in simulations).
• No resampling required: The computation is fully analytical and closed-form, yielding substantial reductions in computational cost.
• Weak distributional assumptions: The technique does not invoke Cramér's condition, permitting discrete regressors and non-identically distributed clusters, which are empirically prevalent.
• Stronger theoretical guarantees: It maintains third-order refinement even under conditions where both the pairs and wild cluster bootstrap only guarantee first-order size control or can fail (e.g., under skewed errors for the wild bootstrap [djogbenou2019asymptotic]).

Conventional methods relying on $t$0-distribution critical values with degrees-of-freedom corrections (e.g., [mccaffrey2003bias, bester2011inference, imbens2016robust, young2016improved]) require either homoscedasticity, normality of errors, or do not deliver high-order refinement unless such restrictions are met.

Simulation Evidence and Empirical Implications

Monte Carlo experiments rigorously compare the proposed method to the $t$1, pairs bootstrap, and wild cluster bootstrap in two canonical designs:

1. A "Bertrand et al." design with binary (discrete) regressors [bertrand2004much], where the cluster-pairs bootstrap is known to under-reject and the $t$2 critical value over-rejects for small $t$3.
2. A design with skewed errors, which stresses the inadequacy of the wild cluster bootstrap for asymptotic refinement under nonzero score skewness.

Strong numerical results confirm that the analytic critical value provides accurate size control and achieves rapid rate-of-convergence to nominal levels, dominating alternatives under both discrete regressors and gross error skewness. In the binary regressor scenario, its finite-sample performance matches the wild bootstrap, while in the skewed error scenario, it strictly dominates the wild bootstrap, which fails to refine size.

Theoretical and Practical Implications

The approach generalizes naturally to regression with non-identically distributed clusters, adaptive randomization schemes, and discrete or mixed regressors where bootstrap methods do not uniformly apply. Its high-order refinement and analytic tractability suggest improved power properties for robust hypothesis testing in high-salience empirical designs such as difference-in-differences or panel models with policy interventions.

The solution also obviates computational difficulties and ambiguity in practical implementation of the bootstrap, providing fully reproducible and deterministic inference—an increasing desideratum in empirical practice. Additionally, the method may be extended or adapted to more complex estimators (e.g., GMM, high-dimensional fixed effects) or to allow for dependence structures beyond classical clustering.

Directions for Future Research

Potential extensions include:

• Adaptations for two-way, network, or spatial clustering structures.
• Application to high-dimensional settings, where clusters may be small but the parameter dimension $t$4 grows with $t$5.
• Robustification to heavy-tailed or outlier-prone cluster distributions by augmenting moment truncation assumptions.
• Integration with post-selection or inference post model selection frameworks relevant in causal inference.

Conclusion

This paper advances cluster-robust inference by providing an analytic, high-order accurate critical value correction for the $t$6-statistic under minimal assumptions. Its theoretical guarantees and empirical performance distinctly improve upon both conventional $t$7-based and resampling-based critical values in challenging applications, rendering it a compelling default for inference with clustered data. The method’s flexibility and extensibility warrant further developments for complex dependence structures and large-scale empirical applications.

---

References

• Gafarov, B., & Ura, T., "Refined Cluster Robust Inference" (2603.24786)
• Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2008). "Bootstrap-based improvements for inference with clustered errors."
• Djogbenou, A. A., MacKinnon, J. G., & Nielsen, M. Ø. (2019). "Asymptotic Theory and Wild Bootstrap Inference with Clustered Errors."
• Bhattacharya, R. N. & Rao, R. R. (2010). "Normal Approximation and Asymptotic Expansions."
• Bertrand, M., Duflo, E., & Mullainathan, S. (2004). "How much should we trust differences-in-differences estimates?"
• McCaffrey, D. F., Bell, R. M. (2003). "Bias reduction in standard errors for linear regression with multi-stage samples."
• Imbens, G. W., & Kolesar, M. (2016). "Robust standard errors in small samples: Some practical advice."
• Young, A. (2016). "Improved, nearly exact, statistical inference with robust and clustered covariance matrices using effective degrees of freedom corrections."
• Bester, C. A., Conley, T. G., & Hansen, C. B. (2011). "Inference with dependent data using cluster covariance estimators."