clustered standard errors

This study addresses the variable reliability of cluster-robust inference methods in cross-sectional and panel data regressions, which often depends on data structure and model specification. The authors propose an integrated evaluation framework to systematically compare the performance of various cluster-robust variance estimators and inference procedures—including analytical and bootstrap approaches—across diverse empirical scenarios. Their analysis demonstrates that while no single method universally dominates, conducting inference through cross-validation using multiple methods substantially enhances result credibility. This framework offers applied researchers a practical guide for selecting more reliable statistical inference strategies tailored to their specific contexts, thereby strengthening the robustness of empirical conclusions and policy recommendations.

Conventional clustered robust inference fails when cluster sizes are non-negligible—e.g., following Zipf’s law—and 77% of empirical studies in the *American Economic Review* and *Econometrica* (2020–2021) violate its implicit equal-size or bounded-size assumptions. Method: This paper establishes the first necessary and sufficient condition for consistency of clustered robust estimators and proposes two new procedures: score subsampling and size-adjusted reweighting. Both methods are theoretically grounded—guaranteeing consistency and uniform size control—and practically implementable, with ready-to-use Stata packages. Results: Monte Carlo simulations demonstrate that the proposed methods strictly maintain nominal test size even where conventional approaches severely distort inference. They constitute the first truly robust and implementable inferential framework for settings with large, heterogeneous cluster sizes.

This paper addresses statistical inference challenges in difference-in-differences (DID) designs with a single treated cluster and a fixed number of control clusters. Under weak assumptions permitting arbitrary unknown intra-cluster dependence, we propose a variance-free t-test that avoids estimating the asymptotic variance. The method requires only a user-specified bound on the relative heteroskedasticity between treated and control clusters; it then constructs customized critical values—either analytically or via numerical optimization—to achieve valid inference at any desired significance level. Unlike conventional approaches, it does not rely on asymptotic normality or large numbers of clusters, thereby substantially improving inference reliability in small-sample and limited-control-group settings. Extensive simulations and empirical applications demonstrate the method’s robustness and high statistical power. A table of commonly used critical values is provided for immediate implementation by applied researchers.

In cluster-randomized trials, conventional methods such as generalized linear mixed models (GLMMs) and generalized estimating equations (GEE) suffer from ambiguous estimands for treatment effects under model misspecification or informative cluster size. This paper proposes a model-robust standardization approach: first constructing marginal estimators simultaneously consistent for both cluster-averaged and individual-averaged treatment effects; deriving variance estimates via the jackknife and developing a formal test for informative cluster size. The method avoids specifying the intra-cluster correlation structure correctly and retains consistency under diverse forms of model misspecification. Simulation studies demonstrate substantially improved estimation accuracy and inferential reliability compared to standard GLMM and GEE approaches. An open-source R package, MRStdCRT, implements the proposed methodology.

In observational causal inference, weighting methods mitigate covariate imbalance but often inflate variance estimates and yield overly conservative standard errors. This paper proposes augmenting weighted regression with main effects of covariates and their interactions with the treatment variable, integrated with residualization and parametric model augmentation to form a unified inferential framework. We establish, for the first time under design-based, model-based, and finite-sample-corrected superpopulation sampling assumptions, that this approach yields asymptotically valid and more precise standard errors. Theory, simulations, and multiple empirical applications demonstrate substantially narrower confidence intervals—on average 15–30% shorter—with improved inferential accuracy and robustness to both exact and approximately balanced weights. The key innovation lies in achieving simultaneous gains in statistical efficiency and asymptotic validity at minimal variance cost.

This study addresses the interpretational ambiguity of weighted estimators when treatment effects are heterogeneous, as their validity hinges critically on the choice of weights. To tackle this issue, the authors propose an estimator that minimizes worst-case bias and construct confidence intervals that are uniformly valid over a broad class of weighting schemes. Their approach integrates minimax bias reduction, bounds from heterogeneity-robust sensitivity analysis, and theoretical characterizations of discrepancies among weighted estimators, thereby enabling inference robust to weight uncertainty. Empirical applications illustrate the method’s utility: in Lakdawala et al.’s event study, findings remain robust across a wide range of weights, whereas in the Project STAR experiment, conclusions prove sensitive even to minor perturbations of baseline weights.

This study addresses the inconsistency in sensitivity analyses for unmeasured confounding that arises when observational studies with clustered treatment assignment are analyzed at different levels—individual versus cluster. Focusing on linear regression models under clustered treatment, the authors propose a correction method based on Pearson’s partial eta-squared. By applying the Mundlak transformation to incorporate cluster means of covariates and parameterizing unmeasured confounding bias through partial R², the approach ensures equivalence between individual- and cluster-level sensitivity analyses. The method explicitly accounts for between-cluster variation in driving bias, thereby reconciling cross-level discrepancies and substantially enhancing the robustness and reliability of causal inference in clustered data settings.

This study addresses the sensitivity of causal effect estimation to model misspecification in longitudinal cluster-randomized and quasi-experimental designs. Within an M-estimation framework, it demonstrates that fixed-effects models yield consistent and asymptotically normal estimates of nonparametrically defined treatment effects, provided the treatment effect structure is correctly specified—even when other model components are arbitrarily misspecified. The work establishes, for the first time, that fixed-effects models are valid for estimating superpopulation marginal effects and reveals their robustness to partial misspecification of the treatment effect structure across diverse longitudinal settings. Through theoretical analysis, simulations, and reanalyses of empirical data, the paper further shows that fixed-effects models outperform mixed-effects models in robustness and reliability when time-invariant confounding exists at the cluster or individual level.

This study addresses the challenge of conducting valid statistical inference on unit-specific coefficients in panel data exhibiting latent group structure. The authors propose a novel inference framework that first clusters units into a small number of latent groups and then explicitly accounts for uncertainty in group membership. Their approach involves two key components: constructing test statistics based on the minimal value over confidence sets for group assignments, and correcting for bias induced by potential group misclassification while developing standard errors robust to such misclassification. Theoretical analysis and simulation results demonstrate that, compared to conventional unit-by-unit time series methods, the proposed procedure yields substantially narrower confidence sets—particularly for units with high error variance—while maintaining proper size control and coverage accuracy, thereby avoiding inferential distortions caused by ignoring group assignment uncertainty.

This study addresses the lack of a unified and interpretable sensitivity analysis framework for causal panel data methods, which hinders quantification of unobserved confounding. It introduces Riesz representation theory into causal panel sensitivity analysis for the first time, proposing a workflow that balances theoretical rigor with practical usability. The framework offers two complementary routes: Route A provides direct sensitivity profiles via bounds on omitted variable bias and partial-R² robustness values, while Route B establishes auxiliary diagnostic benchmarks based on observed covariates. Compatible with diverse estimators—including synthetic difference-in-differences (SDID), matrix completion, and fixed-effects imputation—the approach supports both corrected inference and finite-difference auditing. Applied to California’s tobacco control policy, SDID yields an estimated effect of −15.60 packs per capita (adjusted SE = 9.49, p = 0.051), with low single-digit robustness values indicating reliable conclusions; the method also extends successfully to county-level staggered minimum wage policy analysis.