Linear Regression: Inference Based on Cluster Estimates

This article proposes a novel estimator for regression coefficients in clustered data that explicitly accounts for within-cluster dependence. We study the asymptotic properties of the proposed estimator under both finite and infinite cluster sizes. The analysis is then extended to a standard random coefficient model, where we derive asymptotic results for the average (common) parameters and develop a Wald-type test for general linear hypotheses. We also investigate the performance of the conventional pooled ordinary least squares (POLS) estimator within the random coefficients framework and show that it can be unreliable across a wide range of empirically relevant settings. Furthermore, we introduce a new test for parameter stability at a higher (superblock; Tier 2, Tier 3,…) level, assuming that parameters are stable across clusters within that level. Extensive simulation studies demonstrate the effectiveness of the proposed tests, and an empirical application illustrates their practical relevance.

💡 Research Summary

The paper tackles a pervasive problem in applied economics and social sciences: regression analysis with clustered observations that exhibit within‑cluster dependence. While a large body of literature has focused on robust standard errors under the assumption that the number of clusters G grows to infinity while each cluster size N remains fixed, real‑world data often violate this setting. In many applications a few “dominant” clusters contain a disproportionate share of observations, leading to severe imbalance in cluster sizes. The authors illustrate this phenomenon with Indian industrial survey data, where a single state’s cluster can account for half of the sample, and with other examples from the United States.

Three asymptotic regimes are distinguished: (1) G fixed, N → ∞, where the ordinary least‑squares estimator (LSE) becomes inconsistent; (2) both G and N diverge without restrictions on their relative rates, a regime that is under‑studied and where conventional pooled OLS (POLS) can be unreliable, especially when a few large clusters dominate; (3) N fixed, G → ∞, the setting most existing cluster‑robust methods address. The paper shows that regime (2) is particularly problematic for POLS because the estimator’s bias does not vanish when a small number of large clusters exert outsized influence.

To overcome these limitations the authors propose a “cluster‑wise averaging” estimator. For each cluster g they compute the ordinary least‑squares estimate β̂_g = (X_g′X_g)⁻¹ X_g′Y_g using only observations within that cluster. The final estimator is the simple arithmetic mean across clusters: β̂ = G⁻¹ Σ_{g=1}^G β̂_g. This construction preserves the full within‑cluster dependence structure (captured by the covariance matrix Ω_g) while exploiting independence across clusters. By allowing the eigenvalues of Ω_g to grow at three rates—strong (λ_max(Ω_g)=O(N_g)), semi‑strong (λ_max(Ω_g)=O(h(N_g)) with h(N_g)→∞ but h(N_g)/N_g→0), and weak (λ_max(Ω_g)=O(1))—the authors prove that β̂ remains consistent and asymptotically normal for any mixture of these dependence types, provided G→∞. Even when a few clusters are large and strongly dependent, the averaging step dilutes their influence because the variance of each β̂_g scales as Ω_g/N_g.

The framework is then extended to random‑coefficients (RC) models, where the true parameter vector varies across clusters: β_g = β + u_g, with u_g having mean zero and covariance Σ_u. The same cluster‑wise averaging estimator consistently estimates the common mean β. Leveraging the asymptotic variance of β̂, the authors construct a Wald‑type test for general linear hypotheses H₀: Rβ = r. Under the null, the test statistic follows a χ² distribution with rank(R) degrees of freedom as G→∞, regardless of the underlying dependence regime.

A further contribution is a novel test for parameter stability at a higher hierarchical level (super‑block), such as states or districts that contain many clusters. The null hypothesis posits that all clusters within a super‑block share the same parameter vector. By comparing within‑super‑block and between‑super‑block variability of the cluster‑wise estimates, an F‑type statistic is derived. The test remains valid under severe cluster‑size imbalance because it relies on the same averaging principle and does not require each super‑block to have a large number of observations.

Extensive Monte‑Carlo simulations evaluate the performance of the proposed methods across a range of designs: (i) strong dependence with a few large clusters comprising 10 % of the sample, (ii) semi‑strong dependence with many medium‑sized clusters, and (iii) weak dependence with nearly balanced clusters. In all designs the cluster‑wise averaging estimator exhibits negligible bias, accurate standard‑error estimates, and correct coverage, whereas POLS displays substantial bias and under‑coverage in design (i). The Wald‑type and super‑block stability tests achieve nominal size and higher power than existing cluster‑robust alternatives, especially when cluster sizes are heterogeneous.

Empirical applications reinforce the theoretical findings. Using the Indian Annual Survey of Industries (ASI), the authors estimate the effect of a policy variable on firm outcomes. In the state of Tripura, a single massive cluster dominates the data; the proposed estimator yields stable coefficient estimates and correctly rejects the hypothesis of parameter homogeneity across states. In the Household Consumption Expenditure Survey (HCES), the authors model Engel curves at the household level, treating villages (FSUs) as clusters and states as super‑blocks. The analysis confirms that while Engel‑curve parameters differ significantly across states, they remain stable across villages within each state, validating the super‑block stability test.

In summary, the paper makes three major contributions: (1) a simple yet theoretically sound estimator that remains consistent under arbitrary cluster‑size imbalance and various dependence structures; (2) an extension to random‑coefficients models with a Wald‑type hypothesis test; and (3) a new hierarchical stability test for multi‑level clustered data. By addressing the shortcomings of conventional POLS and cluster‑robust variance estimators in the presence of dominant clusters, the work provides applied researchers with a robust toolkit for inference in complex cross‑sectional settings.