Few-Cluster Corrections and Wild Cluster Bootstrap

Why a Large Panel Can Still Have Few Independent Units

Cluster-robust inference allows errors within a panel unit to be dependent, while relying on independence across units. That distinction makes the number of independent units central to inference. A panel with many observations over time but only a handful of units does not provide the same amount of independent information as a panel with many units. Conventional cluster-robust standard errors can then be too small, and tests that compare a statistic to a standard normal distribution can reject too often.

The issue is not that clustering has become inappropriate. Rather, the usual cluster-robust variance estimator is justified by an approximation that improves as the number of independent clusters grows. With few clusters, finite-sample adjustments and a wild cluster bootstrap can improve inference, but neither is a universal remedy. Cluster count, cluster sizes, leverage, and the design still matter.

For concreteness, consider a regression with unit effects, \(y_{gt}=x_{gt}'\beta+\alpha_g+u_{gt}\). The parameter of interest may be one component of \(\beta\), or a stated linear combination of coefficients. Unit-clustered inference allows \(u_{gt}\) to be heteroskedastic and serially dependent within \(g\). It requires the independent-cluster assumption across \(g\), along with the conditions needed for the regression estimator itself to identify the target parameter. It does not by itself address dependence across units.

Diagnosing the Finite-Cluster Problem

A first diagnostic is to report \(G\), not just the total observation count. Also inspect cluster sizes and the regression’s leverage across clusters. A nominally moderate \(G\) can be less informative if a few units dominate the identifying variation. There is no diagnostic that proves a cluster-robust approximation is accurate; these checks help identify reasons for caution.

The usual unit-clustered sandwich estimator aggregates cluster-level score contributions. In schematic form, its unadjusted (“CR0”) covariance estimate is

where \(\widehat s_g\) is the contribution of cluster \(g\) to the regression score and \(\widehat A\) is the estimator’s bread matrix. Its large-sample justification treats the cluster contributions—not individual rows—as the independent units. When \(G\) is small, the estimated variance and the reference distribution for the test statistic can both be poorly approximated.

Degrees-of-Freedom Corrections

A common finite-cluster adjustment rescales CR0 to produce the CR1 covariance estimate. In a basic linear-regression convention, let \(N\) be the total number of observations and \(K\) the rank of the estimated regressor matrix, including any absorbed effects counted by the implementation. Then

The adjustment increases the covariance estimate when its multiplier exceeds one; the standard error is multiplied by the square root of that multiplier. Inference commonly uses a \(t_{G-1}\) reference distribution for a single coefficient. For a confidence level \(1-\alpha\), the corresponding interval is the estimate plus or minus the \(1-\alpha/2\) quantile of \(t_{G-1}\) times its adjusted standard error.

CR1 and the \(G-1\) reference degrees of freedom are practical corrections, not a guarantee of correctly sized tests. Implementations differ in how they count \(K\), particularly with absorbed fixed effects. Report the convention used. More refined leverage corrections, such as CR2, adjust cluster residual contributions for the effect of fitting the regression. They are often paired with a coefficient-specific Satterthwaite degrees-of-freedom approximation. That approximation need not equal \(G-1\), and can be substantially smaller when a few clusters are influential. These methods improve on a blanket correction in some settings, but still rely on the design and cannot manufacture independent information that the data do not contain.

The Wild Cluster Bootstrap

The wild cluster bootstrap generates a reference distribution by perturbing fitted residuals at the cluster level. Multiplying an entire cluster’s residual vector by one random weight preserves the pattern of dependence within that cluster. For a test of a coefficient, a standard approach imposes the null when forming bootstrap outcomes. This is important: the test should be calibrated under the null hypothesis being tested.

Rademacher weights take values \(+1\) and \(-1\) with probability one-half each. Other cluster-level weight distributions, including a six-point distribution sometimes used with few clusters, are also used. The choice of weights and any leverage adjustment should be stated. In each draw, use a covariance estimate and studentization convention consistent with the intended test. With \(B\) random draws, a common Monte Carlo p-value calculation is \((1+\#\{|t^{*(b)}|\geq |t_{\mathrm{obs}}|\})/(B+1)\); the added one avoids reporting a zero p-value merely because no simulated draw exceeded the observed statistic.

The following sequence summarizes the test. The exact implementation can vary, so a report should specify the restricted fit, weights, studentization, and number of draws.

Confidence Intervals and Practical Limits

A bootstrap confidence interval can be obtained by inverting tests. For each candidate value \(\beta_0\), impose \(H_0:\beta=\beta_0\), run the null-imposed bootstrap test, and retain candidate values whose tests do not reject at level \(\alpha\). The retained set is the \(1-\alpha\) confidence set under that procedure. It may be asymmetric, and its endpoints need not match a symmetric interval based on a single standard error. The interval’s interpretation remains a repeated-sampling one: the procedure is designed to cover the fixed parameter at its stated rate under its assumptions, not to assign a probability to the parameter after observing these data.

Bootstrap inference is not automatically reliable whenever \(G\) is small. With very few clusters, its distribution can be discrete; cluster imbalance, high leverage, or a design in which identifying variation comes from only one or two units can undermine the approximation. The bootstrap must also reflect the actual independent sampling or assignment structure. It does not correct a misspecified regression, establish exogeneity, or make dependent clusters independent. When appropriate, report conventional and adjusted results as sensitivity analyses, explain the number and sizes of clusters, and state the bootstrap choices.

Common Mistake

• Wrong reasoning: “There are hundreds of panel observations, so a normal approximation must be accurate.”
• Correction: Under unit clustering, the independent pieces are the \(G\) unit-level contributions. A large number of observations within a small number of units does not substitute for many independent units. Report \(G\), examine cluster influence, and consider finite-cluster corrections and a carefully specified wild cluster bootstrap.
• Wrong reasoning: “The CR1 multiplier or bootstrap guarantees correct inference.”
• Correction: These are methods for improving finite-sample inference under stated assumptions. Neither repairs a poor identifying design or dependence across supposedly independent clusters, and both can remain unreliable with extremely few or highly influential clusters.