Wild Bootstrap-t Tests With Robust Standard Errors

Why Bootstrap the Robust t Statistic?

A heteroskedasticity-robust standard error avoids assuming that all observations have the same error variance. But the usual robust t reference distribution is an approximation, and that approximation can be inaccurate in some finite samples. A wild bootstrap-t test refines the calibration by generating data under the null hypothesis, recomputing a robust standard error in every bootstrap sample, and using the resulting studentized statistics to assess the observed statistic.

The procedure is not simply a bootstrap of the estimated coefficient. Each bootstrap statistic is centered at the null value and divided by its own bootstrap robust standard error. That studentization matters: the bootstrap distribution represents the test statistic, not just the variation in the estimate.

Model and Bootstrap Construction

Consider the linear regression model \(y_i=x_i'\beta+u_i\), for \(i=1,\ldots,n\), where \(y_i\) is the outcome, \(x_i\) is a vector of regressors, and \(\beta\) is the coefficient vector. The standard setting for this procedure treats the regressors as fixed, or conditions on their observed values. Assume the design matrix has full column rank, \(E(u_i\mid X)=0\), and errors are independent across observations but may have different conditional variances. Suitable finite-moment and leverage conditions are also needed for the approximation; in particular, no single observation should dominate the regression as sample size grows.

These conditions allow heteroskedasticity but do not allow arbitrary dependence across observations. In particular, a wild bootstrap designed for independent observations is not automatically appropriate for clustered or serially dependent errors.

Fit the regression subject to \(r'\beta=q\), giving the restricted estimate \(\hat{\beta}_0\) and restricted residuals \(\hat{u}_{0i}=y_i-x_i'\hat{\beta}_0\). A common choice is the Rademacher multiplier \(v_i\), which independently equals \(+1\) or \(-1\) with probability one-half. It has mean zero and variance one. For each bootstrap replication, generate

where \(\tilde{u}_{0i}\) is a restricted residual, optionally adjusted for leverage. For example, an HC2-style adjustment uses \(\tilde{u}_{0i}=\hat{u}_{0i}/\sqrt{1-h_{ii}}\), with \(h_{ii}\) the diagonal leverage from the regression design. The chosen adjustment should be stated and applied consistently. The restricted fitted value keeps the bootstrap data-generating process under \(H_0\); the multipliers let the bootstrap error size vary across observations instead of forcing a common variance.

Fit the unrestricted regression to each \(y^*\). Let \(\hat{\beta}^*\) be its coefficient estimate and let \(\widehat{V}^*\) be a heteroskedasticity-robust covariance estimate from that bootstrap regression. The bootstrap t statistic is

Use the same robust covariance specification for the observed and bootstrap statistics—for example, HC2 in both. For the observed data, the statistic is \(t_{\mathrm{obs}}=(r'\hat{\beta}-q)/\sqrt{r'\widehat{V}r}\). A scalar coefficient test is the special case where \(r\) selects one coefficient.

From Bootstrap Statistics to a Test

Generate \(B\) bootstrap samples and calculate \(t_1^*,\ldots,t_B^*\). For a two-sided test at level \(\alpha\), compare \(|t_{\mathrm{obs}}|\) with the empirical \(1-\alpha\) quantile of \(|t_1^*|,\ldots,|t_B^*|\). Reject \(H_0\) when the observed absolute statistic exceeds that test-specific critical value. Alternatively, estimate the two-sided p-value by counting bootstrap statistics at least as extreme as the observed one. A common finite-simulation estimate is

The added one avoids reporting a zero p-value merely because no simulated statistic crossed the observed threshold. For a one-sided alternative, use the corresponding tail of the signed bootstrap statistics and its one-sided quantile. These are bootstrap approximations, not exact finite-sample probabilities: their quality depends on the model conditions and on how well the resampling scheme reproduces the statistic’s sampling behavior.

Worked Tests

Why Studentization and Null Imposition Matter

Studentization allows each bootstrap replication to reflect not only a different coefficient estimate but also the uncertainty associated with that particular bootstrap sample. If one instead compared unstudentized coefficient deviations, the resulting distribution would not directly calibrate the robust t statistic being used for the observed test. The covariance estimate must therefore be recomputed in each replication, not held fixed at its observed-data value.

Null imposition is equally important. The test asks how unusual the observed statistic would be if \(H_0\) held. Generating bootstrap outcomes around the unrestricted fitted values generally fails to reproduce that null reference distribution. Generating them around restricted fitted values makes the restriction true in the bootstrap data-generating process, while the unrestricted refit measures how far each sample estimate wanders from that restriction.

A brief justification follows from the construction. Conditional on the observed design, the multipliers have \(E(v_i)=0\) and \(E(v_i^2)=1\). Consequently, the bootstrap disturbance \(\tilde{u}_{0i}v_i\) has conditional mean zero and conditional variance \(\tilde{u}_{0i}^{\,2}\). This retains observation-specific residual scale rather than imposing one common variance. Under regularity conditions, the resulting coefficient fluctuations and their robust standard errors approximate the corresponding null sampling behavior. Studentizing both the observed statistic and every bootstrap draw then permits comparison on the same scale. This argument motivates the approximation; it does not make the procedure exact in finite samples.

Common Pitfalls and Reporting

• Resampling pairs instead of residual signs. That describes a different bootstrap. The wild bootstrap keeps the design fixed and varies residual contributions with multipliers, a feature useful when conditional variances differ.
• Generating outcomes from the unrestricted fit. This does not impose the null. Use restricted fitted values when constructing bootstrap outcomes for a null test.
• Using the observed standard error for every replication. That removes the bootstrap studentization. Recalculate the selected robust covariance estimate in each bootstrap sample.
• Calling the result an exact test. The bootstrap p-value and critical value are simulation-based approximations whose reliability depends on the assumptions and design.
• Ignoring dependence or severe leverage. Independent-observation wild multipliers do not automatically handle clustered errors, and extreme leverage can undermine the approximation. Match the method to the sampling structure and inspect the design.

A clear report states the restriction, the robust covariance variant, how restricted residuals were handled, the multiplier distribution, the number of replications, and whether the result uses a bootstrap p-value or critical value. Include the observed statistic and interpret the conclusion in the outcome’s units. This makes the calibration choices visible and the inference reproducible.