Use Heteroscedasticity-Robust Standard Errors for the Right Problem

Identify what robust changes

In an ordinary least-squares analysis, a heteroscedasticity-robust covariance estimator changes estimated uncertainty while retaining the fitted coefficients from the same model and data. It does not automatically repair nonlinearity, omitted confounding, dependence between observations or data errors.

The statsmodels robust-covariance reference documents several covariance options and inferential settings. Record the exact option, such as an HC variant, rather than writing only robust analysis in the methods.

Check the relevant assumptions

Inspect residual spread and the sampling unit. If observations are correlated within people or sites, an independent-observation heteroscedasticity correction may be insufficient. If the model’s mean structure is wrong, changing the standard error does not make its coefficient answer the intended question.

Small samples and high leverage can affect covariance-estimator performance. Preserve the sample size, number of fitted parameters and chosen small-sample conventions. Do not select a covariance method solely because it produces the desired p-value. Preserve any sampling weights and explain whether the chosen covariance implementation supports their intended meaning.

Compare a hypothetical output

Suppose the same fitted model estimates a coefficient of 2.0 units. Its conventional standard error is 0.5, while the chosen robust standard error is 0.8. Under an illustrative normal approximation, the conventional 95% interval is 2 ± 1.96×0.5, or [1.02, 2.98]. The robust interval is 2 ± 1.96×0.8, or [0.432, 3.568].

The coefficient did not change, but its estimated precision did. A robust standard error is not guaranteed to be larger; its purpose is an alternative covariance calculation under specified assumptions. If the software uses a t reference distribution, use that method’s critical value instead of the illustrative normal value.

Report the decision transparently

Name the fitted model, covariance estimator, reference distribution and any degrees-of-freedom correction. Include the effect estimate and interval in meaningful units. Keep a diagnostic note explaining why this uncertainty method fits the design.

If a sensitivity analysis uses another estimator, compare interpretations without ranking methods by statistical significance. The final report should make clear which problem was addressed and which remain: changing residual spread, correlation, nonlinear means and causal identification are separate issues that require separate reasoning.

Source verification: 6 November 2026. Examples are hypothetical.