Do coefficients change?

Not when only the covariance estimator changes for the same fit.

Are the errors robust to outliers?

Not in the sense of making least-squares coefficients resistant to them.

Do they handle serial correlation?

Basic HC estimators do not; dependence needs an appropriate method.

AI Glossary / Statistics and probability / Intermediate

Heteroskedasticity-consistent standard errors

Definition

Heteroskedasticity-consistent standard errors estimate regression-coefficient uncertainty while allowing error variance to differ across observations. In ordinary least squares, they replace the conventional coefficient covariance calculation with a sandwich estimator based on individual squared residuals. Their validity is generally asymptotic and depends on appropriate assumptions; they do not make an otherwise flawed regression reliable.

What it means

A regression may have small prediction errors for one kind of observation and much larger errors for another. Conventional standard errors based on a single common error variance can then misrepresent how uncertain the coefficients are.
The robust calculation changes the uncertainty attached to the fitted coefficients. For the same ordinary least-squares fit, it does not change the coefficients, fitted values, or residuals. Confidence intervals and significance tests may change because their estimated standard errors change.

How it works

Fit the specified ordinary least-squares model and retain its design matrix X and residuals e. With full column rank, the HC0 covariance estimator is (X'X)^(-1) X' diag(e_i²) X (X'X)^(-1). The outer matrices form the sandwich around a middle matrix that allows each observation its own residual-based variance contribution.
Take the square roots of the covariance matrix's diagonal elements to obtain coefficient standard errors. Off-diagonal elements matter for inference on combinations of coefficients, so the covariance matrix contains more information than separate standard errors alone.
HC variants adjust the residual contributions. HC1 multiplies HC0 by n/(n - k), where k is the number of estimated regression parameters, including an intercept if present. HC2 and HC3 adjust for leverage; HC3 uses e_i²/(1 - h_ii)², where h_ii is the observation's leverage.
These variants address finite-sample behavior differently. None guarantees exact inference in a small sample, particularly when a few observations have very high leverage. Specify which version was used and how test statistics and confidence intervals were calibrated.
For ordinary independent observations, the method permits heteroskedasticity without requiring a fully specified variance function. Consistency still depends on conditions such as an appropriate estimating relationship, sufficient moments, and a design that is not dominated by a few observations.
Dependence requires further attention. Repeated observations within customers may call for cluster-robust inference; serially correlated time-series errors may require a heteroskedasticity-and-autocorrelation-consistent approach. Basic HC standard errors do not automatically address either structure.
Inspect the mean model as well as its variance. Robust uncertainty does not remove confounding, repair measurement errors, establish causality, or protect a coefficient against influential outliers. Robust standard errors and robust regression are different ideas.
A variance model and weighted least squares may improve efficiency when justified, but they change the fitting procedure. Choosing robust covariance estimation instead leaves the least-squares estimate in place.

Three practical examples

A demand analyst estimates a linear relationship when sales variability increases with store size.
An AI evaluation team models continuous performance scores whose residual spread differs across task difficulty.
A service analyst reports HC standard errors for an independent cross-sectional sample while documenting why repeated-customer dependence is absent.

Think of it this way

Think of keeping the same measured location on a map while redrawing its uncertainty boundary using more realistic information. The location stays put. A better boundary does not prove that the map itself represents the right destination.

Worked example

Use a deliberately tiny example to illustrate arithmetic, not dependable inference. Fit y = b*x with no intercept to x = 1, 2, 3 and y = 2, 2, 4. Omitting the intercept is part of this constructed model, not a general recommendation.
The coefficient is sum(x*y)/sum(x²) = 18/14 = 9/7. Residuals are 5/7, -4/7, and 1/7.
With one coefficient, HC0 variance is sum(x_i²*e_i²)/(sum(x_i²))². The numerator is (25 + 64 + 9)/49 = 2, and the denominator is 14² = 196. Thus HC0 variance is 1/98 and its standard error is approximately 0.10102.
HC1 multiplies this variance by 3/(3 - 1) = 1.5, giving a standard error of approximately 0.12372. For comparison, the conventional residual-variance estimate is (6/7)/2 = 3/7, giving coefficient variance 3/98 and standard error 0.17496.
All three calculations retain b = 9/7. Robust standard errors are smaller here, showing that they are not necessarily conservative enlargements. With only three observations, the numbers should not be used to claim reliable coverage or a substantive discovery.

Illustrative case study

Fictional case: Birch Operations models completion time using job size across independent one-off projects. Residual plots show greater spread for larger jobs. The analyst retains the proposed mean model while comparing conventional uncertainty with a specified HC3 calculation.
Before reporting, the team checks whether projects actually share managers or common schedules that could create dependence. It also examines influential large projects and alternative mean specifications.
The final report states the covariance method and explains that it changes coefficient uncertainty, not the fitted time predictions. Individual-project prediction intervals receive separate attention because coefficient standard errors alone do not describe the full variability of a future completion time.

Common misconceptions and mistakes

Common mistakes include assuming robust means larger, treating HC0 and HC3 as interchangeable labels, forgetting the intercept when counting parameters, and using ordinary HC errors for strongly dependent observations.
A robust p-value is not evidence that the functional form is correct. Nor does a narrow coefficient interval guarantee accurate predictions for individual cases or for a different deployment population.

Business use

These standard errors can improve regression reporting when unequal variability is plausible and the sampling assumptions fit. Business readers should see the estimated relationship, its uncertainty, the covariance method, and material design limitations. If a decision changes with reasonable covariance choices, investigate why instead of selecting the most favorable result.

Check your understanding

What remains 9/7 in the example?

Answer: The fitted coefficient.

Why is the HC1 multiplier 1.5?

Answer: There are three observations and one estimated parameter.

Can the tiny example establish reliable confidence-interval coverage?

Answer: No; it illustrates the formula only.

Related terms

Leverage
Weighted least squares
Heteroskedasticity
sandwich estimator
HC0
HC1
HC3
cluster-robust standard errors
HAC estimator

Learn next

Compare conventional and HC covariance estimates for one fixed regression. Then study leverage and dependence so that the chosen uncertainty method matches how the observations were collected.

Leverage

Sources

Statsmodels, HC0: https://www.statsmodels.org/stable/generated/statsmodels.regression.linear_model.RegressionResults.HC0_se.html Statsmodels, HC1: https://www.statsmodels.org/stable/generated/statsmodels.regression.linear_model.RegressionResults.HC1_se.html Statsmodels, HC3: https://www.statsmodels.org/stable/generated/statsmodels.regression.linear_model.RegressionResults.HC3_se.html James Powell, heteroskedasticity notes: https://eml.berkeley.edu/~powell/e240b_sp06/hetnotes.pdf

Previous: Heterogeneous treatment effect Next: Heuristic