Homoscedasticity means that error or residual variance is constant across the comparison being modeled. Heteroscedasticity means that variance changes. A residual plot can show the pattern; Levene’s test and the Breusch–Pagan test address different versions of the question. No single p-value substitutes for identifying the design, variance structure, and consequence for inference.
What the terms describe
In a regression setting, homoscedastic residuals have roughly similar variance across fitted values or predictors. Heteroscedastic residuals may fan outward, narrow, form clusters with different spread, or change variance in a more complex way.
For grouped data, the same language often concerns whether groups have equal population variances. That is related but not identical to asking whether regression error variance changes with predictors. Match the evidence to the question:
| Evidence | Primary question | Typical output |
|---|---|---|
| Residual-versus-fitted plot | Does spread or structure change across fitted values? | Visual pattern |
| Levene’s test | Are group variances equal? | Test statistic and p-value |
| Breusch–Pagan test | Does regression error variance vary systematically with predictors? | LM-type statistic and p-value |
| Robust standard errors | Can covariance estimation remain valid under heteroscedasticity? | Revised standard errors and intervals |
Read the residual plot first
A useful residual plot has fitted values on the horizontal axis and residuals on the vertical axis. Look for:
- a roughly even vertical band around zero;
- a fan or funnel that widens or narrows;
- separate variance bands for subgroups;
- curves or other systematic patterns;
- individual points with unusually large residuals.
An even band is compatible with constant variance, but it does not prove it. A funnel is evidence of changing spread, but it does not say whether the cause is an omitted predictor, an unsuitable outcome scale, clustering, a mean-model problem, or a data-quality issue. Residual plots and influence diagnostics separates these visual clues.
Levene’s test: equality across groups
Levene’s test evaluates the null hypothesis that group variances are equal.[NIST/SEMATECH (2012)] A small p-value is evidence against that equal-variance model under the test assumptions; a large p-value is not proof that all variances are identical.
The center used in the calculation matters. NIST documents versions centered on group means, medians, or trimmed means, with the median-centered Brown–Forsythe form providing robustness for many nonnormal settings.[NIST/SEMATECH (2012)] A report should say which version was used rather than writing only “Levene’s test.”
Sample size also affects interpretation. A small study may have little ability to detect an important variance difference, while a very large study may detect a small departure that has limited practical effect. Inspect the group spreads and their consequences alongside the p-value.
Breusch–Pagan: variance linked to predictors
The Breusch–Pagan test asks whether regression error variance is systematically related to specified predictors.[Breusch (1979)] Its null is homoscedasticity under the test model. A small p-value flags evidence that the modeled variance is not constant; it does not reveal the complete form or cause of the change.
Because the test concerns a regression variance model, it is not interchangeable with Levene’s group-variance test. A good report names the auxiliary variables or fitted-value relationship used and checks the residual plot for the shape the statistic compresses.
What heteroscedasticity changes
Under common linear-model conditions, heteroscedasticity does not automatically make every coefficient estimate biased. It can make conventional standard errors, tests, and confidence intervals unreliable, and it can expose a poorly specified mean model. The consequence therefore depends on what is being estimated and how uncertainty is calculated.
Heteroscedasticity-consistent covariance estimators provide standard errors designed to remain useful under nonconstant variance.[White (1980)] They do not make residual variance constant or repair every model defect. A transformation, explicit variance model, weighted method, cluster-aware method, or revised mean structure may be more appropriate depending on the pattern and design.
A compact interpretation sequence
- Name the setting: grouped variances or regression residual variance.
- Inspect the spread visually.
- Name the exact test and its null.
- Report the test statistic, degrees of freedom where relevant, and p-value.
- Explain how the variance pattern affects standard errors, intervals, or predictions.
- State the response: robust covariance, model revision, transformation, or justified retention.
Assumption checks and sensitivity analysis explains why an assumption result should trigger an evaluation rather than an automatic pass/fail decision. robust standard errors covers the distinction between changing uncertainty estimates and changing the fitted model.
Common reading errors
- Calling any unequal-looking group range “heteroscedasticity” without considering sample size or center.
- Treating a large Levene p-value as proof of equal variance.
- Using Levene’s and Breusch–Pagan as names for the same test.
- Running a variance test without inspecting the residual pattern.
- Assuming robust standard errors remove heteroscedasticity.
- Ignoring dependence, clustering, or a misspecified mean structure.
A defensible reporting sentence
“The residual-versus-fitted plot showed increasing spread at higher fitted values. The Breusch–Pagan test provided evidence against constant variance; inference was therefore reported with a specified heteroscedasticity-consistent covariance estimator, and the mean model was checked for misspecification.” Include the actual statistic, p-value, estimator, and plot context in a real report.