Does it imply nonnormal data?

No. Normal populations can have different variances.

Does a larger sample fix unequal variances?

No; it can improve estimation but does not make population variances equal.

Is heteroskedasticity the same idea?

Yes, usually expressed as unequal conditional error variance in a model.

AI Glossary / Statistics and probability / Intermediate

Heterogeneity of variance

Definition

Heterogeneity of variance means that the populations or conditional distributions being compared have different variances. Their outcomes differ in spread, even if their means are identical. In regression, unequal error variance across predictor values is commonly called heteroskedasticity. Unequal sample variances can suggest heterogeneity, but they do not by themselves establish that the underlying population variances differ.

What it means

Variance measures squared dispersion around a mean. Two services can have the same average completion time while one delivers much less consistently. Heterogeneity describes that difference in variability.
The distinction matters both operationally and statistically. Customers experience the spread of waiting times, not just the average. Statistical procedures that assume a common variance may calculate inappropriate uncertainty when variability differs, especially when group sizes are also unequal.

How it works

First specify the level of comparison. In a simple group comparison, interest may center on each population variance. In a regression, the relevant condition usually concerns the variance of the errors after accounting for the model's predictors. Different raw outcome variances need not imply different conditional error variances.
Calculate descriptive measures separately by group and plot the observations. Standard deviations retain the outcome's units, whereas variances have squared units. Boxplots and residual plots can reveal differences in spread, asymmetry, extreme observations, or patterns that a single number hides.
Formal procedures such as Levene's test assess equality of variances. Variants based on deviations from group medians are often used when distributions depart from normality. A small p-value indicates evidence against equality under the test assumptions; a large p-value does not establish equal variances.
Choose the analysis to suit the question. For two independent means, Welch's t-test estimates uncertainty without assuming equal population variances. Its standard error is sqrt(s1²/n1 + s2²/n2), with an adjusted degrees-of-freedom calculation. It still requires an appropriate sampling structure and consideration of distributional shape and sample size.
In regression, heteroskedasticity-consistent standard errors can address some uncertainty-estimation problems while leaving the ordinary least-squares coefficients unchanged. They do not correct a wrong mean model, dependence between observations, or omitted-variable bias. Weighted least squares is another option when a suitable variance model can be justified.
A transformation may stabilize variance, but it also changes the analysis scale and interpretation. Check whether the transformed question serves the business objective rather than applying a logarithm solely to satisfy a diagnostic.
Report the size and practical meaning of the spread difference alongside any inferential result. Variability itself may be the main operational outcome, rather than merely an assumption to manage.

Three practical examples

Two AI transcription services have similar mean error rates but different variability across recordings.
A delivery analyst compares routes whose typical travel times are similar but whose delays differ substantially.
A forecasting team finds larger residual variation for high-volume products than for low-volume products, suggesting uncertainty should depend on scale.

Think of it this way

Think of two archers whose arrows are centered on the same point. One has a tight cluster and the other a wide scatter. Equal centers do not imply equal consistency, and comparing only the average location misses the difference.

Worked example

Consider three illustrative processing times for each of two groups. Group A has 8, 10, and 12 minutes; group B has 4, 10, and 16 minutes. Both sample means are 10 minutes.
For A, squared deviations from the mean sum to 4 + 0 + 4 = 8. Dividing by n - 1 = 2 gives sample variance 4 minutes squared and sample standard deviation 2 minutes.
For B, squared deviations sum to 36 + 0 + 36 = 72. Its sample variance is 36 minutes squared and standard deviation is 6 minutes. B's variance is nine times A's, while its standard deviation is three times as large.
For an independent-means comparison, the Welch standard error is sqrt(4/3 + 36/3) = 3.65148 minutes. The observed mean difference is zero, so the corresponding t statistic is zero. That says nothing about equality of spread: the mean-comparison question differs from the variance-comparison question.
With only three observations per group, these descriptive differences are not strong evidence about long-run population variability. More representative observations would be needed for a dependable operational conclusion.

Illustrative case study

Fictional case: Amber Support compares response times from two help desks. Their average times are close, but one desk has alternating quiet and overloaded periods. A single pooled standard deviation conceals this pattern.
The analyst plots times by desk and shift, checks whether repeated requests from the same customer create dependence, and investigates staffing differences. Mean comparisons use an approach appropriate to unequal variability, while service reports include separate spread measures and upper percentiles.
Management discovers that consistency is the relevant customer concern. It changes shift coverage and gathers a fresh evaluation sample. The original variance difference is treated as a finding to explain, not as proof that one desk's staff are less capable.

Common misconceptions and mistakes

Avoid equating unequal variances with unequal means, interpreting a variance ratio as a standard-deviation ratio, or using a nonsignificant preliminary test as proof of equality. Do not delete valid extremes merely to make group spreads similar.
Pooling data from distinct conditions can hide heterogeneity. Conversely, raw spread differences may disappear after an appropriate mean model explains important predictors.

Business use

Heterogeneity of variance helps identify inconsistent service, uneven model reliability, and changing uncertainty across operating conditions. Teams should compare variability on meaningful scales and use methods that reflect the sampling design. If reliability drives decisions, monitor spread directly rather than treating it only as a statistical nuisance.

Check your understanding

In the example, are the means different?

Answer: No, both equal 10 minutes.

Is B's standard deviation nine times A's?

Answer: No, it is three times as large.

Would a mean test establish equal variability?

Answer: No; it targets a different hypothesis.

Related terms

Learn next

Compare plots of equal-mean distributions with different spreads. Then study Welch's method and regression residual diagnostics, paying attention to the difference between marginal outcome variability and conditional error variability.

Sources

David Howell, glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html NIST, Levene test: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm NIST, two-sample t-test: https://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm NIST, model residuals: https://www.itl.nist.gov/div898/handbook/pri/section2/pri24.htm

Previous: Hessian Next: Heterogeneous subsamples