Heteroscedasticity

Heteroscedasticity describes a pattern in regression where the scatter of the errors — the gaps between what the model predicts and what actually happened — is uneven across the data instead of staying roughly the same. Picture a chart of predicted spend against actual sales: if the points hug the line tightly for small budgets but fan out into a wide cloud for large budgets, you have heteroscedasticity. The opposite, where the spread stays constant, is called homoscedasticity, and it is one of the quiet assumptions behind ordinary least-squares regression. The word itself joins Greek roots meaning "different" and "scatter." It does not bias the model's central estimates of the slopes, but it does corrupt the part of the output that tells you how sure to be — the standard errors that feed every confidence interval and significance test.

The practical danger is false confidence. When error variance balloons for some observations, the formula that computes standard errors assumes a single steady spread, so it reports intervals that are too tight in some regions and too loose in others. A coefficient might look statistically significant when honest math would call it shaky, or the reverse. In marketing analytics this bites hardest in spend-response models, where small accounts behave predictably and large campaigns swing wildly, and in revenue forecasts where bigger customers carry bigger surprises. Because the bias hides in the uncertainty rather than the headline number, an analyst who never checks the residuals can ship a model that looks crisp and is actually fragile. Heteroscedasticity is a reason to inspect residuals, not just R-squared.

Heteroscedasticity is easy to confuse with its cousins, but each breaks a different assumption. Homoscedasticity is simply its absence — constant error spread — and is what tidy textbook regressions assume. Autocorrelation is a separate flaw where errors are correlated across time or order, common in weekly sales series, and it concerns the relationship between residuals rather than their changing size. Non-normality means the errors do not follow a bell curve at all, which is again distinct from whether their spread is steady. A model can suffer one, two, or all three at once. The reason to keep them apart is that the fixes differ: heteroscedasticity calls for robust standard errors or a transformation, autocorrelation calls for time-series methods, and non-normality may call for a different model family entirely.

Diagnosing heteroscedasticity usually starts with the eyes. Plot the residuals against the fitted values and look for a funnel, cone, or megaphone shape; an even, shapeless band suggests you are fine. Formal tests such as Breusch-Pagan or White's test put a number on it, returning a p-value for whether the variance is constant. Once found, the common remedies are weighted least squares (giving noisier observations less influence), heteroscedasticity-consistent standard errors that stay valid despite the uneven spread, or transforming the response — taking the logarithm of revenue or spend, for instance, which often tames a fan that grows with size. The goal is not to erase the pattern but to stop it from poisoning the confidence numbers you report to a marketing team making budget calls.

Handle heteroscedasticity by checking for it as a routine step, not a rescue mission after a result looks odd. Every time you fit a regression that someone will act on — a media-mix model, a price-elasticity estimate, a lift study — plot the residuals and run a formal test before you trust a single p-value. If the spread is uneven, reach first for robust standard errors, since they keep your significance tests honest without changing the model's meaning. If the funnel grows with the size of the outcome, a log transformation of the target frequently both stabilizes variance and makes the relationship more sensible. Document what you found and what you did, so the next analyst does not mistake your corrected intervals for naive ones.

The discipline is to treat the residual plot as part of reading the model, the way you would read the coefficients. Used well, awareness of heteroscedasticity protects you from over-trusting a slick-looking spend-response curve or declaring a channel "proven" on a confidence interval that was never valid. It pairs naturally with incrementality testing and careful segmentation, because the moment you mix small and large accounts in one regression you should expect their error spreads to differ. The failure is to skip the check, report tidy significance, and let a budget decision ride on a number that the uneven variance had already undermined.

Heteroscedasticity joins Greek roots for "different" and "scatter" and names non-constant error variance in regression, a violation that corrupts standard errors and significance tests.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where heteroscedasticity is a core concern: