MerQuR+Statistics Glossary

Statistics Glossary → Heteroscedasticity

Heteroscedasticityconcept

A situation in a regression or ANOVA model where the variance of the errors (residuals) changes with the predictors or with the fitted value.

Ordinary least squares assumes the error variance is the same for every observation (homoscedasticity). Under heteroscedasticity, for example, the spread of spending grows as income grows, and the residuals-versus-fitted plot shows a funnel shape. The coefficient estimates remain unbiased, but the classical standard errors are wrong, so t and F tests, p values and confidence intervals become unreliable.

The Breusch-Pagan test regresses the squared residuals on the predictors; the White test also adds the squares and cross-products of the predictors to detect more general forms. Remedies include robust (sandwich, HC) standard errors, weighted least squares, a transformation such as the log of the outcome, or generalized linear models that tie the variance to the mean. Equality of group variances (Levene's test) is the same issue in its ANOVA form.

An everyday example

When electricity use is modelled on house size, small flats use very similar amounts, while large villas vary widely depending on pools, heating and air conditioning. Residuals are around ±20 kWh for small homes and ±200 kWh for large ones. Classical standard errors would then understate the uncertainty for large houses.

In MerQur

Offered as 'Model diagnostics: Heteroscedasticity tests (Breusch-Pagan, White)' in two-way ANOVA and ANCOVA; in two-way ANOVA 'Robust SE (HC)' gives Wald F tests with an HC covariance under variance heterogeneity. In multiple linear regression the model diagnostics give the Breusch-Pagan, Koenker and White tests and 'Standard error type' (HC0-HC3) gives heteroscedasticity-robust standard errors; instrumental variables regression offers HC0, HC1 and HC3, quantile regression the nid or ker standard errors for heteroscedastic data, and Tobit regression heteroscedasticity diagnostics and a sandwich standard error.

Analyses where this concept appears

Two-Way ANOVAANCOVAMultiple Linear RegressionInstrumental Variables Regression (2SLS)Quantile RegressionTobit Regression (Censored Regression)

Related concepts

Breusch-Pagan and White testsRobust (sandwich) standard errors (HC0-HC3)Homogeneity of variance assumption