Heteroskedasticity
Also known as · heteroscedasticity · non-constant variance
When the variance of the error term in a regression varies with the regressors. Violates one of the Gauss–Markov assumptions, so OLS Estimator is unbiased but no longer BLUE — standard errors are wrong.
When to use this concept: any time you suspect that residuals fan out (or shrink) with a regressor. Diagnose with the recipe Testing for heteroskedasticity.
Variance of the OLS estimator under homoskedasticity:
Under heteroskedasticity this formula no longer applies — use White's robust standard errors instead.