Heteroskedasticity: OLS Coefficients Stay Unbiased—So Why Can Your Conclusion Be Wrong?

Quick takeaway

Heteroskedasticity does not generally bias OLS coefficient estimates, but it does make the usual standard errors (and therefore t-tests and confidence intervals) unreliable. The fix is usually to keep the estimated coefficients and use heteroskedasticity-robust (sandwich) standard errors for valid inference.

What is heteroskedasticity?

Heteroskedasticity means the variance of the error term changes with covariates: Var(u | X) is not constant. In many real-world datasets the spread of residuals grows or shrinks with certain predictors (e.g., income dispersion rising with age, measurement noise increasing with magnitude).

Why OLS coefficients remain unbiased

Unbiasedness of OLS estimates rests primarily on the assumption E[u | X] = 0 (errors have zero conditional mean). That condition can hold even if Var(u | X) varies with X. So the point estimates β̂ produced by OLS remain (conditionally) unbiased and consistent under heteroskedasticity.

Why inference breaks

The usual OLS standard error formula assumes homoskedasticity (constant variance). Under homoskedasticity:

Var(β̂) = σ² (X'X)^{-1}.

When errors are heteroskedastic the true variance becomes:

Var(β̂) = (X'X)^{-1} X'ΩX (X'X)^{-1},

where Ω is a diagonal matrix of error variances. If you keep using the homoskedastic formula you get wrong standard errors — often underestimating or overestimating uncertainty. That leads to misleading t-statistics, p-values, and confidence intervals: you might falsely declare effects significant (Type I error) or miss real effects (Type II error).

Fast fixes and best practices

• Use heteroskedasticity-robust (sandwich) standard errors. These adjust the estimated Var(β̂) to account for varying error variance and restore valid hypothesis testing.
• Popular robust estimators: HC0, HC1, HC2, HC3 (HC3 often recommended in small samples because it's more conservative).
• Consider clustering if observations are correlated within groups (cluster-robust SEs).
• Run diagnostic tests to detect heteroskedasticity (e.g., Breusch–Pagan, White test) — but even if you skip tests, robust SEs are inexpensive and usually appropriate as a precaution.

Quick code examples

• In Python (statsmodels):

  from statsmodels.api import OLS res = OLS(y, X).fit(cov_type='HC3')

• In R (sandwich & lmtest):

  library(sandwich); library(lmtest) model <- lm(y ~ X) coeftest(model, vcov = vcovHC(model, type = "HC3"))

Practical tips

• Report robust standard errors in applied work when heteroskedasticity is plausible.
• If data are clustered (e.g., students within schools), use cluster-robust SEs rather than simple heteroskedasticity-robust SEs.
• In very small samples be cautious: some robust methods still rely on large-sample approximations; prefer HC3 or permutation/bootstrap approaches when appropriate.

Bottom line

Heteroskedasticity rarely ruins your β̂ — but it can wreck your inference. The simplest reliable approach is to keep the OLS coefficients and use heteroskedasticity-robust (or cluster-robust) standard errors so your t-tests and confidence intervals are trustworthy.

#DataScience #Statistics #MachineLearning