ECON 3209 · Week 12, Lecture 1 · Kerala Agricultural University
Autumn 2026
By the end of this lecture, you will be able to:
In a homoskedastic model,
\[Var(\varepsilon_i \mid X_i) = \sigma^2\]
Under heteroscedasticity,
\[Var(\varepsilon_i \mid X_i) = \sigma_i^2\]
so the spread of the errors changes with the regressors.
Cross-sectional data often contain units of very different scale, and scale differences naturally create unequal variance.
If the zero conditional mean assumption still holds, OLS coefficients remain unbiased.
But under heteroscedasticity: - OLS is no longer efficient among linear unbiased estimators. - Usual homoskedastic standard errors are biased. - t-tests and confidence intervals can be misleading.
The usual textbook variance formula assumes constant error variance.
When that assumption fails, software that uses the homoskedastic formula may understate or overstate uncertainty.
So the biggest danger is often false confidence in hypothesis tests.
Residual plots are informal, but they are often the first useful diagnostic.
Heteroscedasticity is mainly an inference problem unless the variance structure itself is of substantive interest.
✅ Heteroscedasticity means the variance of the error term changes across observations.
✅ It is very common in cross-sectional data such as income, farm size, and expenditure data.
✅ OLS coefficients can remain unbiased, but usual standard errors become unreliable.
✅ Residual plots often reveal funnel or megaphone patterns.
✅ The main danger is incorrect inference, not necessarily biased coefficient estimates.
Lecture 2 — Detection Tests
We will cover: - Breusch-Pagan and White tests - reading residual plots - Python implementation in statsmodels - interpreting p-values and test outputs
ECON 3209 — Kerala Agricultural University