Discuss methods of detecting the presence of heteroscedasticity. Explain the Goldfeld-Quandt test for detecting it, stating the assumptions required to apply this test.
10 marks · directive: Discuss
The core of a good answer
Heteroscedasticity can be detected informally by plotting squared residuals against fitted values or regressors, and formally by tests such as Park, Glejser, Goldfeld-Quandt, Breusch-Pagan and White; the Goldfeld-Quandt test orders observations by the regressor thought to drive the variance, omits c central observations, fits separate OLS regressions to the two end groups and compares λ = (RSS₂/df)/(RSS₁/df) with F[(n − c − 2k)/2, (n − c − 2k)/2], assuming normal, serially independent errors whose variance is monotonically related to that regressor.
An outline written by upsc.help, not an official UPSC answer; every fact checked by a separate review. How we work
What the question asks
Two parts: a survey of detection methods (graphical and formal, with one line on each), then the Goldfeld-Quandt procedure step by step with its test statistic, degrees of freedom and assumptions, and its limitations.
Answer outline
Introduction
Heteroscedasticity means Var(uᵢ) = σᵢ² varies across observations. OLS stays unbiased but is no longer efficient, and the usual standard errors, t and F tests are invalid.
Informal methods
- Nature of the data: cross-section data on units of very different size (household budgets, firm sizes) are prone to it.
- Residual plots: plot eᵢ² against Ŷᵢ or against each Xⱼ; a fan or funnel shape suggests heteroscedasticity.
Formal tests
- Park and Glejser tests: regressions of a function of the residuals on the suspected regressor.
- Breusch-Pagan (1979): regress the squared residuals on the explanatory variables and use a chi-squared statistic.
- White (1980): a heteroscedasticity-consistent covariance estimator that also gives a direct test.
Goldfeld-Quandt test: procedure
- Assume σᵢ² is positively related to one regressor, for example σᵢ² = σ²Xᵢ².
- H₀: homoscedasticity; H₁: variance increases with X.
- 1. Order the n observations by increasing X.
- 2. Omit c central observations; split the rest into two groups of (n − c)/2.
- 3. Fit OLS separately to each group; get RSS₁ (small X) and RSS₂ (large X), each with (n − c)/2 − k degrees of freedom (k parameters including the intercept).
- 4. λ = [RSS₂/((n − c − 2k)/2)] / [RSS₁/((n − c − 2k)/2)] = RSS₂/RSS₁.
- 5. Under H₀, λ ~ F with (n − c − 2k)/2 and (n − c − 2k)/2 d.f.; reject H₀ if λ exceeds the critical F.
Assumptions and limits
- Errors normally distributed and serially independent (needed for the exact F distribution).
- Variance is a monotonic function of a single, known ordering variable.
- c must be chosen: omitting central observations sharpens the contrast but too large c leaves few degrees of freedom.
- It needs a known ordering variable with variance monotonic in it (a quadratic relation can make it wrongly accept homoscedasticity), and it cannot separate heteroscedasticity from misspecification such as an omitted variable.
Conclusion
Graphical inspection followed by a formal test is standard; the Goldfeld-Quandt test is a simple exact F test when the variance is believed to rise or fall with one regressor and errors are normal.
Examples and data to cite
- Goldfeld and Quandt published the test in the Journal of the American Statistical Association in 1965; Breusch and Pagan (1979) and White (1980) followed.
- In household expenditure studies, the variance of food spending rises with income: a classic case for the Goldfeld-Quandt test with income as the ordering variable.
Facts to remember
- Under heteroscedasticity OLS is unbiased but not BLUE; usual standard errors are biased.
- GQ statistic λ = RSS₂/RSS₁ ~ F((n − c − 2k)/2, (n − c − 2k)/2).
- Remedies: weighted least squares, transformation (logs), or White's robust standard errors.