Heteroscedasticity and Scale Test¶
Diagnostic estimation — instantiates Residual-Driven Model Refinement
Tests whether residual spread stays constant or grows with the fitted value or a predictor; scale-dependent variance means the model's error structure — not just its mean — is misspecified.
A model can predict the right average and still be wrong about its own uncertainty. The Heteroscedasticity and Scale Test checks the assumption that residual variance is constant across the range of predictions — homoscedasticity — by asking whether the spread of the residuals fans out (or narrows) with the fitted value or with some predictor. When it does, the model's noise envelope is scale-dependent: its confidence intervals are too tight in one region and too loose in another, and every downstream interval, test, and risk estimate inherits the distortion. What makes this test distinct is that it reads the size of the residuals against level — the second moment — rather than their correlation across order or their location within subgroups.
Example¶
A property-valuation model predicts home prices, and its residuals average out to roughly zero. Plotting the residuals against the predicted price reveals a clear funnel: errors of roughly ±$10k on modest homes widen to roughly ±$120k on the most expensive ones. A formal test regressing squared residuals on the fitted values rejects constant variance. Nothing is wrong with the average prediction — but the model's stated uncertainty is fiction, far too confident on cheap homes and far too vague on expensive ones. The scale pattern points at the fix: model price on a log scale (or use weighted least squares), after which the spread stabilizes and the intervals mean what they claim.
How it works¶
- Look at spread, not sign. Work with squared or absolute residuals, since the question is about magnitude, not direction.
- Plot against level. Examine residual spread versus the fitted value and versus each candidate predictor; a funnel or fan is the signature.
- Test formally. Regress squared residuals on predictors (Breusch–Pagan) or on the fitted values and their squares (White) to get a significance verdict on constant variance.[n1]
- Read the remedy off the pattern. Variance that scales with the mean points to a transform; variance that jumps by group points to group-wise weighting.
Tuning parameters¶
- What variance is tested against — the fitted value, a specific predictor, or group membership. Each targets a different form of scale dependence.
- Test choice — general tests (White) catch more forms of heteroscedasticity but have less power against any one; targeted tests (Breusch–Pagan on a named predictor) are sharper but narrower.
- Remedy stance — whether a finding routes to a variance-stabilizing transform, weighted least squares, or robust standard errors — each treats the symptom differently.
- Grouping — for group-wise variance comparison, how the groups are defined.
When it helps, and when it misleads¶
Its strength is exposing a failure that leaves the mean prediction untouched and therefore hides from ordinary accuracy checks: a model can be unbiased and still mis-state its own reliability everywhere. Catching that repairs every interval and significance test built on top of it.
Its subtle trap is misreading the cause. Heteroscedasticity is frequently a symptom of a mean-model problem — an omitted variable or wrong functional form — rather than genuine non-constant noise, so treating it purely as a variance nuisance patches the surface and leaves the omission. The classic misuse is reaching straight for robust standard errors to make the test statistics behave, without asking what the scaling reveals about missing structure. The discipline is to read the variance pattern as a clue about the model, not just a correction to apply.
How it implements the components¶
uncertainty_and_noise_envelope— it directly characterizes the noise envelope and tests whether it is constant; a scale-dependent envelope is precisely its finding.residual_adequacy_criteria— constant variance (homoscedasticity) is the specific adequacy criterion it holds the residuals to.
It does not test serial or order structure — that is the Autocorrelation and Whiteness Test; it does not localize error to subgroups — that is the Cross-Validated Error-Slice Report; and it does not check distributional shape or tail behavior — that is the Quantile-Quantile Residual Check.
Related¶
- Instantiates: Residual-Driven Model Refinement — it supplies the verdict on whether the model's uncertainty, not just its mean, is correctly specified.
- Sibling mechanisms: Autocorrelation and Whiteness Test · Quantile-Quantile Residual Check · Cross-Validated Error-Slice Report · Residual-versus-Fitted Plot · Model-Revision Experiment Log
Editorial Notes¶
Form Classification¶
Form family: Assessment, Review & Assurance
Rationale: Heteroscedasticity and Scale Test operates as a bounded evaluation of existing evidence or work that produces a finding or disposition because it tests whether residual spread stays constant or grows with the fitted value or a predictor; scale-dependent variance means the model's error structure — not just its mean — is misspecified
Independent corroboration: The frozen evidence defines Heteroscedasticity and Scale Test as 'Tests whether residual spread stays constant or grows with the fitted value or a predictor; scale-dependent variance means the model's error structure — not just its mean — is misspecified', so its operative form is Assessment, Review & Assurance.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Breusch-Pagan and White tests are named statistical diagnostics for non-constant residual variance and model misspecification.
Review resolution: Both reviewers independently assign statistics_experimental_design as the primary originating domain, so that shared primary is retained. Alternate domains are the union of reviewer-identified formative or independently originating lineages; later application settings alone are excluded. The evidence describes one principal historical lineage. It has established independent use across several domains, but that does not make it domain-free. The encyclopedia entry generalizes the established mechanism without creating a new composite lineage.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] The Breusch–Pagan test regresses the squared residuals on the model's predictors and tests whether they explain any of the residual variance; White's test generalizes this to include the fitted values and their squares, catching more forms of non-constant variance at some cost in power. ↩