Does a significant interaction make adjustment impossible?

No; it calls for a model and interpretation that allow the relationship to vary.

AI Glossary / Statistics and probability / Intermediate

Homogeneity of regression

Definition

Homogeneity of regression usually means equality of the covariate regression slopes across groups in a common-slope analysis of covariance model. The expected outcome changes at the same rate with the covariate in every group. The regression lines may have different intercepts: equal slopes do not require identical lines, equal group means, or equal residual variances.

What it means

When comparing groups after adjusting for a baseline characteristic, a common-slope model assumes that the adjustment works in the same way for each group. For example, each additional baseline-score point might be associated with two additional outcome points in both groups.
If that relationship differs by group, there is a group-by-covariate interaction. Then the difference between groups depends on the covariate value, so a single constant adjusted difference can conceal an important pattern.

How it works

Specify the conditional mean relationship. For two groups coded G = 0 and G = 1, the common-slope model is E(Y|X,G) = b0 + b1*X + b2*G. Group slopes both equal b1, while the group difference at any common X equals b2.
An interaction model adds b3*X*G. Its slopes are b1 for group zero and b1 + b3 for group one. Homogeneity of slopes corresponds to b3 = 0 in this linear specification. With several groups, the relevant interaction terms are assessed together.
Plot outcomes against the covariate by group, with fitted relationships and uncertainty. Check whether a linear form is appropriate before focusing only on slope equality. Curvature, outliers, limited covariate ranges, or a few influential observations can affect the apparent comparison.
An interaction test can assess evidence against equal slopes under the model assumptions. Failure to reject does not prove equality; the study may have little power to detect meaningful differences. Consider estimated slope differences and their uncertainty, rather than treating a preliminary p-value as a definitive gate.
If slopes differ meaningfully, retain an appropriate interaction model and compare groups at clearly specified, relevant covariate values. A more flexible model may be needed when the relationship is nonlinear. Unequal slopes do not mean that all covariate-adjusted analysis must stop.
Check overlap. Comparing both groups at a covariate value outside one group's observed range relies on extrapolation. An adjusted mean can look precise in a table while depending on an unsupported extension of the fitted relationship.
Distinguish this assumption from other requirements. Independent observations, an appropriate error model, valid measurement, and a defensible study design remain separate issues. Equal slopes do not guarantee equal residual variance or normal errors.
For causal comparisons, covariate choice and timing also matter. Adjusting for a variable changed by treatment can alter the estimand or introduce bias. A common-slope fit alone does not turn an observational group difference into a causal treatment effect.

Three practical examples

A training study compares final scores after accounting for a baseline assessment and examines whether the baseline relationship differs by program.
An AI benchmark compares systems across task difficulty while checking whether their performance slopes are parallel.
A service study compares processing times across teams after accounting for request size, allowing interaction when size has different associations across teams.

Think of it this way

Think of two ramps with the same steepness but different starting heights. Moving the same distance along either ramp changes height equally. Different starting heights are compatible with equal slopes; different steepness is what breaks the common-slope assumption.

Worked example

Assume illustrative conditional mean models for groups A and B: E(Y|X,A) = 10 + 2X and E(Y|X,B) = 15 + 2X. Both slopes are 2, so the common-slope assumption holds even though the intercepts differ.
Suppose A's average covariate is 3 and B's is 6. Under these stipulated linear models, their corresponding expected group means are 10 + 2 x 3 = 16 and 15 + 2 x 6 = 27. The unadjusted expected difference is 11.
At a common covariate value X = 4, the expected outcomes are 18 and 23, giving an adjusted difference of 5. The change from 11 to 5 reflects the different covariate distributions under the assumed model, not a mathematical contradiction.
Now change B's model to 15 + 3X while retaining A's model. The difference B minus A becomes 5 + X: it is 7 at X = 2, 9 at X = 4, and 11 at X = 6. There is no longer one constant group difference.
These are assumed population mean relationships, not estimated study results. Real analysis would require data, overlap checks, and uncertainty estimates before deciding whether common slopes are plausible or whether an adjusted contrast is interpretable.

Illustrative case study

Fictional case: Birch Learning compares two AI-assisted tutorials after measuring learners' baseline skill. An initial report presents one adjusted average difference. A plot suggests that the tutorial gap may grow with baseline skill.
The analyst fits a group-by-baseline interaction, checks influential learners, and examines whether both tutorials include participants across the relevant skill range. The team reports contrasts at a few prespecified baseline values with uncertainty instead of extrapolating to skill levels absent from one group.
Because participants chose their tutorials, the report describes adjusted associations and discusses remaining differences between participants. It does not claim that testing slope homogeneity establishes causal comparability.

Common misconceptions and mistakes

Common mistakes include assuming equal slopes means equal intercepts, confusing slope homogeneity with variance homogeneity, and declaring equality solely because an interaction test is nonsignificant.
Do not delete legitimate influential observations merely to obtain parallel lines. Also avoid presenting adjusted means outside the observed covariate overlap without explaining the extrapolation and its limitations.

Business use

Checking homogeneity of regression helps determine whether one adjusted group comparison is an adequate summary. When a business decision varies with baseline conditions, an interaction may be more informative than a single overall difference. Reports should identify the comparison values, data support, and uncertainty behind conditional conclusions.

Check your understanding

Do the first two model lines have equal slopes?

Answer: Yes, both equal 2.

What is their difference at X = 4?

Answer: Five outcome units.

After B's slope changes to 3, why must X be stated?

Answer: The group difference becomes 5 + X.

Related terms

Analysis of covariance
Adjusted means
regression slope
interaction
adjusted mean
covariate
extrapolation
common support
effect modification

Learn next

Compare parallel-line and interaction models using plots and conditional contrasts. Then study how covariate overlap, measurement, and causal timing affect the meaning of an adjusted group comparison.

Sources

David Howell, glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html NCSS, one-way ANCOVA: https://www.ncss.com/wp-content/themes/ncss/pdf/Procedures/NCSS/One-Way_Analysis_of_Covariance-ANCOVA.pdf Newcastle University, ANCOVA guidance: https://services.ncl.ac.uk/itservice/research/dataanalysis/advancedmodelling/ancova/

Previous: Holt-Winters forecasting Next: Homogeneity of variance