Interpreting Interactions With Unequal Group Ranges

Why Group Ranges Matter

An interaction model can give each group its own fitted line, but the predictor values represented in the data may differ substantially among groups. One group might be observed mostly at low doses and another mostly at high doses. A fitted line can still produce predictions for both groups at any chosen predictor value, but some of those predictions may lie beyond the values actually observed for one group. Such comparisons rely more heavily on the model’s assumed line shape than on direct data support.

The key diagnostic is range overlap: the extent to which groups have observed predictor values in common. Overlap helps distinguish comparisons made where both groups are represented from comparisons that extend beyond one or more groups’ observed data. It does not, by itself, prove that a comparison is reliable or causal. It clarifies what part of a line comparison is supported by the observed predictor ranges.

Group-Specific Lines and Observed Ranges

For a quantitative response \(Y\), predictor \(x\), and group \(g\), a group-specific linear regression model can be written as

Here, \(a_g\) is the intercept for group \(g\), \(b_g\) is its slope, and \(\varepsilon_{ig}\) is the error for observation \(i\) in that group. The fitted line is \(\widehat{Y}_g(x)=\widehat{a}_g+\widehat{b}_gx\). Its slope describes the fitted change in response for a one-unit increase in \(x\), within that group, under the linear model.

For two groups with ranges \([L_1,U_1]\) and \([L_2,U_2]\), their common range is \([\max(L_1,L_2),\min(U_1,U_2)]\), provided the lower endpoint is no greater than the upper endpoint. If it is greater, the ranges do not overlap. With more groups, the same idea applies: the common range begins at the largest lower endpoint and ends at the smallest upper endpoint.

A linear model’s usual working assumptions include that the mean response is adequately represented by a line within each group and that errors have mean zero and are independent across observations. Constant error variance is commonly assumed for standard regression inference; normal errors may also be used for exact small-sample inference. These assumptions do not make extrapolation equivalent to observing data at the extrapolated values. Rather, outside a group’s observed range, the prediction depends on the assumed linear pattern continuing there.

Finding and Interpreting the Shared Range

Before comparing fitted lines at a predictor value, check whether that value falls in the observed range of every group in the comparison. If so, the comparison is within the observed ranges. If not, at least one fitted outcome is extrapolated. With several groups, a value can lie in the overlap of one pair but outside the range of another group; the relevant overlap depends on which groups are being compared.

Why Comparisons Outside the Overlap Can Mislead

The fitted difference between two groups at a chosen value of \(x\) is obtained by subtracting their fitted lines. For groups 1 and 2,

This expression is a difference between model predictions. Its arithmetic is valid wherever the lines are evaluated, but its empirical support varies with \(x\). Within the groups’ common range, both fitted outcomes are evaluated where observations occurred for each group. Outside that range, at least one line is being extended beyond its group’s observed data.

This distinction matters for both slope and group contrasts. A slope is estimated from how outcomes vary over the predictor values observed in that group. A group observed over a narrow range may provide little information about how its line behaves far outside that range. Similarly, two lines may cross at a value where one group has no observations. The crossing is then a feature of the fitted equations, not evidence that observed responses for the groups actually switch order at that value.

No overlap does not make a regression calculation impossible, and overlap alone does not guarantee a sound comparison. Model-based comparisons outside observed ranges may be defensible when the linear form is well supported by subject-matter knowledge and other evidence. But the conclusion should identify the extrapolation and should not present it as if both groups had been observed at the same predictor values.

Narrow Overlap and Unequal Ranges

A small overlap is different from no overlap. With a small overlap, comparisons inside that interval are interpolations for both groups, but the shared values cover only a limited part of each group’s predictor range. A comparison at one point in the overlap may be meaningful as a model-based contrast there, but it should not automatically be generalized to lower or higher predictor values that are not shared.

Unequal ranges can also make the lines differently anchored by observed data. One group’s slope may be estimated across a broad predictor span while another’s is estimated across a narrow span. Even if both fitted lines are straight, their apparent behavior outside the narrow group’s range is particularly dependent on extending its estimated line. The range check describes this limitation; it does not replace an assessment of model fit or statistical uncertainty.

A Short Justification for the Range Check

The need to check the range follows directly from what a fitted line represents. Within group \(g\), the fitted value at \(x\) is calculated as \(\widehat{a}_g+\widehat{b}_gx\), using the estimated intercept and slope. If \(x\) is outside the observed range, there are no observations from that group at or beyond that end of the range to anchor the line there. The prediction still follows from the fitted equation, but its interpretation requires the line to continue beyond the observed predictor values.

For two groups, subtracting their fitted values gives the group contrast shown earlier. If both groups were observed at \(x\), the contrast is evaluated within both observed ranges. If only one was observed there, the corresponding term in the subtraction is an extrapolated fitted value. Therefore, checking each group’s range before interpreting a difference identifies exactly which part of the comparison relies on extrapolation.

Common Mistakes

• Calling any computed contrast an observed comparison. A fitted equation can produce a contrast outside a group’s range, but that does not mean the group was observed at that predictor value. Name the extrapolation.
• Treating a line crossing as a demonstrated reversal. If the crossing lies outside the groups’ common range, it is a model-based crossing, not evidence of a reversal where both groups were observed.
• Assuming some overlap supports every comparison. A comparison at a value outside even one group’s range still extrapolates for that group, despite overlap elsewhere.
• Assuming overlap guarantees a trustworthy model. Overlap addresses range support, not whether the linear form fits, errors meet the assumptions for inference, or the data support a causal claim.