Checking Covariate Overlap Across Treatment Groups

Why Covariate Overlap Matters

In ANCOVA, a regression line describes how the mean response changes with a covariate within a treatment group. When the covariate values observed in the groups differ, a comparison of those lines may depend on how well the data support the comparison across the covariate’s scale. If one group has little or no data in part of the range observed for another group, a fitted line can still extend there—but that extension is an extrapolation, not a comparison directly backed by observations from both groups.

Checking overlap is therefore a diagnostic step, not a formal test of equal slopes. It helps identify which parts of the covariate scale offer a credible basis for comparing group-specific relationships, and which parts require caution. A broad overlap in ranges is a useful first check, but the locations and density of observations also matter.

Model and Definitions

Let \(Y_{ij}\) be the response for observation \(i\) in treatment group \(j\), and let \(X_{ij}\) be its quantitative covariate. A group-specific linear mean model can be written as

Here \(\alpha_j\) is the group’s mean response at the chosen covariate reference value \(c\), \(\beta_j\) is its slope, and \(\epsilon_{ij}\) is the deviation from the group-specific mean. For the usual linear-model analysis, the conditional mean is assumed linear in \(X\) within each group, errors have conditional mean zero and are independent, and a common error variance is used for conventional pooled inference. Normal errors are additionally used for exact small-sample inference. The overlap check concerns where the covariate values occur; it does not establish these model assumptions.

For two groups with ranges \([L_1,U_1]\) and \([L_2,U_2]\), their range overlap is \([\max(L_1,L_2),\min(U_1,U_2)]\) when the lower endpoint does not exceed the upper endpoint. If it does, their observed ranges do not overlap. With more than two groups, a range shared by all groups is the intersection across every group under consideration. The relevant covariate units should always be stated: an overlap of 5 to 10 years has a different practical meaning from an overlap of 5 to 10 millimetres.

Start With Ranges, Then Examine Distributions

A range comparison answers whether the groups’ minimum-to-maximum intervals intersect and how wide that intersection is. It does not show whether observations are spread throughout the intersection. Two groups can have wide overlapping ranges but have most of their observations in different parts of them. A histogram, density plot, or dot plot by group can reveal that pattern. Group-specific sample sizes, means, standard deviations, and counts within relevant covariate intervals provide useful context for the plots.

For a continuous covariate, exact matches between individuals in different groups are not generally required. The important question is whether both groups have observations reasonably near the covariate values used for a comparison, with enough spread to estimate each group’s slope. A few observations near the extremes may technically create range overlap while providing little support there. There is no universal percentage or minimum number that defines adequate overlap in every application; the degree of support and its practical relevance require judgment.

Range Overlap Can Hide Distribution Gaps

The minimum and maximum are only two features of a distribution. They can be separated by many unobserved covariate values, or a group can have a cluster near one end and little data elsewhere. In those situations, reporting a range intersection alone can give an overly reassuring impression of common support.

Why Extrapolation Changes the Interpretation

A fitted group-specific line can be evaluated at covariate values beyond the range observed for that group. The calculation is mathematically defined, but its interpretation depends on extending the assumed linear relationship into a region with no observations from that group. The data cannot show whether the relationship continues in the same way there.

The consequence of extending a line becomes clearer by examining how a slope error affects a prediction. Suppose a group’s upper observed covariate boundary is \(U\), and a prediction is made at \(x^*>U\). If the estimated slope differs from the true slope by \(d\), the slope-related prediction difference from the boundary to \(x^*\) is

This follows because the fitted change from \(U\) to \(x^*\) uses the estimated slope, while the true linear change uses the true slope; subtracting the two changes leaves the slope difference multiplied by the distance. The farther \(x^*\) is beyond \(U\), the more a given slope discrepancy can affect the extended prediction. This is one reason a comparison far outside a group’s observed covariate range should be treated cautiously.

Lack of overlap does not make a fitted slope calculation impossible, nor does it prove that the fitted model is wrong. It does mean that a cross-group interpretation over unsupported covariate values depends more on model assumptions and less on observations from both groups. If the response relationship bends, changes shape, or differs from the assumed line in the unsupported region, that interpretation may be misleading.

A Practical Overlap Check

A clear overlap assessment combines numerical summaries with a visual display. The following sequence is useful before interpreting slope comparisons:

The supported region is not automatically identical to the range intersection. It may be narrower when observations are sparse near a boundary or when the distributions have gaps. Conversely, for a particular scientific question, a modest number of well-spread observations may be practically informative. The plot and context matter; there is no universal cutoff that turns overlap into a yes-or-no verdict.

Common Mistake

Incorrect reasoning: “The groups’ minimum and maximum covariate values overlap, so slope comparisons are supported throughout that interval.” This conclusion treats every value between the endpoints as if it were similarly represented in both groups.

Correction: The intersection of the ranges is only a first check. Inspect group-specific distributions and covariate counts to see whether both groups have observations across the relevant region. If a group has a gap or only a few observations near part of the interval, describe that weakness. Keep interpretations within regions with meaningful support, and identify any extension beyond a group’s observed range as extrapolation.