Group-Specific Slopes With Indicator Interactions

Why Allow Group-Specific Slopes?

A regression model with a quantitative predictor and a group indicator can allow the groups to have different average outcomes. If the model includes only those two terms, it assumes the groups’ regression lines are parallel: the expected outcome changes at the same rate with the predictor in both groups. That assumption may not fit a situation in which the rate of change itself differs by group.

An indicator interaction allows that rate to differ. The interaction is the product of a group indicator and the quantitative predictor. With this term, the model can represent one line for the reference group and another line for the comparison group, with both different slopes and different intercepts.

The Model and Its Group-Specific Lines

Let \(X\) be a quantitative predictor, and let \(G\) indicate group membership, with \(G=0\) for the reference group and \(G=1\) for the comparison group. For an individual \(i\), fit the model

Here, \(\varepsilon_i\) is the error for individual \(i\). The model describes the conditional mean of \(Y\) given the predictor and group. Substituting each possible indicator value gives the two lines:

Thus, \(\beta_0\) and \(\beta_1\) are the reference group’s intercept and slope. The comparison group’s intercept is \(\beta_0+\beta_2\), and its slope is \(\beta_1+\beta_3\). The coefficient \(\beta_2\) is the difference in intercepts, while \(\beta_3\) is the difference in slopes, comparison group minus reference group.

At a particular predictor value \(x\), the comparison group’s expected response minus the reference group’s expected response is

Therefore, \(\beta_2\) alone is not generally the difference between the groups at every predictor value. It is the difference at \(X=0\). When slopes differ, the expected group difference changes with \(X\).

Assumptions and What the Model Says

The interaction changes the form of the conditional mean model; it does not remove the assumptions needed to use a regression model. The model assumes that, within each group, the expected response is linear in \(X\). In the usual regression setup, the errors have conditional mean zero, \(E(\varepsilon_i\mid X_i,G_i)=0\), and observations are independent. Standard pooled regression inference also commonly assumes a common error variance across predictor values and groups; exact small-sample inference additionally uses a normal-error assumption. Those inference assumptions should be assessed before drawing inferential conclusions.

The groups need observations at predictor values that make their lines estimable. If a group has no variation in \(X\), its slope cannot be estimated from its own observations. Comparisons between groups at predictor values far outside the observed data range are extrapolations and should be treated cautiously. The model describes conditional means; by itself, it does not establish that group membership or \(X\) causes a change in the response.

The coefficients are commonly estimated by least squares: choose the coefficient values that minimize the sum of squared residuals, where each residual is observed \(Y_i\) minus its fitted value. With the interaction in the model, the fitted values can follow different slopes in the two groups.

Why the Substitution Works

The interaction’s role follows directly from multiplying the indicator by \(X\). When \(G=0\), both terms involving \(G\) are zero, leaving only the reference group’s line. When \(G=1\), the indicator term adds \(\beta_2\) to the intercept, and the product term adds \(\beta_3X\) to the predictor part of the line:

This also explains why omitting the interaction imposes equal slopes: without the \(XG\) term, the coefficient of \(X\) remains \(\beta_1\) for both groups. The indicator alone can shift one line up or down, but cannot change its slope.

Changing the Reference Group

The choice of reference group determines which line is represented directly by \(\beta_0+\beta_1X\). Reversing the indicator changes the coefficient values and their interpretations, but does not change the two fitted lines or predictions, provided the model is otherwise equivalent.

Centering the Predictor and Avoiding Misinterpretation

If \(X=0\) is not a meaningful value, the intercept difference \(\beta_2\) may have little direct practical meaning. Centering \(X\) at a useful value can make that coefficient more interpretable. For example, in the blood pressure example, \(X=0\) represents age 50, so \(\beta_2=-4\) describes the fitted Group B minus Group A difference at age 50. Centering changes the intercept-related coefficients, but not the two group-specific slopes or their difference.

A frequent mistake is to interpret \(\beta_2\) as the group difference at all values of \(X\), or to interpret \(\beta_3\) as the outcome difference between groups. The correction is to calculate the group difference at the specified predictor value using \(\beta_2+\beta_3x\); \(\beta_3\) is a difference in slopes, measured in response units per predictor unit. Another mistake is to say the comparison group’s slope is \(\beta_3\). Its slope is \(\beta_1+\beta_3\); \(\beta_3\) alone is the amount by which that slope differs from the reference group’s slope.

This model is useful whenever a quantitative predictor may relate to the response differently across two groups. Reporting both fitted lines, with units for slopes and a clear reference-group definition, makes the interaction interpretable. A coefficient table alone can obscure the fact that its terms combine to form group-specific intercepts, slopes, and fitted differences.