Comparing Adjusted Group Means With Contrasts

Why Compare Adjusted Group Means?

Suppose two clinics serve patients with different age distributions. A comparison of their raw average recovery times combines differences between clinics with differences in the ages of the patients they serve. A regression model can compare the clinics at a common age, separating the modeled group difference from the modeled association between age and recovery time. That comparison is an adjusted group-mean comparison.

“Adjusted” does not mean that a group’s observed data have been changed. It means that the fitted model is used to predict the response for each group at the same specified values of the covariates. Those values define the reference profile for the comparison. The profile matters: a comparison at age 50 is not automatically a comparison averaged over every age in a population.

Model and Definition

For clarity, begin with two groups, A and B, and one quantitative covariate \(x\). Use the indicator \(I(B_i)\), which equals 1 for an observation in group B and 0 for an observation in group A. A common-slope regression model is

Here \(x_0\) is the chosen reference value of the covariate. Centering \(x\) at \(x_0\) makes the intercept the fitted mean for group A at that value. For multiple covariates, the same idea applies: specify a reference profile, and use the fitted model to predict each group’s mean response at that profile.

In the two-group model above, the adjusted mean for A at \(x=x_0\) is \(\beta_0\), and the adjusted mean for B at \(x=x_0\) is \(\beta_0+\beta_1\). Thus the B-minus-A contrast is

The intercept cancels, leaving the group coefficient. With three or more groups, treatment-indicator coefficients are differences from a reference group; comparing two nonreference groups generally requires subtracting their coefficients. Whatever the parameterization, the requested comparison should be stated in terms of the adjusted means and their subtraction order.

Conditions for Inference

A contrast test relies on the regression model being suitable for the data. In the common-slope model, the expected response is linear in the quantitative covariates, and the group differences are constant across covariate values. The errors have mean zero given the predictors, are independent, and have a common variance. For exact small-sample t inference, the errors are also normally distributed. The design matrix must have enough independent information to estimate the coefficients, and the covariate profile should be supported by the observed data rather than being an extreme extrapolation.

A group-by-covariate interaction would allow the group difference to vary with the covariate. If that pattern is needed, a single common adjusted difference is not an adequate summary; the comparison must be stated at a particular covariate value using a model that includes the interaction. In the examples below, the stated diagnostic summaries and study designs are hypothetical and are taken to support the common-slope model.

Testing a Difference Between Adjusted Means

Write the desired adjusted-mean difference as a contrast \(L\) of the regression coefficients. Its estimate is the same linear combination of the estimated coefficients. The standard error must reflect the uncertainty in all coefficients involved, including their covariances. Under the regression assumptions, a contrast can be tested by comparing its estimate with its null value using a t statistic and the model’s residual degrees of freedom.

Under \(H_0\) and the regression error assumptions, this statistic has a t distribution with the fitted model’s residual degrees of freedom. For a two-sided test of equal adjusted means, \(L_0=0\) and \(H_a:L\ne 0\). The p-value is the probability, assuming \(H_0\) and the model, of obtaining a t statistic at least as far from zero as the observed statistic. A confidence interval for the difference is the estimate plus or minus the appropriate t critical value times its standard error.

Worked Comparisons

Why the Reference Profile and Covariance Matter

An adjusted mean is not a raw group average, and it is not automatically an average over a particular population. It is a model-based prediction at the covariate values used to define it. For instance, an age-50 comparison describes expected outcomes at age 50; it does not by itself describe the average difference among all patients of every age. When several covariates are included, identify their reference values as well. A profile that is far outside the observed data can produce a precise-looking but poorly supported extrapolation.

Covariance matters because estimated regression coefficients are generally not independent. For two coefficient estimates \(\widehat{\beta}_1\) and \(\widehat{\beta}_2\), the variance of their difference is

This follows from expanding the variance of a linear combination: the variance terms contribute positively, while the cross-product contributes twice the covariance with a negative sign. In practice, use the contrast standard error from the fitted model or compute it using the full estimated covariance matrix. Adding the two standard errors, or treating the coefficients as independent without justification, generally gives the wrong uncertainty.

Common Pitfall and a Short Workflow

A common mistake is to interpret a group indicator coefficient as an adjusted difference without checking what it compares. Under reference-group coding with a common slope, a group coefficient is that group’s adjusted difference from the reference group at the same covariate profile. It is not necessarily the comparison of interest when both groups are nonreference groups. State the target contrast first, then construct the coefficient combination that represents it.