Why Compare Groups at a Shared Covariate Value?
Treatment groups can have different covariate distributions. For example, one group might begin a study with higher baseline scores than another. Their raw outcome averages then combine two features: the groups’ outcome levels and their different baseline covariate values. An adjusted treatment mean answers a more specific question: what outcome does the fitted model predict for each group if the covariate is set to the same value?
This tutorial uses the additive one-covariate ANCOVA model. The earlier material on treatment effects after covariate adjustment introduced the idea of comparing groups while holding the covariate fixed. Here the calculation is made explicit for each group, and the meaning of the resulting means is clarified.
The Model and the Calculation
Let \(Y_{ij}\) be the outcome for individual \(j\) in treatment group \(i\), and let \(x_{ij}\) be that individual’s covariate value. The additive ANCOVA model is
Here \(\alpha_i\) is the intercept for treatment group \(i\), \(\beta\) is the common covariate slope, and \(\varepsilon_{ij}\) is the error for that observation. The model says that the expected outcome in group \(i\), at covariate value \(x\), is \(\alpha_i+\beta x\). The phrase common slope means the model uses the same \(\beta\) for every treatment group.
For the standard model-based calculations, assume the conditional mean is linear in the covariate, errors have mean zero and a common variance, and observations’ errors are independent. Normal errors, \(\varepsilon_{ij}\sim N(0,\sigma^2)\), are additionally used for the usual small-sample inference procedures; the calculations below are predictions, not tests or intervals. These assumptions concern the model and data collection, not whether the arithmetic can be performed.
After fitting the model, let \(\hat{\alpha}_i\) be the estimated intercept for group \(i\), and let \(\hat{\beta}\) be the estimated common slope. The fitted adjusted treatment mean at \(x^*\) is
This is a prediction on the outcome’s scale and in its units. The same covariate value must be used for every group in the comparison. The value \(x^*\) should also be meaningful for the question and reasonably supported by covariate values in the groups being compared. Predicting far outside the observed range is extrapolation, and the fitted model may not describe that region reliably.
Worked Example: Two Groups at One Shared Covariate Value
Suppose a hypothetical study compares two treatments. The covariate is baseline score, measured in points, and the outcome is follow-up score, also measured in points. The data are deliberately simple so the fitted calculations can be checked directly.
| Group | Baseline \(x\) (points) | Follow-up \(Y\) (points) |
|---|---|---|
| A | 0, 2, 4 | 10, 14, 18 |
| B | 1, 3, 5 | 16, 20, 24 |
Within each group, increasing baseline score by 2 points accompanies an increase of 4 outcome points. Thus the common within-group slope estimate is \(\hat{\beta}=4/2=2\) outcome points per baseline point. The group A averages are \(\bar{x}_A=2\) and \(\bar{Y}_A=14\); for group B they are \(\bar{x}_B=3\) and \(\bar{Y}_B=20\). Using the fitted common slope, the intercept estimates are
Choose the shared covariate value \(x^*=2.5\) baseline points. The adjusted means are
The model therefore predicts follow-up scores of 15 points for group A and 19 points for group B at the same baseline score of 2.5 points. Their adjusted difference, B minus A, is 4 follow-up points. The raw group averages were 14 and 20 points, but those averages came from different average baseline scores. The adjusted means answer a different, explicitly conditional comparison.
How the Common Slope Is Estimated
The common slope is estimated from the covariate-outcome relationships within treatment groups, rather than by treating differences between group averages as if they were covariate information. One useful expression for the least-squares estimate is
Here \(\bar{x}_i\) and \(\bar{Y}_i\) are the covariate and outcome averages within group \(i\). Once this slope is estimated, the fitted group intercept can be written as
That expression ensures that the fitted line for each group passes through that group’s covariate and outcome averages. These formulas give a direct route from group summaries to adjusted means; in an analysis, the same estimates can also be obtained by fitting the additive ANCOVA model.
Worked Example: Three Groups and a Shared Baseline
In a hypothetical study, three treatments are compared using a baseline measure \(x\) and a follow-up outcome \(Y\), both in points. The observed values are:
| Group | Baseline \(x\) (points) | Follow-up \(Y\) (points) |
|---|---|---|
| A | 1, 2, 3 | 7, 9, 11 |
| B | 2, 3, 4 | 13, 15, 17 |
| C | 0, 1, 2 | 3, 5, 7 |
Within each group, the covariate deviations from its group mean are \(-1,0,1\), while the corresponding outcome deviations are \(-2,0,2\). Each group contributes \(4\) to the numerator of the within-group slope formula and \(2\) to its denominator. Across all three groups,
The group means are \((\bar{x}_A,\bar{Y}_A)=(2,9)\), \((\bar{x}_B,\bar{Y}_B)=(3,15)\), and \((\bar{x}_C,\bar{Y}_C)=(1,5)\). Therefore,
At the shared baseline value \(x^*=2\) points, the fitted adjusted means are
Thus, at a baseline score of 2 points, the model predicts follow-up scores of 9, 13, and 7 points for A, B, and C. This value lies within the observed baseline range for every group. For instance, the adjusted difference B minus C is \(13-7=6\) follow-up points. This comparison is between model-predicted group means at the same baseline score, not between the groups’ unadjusted averages.
What the Adjusted Means Compare
The value \(x^*\) gives the comparison a specific covariate context. If the covariate is baseline score, the adjusted means compare the groups as the model predicts them for individuals with the chosen baseline score. They are not necessarily the observed averages in any group, nor do they describe every individual in that group.
Under the common-slope model, the difference between two groups’ adjusted means has a particularly simple form. For groups \(i\) and \(k\),
The shared covariate contribution cancels. Consequently, this fitted group difference is the same at every covariate value in this additive model. The adjusted means themselves do change with \(x^*\) when \(\hat{\beta}\ne 0\), even though their pairwise differences do not. This conclusion follows from the common-slope model; it should not be read as a claim that real groups must have identical covariate relationships.
Adjusted means are also not automatically causal treatment effects. A causal interpretation depends on the study design and the assumptions that justify comparing the treatment groups. The calculation alone supplies model-based comparisons at a shared covariate value; it does not establish that treatment caused the difference.
Worked Example: Choosing and Interpreting the Shared Value
Consider hypothetical baseline and follow-up scores from two treatment groups. Suppose the fitted model gives \(\hat{\alpha}_A=13\), \(\hat{\alpha}_B=16\), and \(\hat{\beta}=-0.5\) follow-up points per baseline point. The observed baseline values are 2 and 4 points in group A, and 3 and 5 points in group B. Choose \(x^*=3.5\) points, which lies inside both groups’ observed ranges.
The adjusted means are calculated using the same covariate value:
At a baseline score of 3.5 points, the model predicts follow-up scores of 11.25 points for A and 14.25 points for B. The adjusted difference B minus A is \(14.25-11.25=3\) follow-up points. The negative slope means the fitted outcome decreases by 0.5 points per additional baseline point in this model; it does not change which shared value is used for the comparison. Selecting \(x^*=3.5\) makes the prediction an interpolation within the observed ranges rather than a prediction beyond the data.
Common Mistake: Using Different Covariate Values
A tempting but incorrect comparison is to predict one group at its own average covariate value and another group at a different group’s average, then call the results adjusted means. For example, comparing \(\hat{Y}_A(\bar{x}_A)\) with \(\hat{Y}_B(\bar{x}_B)\) retains a covariate difference in the comparison. It does not hold the covariate constant.
The correction is to choose one value \(x^*\) and use it for every group: calculate \(\hat{\alpha}_i+\hat{\beta}x^*\) for each \(i\). Also avoid choosing a shared value simply because it produces a convenient result. It should be relevant to the scientific question and supported by the groups’ covariate ranges. A value outside their overlap can require extrapolation, making the comparison depend more heavily on the assumed linear model.
Check Your Understanding
Use the fitted additive ANCOVA model to answer these questions.
- If \(\hat{\alpha}_A=8\), \(\hat{\alpha}_B=11\), and \(\hat{\beta}=1.5\), calculate both adjusted means at \(x^*=4\).
- In words, what does the adjusted mean for group A at \(x^*=4\) represent?
- Why is comparing groups at their separate average covariate values not the same as comparing them at a shared covariate value?
- Under the common-slope model, show why the adjusted difference between two groups does not depend on \(x^*\).
- What concern arises if the selected shared covariate value lies well outside the observed ranges for the treatment groups?