Adjusting a Group Comparison for Baseline Scores

Why Compare Groups at the Same Baseline Score?

Suppose two groups have different average outcomes at follow-up, but they also began with different baseline scores. Their raw difference combines those features: it reflects both which group an observation belongs to and where that observation started. A multiple-regression model can answer a more specific descriptive question: what difference does the fitted model give the groups when their baseline scores are held at the same value?

This comparison is useful when the baseline score is relevant to the outcome. For example, a model might relate a follow-up test score to a student's baseline score and to whether the student was in one of two class formats. The adjusted comparison is not a comparison of the groups' overall averages. It is a comparison of their fitted outcomes at a shared baseline score.

The Additive Model and Its Group Contrast

Let \(Y\) be a quantitative outcome, \(X\) a quantitative baseline score, and \(G\) an indicator for group membership. Code the first group as \(G=0\) and the second group as \(G=1\). An additive linear regression model for the conditional mean is

E(Y\mid G,X)=\beta_0+\beta_G G+\beta_X X.

Here, \(\beta_0\) is the model's intercept, \(\beta_G\) is the group coefficient, and \(\beta_X\) is the baseline-score slope. For a fitted model, write the estimated coefficients as \(b_0\), \(b_G\), and \(b_X\). The fitted outcome for group \(G\) at baseline score \(X=x\) is

\hat{Y}=b_0+b_GG+b_Xx.

Definition: The adjusted group contrast at baseline score \(x\) is the fitted outcome for the \(G=1\) group at \(X=x\), minus the fitted outcome for the \(G=0\) group at that same \(X=x\). In the additive model above, this contrast is \(b_G\), in the units of \(Y\).

The model makes a specific assumption about how the groups compare: the baseline-score slope is the same in both groups. In other words, the model has one \(b_X\) term and no separate group-by-baseline term. Consequently, the fitted group difference is constant across baseline scores. That is a feature of this model's structure, not a general fact about all data.

For the fitted contrast to describe conditional mean outcomes, the model must reasonably represent how the mean outcome relates to group and baseline score. In particular, it assumes a linear relationship with baseline score and a common baseline slope across groups. The calculations below take the fitted equations as given; arithmetic alone does not establish that those assumptions fit real data. Also, the contrast is descriptive: neither adjustment nor the regression equation by itself shows that group membership caused an outcome difference.

Why the Group Coefficient Is the Adjusted Difference

The result follows by substituting the two group codes while keeping the baseline score fixed. At \(X=x\), the fitted outcome for \(G=1\) is \(b_0+b_G+b_Xx\). For \(G=0\), it is \(b_0+b_Xx\). Subtracting the second prediction from the first gives

\begin{aligned} \hat{Y}_{G=1}(x)-\hat{Y}_{G=0}(x) &=(b_0+b_G+b_Xx)-(b_0+b_Xx)\\ &=b_G. \end{aligned}

The intercept and the baseline contribution cancel because both predictions use the same \(x\). This is why \(b_G\) is interpreted as an adjusted group difference, rather than as the difference between group averages. The direction matters: with the coding above, the contrast is group \(G=1\) minus group \(G=0\). Reversing the subtraction would reverse the sign.

Worked Example: Comparing Two Class Formats at the Same Starting Score

Suppose a hypothetical fitted model predicts a final test score \(Y\) from a student's baseline test score \(X\) and class format. Let \(G=0\) represent the standard format and \(G=1\) the workshop format. The fitted equation is

\hat{Y}=42+5G+0.60X.

The outcome and baseline score are both measured in test-score points. The model is additive, so it assigns both formats the same baseline slope, \(0.60\) points per baseline point. Consider students with a baseline score of \(70\).

For the workshop format, substitute \(G=1\) and \(X=70\):

\hat{Y}_{G=1}(70)=42+5(1)+0.60(70)=42+5+42=89\text{ points}.

For the standard format, substitute \(G=0\) and \(X=70\):

\hat{Y}_{G=0}(70)=42+5(0)+0.60(70)=42+0+42=84\text{ points}.

The adjusted contrast is \(89-84=5\) points. Equivalently, it is the coefficient \(b_G=5\). In this fitted model, students in the workshop group are predicted to score 5 points higher at follow-up than students in the standard-format group when compared at the same baseline score. This is a model-based association, not evidence that the workshop caused a 5-point increase.

The model's structure also gives the same contrast at another baseline score. At \(X=80\), the fitted predictions are \(42+5+0.60(80)=95\) points for the workshop format and \(42+0+0.60(80)=90\) points for the standard format; the difference is again 5 points. This invariance follows from the common-slope additive form.

Working Through the Comparison in Context

A clear adjusted comparison identifies the outcome, names both groups in the subtraction order, states the shared baseline value when it is relevant, and includes units. It also makes clear that the predictions come from a fitted model. Merely reporting a coefficient without these details can leave its direction and meaning unclear.

Worked Example: Interpreting a Difference in Blood Pressure

Imagine a hypothetical fitted model for follow-up systolic blood pressure \(Y\), measured in millimetres of mercury (mm Hg). Let \(X\) be baseline systolic blood pressure in mm Hg, \(G=0\) denote clinic B, and \(G=1\) denote clinic A. Suppose the fitted equation is

\hat{Y}=12-3G+0.80X.

Compare the clinics for a baseline score of \(120\) mm Hg. For clinic A, \(G=1\):

\hat{Y}_{A}(120)=12-3(1)+0.80(120)=12-3+96=105\text{ mm Hg}.

For clinic B, \(G=0\):

\hat{Y}_{B}(120)=12-3(0)+0.80(120)=12+0+96=108\text{ mm Hg}.

Taking clinic A minus clinic B gives \(105-108=-3\) mm Hg. Thus, at the same baseline blood pressure of \(120\) mm Hg, the model predicts a follow-up systolic pressure 3 mm Hg lower for clinic A than for clinic B. The negative sign is meaningful because the contrast was defined as A minus B. It does not establish that receiving care at clinic A lowers blood pressure.

The fitted equation has a common baseline slope of \(0.80\) mm Hg of predicted follow-up pressure per 1 mm Hg of baseline pressure. Both predictions use the same baseline value and the same slope, which is why the baseline contribution cancels and the adjusted contrast is the group coefficient, \(-3\) mm Hg.

Worked Example: Calculating a Contrast at a Specified Baseline Score

Consider a hypothetical model for a follow-up reading \(Y\), measured in points. The baseline reading \(X\) is also measured in points. Let \(G=0\) represent in-person instruction and \(G=1\) represent remote instruction. The fitted equation is

\hat{Y}=18+2G+0.75X.

Find the fitted remote-minus-in-person difference for two learners whose baseline score is \(64\) points. For remote instruction, \(G=1\):

\hat{Y}_{\text{remote}}(64)=18+2(1)+0.75(64)=18+2+48=68\text{ points}.

For in-person instruction, \(G=0\):

\hat{Y}_{\text{in-person}}(64)=18+2(0)+0.75(64)=18+0+48=66\text{ points}.

The requested contrast is remote minus in-person, so \(68-66=2\) points. The model predicts a 2-point higher follow-up reading for the remote group at this shared baseline score. As a check, the group coefficient is \(b_G=2\), which matches the directly calculated contrast. As in the other examples, this statement summarizes the fitted association; it does not identify a causal effect of instructional format.

What the Adjustment Does—and Does Not—Say

The purpose of fixing the baseline score is to isolate a particular model-based comparison. If two groups differ in baseline scores, comparing fitted outcomes at a shared score removes the baseline contribution from the subtraction. The result answers a conditional question about the specified model. It does not describe how far apart the groups' overall average outcomes are, since those averages may reflect different baseline-score distributions as well as the group coefficient.

This distinction matters in reporting. “The fitted follow-up outcome is 5 points higher for group 1 than group 0 at the same baseline score” describes an adjusted contrast. “Group 1 improved the outcome by 5 points” suggests change or causation not established by that calculation. To support a causal conclusion requires more than fitting a model and holding one predictor fixed.

Key takeaway: In an additive regression with a binary group indicator and a baseline-score predictor, the group coefficient is the fitted difference between the two groups at the same baseline score. State which group is subtracted from which, give the outcome units, and describe the result as an adjusted association unless a separate causal argument justifies more.

Common Mistakes

Comparing different baseline scores. Subtracting a group 1 prediction at one score from a group 0 prediction at another does not calculate the adjusted contrast at a shared score. Set \(X=x\) to the same value in both predictions.
Reversing the group order. With \(G=1\) minus \(G=0\), the contrast is \(b_G\). If the question asks for \(G=0\) minus \(G=1\), the contrast is \(-b_G\).
Calling the coefficient a causal effect. Adjustment for a baseline score gives a model-based comparison conditional on that score. It does not, on its own, show that group membership produced the difference.
Assuming every model has a constant contrast. The contrast is constant here because the model has a common baseline slope and no group-by-baseline term. That conclusion follows from this specified model form.

Check Your Understanding

Use the additive model and coding described in each question.

A fitted model is \(\hat{Y}=10+4G+0.5X\), where \(G=1\) is group A and \(G=0\) is group B. What is the fitted A-minus-B difference at the same baseline score, and in what units should it be reported?
For the model in question 1, calculate both fitted outcomes at \(X=20\). Confirm that their difference matches the group coefficient.
In a model with \(G=1\) for clinic A and \(G=0\) for clinic B, the group coefficient is \(-2.5\) mm Hg. State the adjusted contrast in the order A minus B.
Why does the baseline-score term cancel when comparing fitted outcomes at the same baseline score?
What additional conclusion would be unjustified if a fitted adjusted contrast were described as proof that group membership caused the outcome difference?