The Common-Slope ANCOVA Model

Why Specify a Common-Slope Model?

In a randomized study, the outcome measured after treatment may be related to the outcome measured before treatment. A baseline measurement can therefore help describe how follow-up outcomes vary among participants. To use it in a regression model, we must specify how baseline relates to follow-up and how treatment assignment enters the model.

The common-slope ANCOVA model includes both the randomized treatment assignment and a continuous baseline outcome as predictors of a continuous follow-up outcome. “Common slope” means the model assigns the same baseline-outcome slope to both treatment groups. It does not mean the groups have the same predicted follow-up outcome: the treatment coefficient allows their fitted lines to have different heights.

The corresponding conditional mean model is \(E(Y_i\mid T_i,X_i)=\beta_0+\beta_1T_i+\beta_2X_i\), when the error has conditional mean zero. For control participants, setting \(T_i=0\) gives a fitted mean of \(\beta_0+\beta_2X_i\). For treatment participants, setting \(T_i=1\) gives a fitted mean of \(\beta_0+\beta_1+\beta_2X_i\). Thus the two groups share the baseline slope \(\beta_2\), while the treatment coefficient shifts one line relative to the other.

Assumptions and What the Coefficients Say

The model is a useful description when the conditional mean of follow-up outcome is adequately represented by a straight-line relationship with baseline outcome in each group, and the two groups have the same slope. It also assumes \(E(\varepsilon_i\mid T_i,X_i)=0\): after accounting for treatment assignment and baseline outcome, errors do not have a systematic positive or negative mean. For conventional regression-based inference, observations are generally treated as independent, and the usual small-sample \(t\)-based procedures additionally rely on normally distributed errors with a common variance. The model’s coefficient interpretations themselves describe the specified conditional mean; they do not by themselves verify these assumptions.

The baseline slope is measured in follow-up-outcome units per baseline-outcome unit. It describes the modelled association between baseline and follow-up, holding treatment assignment fixed. The treatment coefficient is measured in follow-up-outcome units. It compares the model’s treatment and control conditional means at the same baseline value. In a randomized study, it is a baseline-adjusted treatment-assignment contrast; a causal interpretation also depends on the study and outcome data supporting that interpretation. Random assignment does not make an incorrect mean model correct.

One short algebraic check clarifies the phrase “common slope.” At a fixed baseline value \(x\), subtract the fitted control mean from the fitted treatment mean:

The baseline terms cancel, so the difference is the same at every \(x\) under this model. This is a property of the chosen model specification, not a finding that must hold in every real setting. A model with a treatment-by-baseline interaction allows the treatment contrast to vary with baseline; that is a different specification.

Using the Model to Make Adjusted Predictions

A fitted model can be used to obtain a predicted follow-up mean for either assignment at a chosen baseline value. The baseline value must be meaningful for the study data. Each prediction is conditional on the stated value of \(X\); the model does not say that every participant at that baseline value will have exactly the predicted outcome.

Reading the Model Carefully

The coefficients answer different questions. The treatment coefficient compares assignments at a common baseline value; the baseline coefficient describes the shared baseline-follow-up association; and the intercept locates the control line at baseline zero. Keeping these roles separate helps prevent overstatement. In particular, a common slope is an assumption of this model, and a treatment coefficient that is constant across baseline values is implied by that assumption.

The baseline measurement must precede treatment assignment or exposure for it to function as a baseline covariate in this randomized-study setup. An outcome measured after treatment begins is not a baseline outcome merely because it is recorded before the final follow-up. The model also presumes that the baseline-outcome relationship is adequately described as linear over the values being used. A straight line can be a poor description if the conditional mean bends substantially across the observed range.

Another mistake is to describe \(\beta_1\) as the raw difference between the two observed group means. The coefficient comes from the fitted model and compares conditional means at the same baseline value; the raw difference is a separate, unadjusted summary. They may be numerically similar, but their definitions are not interchangeable. Likewise, the model’s constant treatment contrast should not be reported as proof that the true contrast is identical for every baseline value. It is the consequence of fitting a model without an interaction.

ANCOVA is useful because it states clearly how the baseline outcome is being used and what comparison the treatment coefficient represents. Its interpretation depends on the model specification, the quality of the measurements, and the randomized study’s design and conduct. Writing down the coding, units, conditional mean, and coefficient meanings makes the analysis easier to communicate accurately.