Reading Coefficients in an Interaction Model

Why Coefficients Need Context

In a regression model with a treatment-by-baseline interaction, the treatment coefficient does not describe the same contrast at every baseline value. Its meaning depends on how treatment is coded and on the value of the baseline outcome used as the reference. The product-term coefficient explains how the treatment and control baseline slopes differ; it is not itself the treatment contrast.

These distinctions matter when reporting a randomized study. A coefficient can be read correctly only after stating the model, the coding of treatment, and the units of the baseline and outcome variables. The interpretations below concern the fitted conditional mean: they describe the model’s expected outcome for a given treatment assignment and baseline value, not an individual participant’s guaranteed outcome.

The Model and Its Assumptions

Let \(T\) indicate randomized treatment assignment, with \(T=0\) for control and \(T=1\) for treatment. Let \(X\) be a quantitative baseline measurement recorded before assignment, and let \(Y\) be the outcome measured afterward. The treatment-by-baseline regression model is

E(Y\mid T,X)=\beta_0+\beta_T T+\beta_X X+\beta_{TX}(T X).

Here \(\beta_0\), \(\beta_T\), \(\beta_X\), and \(\beta_{TX}\) are model parameters. In a fitted model, their estimates are often written \(b_0\), \(b_T\), \(b_X\), and \(b_{TX}\). The term \(TX\) is the product of the treatment indicator and the baseline value: it equals zero for every control participant and equals \(X\) for a treatment participant.

Model conditions for reading these coefficients: Treatment is coded 0 for control and 1 for treatment; \(X\) is measured before treatment assignment; and the conditional mean is modeled as a straight-line function of \(X\) within each treatment group. The product term allows the two groups to have different baseline slopes. The coefficient meanings follow from this stated model; using standard regression inference also requires appropriate assumptions for the chosen standard errors and study design.

The straight-line condition is a modeling assumption, not a claim that real outcomes must follow perfect lines. If the conditional mean is not adequately described by these lines, the fitted coefficients may provide a poor summary. Randomization supports comparisons by assigned group, but it does not make a misspecified mean model correct.

The coefficient interpretations follow by considering the model separately for control and treatment. Setting \(T=0\) removes both terms involving \(T\). Setting \(T=1\) retains the treatment coefficient and the product term:

\begin{aligned} E(Y\mid T=0,X)&=\beta_0+\beta_X X,\\ E(Y\mid T=1,X)&=(\beta_0+\beta_T)+(\beta_X+\beta_{TX})X. \end{aligned}

This pair of equations is the most direct guide to interpretation. The control line has intercept \(\beta_0\) and slope \(\beta_X\). The treatment line has intercept \(\beta_0+\beta_T\) and slope \(\beta_X+\beta_{TX}\). Thus the treatment coefficient describes the difference between the lines at \(X=0\), while the product-term coefficient describes how much their slopes differ.

What Each Coefficient Says

Interpretation: In this coding, \(\beta_T\) is the treatment-minus-control difference in expected outcome when \(X=0\). The baseline coefficient \(\beta_X\) is the change in expected outcome per one-unit increase in \(X\) in the control group. The product-term coefficient \(\beta_{TX}\) is the treatment-group baseline slope minus the control-group baseline slope: the difference in expected outcome changes per one-unit increase in \(X\).

All three interpretations need units. If \(Y\) is measured in points and \(X\) in baseline points, then \(\beta_T\) is in outcome points, \(\beta_X\) is outcome points per baseline point, and \(\beta_{TX}\) is also outcome points per baseline point. The interaction coefficient is a difference in slopes, so its units are slope units.

The model also implies that the treatment-minus-control difference at a given baseline value \(X\) is \(\beta_T+\beta_{TX}X\). This identity makes the reference role of \(\beta_T\) explicit: when \(X=0\), the product-term contribution is zero. At other baseline values, the product term contributes to the model’s treatment contrast. The interaction coefficient alone is not that contrast.

Worked Example: Symptom Scores and a Zero Baseline

Suppose a hypothetical randomized study models a week-12 symptom score \(Y\) using baseline symptom score \(X\), with both scores measured in points. The fitted model is

\widehat{Y}=18-4T+0.60X-0.15(TX).

The control-group fitted line is \(18+0.60X\), because setting \(T=0\) removes the treatment and product terms. Its baseline slope is \(0.60\) week-12 symptom points per baseline point.

For the treatment group, setting \(T=1\) gives \(18-4+(0.60-0.15)X=14+0.45X\). Its baseline slope is \(0.45\) week-12 symptom points per baseline point.

The treatment coefficient is \(-4\) week-12 symptom points. It is the fitted treatment-minus-control contrast when the baseline score is zero. The product-term coefficient is \(-0.15\) week-12 symptom points per baseline point. It is the difference in slopes: \(0.45-0.60=-0.15\). Thus, for each additional baseline point, the difference between the two fitted slopes is \(-0.15\) outcome points per baseline point. This does not mean that the treatment contrast is always \(-0.15\); that number is the slope difference.

Read the Baseline Slope in the Reference Group

Because \(T=0\) defines the control group, \(\beta_X\) is specifically the control-group baseline slope. It is not automatically a pooled slope or the slope in the treatment group. To obtain the treatment-group slope, the model combines the baseline and product-term coefficients: \(\beta_X+\beta_{TX}\).

The direction of the slope must be interpreted in light of how the outcome is scored. A positive slope means the expected outcome increases as the baseline value increases; whether that represents a better or worse outcome depends on the measurement scale. The slope describes a relationship within the model, not a causal effect of changing a participant’s baseline value.

Worked Example: Baseline Measured in Ten-Point Units

Consider a hypothetical randomized study with a follow-up performance score \(Y\), for which higher scores indicate better performance. Define \(X\) as the baseline score measured in units of ten points: an increase of one unit in \(X\) represents a ten-point increase on the original baseline scale. Suppose the fitted model is

\widehat{Y}=72+3T-2.4X-0.8(TX).

For the control group, \(T=0\), so the fitted line is \(72-2.4X\). The control-group baseline slope is \(-2.4\) follow-up performance points per ten-point increase in the original baseline score.

For the treatment group, the intercept is \(72+3=75\), and the slope is \(-2.4-0.8=-3.2\) follow-up points per ten-point increase in baseline score. Therefore the product-term estimate, \(-0.8\), is the treatment slope minus the control slope: \(-3.2-(-2.4)=-0.8\).

The treatment coefficient, \(3\), is a difference of 3 follow-up performance points at \(X=0\), which here means a baseline score of zero. The interaction coefficient is not a difference of 0.8 follow-up points for every participant; it is a difference in slopes per ten-point baseline increase. Stating the baseline scale is essential to make that interpretation accurate.

The Reference Value Must Be Plausible

The treatment coefficient always refers to the contrast at \(X=0\) under the stated model, but zero may not be a useful or even observed baseline value. If the observed baseline scores are far above zero, interpreting the treatment coefficient as a contrast at zero may describe an extrapolation beyond the data. The algebraic meaning remains correct, but its practical relevance is limited.

This is a reason to inspect the baseline scale and the range of observed values before interpreting the treatment coefficient. It is not a reason to reinterpret the coefficient as an average treatment effect: the model defines it at the reference value, not as an average over participants. The product-term coefficient still describes the difference in baseline slopes, provided the linear model is appropriate over the range being discussed.

Worked Example: A Reference Value Outside the Observed Range

Suppose a hypothetical trial measures a baseline symptom score \(X\) only among participants whose scores range from 5 to 25. The fitted outcome model is

\widehat{Y}=40-1.5T+0.70X+0.20(TX),

where \(Y\) is a follow-up symptom score in points and \(X\) is in baseline symptom points. In the control group, the slope is \(0.70\) follow-up points per baseline point. In the treatment group, it is \(0.70+0.20=0.90\) follow-up points per baseline point. Their slope difference is \(0.90-0.70=0.20\), matching the product-term coefficient.

The treatment coefficient is \(-1.5\) follow-up points at \(X=0\). Since no participant had a baseline score near zero, this is an extrapolated reference contrast rather than a comparison directly supported by observations near that baseline value. It should not be reported as though it were the treatment contrast for the observed participants generally. The product-term estimate still has a clear model interpretation: the treatment-group slope is 0.20 follow-up points per baseline point greater than the control-group slope.

Common Mistakes and a Short Derivation

A frequent mistake is to call \(\beta_T\) “the treatment effect” without specifying the baseline value. Another is to call \(\beta_{TX}\) “the treatment effect for each baseline point.” The first ignores that \(\beta_T\) is defined at \(X=0\); the second confuses a change in the treatment contrast with the contrast itself.

The interpretations can be verified directly. Subtract the control conditional mean from the treatment conditional mean:

\begin{aligned} &E(Y\mid T=1,X)-E(Y\mid T=0,X)\\ &=\{(\beta_0+\beta_T)+(\beta_X+\beta_{TX})X\} -\{\beta_0+\beta_X X\}\\ &=\beta_T+\beta_{TX}X. \end{aligned}

At \(X=0\), this difference is \(\beta_T\). Its change for a one-unit increase in \(X\) is \(\beta_{TX}\), because

\{\beta_T+\beta_{TX}(X+1)\}-\{\beta_T+\beta_{TX}X\}=\beta_{TX}.

This establishes why the product coefficient is a difference in slopes and why it modifies the treatment contrast as baseline changes. It does not establish that the interaction is statistically distinguishable from zero; that is an inferential question requiring its estimate and uncertainty.

Wrong: “The treatment coefficient is the treatment contrast for everyone.” Correction: It is the model’s treatment-minus-control contrast at \(X=0\).
Wrong: “The interaction coefficient is the treatment contrast.” Correction: It is the difference between group-specific baseline slopes, with units of outcome change per baseline unit.
Wrong: “The baseline coefficient is the slope in both groups.” Correction: It is the control-group slope; the treatment-group slope adds the product-term coefficient.
Wrong: “A randomized design makes any coefficient interpretation causal and automatically meaningful.” Correction: Randomization supports comparisons by assignment, but the model, coding, reference value, and observed baseline range still matter.

Key takeaway: With treatment coded 0 for control and 1 for treatment, the treatment coefficient is the contrast at baseline zero, the baseline coefficient is the control-group slope, and the product-term coefficient is the treatment-minus-control difference in slopes.

Check Your Understanding

Use the model and coefficient meanings in this tutorial to answer each question.

In the treatment-by-baseline model, what reference value of \(X\) does the treatment coefficient describe?
If the baseline coefficient is \(-0.5\) outcome points per baseline point and the product-term coefficient is \(0.2\) outcome points per baseline point, what is the fitted baseline slope in the treatment group?
Explain why a product-term estimate of \(0.3\) is not, by itself, the treatment contrast at every baseline value.
A study observes baseline values only from 10 to 30. What practical caution applies to interpreting the treatment coefficient at \(X=0\)?
If baseline is recorded in units of ten points rather than single points, what must be stated when interpreting the product-term coefficient?