Read the Product Coefficient as a Change in Slope

What the Product Coefficient Describes

In a regression model with two quantitative predictors, an interaction allows the relationship between one predictor and the outcome to depend on the value of the other predictor. The product-term coefficient describes that dependence: it tells us how much one predictor’s conditional slope changes when the other predictor increases by one unit.

The key is to read this coefficient as a change in a slope, not as a direct change in the outcome. Its units therefore involve the outcome and both predictors. The sign tells us whether the conditional slope becomes larger or smaller as the other predictor increases.

Model: For quantitative predictors $X$ and $Z$, an interaction model for an outcome $Y$ is $$ E(Y\mid X=x,Z=z)=\beta_0+\beta_1x+\beta_2z+\beta_3xz. $$ Here $\beta_3$ is the product-term coefficient. We assume this equation describes the conditional mean of $Y$ over the predictor values being considered; equivalently, if the model is written with an error $\varepsilon$, we assume $E(\varepsilon\mid X,Z)=0$. This model form is the basis for the interpretations below.

When reading a fitted model, the estimated equation uses estimated coefficients, such as $\hat{\beta}_3$, in place of the unknown model coefficients. The interpretation is parallel: $\hat{\beta}_3$ estimates the change in the conditional slope for a one-unit increase in the other predictor. The algebra describes the fitted model; it does not by itself show that the interaction is statistically distinguishable from zero or that either predictor causes a change in the outcome.

Why the Product Coefficient Is a Change in Slope

Hold $Z$ fixed at a value $z$. The part of the conditional mean that involves $x$ is $\beta_1x+\beta_3xz$. The conditional slope of $X$ at this value of $Z$ is the amount by which the conditional mean changes per one-unit increase in $X$, with $Z$ held fixed. In this linear model it is $\beta_1+\beta_3z$.

Now compare that slope at $Z=z$ with the slope at $Z=z+1$. Subtracting the first from the second gives the change in the conditional slope associated with a one-unit increase in $Z$:

\begin{aligned} \text{Slope of }X\text{ at }Z=z+1 &=\beta_1+\beta_3(z+1),\\ \text{Slope of }X\text{ at }Z=z &=\beta_1+\beta_3z,\\ \text{Change in slope} &=[\beta_1+\beta_3(z+1)]-[\beta_1+\beta_3z]\\ &=\beta_1+\beta_3z+\beta_3-\beta_1-\beta_3z\\ &=\beta_3. \end{aligned}

The terms $\beta_1$ and $\beta_3z$ cancel, leaving the product coefficient. Because the model is symmetric in $X$ and $Z$, the same coefficient also describes the change in $Z$’s conditional slope for a one-unit increase in $X$. The wording of an interpretation should still name which predictor is being treated as the one-unit increase and which predictor’s slope is changing.

Key interpretation: In the model $E(Y\mid X=x,Z=z)=\beta_0+\beta_1x+\beta_2z+\beta_3xz$, $\beta_3$ is the change in the conditional slope of $X$ per one-unit increase in $Z$. Equivalently, it is the change in the conditional slope of $Z$ per one-unit increase in $X$. Its units are outcome units per $X$-unit per $Z$-unit.

Read the Sign and Units Together

If $\beta_3$ is positive, the conditional slope of $X$ increases as $Z$ increases. “Increases” refers to the numerical value of the slope. For example, a slope that changes from $-2$ to $-1$ has increased, although it is still negative. If $\beta_3$ is negative, the conditional slope of $X$ decreases as $Z$ increases. Depending on its starting value, that slope may become less positive, more negative, or change sign.

For units, start with the slope of $X$: it is measured in outcome units per $X$-unit. The product coefficient describes how that slope changes per $Z$-unit, so its units are outcome units per $X$-unit per $Z$-unit. This is not the same as outcome units per $Z$-unit, which are the units of the conditional slope of $Z$.

Worked Example: A Positive Change in the Practice-Time Slope

Suppose a hypothetical fitted model predicts quiz score $Y$, in points, from independent practice time $X$, in hours per week, and coaching time $Z$, in hours per week:

\hat{Y}=52+1.8X+0.6Z+0.40XZ.

The product coefficient is $0.40$. Its units are quiz-score points per practice hour per coaching hour. To interpret it as a change in the practice-time slope, first use the model to write that slope at coaching time $z$:

$$ 1.8+0.40z $$

At $Z=2$ hours per week, the fitted practice-time slope is

$$ 1.8+0.40(2)=1.8+0.80=2.60 $$

quiz-score points per additional practice hour per week. At $Z=3$ hours per week, it is

$$ 1.8+0.40(3)=1.8+1.20=3.00 $$

quiz-score points per additional practice hour per week. The difference is $3.00-2.60=0.40$, exactly the product coefficient. In context, the fitted practice-time slope is $0.40$ quiz-score points per practice hour per week larger for each additional coaching hour per week. This describes the fitted relationship, not a guaranteed outcome for an individual student.

Worked Example: A Negative Change in the Activity Slope

Consider a hypothetical model for a blood-pressure measurement $Y$, in mmHg, using weekly activity $X$, in hours per week, and age $Z$, in decades:

\hat{Y}=132-0.9X+1.4Z-0.20XZ.

The product coefficient is $-0.20$ mmHg per activity hour per week per decade. The conditional slope of activity is $-0.9-0.20z$ mmHg per activity hour per week. At age $Z=4$ decades, the slope is

$$ -0.9-0.20(4)=-0.9-0.8=-1.70 $$

mmHg per activity hour per week. At age $Z=5$ decades, it is

$$ -0.9-0.20(5)=-0.9-1.0=-1.90 $$

mmHg per activity hour per week. The slope has decreased by $-1.90-(-1.70)=-0.20$ mmHg per activity hour per week. Thus, for each additional decade of age, the fitted activity slope is $0.20$ mmHg per activity hour per week lower—that is, more negative. This is an association described by the model; it does not establish that activity causes blood pressure to change or that age causes the association to differ.

Worked Example: Interpreting a Small Positive Product Coefficient

A hypothetical model predicts delivery time $Y$, in minutes, from route length $X$, in kilometres, and traffic level $Z$, measured in traffic-index units:

\hat{Y}=12+3.1X+0.8Z+0.15XZ.

The product coefficient, $0.15$, has units of minutes per kilometre per traffic-index unit. For each one-unit increase in the traffic index, the fitted conditional slope of route length increases by $0.15$ minutes per kilometre. For instance, if the traffic index rises from $Z=6$ to $Z=7$, the route-length slope changes by

$$ [3.1+0.15(7)]-[3.1+0.15(6)] =(3.1+1.05)-(3.1+0.90) =4.15-4.00 =0.15 $$

minutes per kilometre. The coefficient does not say that increasing the traffic index by one unit directly adds $0.15$ minutes to every delivery. It says the fitted association between route length and delivery time is steeper by $0.15$ minutes per kilometre for that one-unit increase in traffic index.

Common Mistakes and Limits

Calling the product coefficient a direct outcome change. The statement “one more unit of $Z$ adds $\beta_3$ outcome units” is generally wrong. The coefficient describes a change in the slope relating the outcome to a predictor; its units include both predictors.
Ignoring the sign of the conditional slope. A negative product coefficient means the slope decreases as the other predictor rises. It does not necessarily mean the slope itself is negative at every predictor value.
Leaving out units. “The slope changes by 0.40” is incomplete. State which slope changes, per one-unit increase in which predictor, and give the combined units.
Turning an association into a causal claim. A product coefficient in a fitted observational regression describes a modeled association. It does not by itself show that changing one predictor would cause the other predictor’s slope to change.
Extrapolating the interpretation without support. The equation gives a linear slope-change pattern, but interpreting it far outside the predictor values represented in the data may be misleading. The coefficient’s algebraic meaning does not ensure that the model is a good description at such values.

Key takeaway: Read the product coefficient as a rate of change in a conditional slope. Name the slope, name the predictor that increases by one unit, include the outcome and predictor units, and interpret the sign as an increase or decrease in that slope.

Check Your Understanding

Answer each question using the product-coefficient interpretation, including units where possible.

A model predicts crop yield in kilograms from sunlight hours and irrigation hours, with product coefficient $-0.3$. What does the sign say about the conditional slope of sunlight hours as irrigation increases?
If the outcome is measured in dollars, $X$ in items produced, and $Z$ in machine-hours, what are the units of the product coefficient?
Why is it incorrect to interpret a product coefficient of $0.5$ as saying that one more unit of $Z$ directly adds $0.5$ outcome units?
A fitted model has a product coefficient of $2$. If $Z$ increases by one unit, by how much does the fitted conditional slope of $X$ change, and in what direction?
Does a nonzero product coefficient from observational data, by itself, prove that changing $Z$ causes the relationship between $X$ and $Y$ to change? Explain briefly.

Read the Product Coefficient as a Change in Slope

What You'll Learn