Interpreting Product Terms in Linear Regression

Why a Product Term Needs Careful Interpretation

In a linear regression, a genotype-by-exposure product term allows the exposure slope to vary with genotype. Its coefficient is not, by itself, the exposure slope for every genotype group. To interpret it, combine it with the exposure main-effect coefficient and specify how genotype and exposure are coded. This tutorial focuses on a quantitative outcome modeled on its original additive scale.

Let $Y$ be a quantitative outcome, $G$ a genotype code, and $E$ an exposure measurement. Begin with binary genotype coding, where $G=0$ denotes the reference genotype group and $G=1$ denotes the comparison group. The exposure may be continuous. If other covariates are included, the interpretations below hold conditional on their values, provided the model does not add further interactions that change the exposure slope.

Definition: The linear model with a genotype-by-exposure product term is $$ E(Y\mid G,E,C)=\beta_0+\beta_GG+\beta_EE+\beta_{GE}(G E)+\boldsymbol{\gamma}^{\mathsf T}C, $$ where $C$ denotes any included covariates. The coefficient $\beta_{GE}$ is the change in the exposure slope for a one-unit increase in $G$. With binary $G$, it is the difference between the exposure slope for $G=1$ and the exposure slope for $G=0$, on the additive scale of $Y$.

The model describes a conditional mean, not an individual outcome. It assumes that the mean outcome is linear in the included terms. Thus, with binary genotype, it allows different intercepts and different linear exposure slopes in the two genotype groups. The regression coefficients have units determined by the variables: $\beta_E$ is outcome units per exposure unit, and $\beta_{GE}$ is the difference in those slopes per one-unit genotype contrast.

Deriving the Genotype-Specific Slopes

For $G=0$, the terms containing $G$ vanish, leaving an exposure slope of $\beta_E$. For $G=1$, the exposure terms are $\beta_EE+\beta_{GE}E$, so the slope is $\beta_E+\beta_{GE}$. Subtracting the first slope from the second gives $\beta_{GE}$. This is the key interpretation: the product coefficient quantifies a difference in slopes, not a slope on its own.

\begin{aligned} \text{Slope at }G=0 &= \beta_E,\\ \text{Slope at }G=1 &= \beta_E+\beta_{GE},\\ \text{Difference in slopes} &= (\beta_E+\beta_{GE})-\beta_E=\beta_{GE}. \end{aligned}

Equivalently, the difference between the genotype-specific mean outcomes at exposure value $E$ is $\beta_G+\beta_{GE}E$. That genotype contrast changes with exposure at rate $\beta_{GE}$. The same coefficient can therefore be described as a difference in exposure slopes or as the change in the genotype contrast per exposure unit. Both descriptions refer to the additive scale of the modeled outcome.

Worked Example: Interpreting Coefficients in Outcome Units

Suppose a hypothetical study models a quantitative biomarker $Y$, measured in units per liter, using a binary genotype $G$ and an exposure $E$ coded in increments of 10 micrograms per cubic meter. The fitted conditional mean is $$ \widehat{Y}=120+2G+1.5E-0.8GE. $$ Assume the model includes no additional covariates. Interpret the terms and compare fitted means at two exposure values.

For the reference genotype group, $G=0$, so $\widehat{Y}=120+1.5E$. Its fitted exposure slope is $1.5$ outcome units per exposure increment, or per 10 micrograms per cubic meter.

For the comparison group, $G=1$, so $\widehat{Y}=122+(1.5-0.8)E=122+0.7E$. Its fitted slope is $0.7$ outcome units per exposure increment. The product coefficient, $-0.8$, is the difference in slopes: $0.7-1.5=-0.8$. In this model, the exposure slope is estimated to be 0.8 outcome units smaller in the comparison group.

At $E=0$, the fitted genotype contrast is $122-120=2$ outcome units. At $E=2$, it is $$ [122+0.7(2)]-[120+1.5(2)]=123.4-123=0.4 $$ outcome units. The contrast has changed by $0.4-2=-1.6$ units over two exposure increments, which agrees with $2(-0.8)=-1.6$. These are fitted conditional means; the coefficients alone do not establish that changing exposure would cause these changes.

Standard Errors for Slopes and Contrasts

A coefficient estimate is subject to sampling uncertainty. Let $\hat{\beta}_E$ and $\hat{\beta}_{GE}$ be the fitted coefficients, and let their estimated variances and covariance be taken from the regression coefficient covariance matrix. For any linear combination $a^{\mathsf T}\hat{\boldsymbol{\beta}}$, its estimated variance is $a^{\mathsf T}\widehat{\operatorname{Var}}(\hat{\boldsymbol{\beta}})a$. Applying this rule to the $G=1$ slope gives a variance that includes both coefficient variances and their covariance.

\begin{aligned} \widehat{\operatorname{Var}}(\hat{\beta}_E+\hat{\beta}_{GE}) &=\widehat{\operatorname{Var}}(\hat{\beta}_E) +\widehat{\operatorname{Var}}(\hat{\beta}_{GE}) +2\widehat{\operatorname{Cov}}(\hat{\beta}_E,\hat{\beta}_{GE}),\\ SE(\hat{\beta}_E+\hat{\beta}_{GE}) &=\sqrt{\widehat{\operatorname{Var}}(\hat{\beta}_E+\hat{\beta}_{GE})}. \end{aligned}

The covariance term matters: the uncertainty in a sum is generally not obtained by adding the two standard errors, nor by ignoring the covariance. For the reference-group slope, the estimate is $\hat{\beta}_E$, with standard error $SE(\hat{\beta}_E)$. For the difference in slopes, the estimate is $\hat{\beta}_{GE}$, with standard error $SE(\hat{\beta}_{GE})$. More generally, the estimated slope at genotype value $g$ is $\hat{\beta}_E+g\hat{\beta}_{GE}$, and its standard error follows from the same linear-combination rule.

Inference conditions: Standard linear-regression $t$ tests and confidence intervals require an appropriate conditional-mean model and a valid estimate of the coefficient covariance matrix. Under independent, normally distributed errors with mean zero and constant variance, the usual regression $t$ reference distribution is exact in finite samples. With large samples, suitable robust covariance estimates can support approximate inference when constant variance or normality is not plausible. Dependence between observations, such as relatedness not handled by the model, also requires an appropriate covariance method.

Worked Example: Testing an Interaction and a Genotype-Specific Slope

In a hypothetical regression with binary genotype and continuous exposure, suppose $\hat{\beta}_E=1.2$ and $\hat{\beta}_{GE}=-0.4$, in outcome units per exposure unit. The estimated variances are $0.04$ and $0.09$, respectively, and the estimated covariance is $-0.03$. Suppose the model has $df=196$, and the assumptions for the usual regression $t$ inference are reasonable. Calculate the comparison-group slope and its uncertainty, then test for a difference in slopes.

The estimated slope for $G=1$ is $$ \hat{\beta}_E+\hat{\beta}_{GE}=1.2+(-0.4)=0.8 $$ outcome units per exposure unit. Its estimated variance is $$ 0.04+0.09+2(-0.03)=0.07, $$ so its standard error is $\sqrt{0.07}=0.2646$ outcome units per exposure unit. A 95% confidence interval, using $t_{0.975,196}\approx1.972$, is $$ 0.8\pm1.972(0.2646)=0.8\pm0.522, $$ or approximately $(0.278,1.322)$ outcome units per exposure unit. In repeated samples under the model and inference procedure, intervals constructed this way would cover the corresponding population slope 95% of the time.

To test whether the exposure slopes differ, use $H_0:\beta_{GE}=0$ against $H_a:\beta_{GE}\ne0$. The standard error for $\hat{\beta}_{GE}$ is $\sqrt{0.09}=0.3$, so $$ t=\frac{-0.4-0}{0.3}=-1.333. $$ With 196 degrees of freedom, the two-sided $p$-value is approximately $0.184$. At $\alpha=0.05$, this does not provide sufficient evidence that the two genotype groups have different exposure slopes. It does not prove the slopes are equal.

A different question is whether the $G=1$ slope itself is zero. For $H_0:\beta_E+\beta_{GE}=0$, the statistic is $0.8/0.2646=3.024$, giving a two-sided $p$-value of about $0.0028$ with 196 degrees of freedom. This tests the comparison-group slope, not whether slopes differ across genotype groups. The two hypotheses must not be conflated.

Product-Term Tests and Additive-Scale Interaction

For binary genotype and a continuous exposure, the null hypothesis of no linear additive-scale interaction is $H_0:\beta_{GE}=0$. Under the model, this is exactly the hypothesis that the exposure slope is the same in the two genotype groups. A usual test divides $\hat{\beta}_{GE}$ by its standard error and compares the result with the appropriate $t$ reference distribution. A confidence interval for $\beta_{GE}$ gives the estimated difference in slopes and its precision in outcome units per exposure unit.

This is a test of a specified model term. It does not show that every possible form of effect modification is absent: for example, curvature in the exposure relationship or a different exposure scale could lead to another model and another interaction question. Nor does a product-term association alone establish causal effect modification. In observational data, causal claims require a causal estimand and assumptions supporting its identification, in addition to a suitable regression model.

Worked Example: Genotype Dosage and a Slope Contrast

Suppose genotype is coded as allele dosage $G=0,1,2$, and a hypothetical fitted model is $$ \widehat{Y}=50+1.1G+0.6E+0.25GE. $$ Assume this model is appropriate for the outcome mean and that dosage is being treated as a linear allele-count scale. The outcome is measured in units per liter, and $E$ is measured in exposure units. Interpret the three genotype-specific slopes and test the slope difference between $G=2$ and $G=0$. Suppose $\hat{\beta}_{GE}=0.25$, its standard error is $0.10$, and $df=300$.

For a general dosage value $g$, the estimated exposure slope is $0.6+0.25g$. Therefore, the slopes for $G=0,1,2$ are $0.6$, $0.85$, and $1.1$ outcome units per exposure unit, respectively. Under the model's linear dosage assumption, each additional allele raises the exposure slope by $0.25$ outcome units per exposure unit.

The slope difference between $G=2$ and $G=0$ is $$ (0.6+2(0.25))-(0.6+0(0.25))=0.50 $$ outcome units per exposure unit. Equivalently, this contrast is $2\beta_{GE}$, because the dosage difference is two. Its estimate is $2(0.25)=0.50$, and its standard error is $2(0.10)=0.20$. The test statistic for no difference is $0.50/0.20=2.50$, with a two-sided $p$-value of approximately $0.013$ on 300 degrees of freedom. A 95% confidence interval, using $t_{0.975,300}\approx1.968$, is $0.50\pm1.968(0.20)$, or approximately $(0.106,0.894)$ outcome units per exposure unit. This evidence concerns a difference between the two dosage endpoints under the stated linear-dosage model.

Centering and the Meaning of Main Effects

The exposure coefficient $\beta_E$ is the exposure slope at $G=0$, while the genotype coefficient $\beta_G$ is the genotype contrast when $E=0$. If zero exposure is outside the observed range or has no useful meaning, recode exposure relative to a meaningful value $e_0$: let $E_c=E-e_0$. The model becomes $$ E(Y\mid G,E_c,C)=(\beta_0+e_0\beta_E)+(\beta_G+e_0\beta_{GE})G+\beta_EE_c+\beta_{GE}GE_c+\boldsymbol{\gamma}^{\mathsf T}C. $$

Centering changes the intercept and the genotype main-effect interpretation: the new genotype coefficient is the genotype contrast at exposure $e_0$, rather than at zero exposure. It does not change the exposure slope at $G=0$, the difference in exposure slopes, or the product coefficient. Centering can make main effects more interpretable and may improve numerical stability, but it does not create or remove the modeled interaction.

Common Mistakes and a Practical Reading

Calling $\beta_{GE}$ the exposure slope for carriers. For binary genotype, the carrier slope is $\beta_E+\beta_{GE}$; the product coefficient is the difference between slopes.
Testing the wrong hypothesis. Testing $\beta_E=0$ asks whether the reference-group slope is zero. Testing $\beta_{GE}=0$ asks whether the genotype groups have equal slopes.
Leaving covariance out of a standard error. For a slope sum, use both variances and twice the covariance. Ignoring covariance can misstate uncertainty.
Reading a coefficient without its coding or units. State the genotype contrast, exposure increment, outcome units, and reference point. Rescaling exposure changes the numerical coefficient's units.
Turning model interaction into a causal claim. A nonzero product coefficient is evidence of a modeled difference in conditional mean slopes; by itself it does not show that changing exposure causes a different outcome response by genotype.

A clear report gives the genotype and exposure coding, the fitted slopes for the relevant genotype values, the difference in slopes with its standard error or interval, and the test of the interaction hypothesis. When covariates are present, say that these are conditional model-based comparisons. This makes the additive scale and the scientific contrast explicit.

Check Your Understanding

Use the model and definitions above to answer these questions.

In a binary-genotype model, what is the exposure slope for $G=1$, and what does $\beta_{GE}$ represent?
Why is the standard error of $\hat{\beta}_E+\hat{\beta}_{GE}$ not generally obtained by adding the two standard errors?
What null hypothesis is tested by a test of the product coefficient in a linear model with binary genotype?
If exposure is centered at $e_0$, which main-effect interpretation changes, and does the product coefficient change?
For dosage $G=0,1,2$, how does the modeled slope difference between $G=2$ and $G=0$ relate to $\beta_{GE}$?