Interpreting the Coefficients
This will depend on if the covariate is entered as a main effect or an interaction.
If you simply enter a categorical predictor and continuous predictor as main effects, your regression coefficients include a conditional mean intercept which represents the reference group, a dummy-coded slope term which is simply an increase/decrease in this conditional mean by each comparison group, and the continuous variable slope term which is a shared slope term which does not vary by group.
If you enter an interaction, the only difference now is that you get an adjusted slope term, which changes based on the categorical variable. Below I fit a model with only main effects and another model with the interaction. The entered variables include:
- Miles per gallon (MPG)
- Engine (VS): either v-shaped or straight
- Horsepower (HP)
I use a two-level categorical variable here for simplicity. The model is fit with the mtcars data in R. The first model is the main effects only (VS and HP) and the second model is the interaction between VS and HP.
#### Main Effects Only ####
fit1 <- lm(mpg ~ factor(vs) + hp,
mtcars)
summary(fit1)
bhat1 <- coef(fit1)
bhat1
#### Main and Interaction Effects ####
fit2 <- lm(mpg ~ factor(vs) * hp,
mtcars)
summary(fit2)
bhat2 <- coef(fit2)
bhat2
Inspecting the coefficients of the model with just the categorical variable and continuous variable entered as main effects only (by printing bhat1):
(Intercept) factor(vs)1 hp
26.96300111 2.57622314 -0.05453412
The intercept (conditional mean of the reference group) is around $26.96$. The conditional mean of the comparison group (VS = 1) is higher, so its intercept is just the addition of the intercept and the factor/categorical variable's coefficient to get the adjusted intercept ($26.96 + 2.58 = 29.54$). The last coefficient is a shared slope term. When the second coefficient is set to zero, the regression line for the reference group is simply constructed as:
$$
y = (\beta_0 \times 26.96) + (\beta_2 \times -0.055) + \epsilon
$$
And if we are constructing the line for the comparison group, it becomes:
$$
y = (\beta_0 \times 26.96) + (\beta_1 \times 2.58) + (\beta_2 \times -0.055) + \epsilon
$$
Notice the first equation can be obtained simply by setting $\beta_1$ to zero, which is how we get our unconditional slope. Contrarily, inspecting the coefficients of the interaction model by printing bhat2, we get these additional terms:
(Intercept) factor(vs)1 hp factor(vs)1:hp
24.49636966 14.50417652 -0.04153284 -0.11656540
The only difference this time is that the last coefficient shifts the slope for the comparison group (this is why it notes VS = 1 in the interaction term). For constructing a line for the first group, we would just get this single intercept and single slope:
$$
y = (\beta_0 \times 24.50) + (\beta_1 \times -0.04) + \epsilon
$$
Or simply identified as:
$$
\text{intercept} = 24.50 \\
\text{slope} = -0.04
$$
For the comparison group, the regression line changes to this:
$$
y = (\beta_0 \times 24.50) + (\beta_1 \times 14.50) + (\beta_2 \times -0.04) + (\beta_3 \times -0.12) + \epsilon
$$
Just like how categorical variable coefficients are simply additions to the intercept, the interaction term here is simply an addition to the slope, so it technically becomes two operations, one for the intercept and one for the slope:
$$
\text{intercept} = 24.50 + 14.50 = 39 \\
\text{slope} = -0.04 + -0.12 = -0.16
$$
Visualizing the Regression
Plots make this distinction more clear. The code for visualizing the regression in R is below:
#### Plot Main Effects Model ####
par(mfrow=c(1,2))
col <- ifelse(mtcars$vs == 0, "black","gray")
plot(mtcars$hp,
mtcars$mpg,
pch=23,
bg=col,
cex=2,
main="HP x MPG by Type",
xlab="HP",
ylab="MPG")
abline(
bhat1[1] + bhat1[2], # adjusted intercept
bhat1[3], # unadjusted slope
col="gray",
lwd=3
)
abline(
bhat1[1], # unadjusted intercept
bhat1[3], # unadjusted slope
col="black",
lwd=3
)
#### Plot Interaction Model ####
plot(mtcars$hp,
mtcars$mpg,
pch=23,
bg=col,
cex=2,
main="HP x MPG by Type",
xlab="HP",
ylab="MPG")
abline(
bhat2[1] + bhat2[2], # adjusted intercept
bhat2[3] + bhat2[4], # adjusted slope
col="gray",
lwd=3
)
abline(
bhat2[1], # unadjusted intercept
bhat2[3], # unadjusted slope
col="black",
lwd=3
)
Here we can see in the first model (left plot) we are simply assuming that the slope is the same between groups (hence the parallel lines), and so the only difference is the mean differences between groups (you can see the gray regression line is higher than the black one because this group has generally larger values). In the second model (right plot), we adjust each slope by group, and so the slope decreases dramatically in the gray points/line.
So to summarize:
- For the main effects-only model, the continuous predictor coefficient is the general slope, and the categorical predictor coefficient is the adjusted conditional mean (intercept).
- For the interaction model, the continuous predictor coefficient is the slope for the reference group, the categorical predictor coefficient is the adjusted conditional mean (intercept) of the comparison group, and the interaction coefficient is the adjusted value of the slope for the comparison group.
