1
$\begingroup$

I create a linear regression model with interaction term in the model say: $$y=a_0+a_1x_1+a_2x_2+a_3x_1*x_2+e$$ where $x_1$ is continuous and $x_2$ is binary. Now, I have couple of questions:

  1. If $x_1$ and $x_2$ both are significant and have positive coefficients, that means that if $x_1$ and $x_2$ increase, then $y$ increases. What can I say if $x_1*x_2$ is statistically insignificant? Does it mean that increase in. Does it mean that for $x_2=1$, the effect of $x_1$ on $y$ is not changed, i.e. increase in $x_1$ still leads to increase in $y$? What if the interaction is statistically significant?

  2. What can I say if the interaction term is statistically insignificant and negative? Does it indicate that if $x_2=1$, increase in $x_1$ actually leads to decrease in $y$?

Thanks.

$\endgroup$
1
$\begingroup$

When you discuss the interaction you don't discuss it in terms of what happens to y but what happens to a1 and a2. If it's positive then as a1 increases a2 increases. If it's negative then as a1 increases a2 decreases. Interactions are about differences in effects. The slopes in regression are the effects and so it's most efficiently discussed in those terms.

Once you consider the foregoing you can understand why many say that main effects don't mean anything when there is an interaction. I don't agree with that as a generalization but certainly there are many cases where it's true.

$\endgroup$
  • 1
    $\begingroup$ "you discuss the interaction ... in terms of what happens to ... a1 and a2" in a model without an interaction. Indeed the interaction model changes the interpretation of a1 and a2, to the point it's probably better to give them different notation: $E_I(y | x_1, x_2) = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \gamma x_1 \cdot x_2$ versus $E_M(y | x_1, x_2) = \alpha_0 + \alpha_1 x_1 + \alpha_2 x_2$ $\endgroup$ – AdamO May 14 '15 at 0:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.