Im trying to understand interaction effects a little better. My question is, if I have a model that's described by:

$y = \beta_{0} + \beta_{0}x_{1} + \beta_{0}x_{2} + \beta_{0}x_{1}x_{2} + \epsilon$

where, $\mathit{x_i\sim N(0,1)}$ and $\epsilon\sim N(0,1)$.

Is the interaction between $x_1$ and $x_2$ commutative?

As in, if I change the order of the variables to the following, does it make a difference... and if so, why?

$y = \beta_{0} + \beta_{0}x_{1} + \beta_{0}x_{2} + \beta_{0}x_{2}x_{1} + \epsilon$

  • 5
    $\begingroup$ Is it really the intention that all the $\beta$'s shall have the same subscipt $_0$, so be identical? $\endgroup$ – kjetil b halvorsen Aug 15 '19 at 15:24
  • $\begingroup$ yes, that's intentional. In the toy model I'm experimenting with I've set all the $\beta$'s to have the same value. However, it's not a necessity that they are all equal. $\endgroup$ – Electrino Aug 15 '19 at 15:29

If $x_1$ and $x_2$ are real numbers, then their multiplication commutes.

You could rewrite your equation like this:

$$y = \beta_0 + \beta_0x_1 + \beta_0 x_2 + \beta_0x_3 +\epsilon$$

$$x_3 = x_1x_2$$

I think this makes it clear that the interaction commutes.


Yes, under the commutativity law for multiplication, $\beta_{0}x_{2}x_{1}$ and $\beta_{0}x_{1}x_{2}$ are the same.

  • $\begingroup$ Follow up question, if $\beta_{0}x_{1}x_{2}$ and $\beta_{0}x_{2}x_{1}$ are commutative, then am I correct in saying that if I measure the interactions $x_{1}:x_{2}$ and $x_{2}:x_{1}$, then I should get the same value (i.e., the interactions are symmetrical?) $\endgroup$ – Electrino Aug 15 '19 at 16:00
  • 2
    $\begingroup$ @Electrino Yes: $x_1x_2$ and $x_2x_1$ are literally the exact same numbers. $\endgroup$ – Dave Aug 15 '19 at 16:39
  • $\begingroup$ thanks for the clarification. The main reason I'm being explicitly clear is because, for some reason, in my model, the interactions aren't exactly symmetrical. I'll post a different question to try get to the bottom of this. $\endgroup$ – Electrino Aug 15 '19 at 18:02
  • $\begingroup$ @Electrino I wonder if it's because of some numerical issue when you switch the multiplication order. When you multiply the $x_1$ and $x_2$ columns to produce $x_3$, are the numbers slightly different than when you reverse the multiplication order? $\endgroup$ – Dave Aug 15 '19 at 18:33

Conventionally speaking, there is no interaction terms in the model. Rewritten as $y=\beta_0(x_1 + x_2 + x_1x_2) + \epsilon$, we see this is simple linear regression of $y$ on $z=x_1+x_2+x_1x_2$ with forcing the intercept term to be zero.

Given that $z$ is symmetric in $x_1$ and $x_2$, it seems clear that the "interaction" is "symmetric" in $x_1$ and $x_2$ in any sense of the word.


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.