I would simply like to know if:




or something completely different and why.

This is not homework. It came up because I'm trying to figure out the bias of omitting an interaction term. When the population model is...

$y=\beta_0 + \beta_1 x_1 + \beta_2 x_2 + x_1 x_2 + \epsilon$

...regressing without the interaction yields:

$E[y|x_1,x_2]=\hat{\beta}_0 + \hat{\beta}_1 x_1 + \hat{\beta}_2 x_2 + E[x_1 x_2|x_1,x_2]$

I expand that conditional expectation at the end to:


But then I'm not sure what to do with $E[x_1|x_1,x_2]$ and $E[x_2|x_1,x_2]$.

Bonus points if you confirm or deny that $cov(x_1,x_2|x_1,x_2)$ is zero, regardless of whether $x_1$ and $x_2$ are independent.

  • $\begingroup$ $X_1$ being measurable wrt $X_1$, $\mathbb{E}[X_1|X_1]=X_1$. $\endgroup$
    – Xi'an
    Jan 16, 2019 at 17:31
  • $\begingroup$ There is no regression coefficient before interaction in your model? $\endgroup$
    – user158565
    Jan 16, 2019 at 17:39
  • $\begingroup$ @user158565 the coefficient is one, for simplicity $\endgroup$
    – suckrates
    Jan 16, 2019 at 17:39
  • $\begingroup$ @Xi'an so $E[x_1|x_1,x_2]=x_1$? $\endgroup$
    – suckrates
    Jan 17, 2019 at 9:32
  • $\begingroup$ Yes and $\text{cov}(X_1,X_2|X_1)=0$ as well. $\endgroup$
    – Xi'an
    Jan 17, 2019 at 10:31

1 Answer 1


In $E[y|x_1,x_2]$ it seems you already said that $E[x_i|x_1,x_2]=x_i$'s by pulling $x_i$ out of the expectation. $E[x_1x_2|x_1,x_2]$ is also $x_1x_2$ because this is like saying what is the expectation of $x_1x_2$, if we know $x_1,x_2$. It's going to be $x_1x_2$. This directly means that $cov(x_1,x_2|x_1,x_2)=0$, since $E[x_1x_2|x_1,x_2]=x_1x_2=E[x_1|x_1,x_2]E[x_2|x_1,x_2]+cov(x_1,x_2|x_1,x_2)=x_1x_2+cov(x_1,x_2|x_1,x_2)$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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