Suppose I have a linear regression form of

$$ \log(Y) = \beta_0 + \beta_1X_2 + \beta_2X_3 + \beta_3X_1Z + \beta_4X_2Z + \epsilon $$

where $X_1, X_2, X_3$ are binary and $X_1$ is omitted as a reference variable. Suppose $Z$ is also binary 0-1. I am wondering how we would be interpret $\beta_1$ and $\beta_2$?

  • $\begingroup$ Note that $\beta_1 = \mathbb{E}[\log Y \mid X_2 = 1, Z = 0] - \mathbb{E}[\log Y \mid X_2 = 0, Z = 0]$ which of course the mean difference in $\log Y$ between a particular treatment ($X_2 = 1$) and control ($X_2 = 0$) when $Z = 0$. If the difference is small then you can interpret it as a percentage change in $Y$ between treatment and control holding Z at 0 ... $\endgroup$ Commented Feb 14, 2023 at 6:22
  • $\begingroup$ Since you have X1 in an interaction you should have it as a main effect. Alternatively, you can take it out of the interaction and add X3. Any good stat software will do this for you. $\endgroup$
    – Peter Flom
    Commented Jan 15 at 17:41

1 Answer 1


You can interpret $\beta_1$ as the percent change in Y when the treatment effect $X_2$ is applied, when you're holding other variables constant, in this case that is Z.

You can scale this interpretation to other $\beta_s$ as well.

Hope this helps.

  • 1
    $\begingroup$ This is incorrect in two ways. (1) Assuming "treatment effect $X_2$ is applied" means that $X_2$ is changed from $0$ to $1,$ then $\log(Y)$ changes by adding $\beta_1+\beta_4 Z,$ not just $\beta_1$ alone. (2) Interpreting that as a proportional change in $\log(Y)$ assumes the change is relatively small: less than $0.1$ in absolute value will do; less than $0.25$ is OK; beyond that it's usually a good idea to distinguish changes in the log from proportional changes. $\endgroup$
    – whuber
    Commented Feb 14, 2023 at 15:37

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.