Suppose I have a regression of the form: $$ Y_{it}=\beta X_{it}+\beta_{2}X_{it}\times D_{i}+\alpha_{i}+\epsilon_{it} $$ In the above, we have information by country-year on $Y_{it}$ and two regressors: $X_{it}$ and $D_{i}$. $X_{it}$ is time varying, but $D_{i}$ is fixed for each country over time. $\alpha_{i}$ measures the country fixed effect.

My question is: can $\beta_{2}$ be identified? As a fixed effects procedure uses only within-country variation, and as $D_{i}$ does not vary over time, how is $\beta_{2}$ separately identified from $\beta?$ I ran the regression and did find that it was identified. Am I missing something?

  • $\begingroup$ They don't have the same values. For countries where D=1, they are the same, but for countries where D=0, they are not. $\endgroup$
    – ChinG
    Commented Mar 3, 2019 at 19:40
  • $\begingroup$ Hi: It does seems identifiable to me intuitively. because the effect of $D_i$ is additive to the effect of X_it. The coefficient on $X_{it}$ is either $\beta$ or $\beta + \beta_{2}$ depending the value of $D_{i}$ so yes, you can get estimates and the $\beta_2$ will representing the change in the slope of the regression line due to a country having a $D_i$ equal to 1. But let's not trust me. Let's wait for another answer.. $\endgroup$
    – mark leeds
    Commented Mar 3, 2019 at 20:03
  • $\begingroup$ I was thinking about this more and perhaps a clearer way to write the model is to use Dit except that Dit=Di for a given i and all t.. To me, this makes the identifiability of the model more straightforward.. – $\endgroup$
    – mark leeds
    Commented Mar 4, 2019 at 18:50
  • $\begingroup$ Hi @markleeds . Yes it will be identified if cross sectional variation. However, if the fixed effects is used, then only within country varition is used, and D does not change within country. So for countries where D=1, the two variables are exactly collinear, and for countries where D=0, the interaction term=0. $\endgroup$
    – ChinG
    Commented Mar 5, 2019 at 17:07
  • 1
    $\begingroup$ Did more research and it seems to be that the use of the term "fixed effects" is referring to a systems model type of fit, so I mis-understood. So I take back everything I said and my apologies for noise. I checked and, if you define it that way and use systemfit in R, it will choke. So, yes, I think it's not identifiable. $\endgroup$
    – mark leeds
    Commented Mar 5, 2019 at 23:25


Your Answer

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

Browse other questions tagged or ask your own question.