2
$\begingroup$

I have panel data of countries and I am estimating a Fixed Effects model similar to

$$ y_{ct} = \underbrace{\theta_t}_{\text{Time FE}} + \underbrace{\theta_c}_{\text{Country FE}} + \beta^1x^1_{ct} + \beta^2x^2_{ct} + \varepsilon_{ct} $$

For every $t$, the variable $x^2_{ct}$ is highly correlated across countries. That is, $x^2_{ct}$ shows high correlation in the cross-sectional dimension. This means that almost all of its variation can be captured by $\theta_t$. Thus, when I fit this model, I get that $\beta^2$ is not significant at all.

However, theory suggest that $x^2_{ct}$ should have a strong causal impact on $y_{ct}$. So the question is: assuming I have controlled for all confounders with $(\theta_t, \theta_c, x^1_{ct})$, is it possible to estimate $\beta^2$ using this setup?

P.S.: for the estimation, I assume two-way clustering in $\varepsilon_{ct}$

Improve this question
$\endgroup$

1 Answer 1

Reset to default
0
$\begingroup$

The significance of a coefficient increases with the variance of the regressor. If $x^{2}_{ct}$ is highly correlated across countries and your data is country level, then you will have low variance. Unsurprisingly, this leads to an insignificant $\beta^2$, whether your have country FE or not. Taking out $\theta_c$ is unlikely to fix this problem.

Of course, if you can control for all relevant country characteristics, then you do not need the country FE. Nevertheless, it is very unlikely that you can.

Especially with cross-country data, regressions without country FE are suspect, because countries themselves can vary across so many dimensions that it is hard to interpret regression coefficients as being causal and it is near impossible to control for all relevant country factors. I would not recommend running a cross-country regression without country FE.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks Bob. When you say that $x_{ct}^2$ will have low variance you mean low variance in the country (cross-sectional) dimension, right? That still leaves room for variation along the time dimension, but my problem is that I believe this variation is already being captured by $\theta_t$ (time FE). In the end, it is a matter of multicollinearity between the dummy variables associated with $\theta_t$ and $x_{ct}^2$. I agree with you, though: neither $\theta_t$ nor $\theta_c$ should be removed because of the likely presence of unobserved confounders. $\endgroup$
    – mbiron
    May 3, 2018 at 15:41
  • $\begingroup$ Yes, I was referring to the cross-country variaion, since this I believe, was the variation you were trying to leverage by removing the country fixed effects. Given the choice, I would rather remove time fixed effects than country fixed effects, but it depends on the context. $\endgroup$
    – Bob
    May 3, 2018 at 16:13

Your Answer

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

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