1
$\begingroup$

I estimate some logistic and OLS models on big survey data (overfitting shouldn't be a problem) where I need to control for a country of a respondent. The country variable is coded as iso3n- 3 digits for every unique country (there are 96 individual countries in my data). Important thing is that I only want to control for a country effect, I'm not interested in the significance or parameters of every individual country.

I wonder what is the theoretical reason to code country as 96 dummies with additional base level. Why couldn't I just treat the nominal country variables as quasi-linear and thus control for its effect on my other variables?

I already tried to estimate two models in R; one with dummies and one with a single nominal "country" variable but estimated parameters of my key independent variable differ. Why is that? I also searched the site but most entries on the subject just assume that dummy coding is needed.

$\endgroup$

1 Answer 1

0
$\begingroup$

I am not entirely sure we are on the same page as regards "quasi-linear", but assuming that this amounts to, say, coding Albania as 1 and Zimbabwe as 96 and, say, Malawi as 48, that would amount to assuming that, in terms of your dependent variable, you expect the effect for Zimbabwe to be twice as high as for Malawi and 96 times as high as for Albania.

Presumably, this is not what you want, as you merely want to allow for separate country effects.

$\endgroup$
1
  • $\begingroup$ Thank you, this is exactly it. Now I see that my question was a bit premature as I eventually come up with the same answer. $\endgroup$
    – Ajzach
    May 29, 2020 at 14:00

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.