Currently I am working on one paper. I have 280 observations and 8 countries in my data set. I have run OLS regression model and all the results seem good. However, I am not convinced by the result of one of the variables and suspected one of the countries for possible influence. then I did run one regression for the whole sample and another regression for without that country. As expected I found the variable to be significant for the whole sample but not for the without x country. I am wondering how could I handle this problem?


2 Answers 2


You probably need to use a mixed effects model with "country" as a random effect. It sounds like your interested in a question where multiple data points have been collected from these 8 countries, but the identity of the country is not what you are actually interested in. You could model this in R with the lmer function in the lme4 package, or with PROC MIXED in SAS.

In R your model would probably look like

lmer(response ~ your predictors + (1|country),data = data)

If data from a single country is driving the significance of your results in a linear regression model, a mixed model should remove that effect. Note that with 8 countries, you'd be at the minimum number of levels to the random effect. You could make your results more robust in general by acquiring data from more countries.

  • $\begingroup$ Thank you for your reply. I am actually looking at the impact of country level variables on firm level performance, so I am interested on the identity of the country to see the impact of the variation of the macro-economic variables on firm performance. I know had it been more countries it would have been more robust. but, my data is limited by criteria and only those 8 countries are considered. With regard to the statistical package, I am using Stata 13. would you mined recommending how to handle the issue on stata? thanks, $\endgroup$
    – user67940
    Feb 2, 2015 at 3:28
  • $\begingroup$ Sorry, don't use stata. It still sounds like you are working in a "mixed model" or "multi-level" or "hiearchical model" framework because your have separate, independent firms clumped into a 8 groups. Your response variables are these separate firm (level 1) and it sounds like your predictors are all/mostly at the country level (level 2). So, while you have what sounds like information on 280 firms, you don't have 280 degrees of freedom because there are correlated by the fact that thet are group by country. In this situation linear regression is not technically valid for inference. $\endgroup$
    – N Brouwer
    Feb 2, 2015 at 14:30

I would run your regression and add an interaction between x and country if you have enough data. In Stata, that would be something like this (assuming your x is continuous):

reg y c.x##i.country, robust

The interactions allow the effect of x to vary by country. You can also add over covariates to the model.

However, if you are interested in the effect of a country-level variable that is constant within each country, it probably makes sense to cluster your standard errors at the country level to avoid running into the Moulton problem. Unfortunately, with only 8 countries you might have a problem since that is very few and this only works as the number of clusters goes to infinity. This also makes your SEs blow up most of the time. So you are between rock and a hard place.

There are also some subtleties about what you are trying to do and SEs. Is it to say something descriptive about y and x in the population of firms from these 8 countries, only some of which are sampled in your data? In that case, maybe areg, vce(cluster country) is better. Or do you want to extrapolate to firms from countries that you don't observe? Or are you trying to say something about your particular firms at another time (say if these 280 firms are not a sample or a census)?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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