I have a data set which includes country-years and I am interested in modeling founding and mortality for corporations in each country-year. I am interested in within- as well as between-country differences of predictors such as GDP, education, population, geographical location, etc. What type of models would you recommend that would get at the above? NOTE: Most foundings research uses negative binomial or Poisson regression but the data are usually focused in one geographical area, unlike my case.

  • $\begingroup$ Can you provide more detail on what your dependent variable is? Is it binary? Also, what is a "country-year?" Don't all countries operate on the same calendar year? Do you have multiple years for each country? $\endgroup$ – StatsStudent Feb 17 '16 at 4:29
  • $\begingroup$ Thanks for getting back to me. I have 4160 observations--80 countries * 52 years where each row is Country.1964 through Country.2015 so USA1964, USA1965, USA 1966, and so on for all 80 countries. DV is the total number of corporation foundings for each country in each year so it will be a count variable. $\endgroup$ – torentino Feb 17 '16 at 5:09

Given the additional information you've provided, I think you can use one of three approaches: (1) generalized estimating equations (GEE), (2) a mixed effects generalized regression model, or (3) a fixed-effects regression model. GEE essentially treats the variance as a nuisance parameter and you simply use it to estimate the response for a unit change in the predictor, averaged over the entire population. This is a good approach when you have no real interested in the correlation between to responses (which it doesn't seem to be of any use to you). The mixed model approach is used to estimate the response for a unit change in the predictor for a specific subject (country). I don't want to repeat what others have already described in terms of the differences between the two, so look here for additional details on the differences between GEE and mixed models. To gain a better understanding of the within-country differences only, I would also encourage you to look into something generally called "fixed-effects" models, from the economics literature (not my field). These models essentially allow you to control for all time-invariant characteristics of the subjects, thereby reducing bias. Check out [Fixed Effects Regression Methods for Longitudinal Data Using SAS][2] by Paul Allison as a good introduction.

Within your models, I think you will want to use a Poisson or Negative Binomial function. It's difficult to say which without seeing your data. I'd recommend you start with a Poisson model and look at the overdispersion statistics to see if a Poisson model is plausible given them. If not, use a more complex negative binomial. You may even need to use zero-inflated Poisson or regression models, depending on your data. I generally recommend starting off simple and then working your way to more complex models as needed.

  • 1
    $\begingroup$ Thanks for your insight. Wouldn't it be better to use multilevel modeling for this sort of dataset? $\endgroup$ – torentino Feb 18 '16 at 16:33
  • $\begingroup$ @torentino, Multilevel modelling is the same as a mixed model. $\endgroup$ – StatsStudent Apr 11 '16 at 0:41

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.