I am analyzing a sample of about 6000 actions carried out by about 500 multinational companies in about 80 countries during a 6 year period. Actions are carried out randomly, and are not longitudinal measurements for the same multinational over time. The actions can be of two types (A or B). My motive is find correlations between a set of predictors and the occurrence of A. As the outcome is binary, I am applying logistic regression.

Each action is conducted by a multinational in a country in a year. Predictor variables are measurements dependent on the multinational and year (e.g. number of employees), and country and year (e.g. GDP) respectively.

This means that because these measurements are based on categorical variables on other levels than the action itself (multinational level and country level), a lot of measurement values in the sample will be equal for multiple actions (observations).

E.g. all actions occurring in the same year and country (although by different multinationals) will have the same GDP value. Similarly actions in a year carried out by the same multinational (but in different countries) will be associated with the same multinational size measurement value.


Multinational Country Year SizeAsNumberOfEmployees GDPOfCountry Choice

MultinationalA CountryA 2008 45 55 A
MultinationalB CountryA 2008 23 55 A
MultinationalC CountryA 2008 99 55 B
MultinationalA CountryB 2008 45 77 B
MultinationalA CountryB 2010 48 83 A
MultinationalB CountryB 2010 28 83 B

So to summarize: predictor variables are measured on a different level than the observations (country, multinational vs. action(obs)) and thus many observations may be associated with equal values that correspond to the groups/levels for which they are measured.

Is this a concern, and if it is how should this be modelled? Is there a need for a multilevel model approach in this case or can I simply use simple logistic regression?


Your data seems to have hierarchical structure and if so multilevel analysis would be appropriate. Some, like Nezlek, would say that if your data has hierarchical structure, than this is a sufficient reason why you should think of it in multilevel way and consider multilevel analysis. There are possibly three random effects: countries, companies and time, with possibly crossed structure (e.g. Raudenbush, 1993). Not accounting for the nested nature of the data could lead to biased results: if you want to infer about higher-levels based only on individual-level analysis, then the results could be biased (so called atomistic fallacy, check e.g. here).

However when choosing model for analysis you should also consider what is your aim and so it is not "yes" or "no" kind of choice and there could be other arguments for or against multilevel analysis.

Check a book by Gelman and Hill (2006) on mixed and hierarchical regression, it gives a nice and readable introduction on multilevel modelling plus tutorial on using lme4 and Bayesian estimation of this models.

  • $\begingroup$ thank you very much for your answer, I realize that modelling choice depends on the nature of data and motive of research, which is I got concerned that I was not doing it optimally given the research case. I have some supplementary questions: 1. Should i use random effects on a hierarchical level if I also have predictor variables measured on that level (e.g. GDP for country)? The reason I ask is that for the numerous examples of random intercept examples I have seen, random effects are alone introduced on levels for which there were no explicit measurements. $\endgroup$ – SuppaiKamo Dec 20 '14 at 11:21
  • $\begingroup$ 2. Most research in my field applying variables on different levels, usually do not apply multilevel modeling (only very recently), but instead add fixed effects on levels with no measurements. In my case all industries within my sample is represented, which calls for fixed effects. However I guess firms are better modelled randomly since the full population is not represented? $\endgroup$ – SuppaiKamo Dec 20 '14 at 11:27
  • $\begingroup$ 1. You can have random slopes for independent variables on certain level - I would say that this possibility is an argument for multilevel model, 2. as you say, if you have all industries fixed effect is probably better, however for time (?), countries and companies random effects could be considered. $\endgroup$ – Tim Dec 20 '14 at 11:54
  • $\begingroup$ I apologize for asking directly, but I have a related question that it would be nice if you had a look at. I can't find an answer anywhere and I could use some help: stats.stackexchange.com/questions/129862/…. The question is about the effect of fixed effects dummies on VIF values on the selected explanatory variables. In my case the fixed effects dummies hyper inflate the variables I need to investigate, but it makes perfect theoretical sense to keep both. $\endgroup$ – SuppaiKamo Dec 20 '14 at 21:19

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.