Logit Model and Cross Country Comparison?

[Munib Inam]

Greetings,

I want to check the difference in determinants of Financial Inclusion(Categorical variable) across 5 countries. The dataset used is the same for all countries (worlbank dataset for over 100 countries), the independent variables (determinants, call categorical) are same across the countries. Can I run 5 different logit regression models and compare the coefficients, hence their differences in them? Or do I include interactive dummies, country*each-determinant?
Remember, all the independent variables are also binary, for example, education has three categories, 1 being primary, 2 secondary and 3 tertiary. Similarly, gender is another determinant (independent variable) which is also binary. And so on.

Any help would be highly appreciated. Thank you!

[Maarten Buis]

Using interaction effects allows you to test whether differences between countries are significant. So that is a good reason for that. Notice however that comparing logit models across countries is controversial. I think there is no problem, but I represent a minority position: http://www.maartenbuis.nl/wp/oddsratio.html

[Munib Inam]

Using interaction effects allows you to test whether differences between countries are significant. So that is a good reason for that. Notice however that comparing logit models across countries is controversial. I think there is no problem, but I represent a minority position: http://www.maartenbuis.nl/wp/oddsratio.html

Hey, thank you so much for your response, Maarten. I've gone through your work on logit models. Quite impressive!

Can you please give me your opinion on the idea of running 5 different logit models, each with the same number of variables drawn from the same dataset, for each country and then comparing their results?

[Maarten Buis]

I find it inconvenient, but that is all.

[Munib Inam]

So the coefficients can be compared? Because I find the interpretation of interactive dummies and so many variables (I have four main variables and 5 countries, so that's 20+ variables) in a single model very inconvenient and cumbersome.

[Maarten Buis]

As I mentioned in #2 the majority position is that you cannot compare the coefficients. This is true regardless of whether you estimate different models or compute interaction effects. I think that is wrong, but that won't help you if the reviewers reject your paper...

If you like different odds ratios for each country, you can have that and still estimate all in one model, so you can test the differences using tricks discussed here: http://www.maartenbuis.nl/publications/ref_cat.html

[Phil Bromiley]

To comment on a little different issue, "Can you please give me your opinion on the idea of running 5 different logit models, each with the same number of variables drawn from the same dataset, for each country and then comparing their results?"

When you run 5 different models of any sort on the same data, you are almost always going to demonstrate the parameter estimates suffer from omitted variables bias. That is, when x comes up significant in model 1 that you omit x from model 2, it suggests model 2 has omitted variables bias. You also can't tell if what is significant in model two is really correct or just a proxy for x. Sometimes, folks do this to select a "best" model based on fit indices like AIC or BIC.

You might consider is a larger model that includes all the variables.

[Munib Inam]

To comment on a little different issue, "Can you please give me your opinion on the idea of running 5 different logit models, each with the same number of variables drawn from the same dataset, for each country and then comparing their results?"

When you run 5 different models of any sort on the same data, you are almost always going to demonstrate the parameter estimates suffer from omitted variables bias. That is, when x comes up significant in model 1 that you omit x from model 2, it suggests model 2 has omitted variables bias. You also can't tell if what is significant in model two is really correct or just a proxy for x. Sometimes, folks do this to select a "best" model based on fit indices like AIC or BIC.

You might consider is a larger model that includes all the variables.

Hey, thank you for your comment.

By same dataset, I meant that the dataset is collected using the same sampling techniques and everything. The dataset is obviously different for each country, which is exactly what I'm looking for in my paper. The dataset from which I've selected my countries is one by Worldbank Global Findex Survey 2014.

For example, using a model with data collected by IMF for one country and comparing it with a model that uses data collected by Worldbank is obviously wrong. So I was wondering if I could use 5 models, each with the same specifications, and then comparing cross country differences based on variations in the co-efficients of these models.

Would I still have an omitted variable bias in such a case?

[Richard Williams]

The phrase "same number of variables" may be confusing. Do you maybe mean same variables, same models, each run on a different country?

[Munib Inam]

The phrase "same number of variables" may be confusing. Do you maybe mean same variables, same models, each run on a different country?

Yes, precisely.

[Munib Inam]

As I mentioned in #2 the majority position is that you cannot compare the coefficients. This is true regardless of whether you estimate different models or compute interaction effects. I think that is wrong, but that won't help you if the reviewers reject your paper...

If you like different odds ratios for each country, you can have that and still estimate all in one model, so you can test the differences using tricks discussed here: http://www.maartenbuis.nl/publications/ref_cat.html

Thanks a lot for this, and all your help!

[Munib Inam]

The phrase "same number of variables" may be confusing. Do you maybe mean same variables, same models, each run on a different country?

Yes, what's your opinion on this?

[Saltanat Akhmadi]

Yes, what's your opinion on this?

Good day,

I have similar concerns about comparison of different logit (or probit) models with the same independent variables for different datasets (in more details, a survey was conducted and the aim is to analyze, how independent variables influence the dependent variable across different issues, and each issue corresponds to separate question in the survey and separate dataset to be used in regression).

[Maarten Buis]

What is your question?

[Saltanat Akhmadi]

What is your question?

Sorry,

My question: is it legitimate to compare the coefficients of two different logit models, if they have the same independent variables, but were run on the different datasets? For exapmle, two different datasets I mean the responses to the same question about different issues/topics in a survey.