GEE interpretation

[Dana Selim]

Dear all,

It’s my first time to use GEE model. I used it for my model, which has a binary outcome. Here is the equation to estimate :
Pr(Ethnic Inclusion_it) = β0 + β1 OilRents_it + β2 CoalRents_it + β3 ForestRents_it+ β4NatgasRents_it + β5 GDPPER_it + β6 oilRents²_it + β7 URBPOP_it + β8 MineralRents_it + ɛ_it

I performed GEE estimation using this code
xtgee y x , i( IDforethnicgroup ) link( probit ) family( binomial ) corr( exchangeable ) robust

I attached the estimation results in a word file.GEE (linkprobit).docx

1-Could anyone help me with the interpretation of any variable ? for example the oil rents.

2-Is there any code that I should apply after this estimation to be able to interpret ?

I'm sorry for the basicness of the question and I'm grateful to anyone who can help me as I couldn't arrive to a clear answer while reading.

[Clyde Schechter]

Microsoft Office document attachments are discouraged here, and your use of them reduces your chances of getting a timely and helpful response. Some of the high-frequency responders on the Forum don't use Office. And among those who do, some are reluctant to download Office documents from strangers because they can contain active malware.

If you read the Forum FAQ you will learn that the preferred way to show Stata output is to paste it directly into the Forum editor, surrounded by code delimiters. The FAQ will also explain how to use code delimiters, if you don't already know.

[Dana Selim]

Dear Dr. Clyde SCHECHTER,

Thank you for clarifying. In fact I tried many times to insert the output directly but I failed, it was too large to fit. I hope you won't have a problem with the picture attached. I checked the FAQ but I couldn't copy my dataex format as I don't have STATA 15 and I couldn't upload it into STATA 14. I hope you won't mind answering my question. I want to know the interpretation of any coefficient for this binary outcome model, using link probit or logit.

Here is the equation to estimate:
Pr(Ethnic Inclusion_it) = β0 + β1 OilRents_it + β2 CoalRents_it + β3 ForestRents_it+ β4NatgasRents_it + β5 GDPPER_it + β6 oilRents²_it + β7 URBPOP_it + β8 MineralRents_it + ɛ_it

xtgee y x , i( IDforethnicgroup ) link( probit ) family( binomial ) corr( exchangeable ) robust

In case you are not able to check the picture. For example if the coefficient of the oil rents is = 0.012, what would be the exact interpretation ?

Thank you in advance.

[Clyde Schechter]

Interpretation of probit coefficients is rather difficult. I don't see a variable called oil rents in your command or output, nor do I see any variable with a coefficient of 0.012. But let's just pretend these things existed. What would they mean? It would mean that associated with a 1-unit increase in oil rents, the inverse normal distribution function of the probability of the outcome variable increasewebuses by 0.012. What that actually means in terms of the probability of the outcome variable itself would be obscure, and would, in any case, depend on the starting value of oil rents from which the 1-unit increase starts. If you had run a logistic regression instead of probit, the coefficient could be interpreted as the logarithm of the odds ratio of the outcome associated with a 1-unit increase in oil rents. Again, translating that into a direct effect on probability of outcome would depend on the starting value.

If you would like a more transparent interpretation of your results, in terms of outcome probabilities themselves, I suggest you run the following command following the regression:

margins, dydx(oil_rent)

Note: replace "oil_rent" with the actual name of the variable whose effect on outcome probability you want.

What this will give you as the average marginal effect of oil rent on outcome probability, averaged over the distribution of all variables in your estimation sample. This figure may or may not be useful. It may be more useful to have the marginal effects at specific values of oil rent. Let's say that an interesting set of values of oil rent are 1 10 and 100, then you could run:

margins, dydx(oil_rent) at (oil_rent = (1 10 100))

The -margins- command is very powerful and enables you to extract readily understandable statistics from obscure models. It is also fairly complicated to use, and I take it you are unfamiliar with it. So I recommend you begin learning about it from the excellent Richard Williams' http://www.stata-journal.com/sjpdf.h...iclenum=st0260. Once you get the basics of -margins- from that, the PDF manual section will help you learn some of the more advanced features and syntax details.

I'm a bit perplexed about the limitations of your computer. If you have an internet connection and it's not behind an excessively onerous firewall, and if you have write permissions to the c:/ado/plus directory, you need only run -ssc install dataex- to get the -dataex- command installed. And I don't see how the output you describe was too large to post directly in the forum. We've had much, much longer things copied in. Perhaps you were trying to paste the image in? But images are poorly behaved on this forum and generally undesirable. The effective way to post Stata output is to copy it from Stata's results window or your log file and paste it directly here between code delimiters. I really don't understand why these things aren't working for you. Oh, well!

Hope this is helpful.

[Dana Selim]

Dear Dr. Clyde SCHECHTER,

I seriously can’t find the right words to say thank you for this rich, detailed, well developed answer. You have been a great help to me. Thanks for sharing with me this useful link of marginal effects command.

Regarding the picture attached, I admit that I tried to paste the output into the “paste from word” button. That’s why I got this message “Maximum number of characters exceeded. It cannot be more than 30000 characters The current number of characters is 46825” and I didn’t think other way could work. In addition, I didn’t understand honestly how to use the dataex option. I thought, for the first while, that I have to upload it to my version. I got it now and I will try it. I’m so sorry again for any inconvenience that I may have caused to STATALIST policy.

Thank you again for your time and for sharing your treasure by this generous explanation. GOD bless every added value you create.

[Hassen Ali]

Dear Clyde, Thank you very much!!
I wish you have an endless happiness and success in your Life!! I have learned lots of statistical models and commands, Interpretation of outputs from your daily posts.
Again, Thank you for your assistance and guidance for us (users especially for beginners like me).
With Best Wishes,Hassen