logistic regression R and Stata – grouping variable

[Bruno Lambert]

Hello,

I mostly use Stata 13 for my regression analysis. I want to conduct a logistic regression on a proportion/number of success. Because I receive errors in Stata I did not expect nor understand (if there are Stata experts who want to know more about the problems I face and can potentially help me solve them, I would be glad to give more details), I want to repeat the analysis in R. In Stata I would use the command: xtlogit DEP_PROP INDEP_A INDEP_B INDEP_C, i(ID). ID is the identifier for each subject. There are eight lines with data for each subject because there are three within factors (INDEP_A, B, C) with two levels each (0 and 1). I can repeat this analysis in R by using the command: glm(DEP_SUC ~ INDEP_A + INDEP_B + INDEP_C, family = “binomial”). DEP_SUC is here a table with the successes and misses per row. Again, there are eight rows for each subject. But while I know how to group these lines in Stata by using the option i(ID ), I do not know what to do in R. I have search for more information about the i() command, but did not find any usefull information.

So, to summarize: I want to find out how three variables (binary) influence a proportion and use logistic regression. In Stata I can group multiple lines per subject using the i( ) command in logistic regression. What is the equivalent in R?

Thank you in advance!

Cross posted at: http://stats.stackexchange.com/quest...on-r-and-stata
https://r-forge.r-project.org/forum/...78&group_id=34

[Nick Cox]

You are asking about R code. It's a grey area, but I don't think that's on-topic here even though you want a translation of Stata,

[daniel klein]

I agree with Nick. It is probably better to tackle the prboems you face in Stata in this forum, so please provide more details on the error you get.

Best
Daniel

[Enrique Pinzon (StataCorp)]

Dear Bruno,

You can use:

xtgee y i.x1 i.x2, family(binomial) link(logit) vce(robust)

The previous command gives you panel/longitudinal logistic population-averaged parameters, which is what you want. The parameters you obtain using Stata are correct. I cannot guarantee they are the same as what software's that do not certify their results would give you.

Stata's glm is cross sectional. I believe this is not what you want. If I am wrong, you can try the new cross-sectional fractional response models in Stata 14 using clustered standard errors:

fracreg logit y i.x1 i.x2, vce(cluster id)

Best,

Enrique

[Bruno Lambert]

Thank you for your answers.
I think I did find a solutions, the threat may be closed.