How to Interpret Regression Estimates?

[Cheng Han]

Dear all,

I have a question regarding the regression output in Stata. I have done a GLM regression on my dataset with dummy coding on continent. What I don't understand is the output from Stata The screenshot of the output is detailed below,

and may I know if the OR is "1 vs 2" or "2 vs 1"? in other words, is the OR representing the ethnicity_1 vs ethnicity_2 or the inverse of it?

Thanks!

[Jenny Williams]

Hello Han,

1. See here for annotated Stata output: https://stats.idre.ucla.edu/stata/ou...sion-analysis/. Note this is for a logistic regression model, but it is basically the same as your glm with bin logit specification.

2. In this case, it should be ethnicity 2 versus ethnicity 1, with ethnicity 1 as the base category. The line "_Iethnicity_1 omitted" lets you know that ethnicity 1 is the automatically omitted base category.

3. You may want to try running the command without the "xi:". It's simpler to run the command without it, and will preserve the value labels in the regression output.

4. You can specify the base level that you want to compare to. For example, you can type "glm es ib6.ethnicity" to specify ethnicity 6 as the base category.

[Cheng Han]

Hi Dr Williams,

Thank you so much! haha... I just can't believe the results I guess hence I doubted the interpretation... But thank you! I think its time to check the data again...

Thanks!

[Jenny Williams]

Yes, always a good idea to check the data!

A correction to my earlier post:

It wasn't immediately clear from your initial post that you were modelling a proportion variable rather than a simple binary outcome (based on your output "note: es has noninteger values").

So, the link that I posed earlier would not be relevant. A logistic model would treat the outcome as 0 versus all other non-zero values, and thus the odds ratio interpretation would hold. However, for your glm model with a proportion outcome, the "odds ratios" estimates should not be interpreted as odds ratios.

Instead, for interpretation of the estimates, see this post by Maarten Buis: https://www.statalist.org/forums/for...291#post105291

See here also: https://www.stata-journal.com/articl...article=st0147

See here also: https://stats.idre.ucla.edu/stata/fa...-a-proportion/

In any case, your model estimates do seem quite large, so it would be good to check the underlying data.