Logit odds ratio

[Daniel Bromley]

I am using xtlogit and have a simple model that performs as we might wish. But the odds ratio for one x variable is 186.8 which seems impossible. All other odds ratios are in line with our intuition (1.02, 1.31, 1.01). What can explain this seeming outlier? More importantly, how do I interpret such a ratio?

[Rich Goldstein]

in my experience this usually occurs because of some type of data anomaly, such as odd scaling of this variable compared to others or a very unbalanced variable - however, mis-specification can also cause something like this

[Richard Williams]

Welcome to Statalist.

How is the variable scaled? What are its mean and standard deviation? Rescale the variable and the numbers will change, so you need to know about the scaling it now has.

Also, what is the standard error like? If it is huge the number may not mean much.

You might just show your output so we can better see what you are talking about.

[Daniel Bromley]

Thanks to both of you. I should have put this in. Here is the xtsum output. This is not the problem.


Variable | Mean Std. Dev. Min Max | Observations
-----------------+--------------------------------------------+----------------
demwb overall | .2372347 .1676211 .02 .754 | N = 196
between | .140461 .0490714 .5043571 | n = 15
within | .0913375 -.022551 .6778776 | T-bar = 13.0667

The variable is an index of "democracy" across 15 Middle East & North African countries. There is not much variation from year to year, but then striking jumps for several of them (just before and just after the revolts of 2010-2011). I think the variable is not very helpful...The robust s.e. is quite large (2.000 +/-) depending on what else is in the model.

I am grateful for your help--I will keep poking, but will probably dispense of the variable. I was hoping it might be helpful in understanding the Arab Spring revolts, but it does not seem so.

[Richard Williams]

The variable only ranges from .02 to .754 with a mean of .24 and an sd of .17. So, a 1 unit increase would be enormous, and the odds ratio is based on a 1 unit increase. I suggest multiplying the variable by 100 and then trying it again. The results will be basically the same but they won't look so weird.