Relatively high odds ratio after firthlogit

[Azriel Alfajri]

Hi everyone,

When i ran -logit- for this data, variable educ==3 was dropped because it predicts success perfectly the dependent variable. I think that was a complete separation problem (from what I've read from several websites), and to handle this issue i should run -firthlogit- instead. My questions:

- Is it -firthlogit- the one that i need to run?
- If it's true, is it normal to have a relatively high odds ratio for educ==3? exp(4.1376) = 62.80

Variable specification:
Dependent: Happy (1/0)
Independent: working_stat, marstat, altruism, religiosity (the first 4)
Control: age, age_sq, sex, educ (0: no school; 1: primary; 2: secondary; 3: tertiary), health, income

Thank you in advance

[Maarten Buis]

The first step should be to look at some tables:

reg happy $xlist I'd dbeta <0.02
tab educ if e(sample)
tab educ
tab educ happy if e(sample)
tab educ happy

In particular look at the tertiary educated. My suspicion is that you have accidentally dropped a lot of those (maybe the if condition) and there are only a very small number left. If that is the case then the solution is to not drop those observations, rather than firthlogit .

[Azriel Alfajri]

I'm sorry sir, i mean the obs. were dropped by the model

when i run tab educ happy if dbeta < 0.02078827, all obs. if educ==3 is happy==1

[Maarten Buis]

I gave you a different set of commands. Regardless, there are no unhappy people with a tertiary education, so tertiary education should have a very very high effect... So high that logit cannot estimate it. That is perfect separation, as you thought. In your case I would combine categories 1 and 2 together, and categories 3 and 4 together in the educ variable. This is not nice, but is probably the best (least bad) way given the small number of unhappy people in your sample

[Azriel Alfajri]

Ah i see, thank you for the insight, sir!