Negative margins value with penalized maximumlikelihood estimates

[Jake Hay]

Hello, I've tried the logit with a binary variable but I found out that the sample size is only 95.
I found out that Firth's Penalized Maximumlikehood methods can reduce small sample bias so that I tried it,
and AIC improved better than logit model,
but when I ran the margins, it shows negative value.


logit experience money if married==1
margins, at (money=4)


firthlogit experience money if married==1
margins, at (money=4)

only difference was firthlogit and then I got a negative margins value(it didn't happen at logit).
Why is that? and how can I solve the problem?

Thanks a lot!

[Marcos Almeida]

Welcome to the Stata Forum / Statalist.

Please read the FAQ, particularly the topic about sharing command/output/data.

That said, I gather this thread may be helpful to you.

[Phil Bromiley]

Firthlogit is user written so it doesn't necessarily conform to all of the features of Stata. You're predicting a probability with logit. I suspect you're predicting something else with firstlogit, but I could be wrong.

[Richard Williams]

firthlogit does not have its own predict command, so the default is for xb, the log odds. You need something like

webuse nhanes2f, clear
firthlogit diabetes i.female
margins female, expression(invlogit(predict(xb)))

[Richard Williams]

Incidentally, margins tells you what is being predicted. If you don't change the defaults, after logit it says things like

Expression   : Pr(diabetes), predict()

After firthlogit,

Expression   : Linear prediction, predict()

People have gotten confused many times because the default prediction was not what they expected. Check the output if something seems odd.