Not significant coefficients in Logit, but in Probit they are

#1
Hello, I would like to ask you for a help. I have a simple model with one explanatory variable and with a binary dependend variable. Explanatory variable is categorical and therefore it is transformed into a dummy variable. Categories/levels of my variable:
1 - days past due in observation period = 0
2 - days past due in observation period are between 1 and 30
3 - days past due in observation period are higher than 31

So I created two dummy variables where a reference group is the worse one (category 3).

I put these two dummy variables into a logistic regression in order to obtain some statistics and in order to check a business logic of the variable.

But prolem is, that I got a really strange results. Coefficients are unusual, standard errors are very large and all of the parameters are not statistically significant. (please, see attached print screen)

Then I got an idea to put the same data into a Probit analysis. And suddenly I got acceptable and valid results (please, also see attached print screen).

I seriously ask you for a help/ideas regarding this issue because I not able to explain why this happens.

Thank you very much in advance.
 

rogojel

TS Contributor
#2
hi,
this looks like the two models would pick different levels of the y - assuming binary logistic regression, I guess one model is looking at the probanility of the event y=1 and the other at the probanility of y=0. It is telling that the cross-tabulation table at the end is the same for both models, - so essentially they behave exactly in the same way.

regards
 
#3
I dont think so. In both cases was modeled event y=1.

So do you have any other suggestions? Why I got so large standard errors in Logit and in Probit did not?

Please, I really need to resolve this problem.
 
#4
Could you provide a simple cross tabulation of y by the three levels of the independent variable? It's useful to check if there are wide disparities in the counts between levels.
 

rogojel

TS Contributor
#5
I dont think so. In both cases was modeled event y=1.

So do you have any other suggestions? Why I got so large standard errors in Logit and in Probit did not?

Please, I really need to resolve this problem.
Hi,
you are right, this can not be the problem. However, the identical cross-tabulation shows that the models are equivalent, so it has to be an interpretation issue. Could you give more output from the program, like equation for Y, odds-ratios, etc?

regards
 

JesperHP

TS Contributor
#6
It could be something with the optimization procedure .... check if you can alter how the likelihood i optimized. The Hessian could be output from optimization procedure and if it is close to being numerically in-invertable standard errors can explode.