Linktest and Fitstat after Logit

[Kim Veloso]

Dear all,

I am trying to predict outcome Y=1 given individual characteristics and run a binary logit regression taking the form below.

logit Y i.age_grp i.sex i.education i.sec3 i.urbrur i.marital [fw=round(weight)]

I then runlinktest andfitstatto assess model fit and predictive power.

McFadden (1977) describes a pseudo R^2 value between 0.2 and 0.4 as an "excellent fit" which suggests that our model has strong predictive power.
However, linktest shows _hatsq as significant which suggests that our model is mis-specified...

Any ideas how this could be solved?

Thank you very much!


. fitstat

. linktest

[Kim Veloso]

I'm sorry to post again here but I could not edit my post to add further information.

I would like to add that the dependent variable here can be considered a rare event.
Only about 2% are positive outcomes. (Example: 957,182 are Y=1 and Y=0 47,859,087)

I am also trying lroc with cut-off point 0.05 because the positive outcome can be considered a rare event(?)

logit Y i.sex i.education i.sec3 i.urbrur i.marital i.age_grp [fw=round(weight)]
lroc //Area under ROC curve = 0.8869
estat classification, cutoff(0.05)

Another follow-up question is: I seem to have trouble with linktest no matter what additional variables I add in the logit regression.

However, once I remove the sampling weights, _hatsq becomes insignificant.

Any thoughts?

Thank you very much.