[Probit] Can I be confident in my results even with a poor Goodness-of-Fit test?

#1
I am running a probit analysis on a data set with about 1000 cases, predicting probability of college graduation by a set of college institutional variables (funding, services, etc).

For several predictor variables, I am getting signficant coeffecients which have signs in the expected directions, but the Pearson's goodness-of-fit test for the model is significant at the .05 level (I am using SPSS).

It appears to be *very* significant, Chi-sq of 11,766.414/919 df, sig = .000

Apparently this means that the model is not performing well, because the expected cell frequencies differ greatly from the observed cell frequencies.

My question: How important is this goodness-of-fit test result? Does this mean that the model is invalid or basically useless (analogous to a low adjusted R-square in linear regression) or is the goodness-of-fit test merely advisable in probit?

Bottom Line: Can I confidently report my results, even with such a poor result on the goodness of fit test?

I have done quite a bit of web searching and searching in SPSS product help and can't find an answer. Any help here would be appreciated. Thanks!

DavidBill
 
#2
It's pretty much understood that when samples are large like yours that the fit statistic will always be significant. It's not usually a problem.
 

Dason

Ambassador to the humans
#3
It's pretty much understood that when samples are large like yours that the fit statistic will always be significant. It's not usually a problem.
This is true. However if the model isn't too complicated it's worth it to take some time to consider the assumptions you're making and to think if there is a better modeling solution.
 
#4
Thanks much for the replies--they definitely ease my mind. I appreciate the suggestion of considering another modeling solution, but I'm not familiar with what alternative statistical models I might use in this situation. I have Googled quite a bit on this, and probit (or logit) are the only suggestions that have surfaced.

To restate: I have a percentage dependent variable, and a mix of categorical and continuous predictors. None of the predictors are percentage data.

If it might be possible to get a quick mention of a couple of other models that could be appropriate for this situation, that would be most appreciated.

DavidBill

PS: I grew up in Ames--nice to hear from (I assume) another IaStater.