Dear Brian:
Two remarks: 1) One of the problems with hypothesis testing is that if you have a very large dataset, than can very precisely measure very small effects, so you are likely to find significant results that are so small that they are not substantively meaningfull. This may or may not be the case in your dataset, but if your data are that large, than all other test will be effected by it to some degree. Anyhow, if you later on analyse the results I would not solely look at the significance levels, but also at whether the size of the effects (for instance odds ratios, or marginal effects) are large enough to be of any real interest. 2) General fit statistics do not tell you what the problem is and how to fix it. Do you have any hypotheses about why the fit is bad? If so, just test those hypotheses directly (and remember my earlier remarks about significance in a very large dataset). Also look at the deviance and the standardised pearson residuals.
Maarten
-----Original Message-----
From: [email protected] [mailto:[email protected]]
Sent: Sun 10/10/2004 10:18 AM
To: [email protected]
Cc:
Subject: st: Osius-Rojek, Stukel's goodness of fit tests for Logistic Regression
Dear Colleagues:
Does anyone know how to implement an Osius-Rojek goodness of fit test (it is a normal approximation to the Pearson Chi-square statistic) for a logistic regression model using Stata commands? How about Stukel's test? I'm constructing a logistic regression model on a very large database and the Hosmer Lemeshow statistic indicates that the model does not "fit", however this is in part due to the sample size of the model. Hosmer and Lemeshow in their textbook, Applied Logistic Regression, recommend two other goodness of fit tests to go with their statistic, Osius-Rojek and Stukel. Any help that you can give me in implementing either of these tests, or any advice on using Stata to assess the goodness of fit in a logistic regression model when the sample size is large would be greatly appreciated. Thanks!
-Brian
<<winmail.dat>>
