turn on suggestions

Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

Find a Community

- Home
- /
- Analytics
- /
- Stat Procs
- /
- Goodness of fit in Binary Logistic Regression

Topic Options

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

11-06-2012 04:32 PM

Hi can any one help clarify doubt in goodness of fit in binary logistic. I created a model to predict the event response and got excellent c score of about .8 and also an attractive ROC carve. I checked the model by scoring the validation dataset and able to capture more than 70% of my responses in top 4 deciles. Also the KS score looks significant. Only thing I am having a trouble with is HL Hosmer Lemshow statistic coming significant means I need to rethink about the model. My question here is can i ignore the HL test and rely more on predictive power? Also, I am using proc logistic here. Appreciate any help on this.

Accepted Solutions

Solution

11-06-2012
04:50 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

11-06-2012 04:50 PM

The Hosmer-Lemeshow test is not my favorite test; it has low power in smaller samples and can show significance for important deviations in vary large ones. I much prefer to look at the observed-predicted plots themselves.

If the HL test is significant, it doesn't say that the model you have is "wrong," it says that it can be "improved." Sometimes "improvement" means a different model, additional variables, or data transformations. However, if the model seems adequate, I may just "declare victory" and move on.

One caution (that doesn't seem to matter here), if the HL test is significant, then it would be inappropriate to claim that nothing is going on (e.g. to "accept the null").

Doc Muhlbaier

Duke

All Replies

Solution

11-06-2012
04:50 PM

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Highlight
- Email to a Friend
- Report Inappropriate Content

11-06-2012 04:50 PM

The Hosmer-Lemeshow test is not my favorite test; it has low power in smaller samples and can show significance for important deviations in vary large ones. I much prefer to look at the observed-predicted plots themselves.

If the HL test is significant, it doesn't say that the model you have is "wrong," it says that it can be "improved." Sometimes "improvement" means a different model, additional variables, or data transformations. However, if the model seems adequate, I may just "declare victory" and move on.

One caution (that doesn't seem to matter here), if the HL test is significant, then it would be inappropriate to claim that nothing is going on (e.g. to "accept the null").

Doc Muhlbaier

Duke