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
- /
- GLM - inverse gaussian distribution

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

07-09-2013 12:27 PM

I am estimating a GLM where the response variable is very skewed. So, I used gamma distribution and log link options. Now, for a comparison purpose, I was trying Normal distribution and Inverse Gaussian distribution. The model give similar coefficients with Normal distribution, but with Inverse Gaussian, the estimations are different, and the predicted values almost blows up (veryyy large). Any idea why this is happening?

Thanks for any help you could provide.

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

07-10-2013 07:30 AM

Without seeing your data and SAS programming statements, it's difficult to tell. However, according to that great statistical reference, Wikipedia, the maximum likelihood estimates of the two parameters of the inverse Gaussian distribution depend on the reciprocals of the observation values. If any of these values is close to zero, this could cause problems in estimation and prediction; if the predictions that "blow up" are associated with observation values close to zero, then this may explain what is happening. I don't know whether adding a constant to the observation values (translating the values away from zero) would improve these predictions.

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

07-10-2013 04:09 PM

I understand it might be hard to say whats going on without looking at the data. I can see the independent variable with the largest range is creating the problem. With inverse gaussian, the estimated coefficient for this variable is quite large causing over predictions.