topic Re: Confidence interval for Odd Ratio in nominal logistic regression in Discussions

Confidence interval for Odd Ratio in nominal logistic regression

mhorva — Thu, 10 Jun 2010 14:41:10 GMT

In a prior post, someone indicated that CI's for the odds ratios were removed... does anyone have guidance on how to hand calculate these?

I have the parameter estimate CI's and can do it easily for when a variable has only 2 levels, but am unsure of the formula when there are 3 or more.

Thanks

Re: Confidence interval for Odd Ratio in nominal logistic regression

mhorva — Thu, 10 Jun 2010 14:53:58 GMT

Similarly, i thought that the OR for a term was exp(b), where b is the parameter estimate.

It appears that JMP is using exp(2 * b) ... i have an explanation why from a JMP class; however, I then dont really understand why every stats txt out there calcs as exp(b)...

Re: Confidence interval for Odd Ratio in nominal logistic regression

mpb — Thu, 10 Jun 2010 20:53:53 GMT

Look at pages 439-440 of the JMP Stat Graph Guide. They discuss and show an example of Unit Odds Ratios and Range Odds Ratios. This distinction may explain what you are seeing.

Re: Confidence interval for Odd Ratio in nominal logistic regression

mhorva — Thu, 08 Jul 2010 15:47:52 GMT

That seems to only be relevant when the predictors are continuous; in my case, the predictors are nominal with 2 or more levels. I dont get unit and range odds ratios in my output.. just 'odds ratios'.

Re: Confidence interval for Odd Ratio in nominal logistic regression

mpb — Thu, 08 Jul 2010 19:50:40 GMT

Near bottom of p439:

Two-level nominal effects are coded 1 and -1 for the first and second levels, so range odds ratios or their reciprocals would be of interest.

This seems to point in the same direction ... If you're still concerned, I'd suggest contacting JMP support and letting us know what they say.

Re: Confidence interval for Odd Ratio in nominal logistic regression

mhorva — Thu, 15 Jul 2010 15:52:09 GMT

Yes, I understand now. JMP is using effect coding whereas I had assumed I was using reference coding due to the value ordering I had set up.

Thank you for the nudge in the right direction. It makes sense now.

I do wish they'd bring back OR CI's though. :)