Hi Jeremy,
Thank you so much for this. From my readings, I was gathering that Discriminant Analysis was becoming less frequently used and that the results were seen as arbitrary. On the other hand, working with probabilities of predictive abilities of a variable, as regression analysis does, seems much more 'accurate' and acceptable.
My outcome variable of anxiety is ordinal (not clinical, borderline clinical, clinical) and my predictor variables are continous/scale data calculated from Likert responses pertaining to frequency of use of certain variables. I seem to have managed to use Multinomial Regression Analysis with these variables, but when I attempt Ordinal Regression Analysis I am faced with some difficulties: 1. I am faced with the following message:
Warnings
There are 70 (66.7%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.
The log-likelihood value is practically zero. There may be a complete separation in the data. The maximum likelihood estimates do not exist.
The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.
and, 2. my predictor variables appear to be broken down into the items that make it up as opposed to keeping it as a continuous variable. My output is nothing like what I would expect from a regression.
Would I be wrong to continue with the Multinomial option? On the other hand, I would prefer the ordinal option - can you direct me to any online guidance for how to conduct/interpret it? I have been using an SPSS book that only appears to include binary and multinomial!
Thanks again,
Nicola
P.S. Helping me seems like hard enough work!
________________________________________
From: Jeremy Miles [[log in to unmask]]
Sent: 22 November 2008 00:00
To: Davies, Nicola
Cc: [log in to unmask]; [log in to unmask]
Subject: Re: Discriminate Analysis Vs Multinomial Regression Analysis
P.S. If you didn't ask so many questions, people like me would have to do work and stuff.
2008/11/21 Jeremy Miles <[log in to unmask]<mailto:[log in to unmask]>>
I wouldn't say that you'll be bombarded with criticism, but ...
Logistic regression does not care whether your predictor variables are normally distributed or not. DFA does.
Personally, I find logistic regression to make much more sense.
If you have three categories of anxiety, are these ordinal, rather than categorical? If so, you might be better off with ordinal logistic regression. Ordered logistic is more powerful, and unless you've got a really big sample, that's what you want.
Jeremy
2008/11/21 Davies, Nicola <[log in to unmask]<mailto:[log in to unmask]>>
Hello Again,
I hope people don't mind me asking so many questions, but as I am analysing my data and reading through my SPSS book I keep finding questions that interest me. I am most certainly learning a great deal from everyone; thank you so much.
I was wondering what your thoughts are on Discriminate Analysis v's Multinomial Regression Analysis? I am examining whether my DV (anxiety level - 3 categories) can be predicted by my IVs (subscales on a questionnaire measuring benchmarks used when assessing health status). I felt that a Discriminate Analysis would be the the best method, but on proceeding with this I have been reading that there is some competition between this method and the regression analysis. Is one seen as more valid in the world of psychology statistics? Will doing a Discriminate Analysis end up getting me bombarded with criticism during my viva?
Best Wishes,
Nicola
--
Jeremy Miles
Learning statistics blog: www.jeremymiles.co.uk/learningstats<http://www.jeremymiles.co.uk/learningstats>
Psychology Research Methods Wiki: www.researchmethodsinpsychology.com<http://www.researchmethodsinpsychology.com>
--
Jeremy Miles
Learning statistics blog: www.jeremymiles.co.uk/learningstats<http://www.jeremymiles.co.uk/learningstats>
Psychology Research Methods Wiki: www.researchmethodsinpsychology.com<http://www.researchmethodsinpsychology.com>
|