As I prepare some work for publication I would like to do an ordinal logistic regression, as opposed to the linear regression which I had originally used (and am much more comfortable with).
I have found a couple of papers online and viewed videos and powerpoint presentations in relation to the write up, but does anybody have any examples of good write ups to hand, be they from a project or a publication? Or even any tips for what to include (and not to include) in the tables and text.
I am a social science researcher (Psychology) if that is more descriptive as to what I would be required to include in my write up.
Make sure you tests for the homogeneity of the odds ratios before you accept the results. One of the major assumptions required for ordinal logistic regression is that the odds ratios are similar across all adjacent categories.
I belive you should raport the results in the same way as you would raport it for linear regression (regression coeficient, SE, statistic value, p value etc.)
Here is the link to a paper that raport logistic regression results. It is using glmm (and is from Ecology ..so not Psychology;))) but I think reporting is roughtly the same.
http://www.staff.amu.edu.pl/~rzwolak/userfiles/2010%20Zwolak%20et%20al_(2).pdf
best,
One of the issues with reporting logistic regression and probit results is the choice of an appropriate "goodness of fit" statistic for the overall model -- as opposed to the individual variables. Here's the title of one article that examines this question and makes some recommendations. There's more literature about this.
STATISTICS IN MEDICINE, VOL. 16, 965Ð980 (1997)
A COMPARISON OF GOODNESS-OF-FIT TESTS
FOR THE LOGISTIC REGRESSION MODEL
D. W. HOSMER,*1 T. HOSMER,2 S. LE CESSIE3 AND S. LEMESHOW1
Thanks for your answers, although my data meets the assumption of homegeneity of the odds ratio, the model fit is pretty poor when calculated by pearson's chi-square, so it's difficult for me to interpret anything meaningful from the model (I have a high percentage of missing cells so I'm not overly surprised).
However, both papers have been great in helping me analyse and write up the results; I have used the tables in the Zwolak paper as a template (with the addition of exponential beta values).
Regards
Tom
Hi Tom, I am not sure what you mean by "a high percentage of missing cells." I am guessing you don't mean missing but sparse. In other words, certain categories that aren't selected depending on the values of certain predictors, or even that are rarely selected at all, right? That would tend to give bad fit. Missingness is a whole different problem, which you should address. Bob
Hi Robert, thanks for your message. Yes, apologies for the rather confusing terminology (although I don't know if what follows will be any better), but I was referring to a large number of cells with zero frequencies i..e. combinations of my predictors at different levels of my DV, as opposed to missing data.
Tom
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis