You might try this tutorial here: http://www.ats.ucla.edu/stat/r/dae/ologit.htm - it is fairly detailed and includes plenty of code and plots.

Much appreciated - just what I'm looking for.

This kind of data (discrete data) follow other distribution than Normal distribution. So, you need to run a regression assuming a Binomial or Poisson distribution of your dependent variable. This runtime is easy by using the PROC GENMOD from SAS System. Best regards.

Thanks

You can use the ordered probit or censored regression, they are among the family of probability models where the techniques of maximum likelihood are used to estimate the outcomes. packages such as stata minitab, e views and matlab can be used in estimation

Hello Ngoni, May I also ask please whether or not there is a cut-point such that you can classify the test results as either pass or fail or with disease or without disease as then you could perform a logistic regression. Just an idea. Good luck with your analysis. Deborah Hilton Statistics Online

http://sites.google.com/site/deborahhilton/

Thanks, Mustapha

Hello Deborah. There is indeed a cut-point that allows the results to be classififed in terms of pass / fail. Thanks, for the suggestion.

Hello Ngoni, in that case refer to the paper on my page that is the diabetes paper in the medical journal of Australia. Cut points were used to classify people as having diabetes or not having diabetes so this maybe a useful example for you to read and maybe this will help you with another idea for analysis in that you could perform a logistic regression along with gender and then your other independent variables in the model. Cheers, Deborah

Hello Ngoni, I've just seen your question and I think my answer above should help in that if you can access this you can see how the predictors are then classified or graded in importance. In terms of having diabetes or not, the strongest predictors were age and weight, followed by blood pressure, family history etc so yes you could determine their degree of importance and also there was confounding/interaction between age and blood pressure, gender and method of diagnosis so that this could also be assessed in the model. It is a while since I published this work, so I have not looked at it for sometime but maybe it is of use. Good luck with your work. Regards, Deborah

31 Jan 2013, 16:04

Estimating probabilities, given categorical outcome data and any list of features, however entangled or correlated, can be done rigorously and consistently using a probability machine. (Reference below). That is, any code that solves the nonparametric regression problem and is provably consistent, will estimate probabilities for any subject or instance. Regression, after all, is the method by which we estimate expectations given outcomes and some list of dependent variables, and the estimated expectation from zero-one outcomes is the probability for that outcome. Probability machines can also be used to provide entirely model-free risk estimates and interaction terms. Routine simulations comparing a probability machine to logistic regression, for example, show that it does as well as LR when the model is known and true, and does very much better when the model is not known, both for subject-specific probability estimates and for risk estimates. Our personal and practical preference is Random Forest run in regression mode, but any R code for a good nonparametric regression learning machine should do just as well. Extension to category data is immediate from the binary, zero-one case.

Ngoni, I think there are very many useful ideas here. I would just refer you to generalized mixed linear models, something that goes back to the late 1980s and 1990s and includes some of the models mentioned here including the Poisson and binomial. These can also be used to model ordinal variables as a linked set of logit models where the probabilities of each response are conditioned on the others. Another variation can do unordered categories. I can't speak to exactly what software has such models incorporated or not, but these models, although not all that common in publications, are hardly unknown. Personally, I use them in HLM, but they aren't all that obscure. It sounds like you might be better off with a single cut point based on the discussion above. Bob