Help needed ! - Multinominal logistic regressing in JMP ( reporting RRR and clustering correction)

[#1]

Greetings,
I'm tryting to run a mutinominal logistic regression (predicting an outcome with more than 2 catagories). I choose the (ref) using value ordering then I go to (analyze > fit model),I then specify the outcome variable and predicting variables. I keep the default nominal regression option and click run. 

My questions are:

1- How do I get odds ratios (OR) or relative risk ratios (RRR) ? I have noticed that odds ratio is an option in binary logistic regression using the red triangle but it is not availabe here. I still read publications where they report RRR in multinominal regression but I can;t find it in JMP
2- How do I adjust for clustering in my model ? I have a big database with a lot of patients that come from the same hospital or facility type and I want to adjust for that in my regression model.
along with the answer please let me know if there is any availabe video tutorial.
I appreciate any help. Thanks

[#2 · Accepted solution]

1. The odds ratios and relative risk ratios are not available directly but you can easily compute them from the fitted model parameters. Click the red triangle and select Save Probability Formulas. Now you have a series of new data columns with the formulas that you need to compute the odds from the probabilities. The Lin[Y] formulas are the linear predictor for the logit (log odds). The Prob[Y] formulas are the conversion of the logit back to probability. You can now use Table > Summary command to get the (mean) probability for groups of interest and then compute any odds ratio that you want.
2. You can add the 'clustering' variables as covariates to the linear predictor just as you did the predictor variables.

[#3]

Thanks for your repose. I still need some details please

1- I did what you said and indeed I got Prob columns, one for each catagory of the outcome variable ( it seems that this is the probablity of that outcome to happen), I also got Lin columns for each outcome (except ref outcome) not sure how Lin is helpful. Table> summery gives me the mean indeed but how do I compute odds ration from it ?. even If i get it, it seems to me this would be looking at all predictors together. In binary logistic regression JMP gives me odds ratio for each predictor with it is P value and confidence interval. how can I get this in Multinominal regression ?

2- I see other poeple working on same/similar databases using Generlized Estimating Equations (GEE) to adjust for clusting in their regression models. Please correct me if I'm wrong but I know that JMP doesnt run GEE. is adding varaibles suspected of clustering (like facility type or hospital ..) to the model have the same function/effect of GEE here ? if yes why do many others bother with GEE then ? also is there anything else I can do in JMP that has similar function to GEE in clustering correction ? 

I'm using is JMP Pro 13.2.0. Thanks again. I really appreciate it !

[#4 · Accepted solution]

1. You would never use the Lin[Y] formulas for your purpose. You would always start with the Prob[Y] formulas. You summarize over the grouping of interest so that the mean represents the expectation (probability). Then you compute the odds and finally the odds ratio. This calculation is done manually, but you get all the probabilities you need from JMP, so there are only two more manual calculations.
2. GEE are a further generalization of the GLM. You are correct that JMP does not provide GEE models, but SAS does. As to their popularity or advantage in this kind of study, I cannot say. I am not an expert in GEE. Perhaps someone else can comment about their advantage or the adequacy of the types of models that are available in JMP and JMP Pro.

[#5]

Thanks for your reply, It doesn't have to be GEE. Any method in JMP that can help adjusting for clustering in my logistic regression models is fine too. In SPSS for example there is GEE and also complex samples option. I prefer to use JMP only if possible. I'm more comfortable with it and it usually has more than I need since I'm not a statistician.  I'm still hoping someone could point me to soultion in JMP if doable. Thank you