There really isn't anything to "treat".  You just report the results for all the levels of race. If there is an interaction between race and distance, then there is an interaction - it makes no sense to say the interaction exists only at one level of race. That would be the equivalent of saying "Me are taller than women, but women aren't shorter than men". In addition, "significance" is a bad guideline for including or excluding a variable. But beware if any of your levels of race have small n.  That could cause problems and you could deal with those by either combining categories or deleting some people (E.g. if you use the standard US Census levels of race, then, in most of the USA, there will be very small numbers of "Native Hawaiian or Pacific Islander" and "American Indian or Alaskan Native".  In some areas there will be small numbers of other categories as well (not a lot of African Americans in Maine, for example). ... View more

It does seem to exhibit some of each; whether that will be problematic is another question - I don't know of good tools for figuring out how great a violation relates to how problematic a regresssion ... View more

Yes. That is how to read them (I am assuming each of the tage variables are coded 0 and 1). If you have actual age, I would recommend using that, instead of the categorical variables for age. ... View more

Have you considered PROC TRANSREG? You can then use the MONOTONE transformation ... View more

That is a good idea, Steve, particularly if there is censoring (which isn't clear from the original post) ... View more

First, the dependent variable does not need to be normally distributed. The errors from the model (as estimated by the residuals) need to be normally distributed.  However, from your description, it is likely that the residuals are also non-normal. Second, your outcome variable is not nominal, it is continuous (length of stay) If that is the case, then there are several options. 1) You could try transforming the DV or the IVs. 2) You could use robust statistics. Just a couple days ago I presented a paper at SGF entitled "Alternative methods of regression when OLS is not right". ... View more

Computing a p value for Cohen's kappa seems rather odd in the first place.  You would never expect the null hypothesis to be true. If it were even close  to true, then there wouldn't even be a point to computing an agreement statistic and even a highly significant kappa may be quite low for any practical purpose. So, doing many kappas does increase the chance of type I error, but if there is even a tiny chance of a type I error, something is very wrong. ... View more

Hi Paige Not sure what happened. I am at SGF and a little busy but i will try to fix it ASAP If you Google "How to ask a statistics question" and my name you can find it Peter ... View more

I don't understand what you are clustering.  If "ID" is, indeed, an ID number, then it seems unlikely (to me) to be usefully included in any cluster analysis.  Unless ID has some meaning beyond just being a code - but most ID numbers, even if they do have such meaning - don't relate linearly to anything. But what are all the variables? What is the ID the identification for? A customer? Is each unique? What are you trying to do? Not "I'm trying to cluster these data" but, in a practical, real-world sense, what is the purpose? I wrote a blog post called How to Ask a Statistics Question that may be useful Message was edited by: Peter Flom ... View more

Can you give more detail? Possibly the data?  When I tried this on a toy example, I got the same results for both models: data junk; input var1  var2  DV $ count; interact = var1*var2; datalines; 0  0 0 100 0  0 1  50 0  1 0  25 0  1 1  75 1  0 0  100 1 0  1  10 1 1 0  10 1 1 1 90 ; proc logistic data = junk;   model DV = var1 var2 interact;   freq count; run; proc logistic data = junk;   model DV = var1|var2;   freq count; run; ... View more