topic Re: repeated measures of scores in Statistical Procedures

repeated measures of scores

mdsaraiv — Fri, 12 Feb 2016 13:42:40 GMT

I am doing analysis of a scores ( 0 to 5) taken repeatedly over time (21 consecutive sectionsequally spaced) in two treatment groups. Since it is not normally distributed which would be the best statistical procedure? I am not statistician, I am epidemiologist. I usually work with repeated measures with normal distribution using mixed models and so.. but I am having trouble to get the best analysis for these scores. I was reading about harsh models but I am not sure if it would be appropriate. Would rasch model appropriated? It would sum up items for individuals but I am not sure if it would work. It is the same measure over time.

Re: repeated measures of scores

Rick_SAS — Fri, 12 Feb 2016 13:47:18 GMT

I'm not an expert, but I think generalized estimating equations (GEEs) are the generalizaton you are looking for. SAS released the GEE procedure in SAS/STAT 13.2.

If you are running an older version of SAS, there are some GEE options in PROC GENMOD. For ideas, read this 2006 paper on "Analyzing Ordinal Repeated Measures Data Using SAS" which shows several examples. An even older paper is this 1997 paper on "Repeated Measures Analysis with Discrete Data Using the SAS System."

Re: repeated measures of scores

mdsaraiv — Fri, 12 Feb 2016 15:37:17 GMT

Thanks so much, I will read and see if makes sense for my data. I will get
back if it works.

##- Please type your reply above this line. Simple formatting, no
attachments. -##

Re: repeated measures of scores

SteveDenham — Mon, 15 Feb 2016 19:48:12 GMT

I would certainly start with PROC GEE, as Rick states, and in particular I would look at this example in the documentation:

http://support.sas.com/documentation/cdl/en/statug/68162/HTML/default/viewer.htm#statug_gee_examples05.htm

More generally, I would use PROC GLIMMIX, and treat the data as coming from a multinomial distribution. However, in GLIMMIX as opposed to GEE, the model will be fit as conditional on the random effect, rather than as a marginal over the random effect (which is time in this case). The marginal log odds ratios will tend to be biased toward the mean compared to the conditional log odds ratios.

Steve Denham