I want to see if there is a differential change over time by group in non-normally distributed data. The questions are on 5-point scales, with the majority (any where from 63-95%) select the lowest response and 0-7% of the sample selects the highest response. I'd rather not switch this to 0/1 response and lose the variability. I generally work in SPSS, but could run analyses in SAS also.
With SPSS, one option would be to use GENLIN with Generalized Estimating Equations (GEE) to account for the paired nature of the data. Via the GUI, click on Analyze > Generalized Linear Models > Generalized Estimating Equations. On the "Type of Model" tab, select Ordinal Logistic (or Ordinal Probit, I suppose, if that is the more usual approach in your discipline).
HTH.
Generalized estimating equations would be a fine way to analyze data from a 5-point scale, but to be clear ... it wouldn't be non-parametric.
If you only have one factor, then you should be able to find a non-parametric procedure in SPSS or SAS for "Friedman's Test". If not, then you could just try a one-way ANOVA with repeated measures on the ranks of the data (they are approximately the same thing). If you have more than 1 factor to examine, then you pretty much need to consider parametric methods ... specifically linear mixed models or preferably generalized linear mixed models, as mentioned above.
Hi,
Your question: How can you run repeated-measures analysis in non-parametric data?
Kindly see the following link:
http://users.iems.northwestern.edu/~ajit/papers/28%29%20Repeated%20Measures%20Munzel-Tamhane02.pdf
Jeff wrote: "Generalized estimating equations would be a fine way to analyze data from a 5-point scale, but to be clear ... it wouldn't be non-parametric."
Agreed.
And while we're on the topic, the expression nonparametric data doesn't make sense (to me) either, even if one is using a nonparametric test. In that situation, in other words, it is the test that is nonparametric, not the data.
Furthermore, as has been discussed in some other threads recently, some tests that are conventionally called nonparametric do actually do estimate population parameters--e.g., the Mann-Whitney U test provides an estimate of the population probability that a randomly selected X is less than a randomly selected Y. So the terminology around "nonparametric" tests is a bit of a mess! ;-)
Yup - should have said repeated measures nonparametric analysis... Friedman's test exists in SPSS, but can't look at both group and time in that type of analysis. I tried the GEE approach with ordinal probit which seems like a decent solution - this is pilot data... I don't really do programming, so I may be stuck with this for now. Will poke around some of Paul Yarnold's links... Thanks everyone!
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS