Felix Bach wrote: > > Hello, thanks for your reply. I think I wrote a bit misunderstanding. > Actually I am only interested if people tend to use more often grammar > type A (actually the interest is if younger children use more often > type B, while older children type A) Thereofre, I would have something > like this > person Type A B C > 1 3 4 3 > 2 2 6 2 > 3 1 9 0 > 4 .... > 5 > ... > 10 > > I assume the soundest way is to use a multinomial regression with > repeated measurements (person ad random factor) but would it be > possible to use an classical anova approach? > Best wishes > Felix
If I've understood you, people use this kind of model all the time. The main concerns are i) the assumptions of ANOVA with count data, ii) the ipsative nature of the data (A+B+C=10). In studies of ipsative data analyses suggest that the the type I error rates etc. for these models are not a problem (there is a Psychological Methods paper on this - I don't have the full ref though). The former can be checked with the standard model assumptions - normality of residuals and so forth (if they are reasonable then the estimates should be close to those the more complex analyses that were suggested). Last, if you really want to test A vs.B between ages then a focussed contrast approach to the interaction effect is likely to be the most powerful option (and might side-step some of the analytical problems of the repeated measures analysis). Thom . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
