I have collected ordinal data (scale from 1-5) on the satisfaction rating of different preventive measures in sports (through questionnaires). I have 400 respondents and five different preventive measures, where some respondents have rated only one preventive measure, others have rated several. I want to compare the difference in satisfaction between groups (each preventive measure being a separate group).
Can the ratings be considered independent when some subjects have rated multiple preventive measures?
I have been recommended a mixed model ANOVA which compares to factors (within subjects and between subjects) but as far as I can tell, I only have a between subjects comparison as time is not a factor.
If this is the case, which statistical analysis should be recommended?
I would start with 5 separate models , one for each preventative measure. That would give you a 'measure of success' of each measure for each prevention - knowing that 200 tried this and it only worked for 2 is valuable.
Whether you can go further depends on whether not responding to a particular preventative means that the respondent did not take that measure. Are the instructions clear in the questionnaire?
Mixed models can handle situations with multiple responses. It depends on how you asked you question. It could be treated as a 'matrix sampling design' where you have imbalance that is not everyone answers the question so that there are potentially 5 outcomes for each respondent but there is imbalance as some only answers 1, 2, 3, or 4 questions. You could then include predictors at level 2 above about the individual in this multivariate multilevel model
http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#multipleresp
Hi Kelvyn, thanks for the input.
The respondents have only rated the preventive measures they used. I want to compare the difference in satisfaction between the different preventive measures, so how would a separate model for each work?
Dear Kasper Petersen,
Only one more observation. ANOVA is a parametric test, data must have Gaussian distribution. If it is not true, I advise you using Friedman's test.
Your sincerely,
Elcio Oliveira
Hi Kasper,
this sounds like you should be using a multinomial (or polytomous) Rasch model. Usually Rasch models are used to evaluate the abilities of persons, controlling for the difficulty of items presented to them. But they can also serve the opposite perspective, which applies in your case, i.e., to assess the rating of items, controlling for the general answer tendency of respondents.
You might use Mplus or some R package like eRm, to solve your model.
Good luck with your analysis,
Lars
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA