Hello everybody, I would greatly appreciate your help with stat. analysis. In my experiment I would like to compare 5 types of foodstuffs produced by 3 types of manufacturers (A,B,C, each type represented by 3 different manufacturers). Plus there were 3 sampling times. Although I have a complete datablock, I think that some nesting of the nuisance (random) variables is fitting, so I thought about using the following model:
foodstuff type (fixed factor) + producer type (fixed factor) + individual producers (nested/random) + sampling (nested/random and probably should be repeated measures/within-subjects).
Seems horribly difficult even without crossing (interaction), but OK, a lot of hard work for me in R and maybe it is not even the best model, but hopefully doable. But guess what, most of the dependent variables (as is common in microbiology) do not have normal distribution (too many zeros).
So, there are already some posts on the (nonexistence of) nonparametric multifactorial ANOVA and I have read some articles on the possibilities (by Oron and Hoff and by Jos Feys). But hell, I'm no statistician to pull it off.
So the options are:
1) Run only the 1-2 dependent variables with normal distribution - but what to do the with the rest?
2) Ignore the non-normality and run the complex model
3) Run nonparametric ANOVA on the 2 fixed factors separately and ignore the random factors
4) Keep pounding my head against the wall :-(
I would really appreciate your help regarding:
-my options
-checking the model I designed, if it makes sense
-suggesting other solutions, preferably together with the appropriate reliable R package
Sorry for a very long post. I'm looking forward to your responses.
My first thought in this situation is rarely a nonparametric model, but some kind go generalised linear model. This can be extended at handle random effects and a range of different outcome types and their probability distributions. Whether this is practical depends on the outcome, but it seems like if you have many zeroes that a count model might be appropriate - alternatively there are various zero-inflated models that might work.
A simpler, albeit less principled approach is to use a transformation of the outcome, but that can be messy if you have excess zeroes.
Thank you very much for suggesting zero-inflated models, e.g. glmmTMB package definetely looks promising!
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R Statistical Package