Hi
I have a dataset that I have recently analysed, and am concerned that my analysis may not pass muster.
I have run a trial looking at the locomotory response of fish to a light stimulus following exposure to a compound at four concentrations. The data consists of the mean tracked movements of fish in each treatment averaged into a) one second time bins and b) 20 second time bins, with the resulting data being far from parametric. I have tried natural log transformation with no success. I am currently comparing movement in the light condition with that in the dark condition within groups using KW followed Mann-Witney. My concern is that the data are not independent observations, so a repeated-measures approach would likely be more appropriate. I have read of some suggesting Friedmans, but it seems this is for block design. I'd be very grateful for any thoughts on the approach I have taken. If you were a reviewer, would you reject it? Any other suggestions on how to analyse would be gratefully received.
I'm not sure what you mean when you state the data are non-parametric. Strictly that's a property of the model not the data. Probably you mean non-normal, but that's pretty vague. What properties of the data suggest you can't use a parametric statistical model?
I'd usually consider a multilevel generalised linear model before rank transformation approaches unless you have very small samples. If you want to a apply the rank transformation approach I would rank the data and apply repeated measures ANOVA.
e.g., see:
https://seriousstats.wordpress.com/2012/02/14/friedman/
I agree with Ajit. The F-test is known to be quite robust against violations of normality. Yet, do inspect the histograms and run the Shapiro test for normality of distributions. The Shapiro test is known to be biased by size, i.e. the larger the size the more likely to yield sign. results (indicating non-normality). If the visual inspection of the histograms shows the distributions are definitely not normal (e.g., very very skewed), run a non-parametric test. You could use the nparLD R package (see attachment for more info), more specifically the f1.ld.f1 (1_between x 1_within) design in this package or the SAS macro in attachment.
You could also use the ARTool: http://depts.washington.edu/madlab/proj/art/ The authors claim it can be used with repeated measures, but I am not sure they are wright.
I would suggest you try all three. They should yield to the same conclusions.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis