Actually, I'm performing a multiple-choice oviposition experiment (with 10 different kinds of substrate) replicated 30 times. At the end, I want to know on which substrate my flies lay more eggs. Given that, I'm dealing with count data (number of eggs laid on each substrate) characterised by non-normal distribution and an overdispersion. Is it better to use Kruskal-Wallis (followed by Dunn post hoc test if a significant difference is noticed) or negative binomial regression (followed by Tukey's post hoc test) ?
In my opinion, since you have 10 independent samples of equal sample sizes and your interest is to see in which substrate the flies lay more eggs, its better to use Kruskal-Wallis test (NON-PARAMETRIC TEST) followed by multiple comparison to separate the means, because the parametric test equivalent to the Kruskal-Wallis test is one way ANOVA. However, if your interest is to see the relationships between the substrate and egg laying, then you can use Poison regression or negative binomial if you have over-dispersion (variance larger than the mean) in your data.
From the description of your experiment, it sounds like the data would probably follow a negative binomial distribution. But since you are conducting an analysis analagous to a one-way anova, it probably won't matter much which you use. Unless there are wacky or outlying values.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics