As the title says it all, how to calculate a sample size for KW test? I'm a bit of a beginner here, should we just calculate for the corresponding one-way Anova and then add 15% or so as I read somewhere or what?
in Sheskin's Hanbook of parametric and nonparametric statistical procedures there is a formula linking the significance level and sample size to the smallest difference in average ranks between any two groups that are tested. If you can express the strength of the effect you are looking for in terms of rank differences you could use the formula to derive a sample size.
If linking the effect size to differences in average ranks is not feasible, the strategy with an additional 15 - 20% samples compared to the corresponding parametric test could be a good way to start.
The problem is that KW is applicable for most any distribution, but for a calculation we need a concrete one. If I had this problem and some educated guess on the distribution were availble ai would try something like bootstrapping. I would generate a sample of a given size with the known distribution, add an effect ov a given size and run the test, then repeat a large number of times. Change the effect size and repeat the whole procedure. Then change the sample size and repeat the whole procedure.