Using Kruskal Wallis vs One Way Anova with small sample size I have an experiment with one factor and five levels. The response is proportion of conversion (ex: 1/11, 1/8...). Doing a power analysis:
power.anova.test(5, between.var =.0004, within.var =.14, n = 20)

I'm estimating between.var by simply doing var(1/11, 1/8, etc.) and estimating within.var by doing var(1/11). Is that right?
If my process is correct:
I'm finding that with an ability to only have sample sizes of around ~20 doing a one way ANOVA doesn't tell me much aka power is extremely low.
What options do I have for my experiment? Is the Kruskal Wallis test more effective with lower sample sizes or do I just need to figure out how to increase my sample size. Are there any other tests I can do?
 A: If the assumptions of ANOVA hold, changing to Kruskal Wallis won't solve the problem of small sample sizes giving low power (nor will changing to any other nonparametric procedure - even ones with ARE of 1 at the normal - like a  permutation test). ANOVA already has the best power for its assumptions and the alternative under consideration with ANOVA.
That is, alternatives will all have the same problem, at least as bad as you have for ANOVA.
One "solution" if you can't do anything about sample size is to design your experiment around a more specific hypothesis than the omnibus one afforded by ANOVA. For example, if you can choose a particular contrast of interest, or perhaps a specific (e.g. ordered) alternative, you can get more power against those specific things, at the expense of power against other alternatives.
A: This process does not seem to be correct, at least as described: these refer to between and within-group variances; so there should be 20 samples/group as described. In general, however, if your data do not comply with ANOVA assumptions (particularly homoscedasticity), rank-based KW may have greater power; permutation t-tests are often recommended for small sample sizes. 
Because these data are not truly continuous, the nonparametric KW may be a more defensible approach, regardless of whether it has a greater power. Also, "does not tell much" vs. "power is extremely low" is a bit misstated: if you detect treatment-related differences, this is all you need to know. 
