2

I have an experiment in which I'm doing, among other things, a One way Anova with one dependent variable and 5 groups. After the ANOVA i'm doing Tukey's post hoc to test for differences between groups.

I want to conduct a power analysis to determine if my power is enough for detecting differences between groups.

I tried g power but it only gives you the power for the omnibus test.

Any ideas on how to do it? Would I end up having a power value for each of the comparissons?

Thanks in advance

Gonzalo Lerner
  • 21
  • 2
  • Hint: think in terms of the minimum difference necessary. With post-hoc tests you will I think be testing whether any single group difference is at least as large as ___, and you'll assume all other differences will be smaller than that. – rolando2 May 08 '18 at 20:06
  • If you are planning and conducting the experiment, you probably have hypotheses in mind. Therefore, you could/should use confirmatory, a-priori [tag:contrasts] (aka [tag:planned-comparisons-test]) instead of rather exploratory post-hoc comparisons. And thus you can conduct power analysis for contrasts. – hplieninger May 09 '18 at 07:57
  • Thank you @rolando2 . I'm not sure if I undestood correctly. Are you saying that I should grab the minnimum difference between 2 groups, do the power analysis for that single t test and then if I have enough power to detect that one I will have powe to detect the other (larger) ones? If that is the case. How to take into account the correction that is made by Tukey post-hoc? – Gonzalo Lerner May 10 '18 at 14:18
  • Thanks @hplieninger i found some info on power analysis of contrasts. In that case I wouldn't correct for multiple comparissons? What if I'm actually interested in the difference between all groups? – Gonzalo Lerner May 10 '18 at 14:45
  • To clarify. The specific problem is that I'm seeing two groups that look pretty different while plotted but are not when compared with Tukey. I want to know if that p>.05 is reflecting equality or i'm underpowered because of my sample size (15 per group). – Gonzalo Lerner May 10 '18 at 14:48
  • 1
    You don't need a correction for contrasts. With five groups, you can test four specific hypotheses; do you *really* have more? Then you're left with post-hoc tests (with correction). – hplieninger May 11 '18 at 07:10

0 Answers0