I have been wanting to get away from using power calculation tools (JMP, Lenth) in favor of Monte Carlo techniques to determine the a-priori power to observe factor level effects in a DOE. I am really new to this and have a few questions. Lets start with what I have:
- A binomial response which I plan to perform a logistic regression to
- List of factors that I think are relevant
- A model (Main Effects + interaction + quadratic terms)
- An experimental design (Taken from JMP or the DoE.wrapper package)
- A threshold requirement (0.6)
- An effect size (ES = 0.1)
I would like to determine the power to observe an effect in the response for each one of the model terms. I have read the excellent response by user gung which I wanted to use to help tailor my own approach. After reading, I had a few specific questions.
- The example uses prior knowledge of the system performance (the rates) that I do not have. How do I continue without this knowledge?
- The solution does not factor in effect size in any way (at least as far as I can tell). How do I factor in effect size to this process?
- The power for each of the terms is coming from the significance of the model term to be non-zero. Is this the same as the power to observe an effect size in the response?
For question 2, I would assume that you would just need to run the model assuming the null performance (p = 0.6) and then re-run the fit assuming the performance of Null + Effect Size (p = 0.6 + ES). I know this last part might be a bit jumbled up. Any help would be greatly appreciated.