Dear community,
I can't seem to find any clear information on power analyses work so I'm throwing my question out onto the internet hoping for answers. :)
Study Details
I am doing a within-subjects design with two independent factors. The dependent variable is binary correct/incorrect.
There are 18 test items.
9 items vs 9 items for Condition A (2 levels)
3 vs 3 vs 3 within each of those 9 for Condition B (3 levels)
I want to be able to detect significance for an Odds Ratio of 1.800, 2-tailed.
In a similar experiment, the lowest mean for a condition was .177.
Question
My question is how I should use GPower to get a minimum sample size?
I've tried z tests -> logistic regression and it gives me a sample of 533, but there is no field to input the number of measures or independent variables, so I'm not sure how to interpret the result.
Or does it mean that I need that many individua
Am I dividing that number by 18 for the 18 test items? (approximately 30 people)
Am I dividing it by the number of conditions? (3x2=6 - approximately 90 people)l participants? (impossible)
Any help would be greatly appreciated!!
Hello Jonathan Serfaty. The method you are using in G*Power was designed for ordinary binary logistic regression where every observation is independent of all other observations. Therefore, it is not appropriate for your situation, where the repeated measures violate that independence assumption. I suspect you will have to use simulation to get your sample size estimate. If you say what statistical software package(s) you use, someone may be able to point to some relevant examples. I hope this helps.
For anyone finding this question in the future - this website is useful: https://jakewestfall.shinyapps.io/crossedpower/
Hello Jonathan Serfaty. You said that your DV is binary (correct vs incorrect). I see no option to select a binary DV in that calculator. And indeed, the article cited at the top of the page says this:
We have also confined ourselves to outcome measures that are nicely behaved, so that the underlying distributions of random effects in our models could be assumed to be normal.
Source:
Article Statistical Power and Optimal Design in Experiments in Which...
Unless you can find a similar calculator that allows for a binary DV, I think you'll have to use simulation.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training