11-26-2012 02:28 PM
I am sure that SAS does not have a specific procedure or option to do the power for Cohen's kappa, but is there a formula that can be programmed to calculate power?
For example, I want to answer the question, "200 cases provide XX% power of obtaining a lower bound confidence interval of 0.60." Since we are dealing with power, that question may need to be adjusted to compare to a hypothesis test rather than a confidence interval, but I am not sure.
I know that my data has high prevalence, so is there a way to do these power calculations for PABAK (prevalence-adjusted bias-adjusted kappa)?
As a side note, I will need to do power calculations for Fleiss' kappa (a multi-reader kappa) in the future. I have not found any literature on this subject. Does anyone know if power calculations exist for Fleiss or even ICC since they are comparable?
11-27-2012 10:32 AM
This is going to be kind of scattered all over the place, but let's see where it goes. The confidence bounds and tests that SAS reports for kappa are based on an assumption of asymptotic normality (which seems really weird for a parameter bounded on [-1,1]). If you are willing to accept all of this asymptotic kind of thing, then you can calculate power based on inverting the formulas in the PROC FREQ documentation, and applying a non-central t to calculate beta, to get 1-beta=power.
Or you could rely on simulation, which is likely a better approach, especially for PABAK.