Is there a straightforward way to conduct a sensitivity power analysis for a Kendall's Tau correlation? I was considering using the sensitivity setting and "Correlation: point biserial model" test of G*Power and then converting that effect size output to Tau, but I'm concerned that may not be accurate since Tau is a non-parametric association. Any help would be greatly appreciated!
I think G power do not support non parametric methods. You may convert tau to person but handle the results with caution as it may be not accurate.
Onipe Adabenege Yahaya
,One of the problems when lawyers try to use ChatGPT and similar large language AI models is that they may cite court cases that don't exist.
Likewise, they may cite R functions that don't exist.
But I'm sure you've used pwr.kendal.tau() in the past and had great success.
Sal Mangiafico ,
Thanks for clarifying; do you know of an existing R function that would work? I normally use SPSS so I'm not very familiar with R.
Hi, Erin E. Kosloski . I did find this one function, that is available through Bioconductor: https://rdrr.io/bioc/genefu/man/power.cor.html . I've never used it though.
The following works in R:
if (!require("BiocManager", quietly = TRUE)){install.packages("BiocManager")}
BiocManager::install(version = "3.19")
BiocManager::install("genefu")
library(genefu)
power.cor(rho=0.5, w=0.1, alpha=0.05, method="kendall")
power.cor(rho=0.5, w=0.05, alpha=0.05, method="kendall")
I don't really understand what w does.
Here is the source code for the function, which may help:
power.cor = function (rho, w, alpha = 0.05, method = c("pearson", "kendall", "spearman"))
{
method
Sal Mangiafico I do not understand w either, but it seems to be crucial and changes the sample size drastically. Anyone an idea?
Hi Sal Mangiafico , thanks for sharing that. I was having difficulty with Bioconductor but ended up using the "presize" package (link below). I used the "prec_cor" function where you choose your confidence level, your type of correlation, and your effect size value and then input either your sample size or your desired width of confidence interval. Since I had a set sample size already (because I was doing a secondary analysis), I put in my n and the output was the lower and upper limits of the confidence interval.
https://cran.r-project.org/web/packages/presize/index.html
Erin E. Kosloski , that package is a great idea. Why don't we usually do sample size calculations this way ? The t-test one makes so much sense, without having to go through the Cohen's d calculation.
- Badges
- Science topic