0
$\begingroup$

I am running a Wilcoxon-signed rank test, and as suggested by King et al., the best effect size to report for this analysis would be the matched-pairs rank biserial correlation.

How is the power (1-B) calculated based on this effect size?

The package "pwr" in R allows calculating the power based on linear correlation coefficients (i.e., Pearson correlation coefficient, "r"). Yet, the matched-pairs rank biserial correlation is very close to the Pearson correlation coefficient. Could I use the latter coefficient in the pwr.r.test() function in R?

Another way would be to use the G*Power software by selecting "t-tests" in the Test Family selection menu and then by selecting "Correlation: Point biserial model" in the Statistical test selection menu. Would this be correct?

Thanks in advance!

King, B. M., Rosopa, P. J., & Minium, E. W. (2018). Statistical reasoning in the behavioral sciences. John Wiley & Sons, 383-384.

$\endgroup$
3
  • $\begingroup$ Perhaps you could calculate power for a paired t-test. In most cases, the Wilcoxon signed-rank test is more powerful, so this would give you a somewhat conservative estimate for power. $\endgroup$
    – num_39
    Mar 15, 2022 at 6:18
  • $\begingroup$ Thank you for your answer :) Yes, that would be one way, however, to calculate it for paired t-tests I would need to use Cohen's-dz which it is not the most appropriate effect-size for non-parametric tests. But as long as I report the limitation it may be good, thanks! $\endgroup$
    – Sara Coppi
    Mar 16, 2022 at 9:27
  • $\begingroup$ The problem here is trying to estimate power without any knowledge of the parameters. If you know something about the parameters, then you have some options. See this answer: stats.stackexchange.com/questions/521477/… $\endgroup$
    – num_39
    Mar 16, 2022 at 9:55

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.