The function 'anopaN2Power()' performs an analysis of statistical power according to the 'ANOPA' framework. See Laurencelle and Cousineau (2023) for more. 'anopaPower2N()' computes the sample size to reach a given power. Finally, 'anopaProp2fsq()' computes the f^2 effect size from a set of proportions.
anopaPower2N(power, P, f2, alpha)
anopaN2Power(N, P, f2, alpha)
anopaProp2fsq(props, ns, unitaryAlpha, method="approximation")sample size;
number of groups;
effect size Cohen's $f^2$;
(default if omitted .05) the decision threshold.
target power to attain;
sample size per group;
a set of expected proportions (if all between 0 and 1) or number of success per group.
for computing effect size $f^2$ is 'approximation' or 'exact' only.
for within-subject design, the measure of correlation across measurements.
anopaPower2N() returns a sample size to reach a given power level.
anopaN2Power() returns statistical power from a given sample size.
anopaProp2fsq() returns $f^2$ the effect size from a set of proportions
and sample sizes.
Note that for anopaProp2fsq(), the expected effect size $f^2$
depends weakly on the sample sizes. Indeed, the Anscombe transform
can reach more extreme scores when the sample sizes are larger, influencing
the expected effect size.
Laurencelle L, Cousineau D (2023). “Analysis of frequency tables: The ANOFA framework.” The Quantitative Methods for Psychology, 19, 173--193. doi:10.20982/tqmp.19.2.p173 .
# 1- Example of the article:
# with expected frequences .34 to .16, assuming as a first guess groups of 25 observations:
f2 <- anopaProp2fsq( c( 0.32, 0.64, 0.40, 0.16), c(25,25,25,25) );
f2
#> [1] 0.1280757
# f-square is 0.128.
# f-square can be converted to eta-square with
eta2 <- f2 / (1 + f2)
# With a total sample of 97 observations over four groups,
# statistical power is quite satisfactory (85%).
anopaN2Power(97, 4, f2)
#> [1] 0.8538199
# 2- Power planning.
# Suppose we plan a four-classification design with expected proportions of:
pred <- c(.35, .25, .25, .15)
# P is the number of classes (here 4)
P <- length(pred)
# We compute the predicted f2 as per Eq. 5
f2 <- 2 * sum(pred * log(P * pred) )
# the result, 0.0822, is a moderate effect size.
# Finally, aiming for a power of 80%, we run
anopaPower2N(0.80, P, f2)
#> [1] 132.501
# to find that a little more than 132 participants are enough.