power analysis- interpreting sample size given established literature

#1
Hello and thank you for clicking,

I am attempting to use relevant effects size from a published study to obtain a suggested sample size for my related study in G*power. However, 1) the previous study only compared 2 groups w/ t-test and I want to compare 5 with one way ANOVA and 2) the suggested sample size (if i was replicating with a t-test) is much larger than the study I gathered the input from.

First question: Should I assume a much lower effect size than obtained with a t-test because I'm adding comparisons?

Second question: Why does G*Power suggest more than double the subjects (than published) when I compute for a simple t-test (the same test they did) using their established effect size? I'm basically just inputing all the numbers from the first study and being told their sample size shouldn't have worked. This finding is making me second guess the suggested Ns when calculating for an ANOVA.

What am I missing? Thank you in advance for any replies.
 

Karabiner

TS Contributor
#2
Why does G*Power suggest more than double the subjects (than published) when I compute for a simple t-test (the same test they did) using their established effect size?
Because G*power does not tell you with which sample size the
established difference will achieve p < 0.05 .
Instead, it tells you with which sample size you have an 80% (or
whatever you asked for) chance to achieve p < 0.05 , if the
effect in the population was of the established size.

By the way, what were the effect size and the sample size
in the study you refer to?

I'm basically just inputing all the numbers from the first study and being told their sample size shouldn't have worked.
Any sample size works, but smaller samples have lower
power i.e. a lower chance to achieve the magical p < 0.05
But lower power is not zero power.

With kind regards

K.
 
#3
I've figured out that the suggested sample size is so much larger because my power was at .95, rather than .8. Their effect size was huge, d= 1.19 with N= 20 (10 per group; animal research). G*power says minimum N=40. However, it still suggests a higher N= 26, when power corrected to .8.

I think I'm getting a clearer understanding.
 

Karabiner

TS Contributor
#4
Their effect size was huge, d= 1.19
Not exactely. This was just a mean difference translated into an
effect size measure (d). A real effect size is what is found in
populations. In samples (like this with n=20), d is not an effect
size, but a combination of a "true" effect and sampling error.
In published research with small samples, that what is called
"effect size" is almost always an overestimation of the true
(uncominated) effect size.

With kind regards

K.