[ANN] PowerAnalyses.jl

Package Announcements
1

After more than a year of having to go back and forth between the R pwr package and G*Power, I finally decided that enough was enough and so there is now a Julia package for power analysis called PowerAnalyses.jl.

The aim of the package is quite straightforward and the same as the aforementioned packages: Given three parameters out of { alpha, power, effect size, sample size }, determine the fourth parameter. For example, to find the required sample size a study, use

julia> using PowerAnalyses

julia> es = 0.5
0.5

julia> alpha = 0.05
0.05

julia> power = 0.95
0.95

julia> n = get_n(OneSampleTTest(two_tails); alpha, power, es)
53.941

As a side note, this package was actually super much fun to make. The way that multiple-dispatch, Distributions.jl and Roots.jl worked together was just too good to be true :partying_face:

Specifically, for the interested people, the pwr package is full of code duplication. For example, compare:

where the code switches between qt, qchisq and other quantile functions. In Julia, that is of course just

quantile(d, beta)

for some d::UnivariateDistribution :exploding_head:

2

@rikh, thanks. If you think it’s appropriate, here’s a link to the Power of a test page, for users to learn more about the subject.

3

Nice. Yes. Good to add.

Another explanation: Statistical tests can tell you whether something is significantly different from some other thing. However, if you don’t have enough data, then there is a big chance that the test will always be not significant. For example, when you do a t-test to compare two samples of both 4 elements, the difference has to be huge for the test to say “the samples are significantly different”.

With a power analysis, you can estimate how big your sample should be (a priori power analysis) to make it less likely that you‘re wasting time on a study which would be very unlikely to find something anyway (to run an underpowered study).

(You could also just use a Bayesian analysis because it is more intuitive/coherent, but I guess that’s a discussion for another thread.)

4

I remember using G*Power…

So glad to never have to touch G*Power again! This is a great package - thanks @rikh!

5

Thanks!
Is there any advantage to adding this to the “blessed repo” HypothesisTests.jl?

6

I‘ve considered and tried basing the types on HypothesisTests.jl. The problem is that those types would make the API for PowerAnalyses much more complex without any clear benefit.

7

I haven’t yet implemented all the G*Power analyses, but yes long-term that should be possible :+1: