Hi all,
I am new to statistic and have a question about calculating the power of test statistic. According to what I have founf so far I can use the webpower package in R and since I want to calculate the power of my logistic regression model, I need to use wp.regression function. Unfortunately I have no idea how I should set some of its variable.
wp.logistic(n = NULL, p0 = NULL, p1 = NULL, alpha = 0.05, power = NULL, alternative = c("two.sided", "less", "greater"), family = c("Bernoulli", "exponential", "lognormal", "normal", "Poisson", "uniform"), parameter = NULL)
The variable that I need help with them are p0 and p1. Their definitions
are as below:
p0= Prob(Y=1|X=0): the probobility of observieng 1 for the outcome variable Y when the predictor X equals 0.
p1= Prob(Y=1|X=1): the probobility of observieng 1 for the outcome variable Y when the predictor X equals 1.
The only thing that I have is p-value of my statistic test. Can I have
your thoughts to see how to calculate p0 and p1?
Hello Nastaran Maus Esfahani
For read:
https://webpower.psychstat.org/wiki/_media/grant/webpower_manual_book.pdf
online version:
https://webpower.psychstat.org/models/reg02/
P0 and p1 are derived from literature or assumptions.
There is no point in a power calculation if you already have the data (and you already tested the hypothesis and you already have a p-value).
Power calculations make sense only to plan experiments (i.e. to get the sample size required to achieve a desired power for an expected/assumed effect).
If you want that for future experiments, p0 and p1 should be chosen to represent what you expect these probabilities to be. They are not calculated. They represent your (resonable) expectation, an "educated guess". This may be based on past experience, similar data, own work or work published by others, or a conclusion by analogy.
Jochen Wilhelm: thanks for this (and for all your answers here that are always insightful). However, although this does not make sense for me, some people perform "post-hoc power analysis" or "retrospective power analysis". It might be the aim of Nastaran Maus Esfahani, but it would not be a very good idea.
Just in case, here is a good reference for Nastaran, explainaing why this approach is flawed:
Article The Abuse of Power: The Pervasive Fallacy of Power Calculati...
Jochen Wilhelm Frédéric Santos Thank you both for your replay.
To clarify, I should say, I have proposed a kernel based association test to see if there is any relationship between my data ( they are genetic variants in each chromosome and disease status. I did the test for data of each chromosome and according to the calculated P-Values, I have found there is a relationship between some chromosomes and disease.
In some paper, I have seen they have compared their method's power to other existing methods and I want to do the same to show my method outperform other methods. I have attached the papers I referred to.
The best way to cope with statistical estimation problems is to follow a key book:
1) http://users.isr.ist.utl.pt/~wurmd/Livros/school/Bishop%20-%20Pattern%20Recognition%20And%20Machine%20Learning%20-%20Springer%20%202006.pdf
2) https://www.cyclismo.org/tutorial/R/probability.html
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis