1
$\begingroup$

Suppose a gene is mutated in 30% of the samples.

I want plot the number of samples required to see 30% of them mutated at various level of power.

I want to do this through a simulation so I generated different sample sizes and for each sample size I generated a 1000 trials of simulated number of mutated samples based on p.

p = 0.3
# from sample size 1:100, I generated 1000 binomial random variable
for (n in 1:100) {
  k = rbinom(1000, n, p)
}

How do I construct the power analysis after I have generated some simulated data?

$\endgroup$
  • 2
    $\begingroup$ Did you want to run some test? Eg, to show that the observed proportion is <.5? Or did you want to get a 95% CI of a given width & that would include .3 95% of the time? $\endgroup$ – gung Mar 14 '15 at 23:15
  • $\begingroup$ I want to see how many samples would I need to see 30% samples mutated at 80% power. $\endgroup$ – kylelyk02 Mar 16 '15 at 21:12
  • $\begingroup$ What does that mean? Do you want to run a binomial test against a null of .5? $\endgroup$ – gung Mar 16 '15 at 21:15
  • 1
    $\begingroup$ i think my null p0=.01 which a sample will be mutated in a gene in a background model, and my p1=.3 $\endgroup$ – kylelyk02 Mar 16 '15 at 21:37
  • $\begingroup$ i think my problem is quite similar to this question except my P0 is .01, so the minimum number of success to reject null should be some number greater than 1 i think. $\endgroup$ – kylelyk02 Mar 16 '15 at 21:58
1
$\begingroup$

You may be able to figure out everything you need to know from my answer here: Simulation of logistic regression power analysis - designed experiments, which is quite comprehensive.

The basic procedure for simulating a power analysis is to:

  1. Simulate data according to your preferred scenario (the alternative hypothesis).
  2. Run the test you intend to use.
  3. Do this many times, and see what percentage of the time your results are significant.
  4. If you want to solve for the required $N$ to achieve a given power, do the above with different $N$s and find the one that yields the power you want.

In your case, you need to apply the binomial test with the null proportion set to $.01$.

Here's an example in R (note that your code had errors):

set.seed(8063)
p = 0.3
# from sample size 1:100, I generated 1000 binomial random variable
k = matrix(NA, nrow=1000, ncol=100)
for(n in 1:100){
  k[,n] = rbinom(1000, n, p)
}

p.mat  = matrix(NA, nrow=1000, ncol=100)
for(i in 1:1000){
  for(n in 1:100){
    p.mat[i,n] = binom.test(x=k[i,n], n=n, p=0.01)$p.value
  }
}
power = c()
power = apply(p.mat, 2, FUN=function(x){ mean(x<.05) })
min(which(power>.8))
# [1] 5
power[4]
# [1] 0.763
power[5]
# [1] 0.82

Brute force search:

tl = NULL
tl[paste("n", 1:10, sep="")] = list(NULL)
pl = tl
for(n in 1:10){
  tl[[n]] = 0:n
  for(h in 0:n){
    pl[[n]][h+1] = binom.test(tl[[n]][h+1], n, 0.01, "g")$p.value
  }
}
msig = lapply(pl, function(x){ min(which(x<.05))-1 })
pow = c()
for(n in 1:10){
  pow[n] = 1-sum(dbinom(0:(msig[[n]]-1), n, .3))
}
names(pow) = paste("n", 1:10, sep=" = ")
min(which(pow>.8))
# [1] 5
pow[4:5]
#   n = 4   n = 5 
# 0.75990 0.83193 
$\endgroup$
  • $\begingroup$ Wouldnt it be binom.test(x=k[i,n], n=n, p=0.01, alternative="greater") in my case? $\endgroup$ – kylelyk02 Mar 17 '15 at 20:06
  • $\begingroup$ @kylelyk02, you could do that if you wanted. $\endgroup$ – gung Mar 17 '15 at 20:22
  • $\begingroup$ i noticed that the power flucuate when sample is low, would it be safer to say power > .8 at power[9] instead of power[5]? $\endgroup$ – kylelyk02 Mar 17 '15 at 20:51
  • $\begingroup$ @kylelyk02, I don't follow that comment, but it occurs to me that your situation is so simple it can easily be found analytically by brute force search. I updated by answer. Note that the analytical results agree very closely w/ the original simulation. $\endgroup$ – gung Mar 18 '15 at 0:27
  • $\begingroup$ here's what i meant by power fluctuate. As you can see there are occasional dips in the power as sample size increases. $\endgroup$ – kylelyk02 Mar 18 '15 at 23:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.