Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Summary: My simulations don't match my power calculation.

(This question has become rather long, so you might not read to the bottom. I think it's a sample size problem.)

I'm running some simulations to help design a study, and I'm running into something that's confusing me.

There are two groups, treated and untreated, and the outcome is crashing.

We have an estimated crash rate of 0.1 in the untreated group, and think we can bring this down to 0.08. A small (and non-feasible) sample size is 100 per group.

So I run power.prop.test() to estimate the power.

power.prop.test(p1=0.1, p2=0.08, n=100)

> power.prop.test(p1=0.1, p2=0.08, n=100)                                                        

     Two-sample comparison of proportions power calculation 

              n = 100
             p1 = 0.1
             p2 = 0.08
      sig.level = 0.05
          power = 0.07122973
    alternative = two.sided

 NOTE: n is number in *each* group 

Then I ran a simulation, because I wanted to explore how often we'd make the wrong decision as to the best treatment.


#Create a data frame called d, populate it with the numbers above.
nTreated <- 100
nUntreated <- 100
probTreated <- 0.1
probUntreated <- 0.08

d <- data.frame(id = 1:10000)
d$nTreated <- nTreated
d$nUntreated <- nUntreated
d$probTreated <- probTreated 
d$probUntreated <- probUntreated

#Generate some random results using rbinom()
d$treatedCrashes <- apply(cbind(d$nTreated, d$probTreated), 1, 
                      function(x)  sum(rbinom(x[1], 1, x[2])))

d$untreatedCrashes <- apply(cbind(d$nUntreated, d$probUntreated), 1, 
                      function(x)  sum(rbinom(x[1], 1, 

#Do fisher's exact test on each replication:
d$fisher <- apply(cbind(d$nTreated - d$treatedCrashes, 
                        d$nUntreated - d$untreatedCrashes, 
                        d$untreatedCrashes), 1, 
                          function(x)  fisher.test(matrix(x, 
#test power
mean(d$fisher < 0.05)

And I get 4.8% power, which is lower than the power.prop.test function said, AND is less than 0.05 - which seems kind of wrong.

>     mean(d$fisher < 0.05)
[1] 0.0478

Is this about the small sample approximation? Is it a stupid coding error? (I don't think it is, but I've often been wrong about that before). Is it something I haven't thought of?

In response to the suggestion that it's because the Fisher's exact test is conditioned on the margins, I reran the model with Barnard's test (in the Exact library). (But I reduced to 1000 replications, as this took 40 minutes).

d$exact <- apply(cbind(d$nTreated - d$treatedCrashes, 
                         d$nUntreated - d$untreatedCrashes, 
                         d$untreatedCrashes), 1, 
                       function(x)  exact.test(matrix(x, nrow=2), to.plot=F, cond.row=T)$p.value)
d$exact <- lapply(d$exact, function(x) x[1][[1]])

mean(d$exact < 0.05)      

> mean(d$exact < 0.05)                      
[1] 0.049                

I have pretty much the same power.

However, I also ran an exact test for power using the power.exact.test() function, also in the Exact library, which gives a very similar level of power:

> power.exact.test(p1=0.1, p2=0.08, n1=100, n2=100,simulation=T, nsim=1000, method="Boschloo")
[1] 0.045

[1] "two.sided"

[1] "Boschloo"

Which makes me think that it's a sample size problem.

share|improve this question
up vote 4 down vote accepted

Fisher's Exact Test is conservative (i.e. the false positive rate for the nominal 0.05 test is actually less than 0.05 when the null hypothesis is true). What you're finding here is no coincidence. To say that a test is "exact" does not mean that it is of the right size, but that the interpretation of the p-value in small samples is correct.

Here's a reference from university lecture notes on biostatistics

share|improve this answer

Fisher's test is conditional on both margins of the table; your simulation is only conditional on one margin.

[You're right, there's more to it:

(1) Your first calculation with power.prop.test uses the asymptotic approximation of the Binomial distribution to the Normal, so won't give exactly the same answer as an exact test.

(2) Any exact test will be conservative because there's only a finite number of possible contingency tables & you can't find a subset to form the rejection region with exactly 5% probability under the null hypothesis, so you have to settle for a bit less (@AdamO's answer).

(3) There are different exact tests formed by conditioning on the grand total, the row total, the column total, or both row & column totals. Choosing a different conditioning scheme means (a) asking a different question of the test, & (b) changing the number of possible contingency tables (so tests that condition most tend to be most conservative - e.g. Fisher's).

(4) There are also different test statistics that can be used, which do not necessarily give the same ordering of possible tables under the null hypothesis.

(5) And when you're simulating, there's simulation error to take into account (@Maarten's answer).]

share|improve this answer
Thanks for that, but I don't think it's the (whole) answer. I've added some details to the question, where I've used a Barnard's test, and it doesn't make much difference. – Jeremy Miles Apr 9 '13 at 23:17

There is randomness in your Monte Carlo simulation (as there should be), so even if the true rejection rate is exactly .05 and there are no errors in your program you'll still be likely to find small deviations from that number. If you were to repeat your Monte Carlo simulation with 10,000 replications many times and the test functioned exactly as it should, then you would still expect to find that (approximately) 95% of these simulations would result in rejection rates between 4.6% and 5.4%. These bounds are based on the 2.5$^{\mathrm{th}}$ and 97.5$^{\mathrm{th}}$ percentiles of a binomial distribution with n=10,000 and p=.05. So the fact that you found a rejection rate of 4.8% in a simulation with 10,000 replications does not provide all that strong evidence against the Fisher's exact test. As AdamO pointed out, you'll in all likelihood find the rejection rate to be less than 5% even if you increase the number of replications, but the current simulation does not have enough replications to reliably find that phenomenon.

share|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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