Fisher’s Exact Test

Hypergeometric cumulative distribution (left, right and two-tailed) including permutation testing.

Description

Matlab's implementation of Fisher's exact test only supports one-tailed tests. Other implementations on the file exchange are reportedly buggy and are far too slow for problems I'm working on. This implementation does 2.5e6 one or two-tailed tests per second and is ideally suited for genome-wide association studies or other large problems where 100s of samples and 10s to 100s of 1000s of tests are being performed.
It has been verified correct for samp les sizes up to 100 and spot checked beyond that up to 1000s of samples.

• MATLAB release: MATLAB 7.11 (2010b)
• Other requirements: Requires a large amount of memory for genome wide assocation studies (>=4Gb)

[James Herman]

I'm really confused as to how to use this function. The help is not clear at all :-(

I've got a contingency table, but it's entirely unclear how I should translate it into the variables: a,M,K,N.

Please help!

[Tobias]

I second James Herman's comment. It would be great if someone could explain how to input a contingency table into this function.

[Michael Boedigheimer]

fexact can use vector inputs and it will calculate contingency tables for you, or it can take the same input as Matlab's hygecdf. For your convenience, from fexact's help, here is a 2x2 contingency table, with a,b,c and d being integer counts
a c
b d

K = a+c;
N = a+b;
M = a+b+c+d

then use fexact( a,M,K,N)
enjoy

[Giuseppe Cardillo]

I compared your function with the mine using this matrix:
x=[7 12; 8 2].*1000000

Using Myfisher22(x), the function written by me, the result was:
Elapsed time is 5.715080 seconds.

Using your
fexact(x(1),sum(x(:)),x(1)+x(3),x(1)+x(2)) the results were:

??? Maximum variable size allowed by the program is exceeded.

Error in ==> fexact>tri2sqind at 325
    k = (1:max_k)';
Error in ==> fexact>getLookup at 211
[k n ind] = tri2sqind(M);
Error in ==> fexact at 142
    P = getLookup( M,N,tail);

Same results reducing the complexity to
x=[7 12; 8 2].*100000 and
x=[7 12; 8 2].*10000

When I tried on x=[7 12; 8 2].*1000 the pc looped and after 15 minutes I switched it off.

Finally I tried it on x=[7 12; 8 2].*100
Elapsed time is 6.019058 seconds; using Myfisher22 on this "simple" matrix
Elapsed time is 0.008264 seconds.

So before to write "Other implementations on the file exchange are reportedly buggy and are far too slow for problems I'm working on."
I think that you should test deeper your function.

Regards

[Michael Boedigheimer]

Giuseppe
I am sorry you are having difficulty with the function. My function is very well tested and as far as I can tell works as described. The problem I am working on is detailed in the documentation of the function. I have hundreds of samples and 10s of thousands or more replicates. In the example you give there are many more samples than that. In your smallest example you have 2900. In that case solving for 1 replicate takes 3s on my computer, but so does solving 10,000 replicates.
y = unidrnd(2,2900,1)-1;
x = unidrnd(2, 2900, 1)-1;
Elapsed time is 2.984087 seconds.
>> x = unidrnd(2, 2900, 10000)-1;
>> tic; z = fexact(x,y); toc
Elapsed time is 2.947621 seconds.

I mean no disrespect to you or your many contributions. I just didn't see how your program can be used to solve the problem I just described.