Fisher test error : LDSTP is too small

Question

input

NN <- c(359,32);JJ <- c(108,13);NNS <- c(103,15);VBN <- c(95,9);RB <- c(63,11);NNP <- c(56,0);VBG <- c(55,10);IN <- c(38,16);VB <- c(20,10);CD <- c(17,6);CC <- c(11,6);DT <- c(11,4);MD <- c(8,5);PRP4 <- c(8,1);PRP <- c(7,4);FW <- c(5,1);VBD <- c(5,3);RBR <- c(4,0);VBP <- c(4,1);VBZ <- c(4,3);WRB <- c(4,2);EX <- c(3,1);NNPS <- c(2,0);WDT <- c(2,3);WP <- c(2,1);PDT <- c(1,1);POS <- c(1,0);RBS <- c(1,0);TO <- c(1,1);UH <- c(0,1)
Finaltable <-
cbind(NN,JJ,NNS,VBN,RB,NNP,VBG,IN,VB,CD,CC,DT,MD,PRP4,PRP,FW,VBD,RBR,VBP,VBZ,WRB,EX,NNPS,WDT,WP,PDT,POS,RBS,TO,UH)
rownames(Finaltable) <- c("tag1","tag2")
Finaltable

chisq.test(Finaltable)


fisher.test(Finaltable)

output

fisher.test(Finaltable) : FEXACT error 7.
LDSTP is too small for this problem.
Try increasing the size of the workspace.

How can I solve this problem without modifying the raw data? Is there any non-parametric test for this comparison?

Answer 1

You can try increasing the workspace argument from its default value, but I don't know if you're going to be able to make it big enough (I gave up at workspace=2e8, which still fails; I ran out of memory at workspace=2e9.) You can also try simulated p-values, e.g. fisher.test(Finaltable,simulate.p.value=TRUE,B=1e7) (for example), but since the p-value is extremely small, you're going to need a huge number of simulations (B) if you want to do more than bound the p-value, which will also be very slow. (For most purposes, knowing that p is <1e-7 is more than enough -- but in some bioinformatics contexts people want to use p as an index of signal strength and/or impose massive multiple-corrections comparisons. I don't really like these approaches, but they're out there ...)