R For循环执行Fisher测试 - 错误消息

Question

My data frame looks like that:

595.00000    18696      984.00200     32185    Group1  
935.00000    18356      1589.00000    31580    Group2            
40.00010     19251      73.00000      33096    Group3            
1058.00000   18233      1930.00000    31239    Group4                
19.00000     19272      27.00000      33142    Group5            
1225.00000   18066      2149.00000    31020    Group6  
....                 

For every group I want to do Fisher exact test.

table <- matrix(c(595.00000, 984.00200, 18696, 32185), ncol=2, byrow=T)  
Group1 <- Fisher.test(table, alternative="greater")

Tried to loop over the data frame with:

for (i in 1:nrow(data.frame))  
 {  
 table= matrix(c(data.frame$V1, data.frame$V2, data.frame$V3, data.frame$V4), ncol=2, byrow=T)    
fisher.test(table, alternative="greater")  
}

But got error message

Error in fisher.test(table, alternative = "greater") :  
FEXACT error 40.  
Out of workspace.  
In addition: Warning message:  
In fisher.test(table, alternative = "greater")  :  
'x' has been rounded to integer: Mean relative difference: 2.123828e-06

How can I fix this problem or maybe do another way of looping over the data?

Answer 1

Your first error is: Out of workspace

?fisher.test
fisher.test(x, y = NULL, workspace = 200000, hybrid = FALSE,
        control = list(), or = 1, alternative = "two.sided",
        conf.int = TRUE, conf.level = 0.95,
        simulate.p.value = FALSE, B = 2000)

You should try increasing the workspace (default = 2e5).

However, this happens in your case because you have really huge values. As a rule of thumb, if all elements of your matrix are > 5 (or in your case 10, because df = 1), then you can safely approximate it with a chi-square test of independence using chisq.test . For your case, I think you should rather use a chisq.test .

And the warning message happens because your values are not integers (595.000) etc. So, if you really want to use a fisher.test recursively, do this (assuming your data is in df and is a data.frame :

# fisher.test with bigger workspace
apply(as.matrix(df[,1:4]), 1, function(x) 
         fisher.test(matrix(round(x), ncol=2), workspace=1e9)$p.value)

Or if you would rather substitute with a chisq.test (which I think you should for these huge values for performance gain with out no significant differences in p-values):

apply(as.matrix(df[,1:4]), 1, function(x) 
         chisq.test(matrix(round(x), ncol=2))$p.value)

This will extract the p-values.

Edit 1: I just noticed that you use one-sided Fisher's exact test . Maybe you should continue using Fisher's test with bigger workspace as I'm not sure of having a one-sided chi-square test of independence as it is already calculated from the right-tail probability (and you can not divide the p-values by 2 as its unsymmetrical).

Edit 2: Since you require the group name with the p-values and you already have a data.frame, I suggest you use data.table package as follows:

# example data
set.seed(45)
df <- as.data.frame(matrix(sample(10:200, 20), ncol=4))
df$grp <- paste0("group", 1:nrow(df))
# load package
require(data.table)
dt <- data.table(df, key="grp")
dt[, p.val := fisher.test(matrix(c(V1, V2, V3, V4), ncol=2), 
                workspace=1e9)$p.value, by=grp]
> dt
#     V1  V2  V3  V4    grp        p.val
# 1: 130  65  76  82 group1 5.086256e-04
# 2:  70  52 168 178 group2 1.139934e-01
# 3:  55 112 195  34 group3 7.161604e-27
# 4:  81  43  91  80 group4 4.229546e-02
# 5:  75  10  86  50 group5 4.212769e-05