When R's Mann-Whitney U test returns extremely distorted p-values

The Mann–Whitney U test (also known as the Wilcoxon rank-sum test) is one of the most popular nonparametric statistical tests. In R, it can be accessed using the wilcox.test function, which has been available since R 1.0.0 (February 2000). With its extensive adoption and long-standing presence in R, the wilcox.test function has become a trusted tool for many researchers. But is it truly reliable, and to what extent can we rely on its accuracy by default?

In my work, I often encounter the task of comparing a large sample (e.g., of size 50+) with a small sample (e.g., of size 5). In some cases, the ranges of these samples do not overlap with each other, which is the extreme case of the Mann–Whitney U test: it gives the minimum possible p-value. In one of the previous posts, I presented the exact equation for such a p-value. If we compare two samples of sizes $n$ and $m$, the minimum p-value we can observe with the one-tailed Mann–Whitney U test is $1/C_{n+m}^n$. For example, if $n=50$ and $m=5$, we get $1/C_{55}^5 \approx 0.0000002874587$. Let’s check these calculations using R:

> wilcox.test(101:105, 1:50, alternative = "greater")$p.value
[1] 0.0001337028

The obtained p-value is $\approx 0.0001337028$, which is $\approx 465$ times larger than we expected! Have we discovered a critical bug in wilcox.test? Can we now trust this function? Let’s find out!

In fact, the wilcox.test has two implementations of the Mann–Whitney U test: the exact one and the approximated one. The choice of a particular strategy is controlled by the exact parameter of this function. Here is an example from the official documentation:

> x <- c(0.80, 0.83, 1.89, 1.04, 1.45, 1.38, 1.91, 1.64, 0.73, 1.46)
> y <- c(1.15, 0.88, 0.90, 0.74, 1.21)
> wilcox.test(x, y, alternative = "greater", exact = TRUE)$p.value
[1] 0.1272061
> wilcox.test(x, y, alternative = "greater", exact = FALSE)$p.value
[1] 0.1223118

In this example, the approximation looks accurate, with a difference ofhis makes it an impra only about $5\%$. When the exact parameter is not specified, it is determined as follows:

if(is.null(exact))
    exact <- (n.x < 50) && (n.y < 50)

Before we start exploring different extreme cases, let us briefly review both methods to estimate the p-value within this test.

The first step of the Mann–Whitney U test is calculating the U statistic:

It is a simple rank-based statistic, which varies from $0$ to $nm$. Once we get the value of $U$, we should match it with the distribution of $U$ statistics under the valid null hypothesis. The distribution can be obtained in two different ways:

• The exact implementation
  The straightforward implementation uses dynamic programming to obtain the exact probability of obtaining the given $U$ statistic for the given $n$ and $m$. The only drawback of this solution is its computational complexity (which is $\mathcal{O}(n^2 m^2)$ for both time and memory). This makes this approach an impractical solution for large samples.
• The approximated implementation
  Fortunately, we have a method of large sample approximation (see [Hollander1973], pages 68–75). Long story short, we approximate the distribution of $U$ statistics with the following normal distribution (in the case of tied values, a small correction is needed):

For the middle values of $U$ (around $nm/2$), this approximation provides reasonably accurate results. However, it performs extremely poorly at the tails. To illustrate the problems, we perform a simple numerical study:

• Enumerate $n$ from $1$ to $50$, set $m$ to $50$ (since $50$ is a critical value, after which wilcox.text switches to the approximated algorithm);
• For each pair of $n,\, m$, we consider two non-overlapped samples $x = \{ 101, 102, \ldots, 100 + n \}$ and $y = \{ 1, 2, \ldots, m \}$;
• For each pair of samples $x$ and $y$, we estimate the exact and approximated p-values using wilcox.test (R 4.3.0). We also evaluate the ratio between two obtained p-values.

This simulation can be performed using the following simple script:

options(scipen = 999)
df <- expand.grid(n = 1:50, m = 50)
df$approx <- sapply(1:nrow(df), function(i) wilcox.test(101:(100+df$n[i]), 1:df$m[i], "g", exact = F)$p.value)
df$exact <-  sapply(1:nrow(df), function(i) wilcox.test(101:(100+df$n[i]), 1:df$m[i], "g", exact = T)$p.value)
df$ratio <- df$approx / df$exact
df

Here are the results:

Thus, for $n=m=50$, the ratio between the exact (the correct one) and approximated (the default one) solutions is about $3.5 \cdot 10^{11}$. Of course, both p-values are pretty small in this case. If you are using the notorious $\alpha = 0.05$, you will most likely avoid such a problem (however, for $n=1,\, m = 50$, the exact true p-value is $0.0196$, while the approximated one is $0.0480$, whish is alarmingly close to $0.05$).

If you prefer using extremely low statistical significance levels and there are non-zero chances of getting extreme values of the $U$ statistic, you may consider making exact = TRUE your default way to call wilcox.test. Such a simple trick may save you from misleading research results.

References

• [Hollander1973]
  Myles Hollander and Douglas A. Wolfe (1973) Nonparametric Statistical Methods. New York: John Wiley & Sons. DOI:10.1002/9781119196037