LearnNonparam

Overview

This R package implements several non-parametric tests in chapters 1-5 of Higgins (2004), including tests for one sample, two samples, k samples, paired comparisons, blocked designs, trends and association. Built with Rcpp for efficiency and R6 for flexible, object-oriented design, it provides a unified framework for performing or creating custom permutation tests.

Installation

Install the stable version from CRAN:

install.packages("LearnNonparam")

Install the development version from Github:

# install.packages("remotes")
remotes::install_github("qddyy/LearnNonparam")

Usage

library(LearnNonparam)

• Construct a test object

  • from some R6 class directly

  t <- Wilcoxon$new(n_permu = 1e6)

  • using the pmt (permutation test) wrapper

  # recommended for a unified API
  t <- pmt("twosample.wilcoxon", n_permu = 1e6)

• Provide it with samples

  set.seed(-1)
  
  t$test(rnorm(10, 1), rnorm(10, 0))

  

• Check the results

  t$statistic

  

  t$p_value

  

  options(digits = 3)
  
  t$print()

  

  ggplot2::theme_set(ggplot2::theme_minimal())
  
  t$plot(style = "ggplot2", binwidth = 1) # or ggplot2::autoplot(t, binwidth = 1)

  

  

• Modify some settings and observe the change

  t$type <- "asymp"
  t$p_value

  

pmts()

Extending

define_pmt allows users to define new permutation tests. Take the two-sample Wilcoxon test as an example:

t_custom <- define_pmt(
    # this is a two-sample permutation test
    method = "twosample",
    statistic = function(x, y) {
        # (optional) pre-calculate certain constants that remain invariant during permutation
        m <- length(x)
        n <- length(y)
        # return a closure to calculate the test statistic
        function(x, y) sum(x) / m - sum(y) / n
    },
    # reject the null hypothesis when the test statistic is too large or too small
    rejection = "<>", n_permu = 1e5
)

For R >= 4.4.0, the quickr package can be used to accelerate statistic. However, this results in repeated crossings of the R-Fortran boundary and makes pre-calculation of constants impossible.

t_quickr <- define_pmt(
    method = "twosample", rejection = "<>", n_permu = 1e5,
    statistic = function(x, y) {
        sum(x) / length(x) - sum(y) / length(y)
    },
    quickr = TRUE
)

In cases where both pre-calculation and computational efficiency are required, the statistic can be written in C++. Leveraging Rcpp sugars and C++14 features, only minor modifications are needed to make it compatible with C++ syntax.

t_cpp <- define_pmt(
    method = "twosample", rejection = "<>", n_permu = 1e5,
    statistic = "[](const auto& x, const auto& y) {
        auto m = x.length();
        auto n = y.length();
        return [=](const auto& x, const auto& y) {
            return sum(x) / m - sum(y) / n;
        };
    }"
)

It’s easy to check that t_custom, t_quickr and t_cpp are equivalent:

x <- rnorm(10, mean = 0)
y <- rnorm(10, mean = 5)

set.seed(0)
t_custom$test(x, y)$print()

set.seed(0)
t_quickr$test(x, y)$print()

set.seed(0)
t_cpp$test(x, y)$print()

Performance

coin is a commonly used R package for performing permutation tests. Below is a benchmark:

library(coin)

data <- c(x, y)
group <- factor(c(rep("x", length(x)), rep("y", length(y))))

options(LearnNonparam.pmt_progress = FALSE)
benchmark <- microbenchmark::microbenchmark(
    pure_R = t_custom$test(x, y),
    quickr = t_quickr$test(x, y),
    Rcpp = t_cpp$test(x, y),
    coin = wilcox_test(data ~ group, distribution = approximate(nresample = 1e5, parallel = "no"))
)

benchmark

It can be seen that C++ brings significantly better performance than pure R, which enables it to even surpass the coin package in its no-parallelization setting. However, all tests in this package are currently written in pure R with no plans for migration to C++ in the future. This is because the primary goal of this package is not to maximize performance but to offer a flexible framework for permutation tests.

References