Closed pneuvial closed 3 years ago
The current implementation of rowWelchTests and rowWilcoxonTests is vectorized in two directions, ie both functions can take as input (1) a m x n data matrix and (2) a n x B matrix of group labels in {0,1}. The output p-values and test statistics are then m x B matrices, and all the computations are vectorized. When B=1 the time requirements are competitive with the functions in matrixTests
.
This is very useful in the sansSouci
package because we are performing calibration by permutation, ie (2) is a matrix of permuted group labels.
Using the current corresponding functions in matrixTests
, this calibration would require looping over B permutations, which would be slower than our current version.
Currently we are using an internal implementation of two-sample tests by row. It would make much sense to rely on the matrixTests package as suggested by @KKPMW a while ago (cf #11)