In linear dimension reduction the data are transformed into components of which only a few are assumed to carry the information of interest, while the remaining ones are considered noise. What makes a component interesting depends on the method. In principal component analysis (PCA) these are the components with large variances, in independent component analysis (ICA) and non-Gaussian component analysis (NGCA) the non-Gaussian components, and in supervised dimension reduction such as sliced inverse regression (SIR) the components which carry information about the response. The package ICtest provides tests and estimates for the number of these interesting, or informative, components, which also explains its name.
PCAasymp(),
PCAlrt() and PCAschott(), the latter also for
p > n), for NGCA and ICA (FOBIasymp()) and for SIR
(SIRasymp()).PCAboot(),
FOBIboot(), ICSboot() and
SIRboot()) and a simulation-based test using non-Gaussian
projection pursuit (NGPPsim()).kSearch(), which turns any of the tests into an
estimate of the dimension using a forward, backward or binary
search.PCAladle(), FOBIladle(),
SIRladle() and the general ladle()) and
predictor augmentation estimators (PCAaug(),
FOBIaug() and SIRaug()), including the
high-dimensional version PCAaugHD().rorth() for
random orthogonal matrices that are uniformly distributed with respect
to the Haar measure.Note that PCAlrt(), PCAaugHD(),
FOBIaug() and SIRaug() are so far available
only in the development version.
The released version can be installed from CRAN:
install.packages("ICtest")The development version can be installed from GitHub:
# install.packages("remotes")
remotes::install_github("klauschN/ICtest")As an example we consider simulated data with three components whose variances are larger than those of the remaining seven, rotated by a random orthogonal matrix.
library(ICtest)
set.seed(1)
n <- 200
p <- 10
Z <- matrix(rnorm(n * p), ncol = p) %*% diag(c(3, 2.5, 2, rep(1, 7)))
X <- Z %*% t(rorth(p))A forward search with the asymptotic test for subsphericity estimates the number of interesting components.
kSearch(X, PCAasymp)
#> Results for estimated k = 3:
#>
#> PCA subsphericity test using cov
#>
#> data: X
#> T = 27.921, df = 27, p-value = 0.415
#> alternative hypothesis: the last 7 eigenvalues are not equal
#>
#> Tested k values and respective p-values:
#> k = 0, p-value = < 2.2e-16
#> k = 1, p-value = < 2.2e-16
#> k = 2, p-value = < 2.2e-16
#> k = 3, p-value = 0.415The ladle estimator gives the same answer, and its criterion can be plotted with ggplot2.
pcaLadle <- PCAladle(X)
pcaLadle$k
#> [1] 3
ggladleplot(pcaLadle, main = "", xlab = "k") +
geom_vline(xintercept = pcaLadle$k, linetype = "dashed") +
scale_x_continuous(breaks = 0:(p - 1)) + theme_bw()
More examples are given in the vignettes, see
vignette(package = "ICtest").
The methods are described in the following papers, among others:
To cite the package itself, see citation("ICtest"). Bug
reports and suggestions are welcome as issues.