Calculate confidence interval for log determinant of covariance matrices
This function uses asymptotic results described by Cai et. al (2016), Theorem 1, to calculate approximate, normal theory confidence intervals (CIs) for the log determinant of one or more sample covariance matrices.
Their results are translated into a CI via the approximation
\log det( \widehat{Σ} ) - bias \pm z_{1 - α/2} \times SE
where \widehat{Σ} is the sample estimate of a population covariance matrix, bias is a bias correction constant and SE is a width factor for the confidence interval. Both bias and SE are functions ofS the sample size, n and number of variables, p.
logdetCI(cov, n, conf = 0.95, method = 1, bias.adj = TRUE)
cov |
a covariance matrix or a (named) list of covariance matrices, all the same size |
n |
sample size, or vector of sample sizes, one for each covariance matrix |
conf |
confidence level |
method |
Three methods are provided, based on Cai et. al Theorem 1 ( |
bias.adj |
logical; set |
Cai et. al (2015) claim that their Theorem 1 holds with either p fixed or p(n) growing with n, as long as
p(n) ≤ n. Their Corollary 1 (method=2) is the special case when p is fixed.
Their Corollary 2 (method=3) is the special case when 0 ≤ p/n < 1 is fixed.
The properties of this CI estimator are unknown in small to moderate sample sizes, but it seems to be the only one available. It is therefore experimental in this version of the package and is subject to change in the future.
The bias term offsets the confidence interval from the sample estimate of \log det( \widehat{Σ} ) . When p is large relative to n, the confidence interval may not overlap the sample estimate.
Strictly speaking, this estimator applies to the MLE of the covariance matrix \widehat{Σ}, i.e., using n rather than n-1 in as the divisor. The factor (n-1 / n) has not yet been taken into account here.
A data frame with one row for each covariance matrix.
lower and upper are the boundaries of the confidence intervals.
Other columns are logdet, bias, se.
Michael Friendly
Cai, T. T.; Liang, T. & Zhou, H. H. (2015) Law of log determinant of sample covariance matrix and optimal estimation of differential entropy for high-dimensional Gaussian distributions. Journal of Multivariate Analysis, 137, 161-172. doi: 10.1016/j.jmva.2015.02.003
data(iris) iris.mod <- lm(as.matrix(iris[,1:4]) ~ iris$Species) iris.boxm <- boxM(iris.mod) cov <- c(iris.boxm$cov, list(pooled=iris.boxm$pooled)) n <- c(rep(50, 3), 150) CI <- logdetCI( cov, n=n, conf=.95, method=1) CI plot(iris.boxm, xlim=c(-14, -8), main="Iris data, Box's M test", gplabel="Species") arrows(CI$lower, 1:4, CI$upper, 1:4, lwd=3, angle=90, len=.1, code=3) CI <- logdetCI( cov, n=n, conf=.95, method=1, bias.adj=FALSE) CI plot(iris.boxm, xlim=c(-14, -8), main="Iris data, Box's M test", gplabel="Species") arrows(CI$lower, 1:4, CI$upper, 1:4, lwd=3, angle=90, len=.1, code=3)
Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.