outl.det: Detection and Display Outliers

In wilsontom/FIEmspro: Flow In-jection Electrospray Mass Spectrometry Processing: \\ data processing, classification modelling and variable selection in metabolite fingerprinting

Description

Wrapper for the detection of sample outliers by computation of Mahalanobis distances using cov.rob from package MASS. The (robust) square root distance from the center is displayed alongside a 2D mapping of the data and its confidence ellipse.

Usage

outl.det(x, method = "classical", conf.level = 0.975, 
            dimen=c(1,2), tol = 1e-7, plotting = TRUE)

Arguments

x	A data.frame or matrix.
method	The method to be used: mve. Minimum volume ellipsoid. mcd. Minimum covariance determinant. classical. Classical product-moment. For details, see cov.rob in package MASS.
conf.level	The confidence level for controlling the cutoff of Mahalanobis distances.
dimen	Dimensions used to plot tolerance ellipse and the data points alongside these two dimensions.
tol	The tolerance to be used for computing Mahalanobis distances (see cov.rob in package MASS)
plotting	A logical value. If TRUE, The Mahalanobis distances against the index of data samples and the tolerance ellipse of the data samples are plotted.

Details

If the number of samples is n and number of variables in a sample is p, the data set must be n > p + 1. In this case, PCA can be used to produce fewer directions of uncorrelated dimensions that explain different dimensions in the data. Due to the inherent difficulties in defining outliers, inclusion of the first few dimensions only is almost always sufficient to compute Mahalanobis distances. However in more complex designs implicating various factors and/or multiple levels, different contributions to the overall variation modelled by PCA may be confounded in such a reduced space. In such situation, the initial dataset must be decomposed into smaller problems to relate potential outlying behaviour.

Value

A list with components:

outlier	List of outliers detected.
conf.level	Confidence level used.
mah.dist	Mahalanobis distances of each data sample.
cutoff	Cutoff of Mahalanobis distances for outliers detection.

Author(s)

Wanchang Lin [e-mail] and David Enot [e-mail]

Examples

## load abr1
data(abr1)
y   <- factor(abr1$fact$class)
x <- preproc(abr1$pos , y=y, method=c("log10","TICnorm"),add=1)[,110:1000]  
## Select classes 1 and 2
tmp <- dat.sel(x, y, choices=c("1","2"))
dat <- tmp$dat[[1]]
ind <- tmp$cl[[1]]


## dimension reduction by PCA
x   <- prcomp(dat,scale=FALSE)$x

## perform and plot outlier detection using classical Mahalanobis distance
## on the first 2 PCA dimensions
res <- outl.det(x[,c(1,2)], method="classical",dimen=c(1,2),
                    conf.level = 0.975)

wilsontom/FIEmspro documentation built on May 4, 2019, 6:28 a.m.