cerioli2010.irmcd.test {CerioliOutlierDetection}R Documentation

Iterated RMCD test of Cerioli (2010)

Description

Given a set of observations, this function tests whether there are outliers in the data set and identifies outlying points. Outlier testing/identification is done using the Mahalanobis-distances based on the MCD dispersion estimate. The iterated reweighted MCD method of Cerioli (2010) is used to ensure the intersection test has the specified nominal size (Type I error rate).

Usage

cerioli2010.irmcd.test(datamat, 
  mcd.alpha = max.bdp.mcd.alpha(n,v), 
  signif.gamma = 0.05, nsamp = 500, 
  nmini = 300, trace = FALSE, 
  delta = 0.025, hrdf.method=c("GM14","HR05"))

Arguments

datamat

(Data Frame or Matrix) Data set to test for outliers (rows = observations, columns = variables). datamat cannot have missing values; please deal with them prior to calling this function. datamat will be converted to a matrix.

mcd.alpha

(Numeric) Value to control the fraction of observations used to compute the covariance matrices in the MCD calculation. Default value is corresponds to the maximum breakpoint case of the MCD; valid values are between 0.5 and 1. See the covMcd documentation in the robustbase library for further details.

signif.gamma

(Numeric) Desired nominal size of the intersection outlier test (e.g., 0.05), i.e., a test that there are no outliers in the data. (This is the γ parameter in Cerioli (2010).) The corresponding α parameter for testing individual observations for outlyingness will be calculated from γ as

alpha = 1 - (1 - gamma)^(1/n).

nsamp

(Integer) Number of subsamples to use in computing the MCD. See the covMcd documentation in the robustbase library.

nmini

(Integer) See the covMcd documentation in the robustbase library.

trace

(Logical) See the covMcd documentation in the robustbase library.

delta

(Numeric) False-positive rate to use in the reweighting step (Step 2). Defaults to 0.025 as used in Cerioli (2010). When the ratio n/ν of sample size to dimension is very small, using a smaller delta can improve the accuracy of the method.

hrdf.method

(String) Method to use for computing degrees of freedom and cutoff values for the non-MCD subset. The original method of Hardin and Rocke (2005) and the expanded method of Green and Martin (2014) are available as the options “HR05” and “GM14”, respectively. “GM14” is the default, as it is more accurate across a wider range of mcd.alpha values.

Details

Calls the finite-sample reweighted MCD (FSRMCD) outlier detection function cerioli2010.fsrmcd.test first to test for the existence of any outliers in the data. If the FSRMCD method rejects the null hypothesis of no outliers in the data, individual observations are then tested for outlyingness using the critical value function returned by cerioli2010.fsrmcd.test with significance γ.

Value

outliers

A matrix of dimension nrow(datamat) by length(signif.gamma) indicating whether each row of datamat is an outlier. The i-th column corresponds to the result of testing observations for outlyingness at significance level signif.gamma[i].

mahdist.rw

a matrix of dimension nrow(datamat) by length(signif.gamma) of Mahalanobis distances computed using the finite-sample reweighted MCD methodology in Cerioli (2010). Even though the distances do not depend on signif.gamma, there is one column per entry in signif.gamma for user convenience.

Author(s)

Written and maintained by Christopher G. Green <christopher.g.green@gmail.com>

References

Andrea Cerioli. Multivariate outlier detection with high-breakdown estimators. Journal of the American Statistical Association, 105(489):147-156, 2010. doi: 10.1198/jasa.2009.tm09147

Andrea Cerioli, Marco Riani, and Anthony C. Atkinson. Controlling the size of multivariate outlier tests with the MCD estimator of scatter. Statistical Computing, 19:341-353, 2009. doi: 10.1007/s11222-008-9096-5

See Also

cerioli2010.fsrmcd.test

Examples

require(mvtnorm, quiet=TRUE)

############################################
# dimension v, number of observations n
v <- 5
n <- 200
simdata <- array( rmvnorm(n*v, mean=rep(0,v), 
    sigma = diag(rep(1,v))), c(n,v) )
# detect outliers
results <- cerioli2010.irmcd.test( simdata, 
    signif.gamma=c(0.05,0.01,0.001) )
# count number of outliers detected for each 
# significance level
colSums( results$outliers )


#############################################
# add some contamination to illustrate how to 
# detect outliers using the irmcd test
# 10/200 = 5% contamination
simdata[ sample(n,10), ] <- array( 
  rmvnorm( 10*v, mean=rep(2,v), sigma = diag(rep(1,v))),
  c(10,v)
)
results <- cerioli2010.irmcd.test( simdata, 
  signif.gamma=0.01 )
mean( results$outliers[,1,drop=TRUE] )

#############################################
# banknote example from Cerioli (2010)
## Not run: 

  require(rrcov) # for CovMcd
  require(alr3)  # banknote data set lives here
  data(banknote, package="alr3")
  # length, width of left edge, width of right edge,
  # width of bottom edge, width of top edge, length
  # of image diagonal, counterfeit (1=counterfeit)

  bnk.gamma <- 0.01
  # genuine banknotes
  # classical mean and covariance
  banknote.real <- banknote[ banknote[,"Y"]==0, 1:6 ]
  cov.cls <- CovClassic( banknote.real  )
  # 1 - (1 - 0.01)^(1/100) quantile of scaled-Beta distribution
  # with m=100 and v=6
  bnk.m <- nrow( banknote.real )
  bnk.v <- ncol( banknote.real ) 
  bnk.alpha <- 1. - ((1. - bnk.gamma)^(1./bnk.m))
  cutoff.cls <- (bnk.m-1.)*(bnk.m-1.)*qbeta( 1. - bnk.alpha, bnk.v/2., 
  (bnk.m - bnk.v - 1.)/2.)/bnk.m
  # Figure 4 (left) in Cerioli (2010)
  plot( getDistance( cov.cls ), xlab="Index number", 
  ylab="Squared Mahalanobis Distance", type="p", 
  ylim=c(0,45)
  )
  abline( h=cutoff.cls )

  # reweighted MCD, maximum breakdown point case
  cov.rob <- CovMcd( banknote.real, 
  alpha=floor((bnk.m + bnk.v + 1.)/2.)/bnk.m, nsamp="best" )
  # cutoff using chi-squared individually
  cutoff.rmcdind <- qchisq(1. - bnk.gamma, df=bnk.v)
  # cufoff using simultaneous chi-square
  cutoff.rmcdsim <- qchisq(1. - bnk.alpha, df=bnk.v)
  # scaled-F cutoff using FSRMCD
  # cutoff value is returned by critvalfcn for observations
  # with weight=0
  tmp.fsrmcd <- cerioli2010.fsrmcd.test( banknote.real, 
  signif.alpha=bnk.alpha )
  cutoff.fsrmcd <- unique(tmp.fsrmcd$critvalfcn( bnk.alpha )[tmp.fsrmcd$weights==0])
  # Figure 4 (right)
  plot( getDistance( cov.rob ), xlab="Index number",
  ylab="Squared Robust Reweighted Distance", type="p",
  ylim=c(0,45)
  )
  abline( h=cutoff.rmcdind, lty="dotted" )
  abline( h=cutoff.rmcdsim, lty="dashed" )
  abline( h=cutoff.fsrmcd, lty="solid" )
  legend( "topright", c("RMCD_ind","RMCD","FSRMCD"), 
    lty=c("dotted","dashed","solid") )

  # forged banknotes
  # classical mean and covariance
  banknote.fake <- banknote[ banknote[,"Y"]==1, 1:6 ]
  cov.cls <- CovClassic( banknote.fake  )
  # 1 - (1 - 0.01)^(1/100) quantile of scaled-Beta distribution
  # with m=100 and v=6
  bnk.m <- nrow( banknote.fake )
  bnk.v <- ncol( banknote.fake ) 
  bnk.alpha <- 1. - ((1. - bnk.gamma)^(1./bnk.m))
  cutoff.cls <- (bnk.m-1.)*(bnk.m-1.)*qbeta( 1. - bnk.alpha, bnk.v/2., 
    (bnk.m - bnk.v - 1.)/2.)/bnk.m
  # Figure 5 (left) in Cerioli (2010)
  plot( getDistance( cov.cls ), xlab="Index number", 
    ylab="Squared Mahalanobis Distance", type="p", 
    ylim=c(0,45)
  )
  abline( h=cutoff.cls )
  
  
  # reweighted MCD, maximum breakdown point case
  cov.rob <- CovMcd( banknote.fake, 
    alpha=floor((bnk.m + bnk.v + 1.)/2.)/bnk.m, nsamp="best" )
  # cutoff using chi-squared individually
  cutoff.rmcdind <- qchisq(1. - bnk.gamma, df=bnk.v)
  # scaled-F cutoff using FSRMCD
  # cutoff value is returned by critvalfcn for observations
  # with weight=0
  tmp.fsrmcd <- cerioli2010.fsrmcd.test( banknote.fake, 
    signif.alpha=bnk.alpha )
  cutoff.fsrmcd <- unique(tmp.fsrmcd$critvalfcn( bnk.alpha )[tmp.fsrmcd$weights==0])
  cutoff.irmcd <- unique(tmp.fsrmcd$critvalfcn( bnk.gamma )[tmp.fsrmcd$weights==0])
  # Figure 5 (right) in Cerioli (2010)
  plot( getDistance( cov.rob ), xlab="Index number",
    ylab="Squared robust reweighted Distance", type="p",
    ylim=c(0,150)
  )
  abline( h=cutoff.rmcdind, lty="dotted" )
  abline( h=cutoff.fsrmcd, lty="dashed" )
  abline( h=cutoff.irmcd, lty="solid" )
  legend( "topright", c("RMCD_ind","FSRMCD","IRMCD"), 
    lty=c("dotted","dashed","solid") )


## End(Not run)

#############################################
# example of how to ensure the size of the intersection test is correct
## Not run: 
  n.sim <- 5000
  simdata <- array( 
    rmvnorm(n*v*n.sim, mean=rep(0,v), sigma=diag(rep(1,v))),
    c(n,v,n.sim)
  )
  # in practice we'd do this using one of the parallel processing
  # methods out there
  results <- apply( simdata, 3, function(dm) {
    z <- cerioli2010.irmcd.test( dm, 
      signif.gamma=0.01 )
    # true if outliers were detected in the data, false otherwise
    any(z$outliers[,1,drop=TRUE])
  })
  # count the percentage of samples where outliers were detected;
  # should be close to the significance level value used (0.01) in these
  # samples for the intersection test
  mean(results)

## End(Not run)

[Package CerioliOutlierDetection version 1.1.9 Index]