Tuesday 03.04.2018 18:30 - 19:30 CRoNoS MDA 2018 Parallel Session E – CRONOSMDA2018
Tuesday 03.04.2018 18:30 - 19:30 Parallel Session E – CRONOSMDA2018
CI031 Room Triton 1-2 SPRING COURSE SESSION IV Chair: Anne Ruiz-Gazen
C0188: Multivariate outlier detection With ICS (Part 4)
Presenter: Anne Ruiz-Gazen, University Toulouse 1 Capitole, France
After a practical introduction of the general use of R for multivariate data analysis,the objective of the course is to present the Invariant Coordinate
Selection (ICS) method as a tool for multivariate outlier detection. ICS was proposed in 2009 and shows remarkable properties for revealing data
structures such as outliers or clusters. It is based on the simultaneous spectral decomposition of two scatter matrices and leads to an ane invariant
coordinate system where the Euclidian distance corresponds to a Mahalanobis Distance (MD) in the original system. However, unlike MD, ICS
makes it possible to select relevant components. This proves useful for detecting outliers lying in a small dimensional subspace for data sets in large
dimensions. This context appears in particular in high reliability standards elds such as automotive, avionics or aerospace. In this context, ICS can
be useful for detecting anomalies with a small proportion of false positives. The method will be illustrated on several artiﬁcial and real data sets
using the recent R packages ICSOutlier and ICSShiny. The package ICSOutlier allows to choose scatter matrices, automatically select the most
relevant components, calculate an outlierness index and identify potential outlying observations. The ICSShiny package provides a user-friendly
application for ICS in particular for outlier detection.
CP001 Room Triton 3 POSTER SESSION Chair: Maria Elena Fernandez Iglesias
C0154: Resampling approaches for multivariate data: Theory, R-package and applications
Presenter: Sarah Friedrich, University of Ulm, Germany
Co-authors: Markus Pauly
In many experiments in the life sciences several endpoints, potentially measured on different scales, are recorded per subject. Classical MANOVA
models assume normally distributed errors and homogeneity of the covariance matrices, two assumptions that are often not met in practice.
Inference is further complicated, if covariance matrices are singular. We propose a test statistic for factorial MANOVA designs which incorporates
general heteroscedastic models with possibly singular covariance matrices. Different bootstrap techniques are used in order to derive inference
even for small samples. The methods are implemented in the R package MANOVA.RM. The package allows for different factorial designs with
nested and crossed factors and comes with a plotting routine and a graphical user interface. We present the main functionalities of the package and
use it to analyse a practical data set.
C0173: Periodic bivariate count time series models
Presenter: Magda Monteiro, University of Aveiro, Portugal
Co-authors: Isabel Pereira, Manuel Scotto
In real life, there are several count time series that exhibit periodicity and are also related to each other allowing there joint modelling as a bivariate
time series. Examples can be found in different ﬁeld areas such as environmental (monthly number of ﬁres in two neighbour counties), tourism
(monthly number of guests from neighbours hotels), labour market (monthly number of short and long-term unemployed in a county) among others.
As in other types of time series, the modelling of these time series can be done using different approaches of which we highlight models based on
thinned operations and bivariate dynamic factor models. We will present a comparative study of a periodic bivariate integer-value autoregressive
(PBINAR) model and a bivariate dynamic factor model under the context of forest ﬁres application.
C0174: Factor model estimation by composite minimization
Presenter: Matteo Farne, University of Bologna, Italy
Co-authors: Angela Montanari
The problem of factor model estimation in large dimensions is addressed under the low rank plus sparse assumption. Existing approaches based on
PCA like POET estimator fail to catch low rank spaces characterized by non-spiked eigenvalues, as in this case the asymptotic consistency of PCA
defaults. UNALCE, an alternative approach based on the minimization of a low rank plus sparse decomposition problem, is shown to produce the
covariance estimate with the least possible dispersed eigenvalues among all the matrices having the same rank of the low rank component and the
same support of the sparse component. Consequently, if dimension and sample size are ﬁxed, loadings and factor scores estimated via UNALCE
provide the tightest possible error bound. The result is based on the sample eigenvalue dispersion lemma. The effectiveness of UNALCE factor
estimates is ﬁnally explored in an exhaustive simulation study, which clariﬁes that the gain of UNALCE is larger as the latent eigenvalues are less
spiked and the sparse component is more sparse.
C0179: Impact of rainfall on the greater ﬂow peaks in Esva river Basin in NW of Spain and relations to land use changes
Presenter: Maria Elena Fernandez Iglesias, University of Oviedo, Spain
Co-authors: Gil Gonzalez-Rodriguez, Jorge Marquinez, Maria Fernandez-Garcia
The river channel in the Esva basin, a coastal catchment of the Cantabrian region (Northwest of Spain), has experienced a slight active channel-
width decrease from 1957 to 1985 (<1%) and more important decrease (close to 13%) from 1985 to 2003. This trend is well related to the main
changes in the forest cover, which also increases slightly from 1957 to 1985 and more importantly after 1985 until now. Models for different
scenarios were applied for the daily datasets before and after 1985 with the aim of identifying whether the changes in the land uses might inﬂuence
in the hydrological response of the river to the rainfall. The preliminary results focused on ﬂow peak events conclude a light increase in this time
reaction rainfall-ﬂow in agreement with the land usage changes.
ECOSTA ECONOMETRICS AND STATISTICS. All rights reserved. 5