Bibliographic Metadata

Title
Exploratory tools for cellwise outlier detection in compositional data with structural zeros / von Lukas Beisteiner
Additional Titles
Explorative Methoden für die Erkennung von zellweisen Ausreißern in Kompositionsdaten mit strukturellen Nullen
AuthorBeisteiner, Lukas
CensorTempl, Matthias
PublishedWien, 2016
Descriptionv, 86 Blätter : Diagramme
Institutional NoteTechnische Universität Wien, Diplomarbeit, 2016
Annotation
Abweichender Titel nach Übersetzung der Verfasserin/des Verfassers
LanguageEnglish
Document typeThesis (Diplom)
Keywords (EN)compositional data analysis / robust statistics / outlier detection
URNurn:nbn:at:at-ubtuw:1-4879 Persistent Identifier (URN)
Restriction-Information
 The work is publicly available
Files
Exploratory tools for cellwise outlier detection in compositional data with structural zeros [1.16 mb]
Links
Reference
Classification
Abstract (English)

The analysis of compositional data using the log-ratio approach is based on ratios between the compositional parts. Zeros in the parts thus cause severe difficulties for the analysis. Log-ratio transformations represent the compositional information into new coordinates. Outliers within these coordinates may be detected, however it remains unclear which particular parts of the composition led to the deviating ratios in question. To address this issue, the thesis presents four exploratory tools for identifying cellwise outliers in compositional data sets with structural zeros. In order to deal with structural zeros the proposed methods use robust imputation methods or split the data into subcompositions determined by their zero patterns. Ratios between parts are analyzed using an isometric log-ratio transformation or by observing pairwise log-ratios. Combining the results from robust regression analysis and robust distance calculations the approaches deduce row- and cellwise outliers within the original sample space. All four methods are applied on the household expenditure data from Albania and then compared. A close-to-reality simulation study is conducted to assess the performance of the different outlier detection algorithms.

Stats
The PDF-Document has been downloaded 47 times.