A comparison of robust nonlinear regression methods by statistical learning

0491783 - ÚI 2019 IT eng A - Abstract
Kalina, Jan - Peštová, Barbora
A comparison of robust nonlinear regression methods by statistical learning.
ISNPS 2018. Book of Abstracts. Salerno, 2018 - (La Rocca, M.; Liseo, B.; Parella, M.; Salmaso, L.; Tardella, L.). s. 42-42. ISBN 978-88-61970-00-7.
[ISNPS 2018. Conference of the International Society for Nonparametric Statistics /4./. 11.06.2018-15.06.2018, Salerno]
Institutional support: RVO:67985807 ; RVO:67985556
Keywords : metalearning * robust estimation * nonlinear regression * nonlinear regression quantiles * heteroscedasticity
OECD category: Statistics and probability
https://drive.google.com/file/d/13Sqxpj5A0oHiNn4jLBGSUPpSmlYvFX-0/view

Various estimators for the standard nonlinear regression model are compared with a focus on methods which are robust to outlying measurements in the data. The main contribution is a metalearning study which has the aim to predict the most suitable estimator for a particular data set. Here, various versions of the nonlinear least weighted squares estimator are compared with nonlinear least squares, nonlinear least trimmed squares and a nonlinear regression median, where the last is a special case of nonlinear regression quantiles. The metalearning study is performed over a data base of 24 economic data sets. The nonlinear least weighted squares estimator is able to yield the best result for the most data sets. The metalearning study gives advice how to select appropriate weights for the nonlinear least weighted squares, particularly it reveals tests of normality and heteroscedasticity to play a crucial role in finding suitable weights.
Permanent Link: http://hdl.handle.net/11104/0285411