[Back]

N. Neykov, P. Filzmoser, P. Neytchev:
"Robust variable selection in joint modeling of location, scale and shape for high dimensional data through trimming (CS6)";
Talk: ICORS2014 - 14th International Conference on Robust Statistics 2014, Martin-Luther-Universität Halle-Wittenberg, D; 2014-08-10 - 2014-08-15; in: "ICORS14 - Conference Guide & Book of Abstracts", (2014), 39.

Generalized linear and additive models (GAMLSSs) for location, scale and shape
were discussed by Rigby & Stasinopoulos (2005) as a class of statistical models for regression
problems with univariate response. A key feature of GAMLSSs is that every parameter of the
conditional response distribution can be modeled by its own predictor and an associated link
function. The GAMLSSs estimation is based on the penalized likelihood whereas the gener-
alized version of AIC and BIC is used for variable selection. In order to avoid the GAMLSSs
de ciencies in presence of high dimensional data settings Mayr et al. (2012) developed the
gradient boosting algorithm gamboostLSS to handle the computations within the GAMLSSs
framework. However, the proposed boosting estimators can be very sensitive to outliers in
the data, especially to outliers in the covariates (leverage points). In order to overcome these
disadvantages, the usage of the boosting estimators based on trimming are recommended to
estimate the unknown parameters in a robust way. The superiority of this approach is illus-
trated by examples in a simulation study. As a prominent measure of robustness, the nite
sample breakdown point of the considered estimators are characterized in this setting using
the notion of d-fullness following Neykov et al. (2014).

Penalized Maximum Trimmed Likelihood Estimator Breakdown point GLMs

http://statistik.wiwi.uni-halle.de/icors2014/