INVESTIGADORES
YOHAI Victor Jaime
congresos y reuniones científicas
Título:
High breakdown point and efficient estimates for generalized linear models
Autor/es:
ANDREA BERGESIO; VÍCTOR J. YOHAI
Lugar:
Londrés
Reunión:
Conferencia; 3rd International Conference of the ERCIM Working Group on COMPUTING & STATISTICS.; 2010
Institución organizadora:
ERCIM Working Group on COMPUTING & STATISTICS.
Resumen:
We present a class of estimators for generalized linear models which combine high breakdown point and high efficiency. The proposed estimators are one step weighted M estimators, where the weights penalize leverage points. The initial estimator is a projection estimator similar to those defined for the linear model. The final estimator is obtained by performing one step of the scores algorithm for computing weighted M-estimates. We prove that the proposed estimators have a breakdown point as close to 0.5 as desired. Moreover, we also prove that they are asymptotically normal with the same asymptotic covariance matrix as the fully iterated weighted M-estimate. A Monte Carlo simulation shows that these estimators perform favorably for finite sample size compared with other robust estimators.
