Computer intensive methods for Generalized Linear Mixed Models
Seminari GRASS - 13 de desembre 2006
DATE: 13 de desembre 2006 a les 11 hores
SPEAKER: Josep Anton Sánchez
IDIOM: Català
PLACE: Aula C5016, Edifici C5, Campus Nord, UPC
SUMMARY:
Generalised Linear Mixed Models allow to deal with non-normal response, like binary or count data. These models take into account dependent structure of grouped data coming from repeated measurements or longitudinal data. Although this methodology has a wide range of application, estimation procedures usually rely on numerical approximations (Laplace, PQL). Therefore, inference from classical approaches is sometimes inaccurate or has lack of power.
An R library which implements generation of bootstrap data and efficient fit of GLMM via PQL is presented. Several options can be used to generate the resampled data: parametric or empirical resampling of random effects and generation according to the theoretical distribution from the linear predictor, reconstruction of data from different residual resampling with the corresponding truncation, with or without the nested structure between random effects and residuals. Fitting the resampled data, confidence intervals and empirical p-values based on the bootstrap methodology can be obtained, for fixed effects and for variance components.
With the same library, a simulation study has been developed in order to compare the performance of different configurations of bootstrap procedures in presence of misspecification of distribution of random effects and/or residuals for logistic and Poisson regression.
SPEAKER: Josep Anton Sánchez
IDIOM: Català
PLACE: Aula C5016, Edifici C5, Campus Nord, UPC
SUMMARY:
Generalised Linear Mixed Models allow to deal with non-normal response, like binary or count data. These models take into account dependent structure of grouped data coming from repeated measurements or longitudinal data. Although this methodology has a wide range of application, estimation procedures usually rely on numerical approximations (Laplace, PQL). Therefore, inference from classical approaches is sometimes inaccurate or has lack of power.
An R library which implements generation of bootstrap data and efficient fit of GLMM via PQL is presented. Several options can be used to generate the resampled data: parametric or empirical resampling of random effects and generation according to the theoretical distribution from the linear predictor, reconstruction of data from different residual resampling with the corresponding truncation, with or without the nested structure between random effects and residuals. Fitting the resampled data, confidence intervals and empirical p-values based on the bootstrap methodology can be obtained, for fixed effects and for variance components.
With the same library, a simulation study has been developed in order to compare the performance of different configurations of bootstrap procedures in presence of misspecification of distribution of random effects and/or residuals for logistic and Poisson regression.
Share: