Você está aqui: Página Inicial / Seminários / 27/10/2017 - Multivariate Longitudinal Compound Poisson Regression Modeling - Palestrante: Afrânio M. C. Vieira (UFSCar)

27/10/2017 - Multivariate Longitudinal Compound Poisson Regression Modeling - Palestrante: Afrânio M. C. Vieira (UFSCar)

Quando 27/10/2017
das 13h00 até 14h30
Adicionar evento ao calendário vCal
iCal

Seminário Conjunto UFSCar/ICMC – 27/10/2017 - 14h00

LOCAL: Sala 38, DEs-UFSCar

PALESTRANTE: Prof. Dr. Afrânio M. C. Vieira, DEs-UFSCar

TÍTULO: Multivariate Longitudinal Compound Poisson Regression Modeling

RESUMO: On studies about the blend of volatile organic compounds, released due to the damage caused by herbivorous insects, usually produce a number of mass concentration of theses compounds (say 30, 50 or higher), correlated each other, what partially describes the metabolic route stimulated by the damage stimulus. These measures are asymmetric with a probability to occur zero values. Furthermore, the stimuli are experiment-based and longitudinal measures are taken, trying to identify significant factors and/or time point that are scientifically relevant.
In this work will be presented a proposal of multivariate data analysis, based on generalized linear mixed models where the multivariate response will be assumed as Tweedie's compound Poisson distribution. The modeling strategy adopted was all-pairwise mixed models as well a pseudo-likelihood approach to reconstruct the covariance response matrix, already adjusted for non normality, experimental design structure and longitudinal correlations. It means that traditional multivariate analysis as PCA, factor and cluster analysis can be produced, bringing light to a complex data structure where, usually, univariate analyses are made. The package pair.mglmm for R computing environment was produced based on cplm and doParallel packages. These packages allow fitting compound Poisson mixed models using parallel computing on multicores processors. Simulations where done to evaluate the computational effort and it can be concluded that even for higher number of multivariate responses, the modeling is feasible, where the full multivariate approach isn't possible. These models were applied on a cotton plant study, where 25 chemical compound mass where measured on 4 time points, under a completely randomized experimental design with 5 treatments. The results were promising in the biological control research area as well as methodological approach for biostatisticians.


Estão todos convidados!

registrado em: