Soladis aux NCS2016 d'octobre à Cambridge

Soladis aux NCS2016 d'octobre à Cambridge

Soladis participera aux conférences Non-clinical Statistics du 4 au 6 octobre à Cambridge, l'occasion pour nos collaborateurs de présenter des résultats de recherche, et s'enrichir des nouvelles découvertes présentées par d'autres chercheurs.

Pour ceux et celles qui prévoient d'y assister, voici un petit avant-gout du poster qui sera présenté par Balazs Banfai, collaborateur de notre filiale suisse SOLADIS GmbH, qui a procédé à une évaluation des différents approches possibles en termes de modélisation d'études longétudinales.


Evaluation of modeling approaches for longitudinal and hierarchical designs (abstract)

by Balazs Banfai

Complicated models can arise in longitudinal studies in proteomics, cytometry or behavioral experiments. Repeated measures call for mixed effects models; regularization may be needed for collapsing the peptide data to protein information; and cytometry and behavioral readouts show non-normal distributions. These experiments usually contain treatment effects, covariates for different cohorts, and factors of experimental conditions. Integrating these modeling approaches may prove to be complex in practice.

The commonly used R packages for mixed models (nlme, lme4), beta regression (betareg), and regularization (glmnet) are compared to the generalized additive models, particularly the GAMLSS framework. In GAMLSS random effects, non-normal response variables and regularization can be handled in a convenient fashion.

The results of the frequentist modeling approaches are contrasted to Bayesian hierarchical models, focusing on model construction, using historical data for prior selection, and technical problems of running the models.

In pharmaceutical research high-throughput analyses are needed, with data management preceding and bioinformatics analysis following the statistical modeling step. Therefore it is important that modeling can be integrated in an analysis pipeline with as much automatization as possible. A comparison is made between the two approaches focusing on modeling performance, and the ease of integration and evaluation. Simulated and real-world examples are presented to show the strengths and weaknesses of the different approaches.