Simple reparameterization to improve convergence in linear mixed models

Authors

  • Gregor GORJANC Univ. of Ljubljana, Biotechnical Fac., Dept. of Animal Science, Groblje 3, SI-1230 Domžale, Slovenia
  • Tina FLISAR Univ. of Ljubljana, Biotechnical Fac., Dept. of Animal Science, Groblje 3, SI-1230 Domžale, Slovenia
  • Jose Carlos MARTÍNEZ-ÁVILA Departamento de Mejora Genética, Instituto Nacional de Investigación Agraria, Carretera de La Coruña, km 7, 28040 Madrid, España
  • Luis Alberto GARCÍA-CORTÉS Departamento de Mejora Genética, Instituto Nacional de Investigación Agraria, Carretera de La Coruña, km 7, 28040 Madrid, España

DOI:

https://doi.org/10.14720/aas.2010.96.2.14699

Keywords:

statistics, mixed model, Bayesian analysis, McMC, reparameterization, convergence

Abstract

Slow convergence and mixing are one of the main problems of Markov chain Monte Carlo (McMC) algorithms applied to mixed models in animal breeding. Poor convergence is to a large extent caused by high posterior correlation between variance components and solutions for the levels of associated effects. A simple reparameterization of the conventional model for variance component estimation is presented which improves McMC sampling and provides the same posterior distributions as the conventional model. Reparameterization is based on the rescaling of hierarchical (random) effects in a model, which alleviates posterior correlation. The developed model is compared against the conventional model using several simulated data sets. Results show that presented reparameterization has better behaviour of associated sampling methods and is several times more efficient for the low values of heritability.

Downloads

Published

27. 12. 2010

Issue

Section

Original Scientific Article

How to Cite

GORJANC, G., FLISAR, T., MARTÍNEZ-ÁVILA, J. C., & GARCÍA-CORTÉS, L. A. (2010). Simple reparameterization to improve convergence in linear mixed models. Acta Agriculturae Slovenica, 96(2), 69–73. https://doi.org/10.14720/aas.2010.96.2.14699