Variational approximations of multivariate generalised linear mixed models

David Hughes, University of Liverpool. Part of the statistics seminar series.

Increasingly in clinical research, information is collected about a wide range of clinical markers thought to be indicators or disease progression.Typically, each marker may be analysed separately, although this potentially ignores their correlation. For this reason multivariate modelling would be desirable, such as a multivariate mixed model.

However, the computational complexity of such models limits use to relatively small datasets. In this talk I propose Mean Field Variational Bayes as a possible solution to this problem. This approach allows computationally efficient scaling to large datasets and high numbers of markers. I will describe our method and illustrate its use in clinical datasets of different sizes, from diverse clinical conditions including primary biliary cirrhosis, liver cancer and diabetic retinopathy.