1: \begin{abstract}

2: The effects of different parametrizations on the convergence of

3: Bayesian computational algorithms for hierarchical models are well

4: explored. Techniques such as centering, noncentering and partial

5: noncentering can be used to accelerate convergence in MCMC and EM

6: algorithms but are still not well studied for variational Bayes (VB)

7: methods. As a fast deterministic approach to posterior approximation,

8: VB is attracting increasing interest due to its suitability for large

9: high-dimensional data. Use of different parametrizations for VB has not

10: only computational but also statistical implications, as different

11: parametrizations are associated with different factorized posterior

12: approximations. We examine the use of partially noncentered

13: parametrizations in VB for generalized linear mixed models (GLMMs). Our

14: paper makes four contributions. First, we show how to implement an

15: algorithm called nonconjugate variational message passing for GLMMs.

16: Second, we show that the partially noncentered parametrization can

17: adapt to the quantity of information in the data and determine a

18: parametrization close to optimal. Third, we show that partial

19: noncentering can accelerate convergence and produce more accurate

20: posterior approximations than centering or noncentering. Finally, we

21: demonstrate how the variational lower bound, produced as part of the

22: computation, can be useful for model selection.

23: \end{abstract}