1: \begin{abstract}

2: Although there is much recent work developing flexible variational methods for Bayesian computation, 

3: Gaussian approximations with structured covariance matrices are often preferred computationally in high-dimensional settings.  

4: This paper considers approximate

5: inference methods for complex latent variable models 

6: where the posterior is close to Gaussian, but with 

7: some skewness in the posterior marginals.

8: We consider skew decomposable graphical 

9: models (SDGMs), which are based on the closed skew normal family

10: of distributions, as

11: variational approximations.  These approximations can

12: reflect the true posterior conditional independence structure and capture posterior skewness. 

13: Different parametrizations are explored for this variational family, 

14: and the speed of convergence and quality of the approximation 

15: can depend on the parametrization used.  

16: To increase flexibility, 

17: implicit copula SDGM approximations  

18: are also developed, where elementwise transformations of 

19: an approximately standardized SDGM random vector are considered. 

20: Our parametrization of the implicit copula approximation 

21: is novel, even in the special case of a Gaussian approximation.

22: Performance of the methods is examined in a number of real examples involving generalized linear

23: mixed models and state space models, and

24: we conclude that our copula approaches are most accurate, 

25: but that the SDGM methods are often nearly as good and have lower

26: computational demands.

27: 

28: \smallskip

29: \noindent \textbf{Keywords:} Closed skew normal distribution; Copula variational approximation; Decomposable graphical model; Importance sampling.

30: 

31: \end{abstract}

32: