Composite likelihood estimation methods based on low-dimensional margins (Harry Joe)
Composite likelihood estimation methods are based on sums of low-dimensional marginal or conditional log-likelihoods, when the joint multivariate probability or density is computationally too difficult. A common special case is pairwise likelihood. We have studied composite likelihood methods for (a) multivariate models for clustered/longitudinal data and (b) models with a latent Gaussian process for times series data. Based on our analyses, through theory and numerical examples, we can give some recommendations on the choice of weights of different margins to achieve a good balance of statistical efficiency and computational time. Some of the many applications of composite likelihood methods will be mentioned.