Composite Likelihood: Some Biomedical Applications

Abstract

In the past several decades, composite likelihood has drawn a good deal of attention as a tool for statistical inference, although the idea was carried out under different names such as pseudo-likelihood and independent likelihood. This approach is particularly appealing when the full log-likelihood function is difficult to compute and/or far from being normal. For the latter, the adequacy of asymptotic approximation for the maximum likelihood estimate maybe in doubt especially when there are additional nuisance parameters to deal with. Composite likelihood approach has the additional advantage for being more robust in that fewer assumptions, compared to the full likelihood approach, are needed to carry out the inference. In this talk, the points noted above are illustrated through a series of examples we have encountered in the past. In addition, we will present in details how composite likelihood approach maybe applied to case-control studies with ordinal disease categories. A genetic association study of schizophrenia will be presented for illustration.