Greetings, We want to fit lme/gls models with a compound symmetry variance structure, but where the ICC takes different values across different groups of clusters. (Specifically a two arm clustrered trial with different ICC in each arm) Equivalent models can be fitted with heteroscedastic variance functions, but these have positivity constraints which we want to avoid and we would prefer to use the correlation structure directly. Is there a way to do this? Writing a new variance structure seems rather non-trivial - but maybe I haven't tried hard enough.. Any ideas? Steve. Dr Steve Roberts steve.roberts at manchester.ac.uk Senior Lecturer in Medical Statistics, CMMCH NHS Trust and University of Manchester Biostatistics Group, 0161 275 5192/5764 / 0161 276 5785