Modeling data from cluster randomized trials with a small number of big clusters and a random-split method