Four simultaneous Markov chains Monte carlo (MCMC) were run for 10,000,000 generation sampling every 1000 genertion. After examining the MCMC convergence using tracer (RAmbaut and Drummond, 2007), the initial 2,000,000 generation from each ran were discarded from the analysis as burn-in while the remaining trees were used to construct a consensus tree with posterior probabilities (PP) assessing the statistical nodal support. Two partitioning schemes were tested, one unlinking the model parameters between nuclear and plastid regions, the other unlinking all markers (i.e three partitions). The best-fit model was chosen by Bayes factor test based on the harmonic mean of the respective long likelihoods (Kass and Raftery, 1995).