A: For low numbers of introductions (5 of the 20 hosts), there is no difference between methods in the posterior log-likelihood distribution. B: Higher numbers of introductions (15 of the 20 hosts), performance of MCMC with a random tree as initialization of the history host is inferior to either p(MC 3 ), neighbour-joining tree initialization of the history host or the combination of both. Moreover, the simulated outbreak has a log-likelihood (the vertical black line) that is higher than the log-likelihood distribution of MCMC with a random tree as initialization. The latter gives the highest likelihood distribution and is chosen as default option in all analyses. ‘random’ is random tree initialization, ‘nj’ is neighbour-joining tree initialization, ‘2’ is MCMC and ‘3’ is p(MC 3 ). The black lines are the log-likelihood values of the simulated outbreaks. (TIF)