A likelihood-based comparison of the fitted models.

Log-likelihood (L) and number of estimated parameters (P) are given for the different models. The model selection criterion known as AIC (for Akaike Information Criterion) is computed as AIC = −2L+2P. A likelihood ratio test comparing each of the two non-relapse models SEIQS and SEIRS to the best relapse model (SEIH3QS) rejects these models (based on a chi-square test with 5 and 7 degrees of freedom respectively). The models with multiple dormant classes perform better than the one with a single class. For parsimony, the model SEIH3QS with three such classes is preferred over the model SEIH6QS with 6 classes (see Methods, Supplementary Material, for a description of the models).