Explaining rapid reinfections in multiple-wave influenza outbreaks: Tristan da Cunha 1971 epidemic as a case study

Camacho, Anton; Ballesteros, Sébastien; Graham, Andrea L.; Carrat, Fabrice; Ratmann, Oliver; Cazelles, Bernard

The maximized likelihood provides a first quantitative answer to the objective question: 'How likely is it that the stochastic process resulted in the observed epidemic?'. The AIC c [21] is then a related measure of the expected predictive capability of the model that penalizes model complexity. The rescaled AIC c values, presented in figure 3 , allow for an immediate ranking of the competing models and show that the Win mechanism best explains the data. Following the rough rule of thumb of Burnham & Anderson [22] , the AoN hypothesis also receives substantial support (DAIC c 2). In contrast, the 2Vi and Mut models have considerably less support (DAIC c . 7), whereas the InH and PPI models have essentially no support (DAIC c p 10). This rule of thumb has proved to be efficient in many practical situations [22, 23] and can formally be justified by computing the evidence ratio of each competing model (electronic supplementary material, text S3). However, it has also been shown that one should be cautious regarding the systematic use of this rule of thumb when applying AIC corrections [23] , which motivates a more detailed analysis of DAIC c values.