Qualité de l'ajustement dans des modèles bayésiens génératifs
Approximate Bayesian Computation (ABC) methods use simulations to calibrate and select models. Those methods are adapted to complex models when likelihood functions are intractable as in population genetics. ABC starts by simulating pseudo-data from different proposed models, then statistical methods are applied to infer parameters or select models. Goodness-of-fit (GOF) algorithms try to check the adequacy between data and selected models. We follow the approach introduced in Lemaire et al. (2016) which is based on hypothesis-testing. The proposed statistical test has a variable and sometimes quite poor power. Here, we present an other goodness-of-fit measure that combines their approach with the novelty detection algorithm Local Outlier Factor (LOF) Breunig et al. (2000). This combination turns out to give good power for toy models as well as for complex models from population genetics.