Genome wide association studies address the localization and identification of causal mutations responsible for common, complex human genetic diseases. Nevertheless, this task has been revealed to be a formidable challenge because of the huge amount and the complexity of the data to analyze. At the frontier between machine learning and statistics, probabilistic graphical models, such as hierarchical Bayesian networks, are potentially powerful tools to tackle this issue. In this research work, we evaluate a novel method based on forests of hierarchical latent class models. We show the relevance of using this class of models for the purpose of genetic association studies. We correct for multiple testing and cope with cardinality heterogeneity amongst the model's latent variables. For this purpose, we design a layer-wise permutation procedure. We empirically prove, using both simulated and real data, the ability of the model's latent variables to capture indirect genetic associations with the disease. Strong associations are evidenced between the disease and the causal genetic marker's ancestor nodes in the forest. At the opposite, very weak associations are obtained regarding the causal genetic marker's non-ancestor nodes.