We all know that correlation is not causality, and that cofounding factors can easily mislead us to spurious conclusions. Even when all cofounding factors are observed, we know that many causal DAGs (Directed Acyclic Graphs) structures can lead to the same likelihood; this is the so-called Markov equivalence. In simple cases, we can reduce the Markov equivalence class to a single DAG structure by using intervention experiments like clinical randomized trial. But in complex situations, even intervention experiments might not allow to infer with certainty all causal relationships. Moreover, in practice, we are often confronted to mixture of observation and intervention experiments. In this talk, we start by presenting the notion of PDAG (Partially Directed Acyclic Graphs) as a representation of the Markov equivalence class of a DAG. We will start by PDAG in the presence of observation experiments only, and then generalize to a mixture of observation and intervention experiments. We will then see how it is possible to compute the likelihood of a PDAG assuming that the underlying variables are connected to each other through generalized linear models (e.g. linear regression, logistic regression, survival, etc.). Finally, we will explain how it is possible to use the BIC criterion and MCMC (Markov Chain Monte-Carlo) in order to explore the PDAG space and derive a posterior distribution.