Determining the Direction of Causal Influence in Large Probabilistic Networks: A Constraint-Based Approach

Determining the direction of causal influence from observational data only is essential in many applications, such as the reconstruction of genetic networks from microarray data. As opposed to many probabilistic network inference algorithms which were designed to induce just statistical models of the data, Conditional Independence (CI) based algorithms are theoretically able to infer true causal models from observational data only. But unfortunately, the small sample sizes available from current microarray experiments render the determination of causal direction highly inaccurate. Here we show that this essential aspect of CI-based algorithms can be significantly improved by double-checking certain key statistical tests and by reconciling potential inconsistencies using a simple constraint propagation scheme.