Facial expressions convey a speaker's emotional state, facilitating the prediction and interpretation of their thoughts and behaviours. Interactive feedback during social interactions provides statistical evidence for the basis of a causal percept, which allows understanding of conversations. We aimed to determine whether emotional expression affects sensitivity to contingent relationships and whether this sensitivity is guided by the statistical evidence for causality. In Experiments 1-3, we tested happy and sad facial expressions and non-emotional control stimuli (e.g. shapes) and varied contingent emotional expressions (negative, zero and positive contingency) as well as outcome frequency (low, moderate and high). Participants' judgements of contingency were based on a probabilistic learning process rather than simple pairing or prior knowledge, and they perceived a weaker sense of causality with sad faces than either happy faces or non-emotional control stimuli. Finally, in Experiment 4, we tested threat-related angry and fearful faces alongside happy faces. The results showed that participants could learn the statistical contingent relationships with faces but still perceived a weaker sense of causality with angry, fearful faces compared to happy faces. Overall, the results suggest that learning was guided by statistical evidence, but aversive expressions (those with negative valence) were less effective. We discuss this result in relation to the stimulus properties (i.e. salience) of faces, the content of emotive expressions and how these impact learning.