Since causal evidence is often ambiguous, models of causal learning should be able to represent uncertainty over causal hypotheses. Uncertainty is especially important in retrospective revaluation (the re-evaluation of ambiguous evidence in light of subsequent learning). We examine how a Bayesian model and an associative model (the modified SOP model of Dickinson & Burke, 1996) deal with this uncertainty. We tested the predictions of the models in an experiment with retrospective revaluation of preventive causes. Results were consistent with the predictions of the Bayesian model, but inconsistent with the predictions of the modified SOP model.