Contingency discourse relations play an important role in natural language understanding. In this paper we propose an unsupervised learning model to automatically identify contingency relationships between scenario-specific events in web news articles (on the Iraq war and on hurricane Katrina). The model generates ranked contingency relationships by identifying appropriate candidate event pairs for each scenario of a particular domain. Scenario-specific events, contributing towards the same objectives in a domain, are likely to be dependent on each other, and thus form good candidates for contingency relationships. In order to evaluate the ranked contingency relationships, we rely on the manipulation theory of causation and a comparison of precision-recall performance curves. We also perform various tests which bring insights into how people perceive causality. For example, our findings show that the larger the distance between two events, the more likely it becomes for the annotators to identify them as non-causal.