We design a novel public goods game (PGG) and a corresponding laboratory experiment to study individuals' beliefs and contributions to a public good when subjects can endogenously affect both the probability and the magnitude of the possible outcomes in both the gains and loss domains. We find that overall contributions to the public good are greater in the loss than in the gains domain, and this is mainly driven by the differences in contributions to the probability account across domains: subjects contribute significantly more to the probability account in the loss domain than in the gains domain, while contributions to the magnitude account do not differ across domains. We find that this effect is persistent over rounds, and that free riding in the loss domain is the lowest. We further conducted treatment with only one public account in both domains to understand the results in the presence of both public accounts. A possible mechanism that explains this disproportionate investment in reducing the probability of collective losses rather than their magnitude, is loss aversion.