Global Sensitivity Analysis in Integrated Assessment Modeling

Integrated assessment modeling studies the nexus between the systems of climate and economy, both known for their high complexity and vast uncertainty. One key question is how sensitive climate policy inference produced by such models is to the uncertainty in their initial assumptions and the results of calibrations. Despite the broad literature on the topic---rich in both single-model and multi-model sensitivity analyses---universal, well-established practices for of analyzing the uncertainty in model's outcome are still missing. In this paper we argue for structured global sensitivity analysis (GSA) as an indispensable routine in climate-economic modeling.We apply a high-efficiency GSA method based on polynomial chaos expansions to the most commonly employed integrated assessment model (IAM), DICE. Our analysis provides two key insights. First, the subjective preselection of a subset of parameters of interest might omit the most influential ones. Second, the opposite strategy---one of pooling all parameters together and considering the model as a "black box"---might distort sensitivity indices. The best practice is thus to consider all exogenous parameters but adjust their set such that the fundamental relations within the model's structure are respected. The methodology of efficient GSA provides a clear, comprehensive decomposition of the uncertainty in model's output while minimizing computational costs, and hence is easily applicable to IAMs of higher complexity