Abstract:
|
The climate change detection and attribution literature aims to quantify the extent to which increasing greenhouse gas emissions are responsible for observed changes in the climate. One of the major challenges of this problem is estimating the covariance matrix for natural variability, as different estimation choices are known to meaningfully impact the results. In addition, traditional approaches often produce uncertainty intervals that exhibit under-coverage, indicating an underestimation of the true uncertainty. In this work, we propose a flexible spatial parameterization of the covariance matrix within a Bayesian hierarchical framework in order to propagate the uncertainty induced from estimating the covariance matrix to the attribution parameter. We show using climate model simulations that this approach achieves better coverage rates than approaches that do not take this uncertainty into account. Our parameterization also allows for covariance estimations to be pooled between different climate models, allowing for the uncertainty from natural variability to be distinguished from the uncertainty induced from the use of climate models.
|