Analysis of climate sensitivity via high-dimensional principal component regression

Abstract

Uncertainty in the sensitivity of global mean air temperature to changes in atmospheric CO₂ concentration arises from factors controlling the response of cloud amounts. Here, we present an analysis of these factors within an ensemble consisting of 164 variations of the NCAR Community Atmosphere Model version 3.1, each of which differs slightly in their treatment of convection and cloud radiative properties. The problem is to uniquely relate observable aspects of model output, which are inherently high dimensional such as the 64,000 spatially mapped values of 11 fields generated from each CAM3.1 model to a scalar prediction of each model’s sensitivity to CO₂ forcing. We approach this problem using principal component regression and provide maps of observable predictors. Furthermore, based on this approach, we develop prediction intervals and show that they closely match their advertised coverage.

Keywords:

• Atmospheric models
• global warming experiments
• regression modeling
• principal component selection
• uncertainty quantification