Received: 26 Jan 2016 – Accepted for review: 26 Feb 2016 – Discussion started: 04 Mar 2016
Abstract. Climate data is highly correlated through the physics and dynamics of the atmosphere. Model evaluation often involves averages of various quantities over different regions and seasons making it difficult from a statistical perspective to quantify the significance of differences that arise between a model and observations. Here we present a strategy that makes use of a set of perfect modeling experiments to quantify the effects of these correlations on model evaluation metrics. This information is incorporated into Bayesian inference through a precision parameter with informative priors. These concepts are illustrated through an example of fitting a line through data that includes either uncorrelated or correlated noise as well as to the calibration of CAM3.1. The concept of a precision parameter may be applied as a strategy to weight different climate model evaluation metrics within a multivariate normal framework. From the example with CAM3.1, the precision parameter plays a central role in rescaling the estimated parametric uncertainties to better accommodate modeling structural errors.
This preprint has been withdrawn.
How to cite. Jackson, C. S. and Huerta, G.: Empirical Bayes approach to climate model calibration, Geosci. Model Dev. Discuss. [preprint], https://doi.org/10.5194/gmd-2016-20, 2016.
Climate data is highly correlated which can make it difficult from a statistical perspective to quantify the significance of differences that arise between a model and observations. Here we explore a common device in Bayesian inference for assessing the statistical significance of a fit between a model and data and suggest how this approach may be applied to the calibration of a climate model.
Climate data is highly correlated which can make it difficult from a statistical perspective to...