Stilt: Easy Emulation of Time Series AR(1) Computer Model Output in Multidimensional Parameter Space

Authors

• Roman Olson, Yonsei University, Center for Climate Physics, Pusan National University
• Kelsey L. Ruckert, Penn State University
• Won Chang, Department of Mathematical Sciences
• Klaus Keller, Penn State University
• Murali Haran, Penn State University
• Soon-Il An, Yonsei University

Date of this Version

12-2018

Document Type

Article

Citation

The R Journal (December 2018) 10(2); Editor: John Verzani

Comments

Copyright 2018, The R Foundation. Open access material. License: CC BY 4.0 International

Abstract

Statistically approximating or “emulating” time series model output in parameter space is a common problem in climate science and other fields. There are many packages for spatio-temporal modeling. However, they often lack focus on time series, and exhibit statistical complexity. Here, we present the R package stilt designed for simplified AR(1) time series Gaussian process emulation, and provide examples relevant to climate modelling. Notably absent is Markov chain Monte Carlo estimation – a challenging concept to many scientists. We keep the number of user choices to a minimum. Hence, the package can be useful pedagogically, while still applicable to real life emulation problems. We provide functions for emulator cross-validation, empirical coverage, prediction, as well as response surface plotting. While the examples focus on climate model emulation, the emulator is general and can be also used for kriging spatio-temporal data.