Dynamic Bayesian Networks for Evaluation of Granger Causal Relationships in Climate Reanalyses

Abstract

Bayesian structure learning provides a principled approach to quantifying uncer-6 tainty in estimated network structures for relationships between climate modes 7 • Dynamic Bayesian networks estimated from NCEP/NCAR and JRA-55 reanal-8 ysis data show broad overall consistency 9 • Structural differences in high posterior credibility associations may be indicative 10 of biases relevant for subsequent model evaluation

Causal Links Between Arctic Sea Ice and Its Potential Drivers Based on the Rate of Information Transfer

“…• The Liang-Kleeman rate of information transfer allows quantification of the directional dependence between Arctic sea ice and its drivers • Recent and future changes in Arctic sea ice are mainly driven by air and sea-surface temperatures and ocean heat transport • The influence of Arctic sea ice on air temperature and ocean heat transport progressively decreases through the twenty-first century However, the presence of a correlation between one variable and another does not firmly demonstrate a causal influence between these variables. In order to identify such a causal link, causal inference frameworks can be used (Granger, 1969;Krakovska et al, 2018;Liang, 2014;Runge et al, 2019;Sugihara et al, 2012) and have been applied to climate studies (e.g., Deza et al, 2015;Harries & O'Kane, 2021;Kretschmer et al, 2016;Mosedale et al, 2006;Tsonis et al, 2015;Vannitsem & Ekelmans, 2018). The Liang-Kleeman information flow method (Liang & Kleeman, 2005) is particularly interesting because it allows identifying of the direction and magnitude of the cause-effect relationships between variables (Liang, 2014(Liang, , 2021.…”

Bayesian Structure Learning for Climate Model Evaluation

“…We show graphs where the "child" is the node associated with a given index at t = 0, whose "parents" are any node for a given index at lags t = 1, …, 6 months for edges with an estimated posterior weight greater than 0.5. Panel (a) shows the format used in Harries and O'Kane (2021). In panel (b) a reduced representation of the same directed acyclic graph is shown where it is assumed the edge exists only between parent and child.…”

“…We use this sampling algorithm for all of the results presented in this study. The required closed-form expressions for the prior and posterior densities for the parameters of the node conditional distributions, and the resulting marginal likelihoods or local scores, for the models used are detailed in Appendix B of Harries and O'Kane (2021).…”

“…development displaced in time by well over a decade. Harries and O'Kane (2021) provide an extensive comparison and discussion of the two reanalysis products.…”

Causal links between Arctic sea ice and its potential drivers based on the rate of information transfer

“…• The Liang-Kleeman rate of information transfer allows quantification of the directional dependence between Arctic sea ice and its drivers • Recent and future changes in Arctic sea ice are mainly driven by air and sea-surface temperatures and ocean heat transport • The influence of Arctic sea ice on air temperature and ocean heat transport progressively decreases through the twenty-first century However, the presence of a correlation between one variable and another does not firmly demonstrate a causal influence between these variables. In order to identify such a causal link, causal inference frameworks can be used (Granger, 1969;Krakovska et al, 2018;Liang, 2014;Runge et al, 2019;Sugihara et al, 2012) and have been applied to climate studies (e.g., Deza et al, 2015;Harries & O'Kane, 2021;Kretschmer et al, 2016;Mosedale et al, 2006;Tsonis et al, 2015;Vannitsem & Ekelmans, 2018). The Liang-Kleeman information flow method (Liang & Kleeman, 2005) is particularly interesting because it allows identifying of the direction and magnitude of the cause-effect relationships between variables (Liang, 2014(Liang, , 2021.…”