A0592
Title: Quantifying causal relationships from climate observations using spatiotemporal stochastic interventions
Authors: Samuel Baugh - Pennsylvania State University (United States) [presenting]
Abstract: Physical dynamics indicate a causal relationship between greenhouse gas concentrations and a changing climate, but estimating the exact magnitude of expected warming is a notoriously difficult problem. Inferring the effect from observations alone can be framed as an observational causal inference problem; however, in addition to the classical challenge of accounting for unobserved counterfactual outcomes, causal inference in the climate system is particularly difficult due to the fact that there is essentially a single observation. The aim is to propose addressing these challenges by extending the stochastic intervention framework to continuous, physically informed spatiotemporal processes. By specifying the distributional form of the stochastic intervention, this method can consistently estimate the spatially-varying causal effect from observations. As these distributions cannot be estimated from observations alone without unrealistically strong assumptions, a framework is proposed for additionally incorporating prior information from climate model simulations to constrain the estimation. The robustness of the resulting method is assessed through sensitivity analyses and validation studies using withheld climate model data.
