Estimating Causal Attribution of Anthropogenic Forcing on High-Temperature Extremes Using a Latent Gaussian Spatial Model

Arnab Hazra; Ritik Roshan Giri

arxiv: 2604.23438 · v1 · submitted 2026-04-25 · 📊 stat.AP · stat.ME

Estimating Causal Attribution of Anthropogenic Forcing on High-Temperature Extremes Using a Latent Gaussian Spatial Model

Ritik Roshan Giri , Arnab Hazra This is my paper

Pith reviewed 2026-05-08 06:51 UTC · model grok-4.3

classification 📊 stat.AP stat.ME

keywords causal attributionextreme temperaturesanthropogenic forcingbivariate GEVlatent Gaussian modelspatial statisticsCMIP6Max-and-smooth

0 comments

The pith

A bivariate extreme value model with latent spatial effects isolates the causal contribution of human forcing to US temperature extremes by comparing factual and counterfactual simulations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a statistical method to measure how much human activities have increased the intensity and likelihood of record high temperatures. It compares annual maximum temperatures from climate model runs that include greenhouse gases and aerosols against identical runs that omit them. A bivariate generalized extreme value distribution captures the joint behavior of these two worlds, while a latent Gaussian layer with intrinsic conditional autoregressive structure accounts for spatial dependence across a grid. Approximate Bayesian inference via the Max-and-smooth algorithm produces posterior distributions for the differences in return levels, which are interpreted as the causal effect of anthropogenic forcing. The resulting maps show where this effect is credibly nonzero and how it has changed over time.

Core claim

By jointly modeling annual temperature maxima from factual and counterfactual CMIP6 runs inside a bivariate generalized extreme value distribution whose location and scale parameters vary spatially according to a latent Gaussian process with multivariate intrinsic CAR priors, the differences in return levels between the two worlds can be attributed to anthropogenic forcing; posterior inference with the Max-and-smooth Laplace approximation then yields spatially resolved estimates of this causal effect together with its temporal trend and credible regions of statistical significance.

What carries the argument

The difference in return levels of annual maxima between factual and counterfactual worlds, obtained from a bivariate GEV whose spatially varying coefficients are given a latent Gaussian prior with multivariate intrinsic CAR dependence.

If this is right

Posterior means and credible intervals for the change in return levels of annual temperature maxima attributable to human influence.
Maps of the temporal trend in that causal effect across the modeling period.
Credible regions identifying locations where the anthropogenic effect on extremes is statistically distinguishable from zero.
A scalable workflow that produces these quantities for gridded data without requiring full MCMC on the original likelihood.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same latent-Gaussian bivariate structure could be applied to other extremes such as precipitation or wind speed once suitable counterfactual runs exist.
If the credible regions remain stable under different climate-model ensembles, the method supplies a template for attribution studies that regulators could use to identify priority adaptation areas.
The Max-and-smooth approximation makes the framework computationally feasible for global grids or multi-model ensembles, opening the door to routine annual updates of extreme-event attribution products.

Load-bearing premise

That the only systematic difference between the factual and counterfactual climate-model runs is the presence of anthropogenic forcing, so that any difference in modeled return levels can be treated as a direct causal effect.

What would settle it

Finding that the estimated causal-effect maps change substantially when additional natural forcings are restored to the counterfactual runs or when the same procedure is applied to observational temperature records instead of the IPSL-CM6A output.

Figures

Figures reproduced from arXiv: 2604.23438 by Arnab Hazra, Ritik Roshan Giri.

**Figure 1.** Figure 1: Sample correlation matrix based on residuals corresponding to all seven latent view at source ↗

**Figure 2.** Figure 2: The curve corresponding to the transformation of the shape parameter, view at source ↗

**Figure 3.** Figure 3: Grid cell-wise posterior means (left) and posterior standard deviations (right) of view at source ↗

**Figure 4.** Figure 4: Grid cell-wise posterior means (left) and posterior standard deviations (right) of view at source ↗

**Figure 5.** Figure 5: Grid cell-wise posterior means (top-left) and posterior standard deviations (top view at source ↗

**Figure 6.** Figure 6: Estimated 95% credible regions C O u+ corresponding to the joint exceedance levels, u = 0.35 (left) and u = 0.65 (right). We employ a latent Gaussian model (LGM), a special class of Bayesian hierarchical model, to correctly specify the marginal return levels and the uncertainties associated with the latent variables and hyperparameters. In the data layer, we model annual temperature maxima using GEV distri… view at source ↗

read the original abstract

Climate change has become a significant global concern due to its capacity to cause substantial disruption to daily life by increasing the frequency and intensity of extreme weather events. Given the rising trend of human interventions in the climate system over recent decades, this study aims to quantify the relative contribution of anthropogenic forcing to the increasing likelihood of climate extremes, with a particular emphasis on high-temperature extremes. Our analysis focuses on annual temperature maxima from the IPSL-CM6A model in the CMIP6 experiment. We propose a novel causal inference framework that focuses on differences in return levels derived from annual temperature maxima between the factual and counterfactual worlds. While jointly modeling the annual maxima from the two worlds using a bivariate generalized extreme value distribution, we model the spatially-varying coefficients using a latent Gaussian framework. Specifically, given that the data are available over a $1^\circ \times 1^\circ$ grid, we employ the multivariate intrinsic conditional autoregressive model for the latent layer in the proposed hierarchical model, ensuring proper posterior distributions. We implement a recently developed highly-efficient approximate Bayesian inference technique, `Max-and-smooth', that uses a Laplace approximation of the likelihood and then performs Gibbs sampling based on the approximate posterior. The results include posterior estimates of the causal effect of anthropogenic forcing on high-temperature extremes, along with the trends in this effect, over the factual world. Furthermore, we estimate credible regions for a significant causal effect to facilitate hotspot detection across the mainland United States.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper fits a bivariate GEV to factual-counterfactual temperature maxima from one CMIP6 run, layers on a multivariate intrinsic CAR for spatial smoothing, and uses max-and-smooth to map credible regions of anthropogenic effect on US extremes, but the single-model design undercuts the causal claim.

read the letter

The core contribution is a hierarchical model that treats annual maxima from the factual and counterfactual IPSL-CM6A runs as a bivariate GEV, lets the location and scale parameters vary spatially through a multivariate intrinsic CAR latent field, and then applies max-and-smooth Laplace approximation plus Gibbs sampling to get posterior maps of the difference in return levels. The output includes trend estimates and credible regions that flag hotspots where the anthropogenic signal is credibly positive. That spatial credible-region step is a practical addition for regional attribution work and the inference method scales to the 1-degree grid without full MCMC. The CAR prior gives the expected spatial smoothness, and the bivariate GEV directly encodes the joint distribution of the two worlds, which is a clean way to define the causal contrast on extremes. The paper does this on real CMIP6 output and produces maps that could be compared to observational attribution studies. The main limitation is the exclusive use of IPSL-CM6A. Differences in return levels between its factual and counterfactual runs mix the forcing signal with whatever biases that single model has in simulating extremes, variability, or land-atmosphere coupling. No ensemble spread or synthetic-data recovery checks are described in the abstract, so it is unclear how much of the reported credible regions would survive if another CMIP6 model were substituted. The causal interpretation therefore rests on an untested assumption that model structural error is negligible relative to the forcing effect. This is for people already working on extreme-event attribution who need a spatial Bayesian pipeline and are willing to add multi-model checks themselves. It is not yet ready to stand alone as a general attribution result. I would send it to peer review because the modeling framework is coherent and the computational approach is worth seeing in full, but the referees will need to press on validation and the single-model exposure.

Referee Report

2 major / 2 minor

Summary. The paper develops a hierarchical model for causal attribution of anthropogenic forcing on high-temperature extremes. It jointly models annual maxima from factual and counterfactual IPSL-CM6A CMIP6 runs via a bivariate GEV distribution, places a multivariate intrinsic CAR prior on spatially varying coefficients within a latent Gaussian framework, and uses Max-and-smooth approximate Bayesian inference to obtain posterior estimates of return-level differences (the causal effect) together with credible regions for hotspot detection over the contiguous United States.

Significance. If the central attribution results are robust, the work supplies a spatially resolved, uncertainty-quantified framework for extreme-event attribution that could support regional hotspot identification. The combination of bivariate extremes modeling with latent Gaussian spatial smoothing and efficient Laplace-based inference is a methodological contribution for gridded climate data.

major comments (2)

[Data and model specification (Section 2)] The entire analysis rests on a single CMIP6 model (IPSL-CM6A). This is load-bearing for the causal claim because return-level differences between factual and counterfactual runs are interpreted as the effect of anthropogenic forcing; any structural bias in IPSL-CM6A’s simulation of extremes or its response to forcing propagates directly into the posterior credible regions. No multi-model ensemble, inter-model comparison, or sensitivity check is described that would isolate the forcing signal from model-specific error.
[Inference and results (Sections 3–4)] No synthetic-data validation or posterior predictive checks for the bivariate GEV + ICAR specification are reported. Without these, it is impossible to verify that the joint distribution and spatial dependence are adequately captured, which directly affects the reliability of the credible regions for the causal effect.

minor comments (2)

The abstract states that trends in the causal effect are estimated, but the model description does not specify whether time-varying coefficients or a separate trend term is included in the latent layer.
Notation for the causal effect parameter (difference in return levels) should be defined explicitly with reference to the GEV location/scale parameters.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments identify important aspects of model choice and validation that we address below with planned revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Data and model specification (Section 2)] The entire analysis rests on a single CMIP6 model (IPSL-CM6A). This is load-bearing for the causal claim because return-level differences between factual and counterfactual runs are interpreted as the effect of anthropogenic forcing; any structural bias in IPSL-CM6A’s simulation of extremes or its response to forcing propagates directly into the posterior credible regions. No multi-model ensemble, inter-model comparison, or sensitivity check is described that would isolate the forcing signal from model-specific error.

Authors: We acknowledge that the analysis relies on a single CMIP6 model (IPSL-CM6A). Our framework defines the causal effect as the difference in return levels between the factual and counterfactual simulations within the same model, isolating the impact of anthropogenic forcing under that model's physics and dynamics. This single-model design was selected to focus on the novel integration of bivariate GEV modeling, multivariate ICAR spatial structure, and Max-and-smooth inference for gridded extremes data. We agree that model-specific biases could influence the results and that multi-model comparisons would enhance generalizability. In the revision, we will add an expanded discussion of this limitation, including known characteristics of IPSL-CM6A temperature extremes and references to broader CMIP6 ensemble literature, while noting that a full multi-model extension lies beyond the scope of the current methodological contribution. revision: partial
Referee: [Inference and results (Sections 3–4)] No synthetic-data validation or posterior predictive checks for the bivariate GEV + ICAR specification are reported. Without these, it is impossible to verify that the joint distribution and spatial dependence are adequately captured, which directly affects the reliability of the credible regions for the causal effect.

Authors: We agree that explicit validation of the bivariate GEV and ICAR components is necessary to support the reliability of the posterior credible regions. Although preliminary internal diagnostics were conducted during model development, they were not reported in the original submission. In the revised manuscript, we will include a dedicated subsection (likely in Section 3) presenting synthetic-data experiments that recover known parameters under the bivariate GEV + ICAR structure, along with posterior predictive checks comparing observed and replicated return levels, spatial correlation patterns, and marginal tail behavior on the real data. revision: yes

Circularity Check

0 steps flagged

No circularity: causal effect defined as post-fit difference between independently modeled factual/counterfactual return levels

full rationale

The paper's central quantity is the difference in return levels obtained by fitting a bivariate GEV to annual maxima from the factual and counterfactual IPSL-CM6A runs, followed by spatial smoothing of coefficients via a multivariate ICAR latent Gaussian layer and Max-and-smooth inference. This difference is computed after model fitting rather than being algebraically forced by the likelihood or prior; the bivariate GEV and ICAR components are standard external tools applied to the two simulation worlds. No self-definitional equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided description. The attribution therefore remains an empirical contrast between two external model outputs, not a tautology internal to the paper's own derivations.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

Only abstract available; ledger entries are inferred from described components. The model relies on standard extreme-value and spatial statistics assumptions plus the validity of CMIP6 factual/counterfactual runs as proxies for causal worlds.

free parameters (2)

Spatially-varying GEV parameters and causal effect coefficients
The hierarchical model treats location, scale, and shape parameters plus the anthropogenic effect as spatially varying and estimated from data.
Hyperparameters of the multivariate intrinsic CAR prior
Precision matrix and variance components of the latent Gaussian layer are fitted or given priors.

axioms (2)

domain assumption Bivariate GEV distribution adequately models joint extremes of factual and counterfactual annual maxima
Invoked when jointly modeling the two worlds to extract return-level differences.
domain assumption Multivariate intrinsic CAR provides proper spatial dependence for the latent coefficients on 1° grid
Used to ensure posterior propriety in the hierarchical model.

pith-pipeline@v0.9.0 · 5566 in / 1503 out tokens · 47789 ms · 2026-05-08T06:51:03.185089+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 2 canonical work pages

[1]

4 Noel Cressie.Statistics for spatial data

Springer, 2005. 4 Noel Cressie.Statistics for spatial data. John Wiley & Sons, 2015. 8 Noel Cressie and Christopher K Wikle.Statistics for spatio-temporal data. John Wiley & Sons, 2011. 3 Anthony C Davison and Rapha¨ el Huser. Statistics of extremes.Annual Review of Statistics and Its Application, 2(1):203–235, 2015. 2, 3, 4 Anthony C. Davison, Simone A. ...

work page doi:10.1214/11-sts376 2005
[2]

Spatial causal inference in the presence of unmeasured confounding and interference.arXiv preprint arXiv:2303.08218, 2023

26 Georgia Papadogeorgou and Srijata Samanta. Spatial causal inference in the presence of unmeasured confounding and interference.arXiv preprint arXiv:2303.08218, 2023. 2 Ganapati P Patil. Digital governance, hotspot geoinformatics, and sustainable develop- ment: A preface.Environmental and Ecological Statistics, 17(2):133, 2010. 3 30 Sara C Pryor and Reb...

work page arXiv 2023
[3]

A hierarchical max-stable spatial model for extreme precipitation.The Annals of Applied Statistics, 6(4):1430, 2012

19 Brian J Reich and Benjamin A Shaby. A hierarchical max-stable spatial model for extreme precipitation.The Annals of Applied Statistics, 6(4):1430, 2012. 26 Brian J Reich and Benjamin A Shaby. A spatial Markov model for climate extremes. Journal of Computational and Graphical Statistics, 28(1):117–126, 2019. 3 Brian J Reich, Shu Yang, Yawen Guan, Andrew...

2012
[4]

Hierarchical modeling for extreme values observed over space and time.Environmental and Ecological Statistics, 16(3):407–426, 2009

4 Huiyan Sang and Alan E Gelfand. Hierarchical modeling for extreme values observed over space and time.Environmental and Ecological Statistics, 16(3):407–426, 2009. 3 Martin Schlather. Models for stationary max-stable random fields.Extremes, 5(1):33–44,

2009
[5]

Bivariate extreme statistics.Annals of the Institute of Statistical Mathematics, 11(2):195–210, 1960

26 Masaaki Sibuya et al. Bivariate extreme statistics.Annals of the Institute of Statistical Mathematics, 11(2):195–210, 1960. 6, 9 Alec Stephenson and Chris Ferro. evd: Functions for extreme value distributions.R package version, 2:3–3, 2018. 6 Peter A Stott, Nikolaos Christidis, Friederike EL Otto, Ying Sun, Jean-Paul Vanderlinden, Geert Jan Van Oldenbo...

1960

[1] [1]

4 Noel Cressie.Statistics for spatial data

Springer, 2005. 4 Noel Cressie.Statistics for spatial data. John Wiley & Sons, 2015. 8 Noel Cressie and Christopher K Wikle.Statistics for spatio-temporal data. John Wiley & Sons, 2011. 3 Anthony C Davison and Rapha¨ el Huser. Statistics of extremes.Annual Review of Statistics and Its Application, 2(1):203–235, 2015. 2, 3, 4 Anthony C. Davison, Simone A. ...

work page doi:10.1214/11-sts376 2005

[2] [2]

Spatial causal inference in the presence of unmeasured confounding and interference.arXiv preprint arXiv:2303.08218, 2023

26 Georgia Papadogeorgou and Srijata Samanta. Spatial causal inference in the presence of unmeasured confounding and interference.arXiv preprint arXiv:2303.08218, 2023. 2 Ganapati P Patil. Digital governance, hotspot geoinformatics, and sustainable develop- ment: A preface.Environmental and Ecological Statistics, 17(2):133, 2010. 3 30 Sara C Pryor and Reb...

work page arXiv 2023

[3] [3]

A hierarchical max-stable spatial model for extreme precipitation.The Annals of Applied Statistics, 6(4):1430, 2012

19 Brian J Reich and Benjamin A Shaby. A hierarchical max-stable spatial model for extreme precipitation.The Annals of Applied Statistics, 6(4):1430, 2012. 26 Brian J Reich and Benjamin A Shaby. A spatial Markov model for climate extremes. Journal of Computational and Graphical Statistics, 28(1):117–126, 2019. 3 Brian J Reich, Shu Yang, Yawen Guan, Andrew...

2012

[4] [4]

Hierarchical modeling for extreme values observed over space and time.Environmental and Ecological Statistics, 16(3):407–426, 2009

4 Huiyan Sang and Alan E Gelfand. Hierarchical modeling for extreme values observed over space and time.Environmental and Ecological Statistics, 16(3):407–426, 2009. 3 Martin Schlather. Models for stationary max-stable random fields.Extremes, 5(1):33–44,

2009

[5] [5]

Bivariate extreme statistics.Annals of the Institute of Statistical Mathematics, 11(2):195–210, 1960

26 Masaaki Sibuya et al. Bivariate extreme statistics.Annals of the Institute of Statistical Mathematics, 11(2):195–210, 1960. 6, 9 Alec Stephenson and Chris Ferro. evd: Functions for extreme value distributions.R package version, 2:3–3, 2018. 6 Peter A Stott, Nikolaos Christidis, Friederike EL Otto, Ying Sun, Jean-Paul Vanderlinden, Geert Jan Van Oldenbo...

1960