A Bayesian Approach for Earthquake Impact Modelling
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "A Bayesian Approach for Earthquake Impact Modelling".
Jane: Here's the gist of the paper. A Bayesian approach to earthquake impact modeling is developed using a generalized Bayesian framework and a scoring-rule-based posterior to generate estimates across mortality, population displacement, and building damage The proposed method achieves results comparable to two leading impact estimation tools in the prediction of total mortality when tested on a set of held-out past events Model Components Disaster impact models typically consist of three main components: the hazard, the vulnerability, and the exposure The authors work with the modified Mercalli intensity (MMI), which is available for earthquake events globally via ‘ShakeMaps’ produced by the United States Geological Survey (USGS) The vulnerability is expressed as a linear regression on various covariates to express the vulnerability, such as standardized terms like Vs30, PopDens, SHDI, GNIc, EQFreq, FirstHaz, Night and (FirstHaz × Night) A latent variable D is constructed by combining the hazard and vulnerability components to obtain a latent variable DImpact for each impact type The resulting impact probabilities are then derived using a normal cumulative density function (CDF) Likelihood-Free Inference Explicit calculation of the likelihood p(y1:N θ) is not practical for this model The authors employ a likelihood-free approach where a given set of parameters is evaluated by the similarity between the observed data and a sample from the model under this parameterisation They use a scoring-rule based approach to generalized Bayesian likelihood-free inference, adapting an idea from Pacchiardi et al.
2024: The loss function is defined by setting the loss function to zero if the average score falls below a threshold δ, otherwise it is infinite The energy score is used as a scoring rule, which accommodates multivariate data and reduces to the absolute deviation in the univariate case A Sequential Monte Carlo (SMC) approximate Bayesian computation (ABC) algorithm is used, which provides benefits such as a larger scope for parallelisation Data Handling and Prior Elicitation The model is fitted using a dataset of 167 real earthquake events spanning from 2011 to 2023, covering 76 countries Data for exposure, hazard, and vulnerability are subject to numerous sources of uncertainty To address data scarcity, a single global model is trained and local vulnerability differences are captured via covariates like gross sub-national income per capita and earthquake frequency Prior distributions are constructed in two stages, starting with noninformative priors constrained by physical interpretation The higher-level prior tests whether impact probabilities fall within reasonable bounds, assigning a prior probability of zero to parameterisations that result in any probabilities outside the bounds Validation and Performance The model is validated using simulated earthquake events and also tested on a set of past earthquake events The model performs relatively well for mortality and displacement, with 80% of the 90% credible intervals containing the observed value Comparing to two leading impact estimation tools, GDACS and PAGER, ODDRIN finds improved performance in the prediction of total mortality over a set of held-out events relative to GDACS and comparable performance to PAGER Performance for building damage was weak, but it appears relatively strong and well-calibrated for mortality and displacement ODDRIN presents a strong foundation for an open-source, statistically validated disaster impact estimation tool The methodology is flexible and could be applied to a range of other hazards such as heatwaves, floods, and tropical cyclones The model currently assumes independence of the error terms between grid cells, which may need to be addressed via the introduction of correlation between the grid-level error terms The use of more impact data, for example using the newly developed Montandon - Global Crisis Data Bank database, may provide better quality displacement and building damage impact data The underlying model is flexible, and could be extended or refined in numerous directions References
Title and authors: Tom: Let's quickly go over the title and who wrote this, so we know what we’re digging into here.
Jane: The paper is titled "A Bayesian Approach for Earthquake Impact Modelling" and it comes from Max Anderson Loake, Hamish Patten, and David Steinsaltz.
Lu: The authors are from Oxford and the UN Food and Agriculture Organization in Italy, which suggests a mix of statistical theory and real-world disaster data.
Meng: It’s interesting that they’re bringing together this generalized Bayesian approach with a likelihood-free framework for inference, which is often needed when you don't have a simple formula for what the data should look like.
Tom: Right, so it starts by saying they develop a novel impact estimation tool to generate earthquake impact estimates across mortality, population displacement, and building damage.
Jane: That’s the core idea: they use this Bayesian method to figure out those three different outcomes after an earthquake.
The paper's summary: Jane: Now, looking at the summary section, what they’re actually doing is modeling disaster impact using hazard, vulnerability, and exposure.
Tom: That’s the standard structure for these kinds of models, right? Hazard is how likely an earthquake is to happen; vulnerability is how weak a place or person is; and exposure tells you what's in that area.
Lu: They use the modified Mercalli intensity or MMI for their hazard component, which they get from ShakeMaps produced by the USGS globally.
Jane: And for vulnerability, they express it as a linear regression using things like Vs30—the average shear wave velocity—population density, SHDI, gross national income per capita, and earthquake frequency.
Tom: So they’re taking these specific physical and economic terms as inputs to describe how vulnerable an area is.
Lu: They then construct this latent variable D Impact for each impact type by combining the hazard and vulnerability components. This latent variable D Impact reflects an increasing function of damage if the exposure was held constant, which is what they mean by that in their symbol definition.
Jane: And from that latent variable, they derive the actual impact probabilities using a normal cumulative density function or CDF.
The paper's improvements: Tom: So, moving onto the improvements section of this paper, it mentions a few key things they did to make it better than existing methods.
Jane: One big improvement is their use of a likelihood-free approach where they evaluate parameter sets by looking at how similar the observed data is to samples from their model under those parameters.
Lu: They specifically adapt an idea from Pacchiardi et al. in a scoring-rule based approach for generalized Bayesian likelihood-free inference, and they use sequential Monte Carlo ABC, or ABC-SMC.
Tom: That sounds like a way to handle the lack of an explicit likelihood function by using a scoring rule instead of calculating the exact probability density function directly.
Meng: From my side, I’m interested in that part about ABC-SMC because it suggests they can get benefits like a larger scope for parallelisation when running these kinds of simulations.
Jane: They also mention they constructed prior distributions in two stages to handle uncertainty better. The first stage uses noninformative priors constrained by physical interpretation, and the higher level prior tests whether the impact probabilities fall within reasonable bounds, assigning a prior probability of zero to parameterizations that result in probabilities outside those bounds.
Conclusion: Tom: So wrapping up with the conclusion of "A Bayesian Approach for Earthquake Impact Modelling," they say this model performs relatively well for mortality and displacement, with eighty percent of the ninety percent credible intervals containing the observed value.
Jane: They also note that when compared to two leading impact estimation tools, GDACS and PAGER, ODDRIN finds improved performance in the prediction of total mortality over a set of held-out events relative to those two.
Lu: Regarding building damage, they admit that the performance for building damage was weak, but they suggest it appears relatively strong and well-calibrated for mortality and displacement.
Meng: If you’re thinking about deploying this at scale, I have to ask about running the model—how much computational overhead is there with ABC-SMC compared to a simpler method?
Jane: That’s a good question for the deployment side. Overall, they conclude that ODDRIN presents a strong foundation for an open-source, statistically validated disaster impact estimation tool.
Tom: They also mention it’s flexible enough to be applied to other hazards like heatwaves or floods, and they flag that they need to address the assumption of independence between error terms between grid cells, which might require introducing correlation between those error terms.
Lu: And on the future work front, they suggest that using more impact data, like something from the Montandon - Global Crisis Data Bank database, could provide better quality displacement and building damage impact data.
Jane: So to summarize this paper, it’s a flexible Bayesian framework that uses a scoring rule-based method to estimate mortality and displacement well, with some promise for future refinement.
Meng: Yeah, it’s promising because it puts the uncertainty quantification into a very formal statistical structure instead of just being an empirical curve fit.
Tom: Right, so they’ve given us a solid tool based on this Bayesian approach to earthquake impact modeling in "A Bayesian Approach for Earthquake Impact Modelling".
Max Anderson Loake, Hamish Patten, David Steinsaltz
Department of Statistics, University of Oxford · Statistics Division, United Nations Food and Agriculture Organization
stat.AP
Submitted: 2024-12-20
Updated: 2025-01-14
Comments: 24 pages, 21 figures and 5 tables
Code: https://github.com/hamishwp/ODDRIN
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 93/100
The gist: A Bayesian approach to earthquake impact modeling is developed using a generalized Bayesian framework and a scoring-rule-based posterior to generate estimates across mortality…
Key concepts
- Hazard
- The hazard component models the likelihood of an earthquake occurring in a specific area during a given time period. This is often represented by modified Mercalli intensity (MMI) data from sources like ShakeMaps, which describes the expected shaking intensity.
- Vulnerability
- Vulnerability describes how susceptible a location or structure is to harm. It is calculated using linear regression on various factors such as soil conditions (Vs30), population density, economic indicators (GNIc), and earthquake frequency.
- Likelihood-Free Inference
- Since calculating the exact probability of observed data given model parameters is too difficult, this method uses a scoring rule. It compares the observed data to samples generated by the model under different parameter settings to approximate which settings are most likely correct.
- Sequential Monte Carlo (SMC) ABC
- This is a computational technique used for likelihood-free inference. It runs simulations sequentially, allowing for efficient sampling of complex models. This helps in estimating impact probabilities and handles multivariate data effectively.
Terminology
Summary
A Bayesian approach to earthquake impact modeling is developed using a generalized Bayesian framework and a scoring-rule-based posterior to generate estimates across mortality, population displacement, and building damage The proposed method achieves results comparable to two leading impact estimation tools in the prediction of total mortality when tested on a set of held-out past events
Model Components
Disaster impact models typically consist of three main components: the hazard, the vulnerability, and the exposure The authors work with the modified Mercalli intensity (MMI), which is available for earthquake events globally via ‘ShakeMaps’ produced by the United States Geological Survey (USGS) The vulnerability is expressed as a linear regression on various covariates to express the vulnerability, such as standardized terms like Vs30, PopDens, SHDI, GNIc, EQFreq, FirstHaz, Night and (FirstHaz × Night) A latent variable D is constructed by combining the hazard and vulnerability components to obtain a latent variable DImpact for each impact type The resulting impact probabilities are then derived using a normal cumulative density function (CDF)
Likelihood-Free Inference
Explicit calculation of the likelihood p(y1:N θ) is not practical for this model The authors employ a likelihood-free approach where a given set of parameters is evaluated by the similarity between the observed data and a sample from the model under this parameterisation They use a scoring-rule based approach to generalized Bayesian likelihood-free inference, adapting an idea from Pacchiardi et al. [2024] The loss function is defined by setting the loss function to zero if the average score falls below a threshold δ, otherwise it is infinite The energy score is used as a scoring rule, which accommodates multivariate data and reduces to the absolute deviation in the univariate case A Sequential Monte Carlo (SMC) approximate Bayesian computation (ABC) algorithm is used, which provides benefits such as a larger scope for parallelisation
Data Handling and Prior Elicitation
The model is fitted using a dataset of 167 real earthquake events spanning from 2011 to 2023, covering 76 countries Data for exposure, hazard, and vulnerability are subject to numerous sources of uncertainty To address data scarcity, a single global model is trained and local vulnerability differences are captured via covariates like gross sub-national income per capita and earthquake frequency Prior distributions are constructed in two stages, starting with noninformative priors constrained by physical interpretation The higher-level prior tests whether impact probabilities fall within reasonable bounds, assigning a prior probability of zero to parameterisations that result in any probabilities outside the bounds
Validation and Performance
The model is validated using simulated earthquake events and also tested on a set of past earthquake events The model performs relatively well for mortality and displacement, with 80% of the 90% credible intervals containing the observed value Comparing to two leading impact estimation tools, GDACS and PAGER, ODDRIN finds improved performance in the prediction of total mortality over a set of held-out events relative to GDACS and comparable performance to PAGER Performance for building damage was weak, but it appears relatively strong and well-calibrated for mortality and displacement
Conclusion
ODDRIN presents a strong foundation for an open-source, statistically validated disaster impact estimation tool The methodology is flexible and could be applied to a range of other hazards such as heatwaves, floods, and tropical cyclones The model currently assumes independence of the error terms between grid cells, which may need to be addressed via the introduction of correlation between the grid-level error terms The use of more impact data, for example using the newly developed Montandon - Global Crisis Data Bank database, may provide better quality displacement and building damage impact data The underlying model is flexible, and could be extended or refined in numerous directions
References
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive likelihood-free inference using ABC-SMC allows for updating predictions as impact observations become available by employing a joint model between impact types to produce conditional distributions.
-
The system can generate gridded spatial maps of estimated impact and provide Bayesian quantification and interpretation of uncertainty, which is an advantage over existing empirical approaches that struggle with spatial mapping.
-
The model can handle multi-shock earthquake events directly, unlike some prior analytical models, and the joint model allows predictions to be updated as new impact observations become available.
-
The tool can produce a gridded spatial map of the estimated impact, which has previously not been available for empirical models, enabling direct handling of spatial extent across disaster zones.
-
The system provides robust uncertainty quantification through a general Bayesian framework, allowing users to answer questions within a framework that is
more intuitive and well-suited for decision-making,
such asthe probability that an event’s mortality exceeds a certain threshold.
Abstract
Immediately following a disaster event, such as an earthquake, estimates of the damage extent play a key role in informing the coordination of response and recovery efforts. We develop a novel impact estimation tool that leverages a generalised Bayesian approach to generate earthquake impact estimates across three impact types: mortality, population displacement, and building damage. Inference is performed within a likelihood-free framework, and a scoring-rule-based posterior avoids information loss from non-sufficient summary statistics. We propose an adaptation of existing scoring-rule-based loss functions that accommodates the use of an approximate Bayesian computation sequential Monte Carlo (ABC-SMC) framework. The fitted model achieves results comparable to those of two leading impact estimation tools in the prediction of total mortality when tested on a set of held-out past events. The proposed method provides four advantages over existing empirical approaches: modelling produces a gridded spatial map of the estimated impact, predictions benefit from the Bayesian quantification and interpretation of uncertainty, there is direct handling of multi-shock earthquake events, and the use of a joint model between impact types allows predictions to be updated as impact observations become available.
Sources
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data