MRIA: Multi-Regional Impact Assessment

• MRIA is a family of analytic frameworks that assess differences in regional impacts by explicitly modeling inputs, outputs, and uncertainty.
• It integrates methods like input–output analysis, linear programming, and network formulations to allocate economic, environmental, and clinical impacts across regions.
• MRIA delivers actionable insights for disaster management, climate forcing, and clinical trials by capturing regional heterogeneity and propagating uncertainty.

Multi-Regional Impact Assessment (MRIA) refers to analytic frameworks that evaluate how impacts vary across multiple regions and how those impacts are linked through common structures such as supply chains, hazard–exposure–vulnerability–loss pipelines, regional covariate distributions, or multimodal evidence streams. In recent literature, MRIA appears both as an explicit model class—for example, a multi-sector, multi-regional, supply-and-use based macroeconomic model solved by linear optimization for disasters—and as a broader methodological principle for environmentally extended input–output analysis, regional climate forcing, multi-regional clinical trials, and ZIP-code-level disaster intelligence (Perumal et al., 1 Aug 2025, Bademi et al., 20 Aug 2025, Zhang et al., 16 May 2026, Xiao et al., 30 Jan 2026).

1. Domain scope and regional units

Published MRIA applications use different regional partitions, sectoral resolutions, and impact objects. In disaster macroeconomics, the regional system can be 12 NUTS-2 regions of the Netherlands plus a rest-of-world region, or 163 countries/regions and 65 sectors in a GTAP-based global model. In environmentally extended multiregional input–output analysis, the regional system can be 52 regions consisting of 50 U.S. states, the District of Columbia, and Puerto Rico. In clinical trials, region is a pre-defined grouping variable such as Asia vs non-Asia or Japan vs rest-of-world. In crisis sensing, the operative regional units can be USPS ZIP code polygons. In climate applications, regions may be the CORDEX South Asia domain at 50 km resolution or a global 2.5°×2.5° grid. In regional risk assessment generalized to MRIA, the regional object can be the joint exposure-class assignment over all assets (Perumal et al., 1 Aug 2025, Mishra et al., 27 Apr 2026, Bademi et al., 20 Aug 2025, Zhang et al., 16 May 2026, Xiao et al., 30 Jan 2026, Sanjay et al., 2020, Estrada et al., 2021, Wu et al., 8 May 2026).

Application area	Regional unit	Primary mechanism
Disaster macroeconomics	NUTS-2 regions plus rest of world	three-step linear programming on multi-regional supply and use tables
Environmentally extended MRIO	52 U.S. regions	Leontief inverse and state-level GHG satellite account
Multi-regional clinical trials	Pre-defined regional grouping variable	regional treatment-effect heterogeneity and dynamic borrowing
Climate forcing	CORDEX South Asia / 2.5°×2.5° grid	downscaled ensembles, pattern scaling, and exceedance probabilities
Rapid disaster intelligence	207 ZIP code polygons	multimodal retrieval-augmented generation

Across these literatures, the common feature is explicit regionalization of both inputs and outputs. What changes is the object being propagated: final demand shocks, greenhouse-gas emissions, probabilistic exposure classes, treatment effects, or evidence about peak flood extent and structural damage. This suggests that MRIA is best understood as a family of region-explicit assessment strategies rather than a single fixed algorithm.

2. Input–output, optimization, and network formulations

One major lineage of MRIA is multiregional input–output analysis. In the U.S. offshore wind study, a multiregional input–output table is constructed with 52 regions and 101 sectors, and total output is obtained from the Leontief system

$\Delta x = (I-A)^{-1}\Delta y,$

with environmental impacts computed by

$\Delta E = ef * \Delta x.$

This formulation is used to distinguish in-state effects from out-of-state spillover effects, and to allocate both economic throughput and production-based greenhouse-gas emissions across states and sectors. In that application, five planned projects require \$16.3 billion in capital investment and generate \$27.6 billion in direct and indirect economic impacts, while construction-phase emissions are assessed through a new multiregional greenhouse-gas dataset (Bademi et al., 20 Aug 2025).

A second formulation makes MRIA explicitly supply-constrained. The updated disaster model is a multi-sector, multi-regional, supply-and-use based macroeconomic model solved by linear optimization. Production capacity is bounded by

$X_{r,s} \le \theta_{r,s} X_{r,s}^0,$

and disaster-induced interregional trade is bounded by

$t_{r',r,p} \le \phi_{r',r,p} T_{r',r,p}^0.$

The solution is organized as three sequential linear programs: minimizing rationing, then minimizing production and disaster trade given minimal rationing, and then computing the production equivalent of rationing. The total economic impact combines efficient production losses, inefficient production, and the production equivalent of rationed products. Within that framework, enhancing production capacity alone is inadequate; regional trade flexibility must also be improved to mitigate disaster impacts (Perumal et al., 1 Aug 2025).

Network formulations add a topological interpretation to MRIO-based MRIA. The Chinese MRIOT study represents the economy as a weighted and directed network whose nodes are province–sector combinations and whose edges are monetary transactions. It reports that inter-provincial trade increased from 9.6 trillion CNY in 2007 to 12.8 trillion CNY in 2012, while intra-provincial trade increased from 35.5 trillion to 68.4 trillion, and it analyzes weighted assortativity, clustering, modularity, and weighted PageRank to identify influential province-sectors and geographically structured communities. The results indicate increasing regional fragmentation and stronger within-province cohesion, which is directly relevant when MRIA is used to infer where shocks are likely to remain localized and where they may propagate broadly (Wang et al., 2021).

A closely related global application analyzes gas disruptions in Qatar through an MRIO model plus linear programming. There, a localized disruption to QAT-GAS reduces capacities on the corresponding export edges, and adaptive scenarios allow trade reallocation and production expansion among top gas producers. The study finds that the largest aggregate impacts are observed in India, China, and South Korea, and that trade reallocation partially mitigates losses while production expansion improves supply conditions unevenly. The methodological point is that network structure governs both vulnerability and recovery potential (Mishra et al., 27 Apr 2026).

3. Uncertainty decomposition and climate forcing

Another important strand of MRIA treats regional impacts as outputs of uncertain hazard–exposure–vulnerability–loss chains. In the exposure-uncertainty framework, the core observation is that exposure characterization is not deterministic: missing attributes are imputed probabilistically, creating a random exposure class $\mathcal{E}$. The central decomposition is the law of total variance,

$$\mathrm{Var}(Y) = \mathbb{E}\!\left[\mathrm{Var}(Y\mid \mathcal{E})\right] + \mathrm{Var}\!\left(\mathbb{E}[Y\mid \mathcal{E}]\right),$$

where the first term is baseline variance and the second is exposure information variance. At the regional level, for

$RL = \sum_{b=1}^{B} L_b,$

the same decomposition separates uncertainty from incomplete exposure information from uncertainty due to hazard, fragility, and loss characterization. The framework uses calibrated class probabilities, temperature scaling for probabilistic classifiers, and Monte Carlo simulation to estimate total variance, baseline variance, exposure information variance, and bias. Although implemented for seismic bridge risk, the paper states that the concepts and methods are directly applicable to MRIA across hazards, regions, and infrastructure types (Wu et al., 8 May 2026).

Climate-focused MRIA uses regionally explicit forcing ensembles rather than supply-chain propagation. The CCCR–IITM regional climate datasets provide dynamically downscaled projections over the CORDEX South Asia domain at 50 km resolution, with daily, monthly, and seasonal outputs, historical simulations, and future projections under RCP2.6, RCP4.5, and RCP8.5. The ensemble is described as useful for impact assessment studies and for quantifying uncertainties in regional projections, and comparability with other CORDEX regions makes it naturally suited for MRIA that needs globally consistent regional climate inputs. For annual mean temperature over India, the Reliability Ensemble Average yields, for example, $4.19 \pm 0.46^\circ\text{C}$ warming by the 2080s under RCP8.5 (Sanjay et al., 2020).

AIRCC-Clim provides a computationally lighter climate layer for MRIA. It emulates 37 CMIP5 AOGCMs, produces monthly and annual temperature and precipitation on a 2.5°×2.5° grid, and computes risk measures such as exceedance probabilities and times of threshold exceedance. Its regional module is based on pattern scaling,

$$P(i,j,t,E,y,s) = T(t,E)\,p(i,j,y,s) + \varepsilon(i,j,t,s),$$

and its risk layer evaluates probabilities such as $\Delta E = ef * \Delta x.$0 or $\Delta E = ef * \Delta x.$1 exceeds a user-defined threshold$\Delta E = ef * \Delta x.$2 This makes it possible to propagate emissions-scenario uncertainty, climate-sensitivity uncertainty, and model-structure uncertainty into regional impact models (Estrada et al., 2021).

4. Clinical-trial MRIA and regional treatment effects

In multi-regional clinical trials, MRIA centers on regional treatment-effect heterogeneity rather than material flows. One structured approach formulates four questions: whether regional heterogeneity exists; which baseline covariates are unevenly distributed across regions; which baseline covariates modify the treatment effect; and how treatment effects vary along key effect modifiers across regions. The implementation combines individualized treatment-effect pseudo-outcomes, permutation-based global tests, and conditional random forests for multivariate variable importance. Region is treated as a categorical grouping variable, but the framework distinguishes region-associated effect modifiers $\Delta E = ef * \Delta x.$3 from non-modifying composition variables $\Delta E = ef * \Delta x.$4, balanced modifiers $\Delta E = ef * \Delta x.$5, and unobserved factors $\Delta E = ef * \Delta x.$6. The intended use is exploratory and regulatory: MRCTs are powered for overall efficacy, not for confirmatory evidence on regional differences (Zhang et al., 16 May 2026).

A more formal treatment for time-to-event outcomes uses restricted mean survival time (RMST) as the treatment-effect estimand and standardizes regional comparisons to a common target covariate distribution $\Delta E = ef * \Delta x.$7. Region-specific effects are written as

$\Delta E = ef * \Delta x.$8

with calibration weighting or inverse probability of sampling weighting used to eliminate exogenous imbalance in inessential traits while preserving intrinsic regional differences. The paper develops weighted Kaplan–Meier, G-formula, Hajek, and augmented estimators, plus a Wald test for $\Delta E = ef * \Delta x.$9. In the PLATO case study, naive analysis suggested a US vs non-US discrepancy, whereas weighted analyses removed evidence of heterogeneity and yielded a global RMST difference favoring ticagrelor (Hua et al., 2024).

Bayesian MRIA in this setting uses dynamic borrowing. The robust MAP approach models the informative prior as a mixture of exchangeable regional–global effects and a weakly informative component, and the paper develops a closed-form approximation to the posterior by expressing the MAP prior as a finite mixture of power priors. This yields a posterior distribution over effective sample size, so borrowing can be interpreted in terms of how many external patients are effectively incorporated into the regional analysis. The stated advantage over a robust power prior is conceptual as well as computational: MAP averages over a continuum of heterogeneity levels instead of fixing a single discount level, while the closed-form approximation makes prior calibration for desired operating characteristics computationally feasible (Zhang et al., 5 Jan 2026).

5. Multimodal and ZIP-level rapid impact assessment

MRIA can also be implemented as regional evidence synthesis in real time. CrisiSense-RAG operationalizes this idea for Hurricane Harvey by issuing 207 ZIP-code queries over Greater Houston and estimating two regional metrics: peak flood extent as percentage of ZIP area inundated and structural damage severity as mean property damage extent per building scaled to 0–100. The system integrates NOAA aerial imagery, filtered social media, 311 emergency calls, precipitation gauges, and FEMA prior knowledge. Text retrieval uses hybrid dense–sparse retrieval with FAISS, BM25, reciprocal rank fusion, and a cross-encoder reranker; imagery retrieval uses CLIP-based retrieval and geospatial filtering. A split-pipeline architecture then separates a Text Analyst from a Visual Analyst, and asynchronous fusion logic gives priority to real-time social evidence for peak flood extent while treating imagery as persistent evidence of structural damage (Xiao et al., 30 Jan 2026).

The system is explicitly designed to handle temporal asynchrony: post-event imagery often reflects flood recession rather than maximum extent. Flood extent therefore follows a rule that can ignore recession imagery when the visual stream conflicts with documented peak conditions, whereas damage severity uses the maximum of text- and image-derived severity because structural damage is more persistent. On Hurricane Harvey, the framework achieves a flood extent MAE of 10.94% to 28.40% and a damage severity MAE of 16.47% to 21.65% in zero-shot settings. Prompt-level metric grounding is reported as critical: semantic alignment improves damage estimates by up to 4.75 percentage points (Xiao et al., 30 Jan 2026).

6. Assumptions, limitations, and recurring issues

Across its economic variants, MRIA commonly inherits fixed-coefficient and static-structure assumptions. The disaster optimization model is static or quasi-static, omits explicit inventories, and represents logistics through aggregate trade bounds rather than detailed infrastructure networks. The offshore-wind EE-MRIO assumes fixed technical coefficients, no capacity constraints, homogeneous outputs within sectors, and no induced household-consumption effects; it also assumes future domestic turbine manufacturing, excludes end-of-life impacts, and does not provide a formal uncertainty analysis. In the Chinese MRIOT network study, inter-provincial flows are estimated via a gravity model, so cross-province propagation is only as reliable as those estimates. The Qatar gas analysis is likewise static, contains no price adjustments or endogenous behavioral responses, and can admit multiple feasible LP solutions (Perumal et al., 1 Aug 2025, Bademi et al., 20 Aug 2025, Wang et al., 2021, Mishra et al., 27 Apr 2026).

Uncertainty-oriented MRIA introduces its own assumptions. The exposure-uncertainty framework discretizes continuous attributes into exposure classes, assumes hazard independence from exposure realizations, and aggregates regional loss as a sum of asset-level losses, so nonlinear network effects require additional modeling. AIRCC-Clim provides only monthly and annual temperature and precipitation at 2.5°×2.5° resolution and is not designed for extremes or variables dominated by high-frequency internal variability. The South Asia CORDEX analysis explicitly notes that precipitation projections remain highly uncertain and that downscaled seasonal averages are not always improved relative to GCMs (Wu et al., 8 May 2026, Estrada et al., 2021, Sanjay et al., 2020).

Clinical and crisis-sensing MRIA face different constraints. MRCTs are typically not designed to provide confirmatory evidence on regional differences, so regional heterogeneity analyses remain exploratory and vulnerable to small regional sample sizes; even sophisticated workflows are associational unless supported by stronger mechanistic evidence. In the ZIP-level RAG system, uneven social-media density, geolocation uncertainty, temporally misaligned imagery, incomplete ground-truth coverage, and the absence of calibrated uncertainty quantification all create region-specific bias risks. These limitations do not negate MRIA’s usefulness, but they show that regionalization improves inference only when the mechanisms generating heterogeneity are modeled explicitly and audited carefully (Zhang et al., 16 May 2026, Xiao et al., 30 Jan 2026).

Taken together, the literature presents MRIA as a class of region-explicit assessment frameworks that link local observations to multi-regional consequences. Whether the underlying machinery is a Leontief inverse, a linear program, a law-of-total-variance decomposition, a weighted RMST estimator, a robust MAP prior, or a multimodal RAG pipeline, the central task remains the same: to quantify how impacts differ across regions, how those differences are generated, and how uncertainty about regional structure should be propagated into substantive conclusions.