Regional loss estimation or catastrophe (cat) models combine hazard and infrastructure data with models of physical vulnerability to estimate natural disaster losses for a region. Their development took hold in the 1980s with the National Earthquake Hazards Reduction Act, widespread availability of GIS, and the founding of cat modeling companies such as AIR Worldwide in 1987 and Risk Management Software (RMS) in 1989(e.g., NRC 1989;King et al. 1997) (AIR Worldwide and RMS are now Verisk and Moody's Insurance Solutions, respectively). Since then, a field of research and practice has focused on expanding and improving those original models to include more hazards, more infrastructure (asset) types, more types of loss, better infrastructure data, better hazard and physical vulnerability models, and analysis with and without infrastructure design changes and retrofits. These models have been enormously useful to both the insurance industry and the public sector by addressing questions such as: How much damage and economic loss is expected in future natural hazard events? Where will it occur? What type of damage will it be? How likely is it? How would specific design changes or retrofits help reduce losses?
Nevertheless, there remains a great distance between understanding the risk and the risk reduction effects of different interventions on the one hand, and the real-world implementation of risk reduction policies on the other. This gap is in part because of the many individuals and organizations involved in the implementation of risk interventions. These stakeholder decisions depend not only on the effect of interventions, but also (1) the costs, auxiliary benefits, and non-risk-related concerns of each intervention; (2) each stakeholder's goals, alternatives, decision process, constraints, timelines, and risk perceptions; and (3) the interactions among all these factors. That is, additional questions remain, such as: What public policies and private sector interventions are likely to gain the required support and actually be implemented? How would each stakeholder 1 Professor, Dept. of Civil and Environmental Engineering, Univ. of Delaware, Newark, DE 19716 (corresponding author). ORCID: https:// orcid.org/0000-0002-6061-5985. Email: rdavidso@udel.edu fare under each policy and hazard event? How do policies and stakeholder actions interact?
To help address these questions, we present the Stakeholder-Based Tool for the Analysis of Regional Risk (STARR). STARR is a computational framework, i.e., a collection of interacting mathematical models, designed to capture stakeholder interactions and inform the creation and analysis of government policies for regional disaster risk management. It includes a few key features. First, it extends and differs from regional loss estimation models like Hazards US (HAZUS) (FEMA 2022), Interdependent Networked Community Resilience Modeling Environment (IN-CORE) (van de Lindt et al. 2023), the Florida Public Hurricane Loss Model (FPHLM) (Hamid et al. 2011), and Regional Resilience Determination (R2D) (Deierlein et al. 2020;McKenna et al. 2024) by explicitly modeling stakeholder decision-making that can affect outcomes on many different dimensions. In STARR, each stakeholder is represented with their own goals, alternatives, constraints, decision processes, and risk.
Second, the outcomes are computed separately for each stakeholder to enable evaluation of various policies from each perspective. Thus, win-win solutions can be identified in which all stakeholders are better off, leading to a higher likelihood of implementation. Third, the framework allows evaluation and comparison of multiple types of public policy interventions in a single framework (e.g., subsidized insurance, retrofit grants, and property acquisition). Different interventions have different effects on the risk as well as different costs and benefits for different stakeholders, and they can interact in surprising ways, so it is valuable to consider them all together. Fourth, STARR is flexible and modular, allowing many analysis variations to accommodate a range of assumptions, questions, and specific uses. Fifth, it is stochastic to represent the uncertainties, especially in hazard occurrence, that make disaster risk management so challenging, and it is dynamic to capture changes in the built environment and economic and societal context over time.
Finally, although it would be possible to extend the framework to consider other hazards and stakeholders (section "Future Directions"), STARR currently focuses on hurricanes and on households and housing. STARR centers on households and housing because of the critical role they play in economic prosperity and postdisaster recovery. Housing is often a household's largest asset and a primary mechanism of preserving intergenerational wealth, and it holds functional and affective meaning for residents (Schuetz 2020). Homes also provide stability for the residents who supply labor for local businesses and consume goods and services.
The STARR computational framework was designed to support three main types of uses. First, it can be used to support policymaking by facilitating development, evaluation, and comparison of possible disaster risk management policies, including descriptions of the decisions insurers and households are likely to take in response, and importantly, outcomes for each stakeholder (government, insurers, and households), which allows the identification of win-win strategies and consideration of which policies are likely to be implementable in practice.
Second, the framework facilitates understanding of the dynamic system of regional hurricane risk management, including interactions among stakeholder actions and the effects of changes in the context or assumptions. Although many previous studies have examined one or two types of decisions or losses, without a framework like STARR that captures the entire system across a region and over time, it is difficult to understand the many sometimes unintended but often consequential interactions among its components.
Third, it can guide future research, demonstrating the interrelation among research advances, building on previous research, comparing model results, and identifying lingering gaps in knowledge. Importantly, because STARR includes equal emphasis on decisionmaking and physical risk, it can provide a vehicle to facilitate tight integration of social science, physical science, and engineering research.
The STARR computational framework has been developed over two decades with new features gradually being added and the framework being applied in case studies in the eastern half of North Carolina. Table 1 summarizes those earlier versions of the framework (which was not previously named) and the key features each included. What we now refer to as STARR includes all the functionality of the previous versions. The aim of the current paper is to present STARR in a way that is accessible for readers and users, illuminates the relationships among the previous versions, clarifies the possible settings within the current version presented herein, and identifies possible future variations. We reference the previous papers for more detail about particular framework settings.
The next section provides an overview of the STARR computational framework, its intended uses, outputs, and key features. Each of the modules within STARR is described in turn, highlighting its purpose, inputs, outputs, and currently available settings. The solution method is described then, followed by discussion of the intended use of STARR and future directions.
Fig. 1 illustrates the STARR computational framework. It consists of seven interacting modules. The stakeholder modules (blue shaded boxes in Fig. 1) describe the decision-making of government agencies, insurers, and households. The context modules (green shaded boxes) describe the natural, built, and economic environments in STARR is dynamic, implemented by stepping through time year to year and incorporating a stochastic representation of the evolving hazard. Each year, the stakeholders first make their interacting decisions based in part on state variables describing the community in terms of number, location, and types of homes and households, hazard risk, and economic conditions (e.g., jobs). Hurricane events causing damage, loss, and disruption may or may not occur in each time step. We then update the context variables to reflect changes due to the decisions (e.g., homes built, homes retrofitted, and people not working) and hazard events (e.g., homes destroyed).
The specific outputs of STARR depend on the particular variations adopted in each module, but generally the final STARR results (open box in Fig. 1) include the following: (1) recommended government policy decisions (normative), (2) expected insurer and household decisions likely to be taken in response (descriptive), (3) outcomes each stakeholder is expected to experience over time, and (4) strength of the economy. All results include uncertainty in hurricane occurrence and can be disaggregated over space and/or time.
A game-theoretic construct is implemented to model the interactions of the three stakeholder groups. The stakeholder types interact in a nested dynamic Stackelberg (leader-follower) game (red boxes in Fig. 1). In the outer game, the government is the Stackelberg leader. With knowledge of how the equilibrium of the inner game will respond to policy, government actors determine what incentives to offer and/or regulations to impose on insurers and households so as to achieve their objectives (Tirole 1988). With the government policies in place, in the inner game, one or more primary insurers make best-response choices given household demand as affected by government policy and reinsurance pricing. Household demand is an aggregation of individual household choices of available actions in response to insurance market prices, government interventions, risk, and economic and community conditions. The intent of the government module is to recommend policies that improve or even optimize government's social welfare objectives.
The insurer and household modules describe how they are expected to act in response. This reflects a philosophy that it will be more effective to develop a system that recognizes the natural behavioral self-interested inclinations of insurers and households rather than try to alter their decisions and behaviors to conform to system-optimized recommendations. The context modules-hazard, building-household inventory, damage and loss, and regional economy-are similar to those in available regional loss estimation models like HAZUS, IN-CORE, the FPHLM, or R2D, but with a few particular specifications required to make them interact with the decision modules.
This module describes the joint inventory of housing units and households that live within them. The outputs are denoted X imcvjt , which includes number of buildings in location i, of type m, resistance c, value v, and inhabited by household type j in time t. More specifically, each housing unit in the STARR framework is located in an area unit i and is classified with a building type m, building resistance c, value v, and household type j. The hazard is assumed to be uniform within each area unit i. Building types m are defined based on easily observable, architectural features (e.g., number of stories or roof shape) that are not used to define structural retrofits. Building resistance c, by contrast, is based on structural details (e.g., roof-to-wall connection type and roof covering nail schedule) important for determining structural vulnerability. It could be defined as a scalar for a building-level damage model or a vector of component resistances for a component-based damage model. Structural retrofits are represented as a change from one building resistance c to another c 0 .
Household types j may be defined based on demographic attributes (e.g., age and income) for use in the Household Decisions Module and to allow disaggregation of results by household type. If the household insurance purchase decisions vary with income, for example, or if the analyst would like to disaggregate results by income class, household types should include income information. The time t index is included to capture changes in the joint buildinghousehold inventory over time due to hurricane damage, retrofits, or new construction.
The Building-Household Inventory, X imcvjt , used in STARR is similar to the residential inventory in previous regional loss estimation models (e.g., HAZUS and FPHLM) with the addition of the households of different types that live within them. The output of the Building-Household Inventory Module, X imcvjt , is used as input for the Damage and Loss, Household Decisions, and Regional Economy Modules (Fig. 1). Table 2 summarizes, for key features of the inventory, the settings that have been used to date in STARR. ASCE 04025031-3 Nat. Hazards Rev. Nat. Hazards Rev., 2025, 26(3): 04025031 ${copyrightStatement}
The Hazards Module describes the long-term probabilistic regional hurricane hazard. It outputs a set of long-term Y-year scenarios s, each with an occurrence probability P s such that exactly one scenario s occurs. Each scenario s is defined by tY time steps. In each time step, a hurricane scenario h either occurs or does not occur.
For each hurricane scenario h, a map of the hazard metrics in every location i is included.
Because the STARR framework is intended to support regional policy decisions, the analysis requires a probabilistic representation of the hazards rather than usage of a single or a few specified events. The hazard should be represented in a scenario-based approach (rather than performing independent probabilistic hazard analyses at each location) in order to capture the spatial correlation of losses, which is important in determining the variability of regional losses (Crowley and Bommer 2006). Hazards represented could include wind, coastal flooding (including storm surge, tides, and wave setup), inland flooding, wind-driven and impinging rain, wind-borne debris, hurricane-spawned tornadoes, and/or the effects of climate change.
The hazard in the STARR framework is represented as a set of hurricane events, h, each with an annual probability of occurrence P h , that are simulated to generate a collection of long-term Y-year scenarios, s. It is important to represent the hazard as a set of longterm scenarios, s, because multiple hurricanes within the same year or a series of hurricanes in quick succession over a few years can create very different outcomes for a stakeholder than the same hurricanes evenly spaced over a long time period. The t time steps per year y in each scenario s are defined so that we can reasonably assume no two hurricane events h occur in the same time period (t of year y) and the probability a hurricane event occurs in one time period is equal across time. Each scenario has an occurrence probability P s such that exactly one scenario s occurs. Finally, the hurricane impact metrics (e.g., wind speed and maximum inundation depth) for each hurricane event h are provided for each area unit i, where the hazard is assumed to be uniform within each area unit.
Table 3 summarizes, for key features of the hazard, the settings currently available in STARR.
In general, any available hazard models [e.g., weather research and forecasting (WRF) (Skamarock et al. 2008)] can be used as long as they produce the output required by the other modules and the final STARR results. The hazards included, and the metrics used to describe them, should be consistent with what the Damage and Loss Module requires.
The purpose of the Damage and Loss Module is to determine the physical damage and economic loss for each building in each hurricane scenario h. Inputs include Building inventory, X imcv , and set of hurricane scenarios h. Outputs are loss L imcvh associated with each building of type m, resistance c, and value v in location i in hurricane scenario h. More specifically, the Damage and Loss Module estimates the physical damage and economic loss for each building in the Building-Household Inventory for each hurricane scenario h under all possible resistance levels (i.e., with and without different retrofits). STARR uses the damage and loss model presented by Peng et al. (2013) and Peng (2013), a component-based model that estimates wind damage based in part on an early version of the FPHLM as described by Gurley et al. (2005) and Pinelli et al. (2004Pinelli et al. ( , 2008)), and flood damage based on results from the component-based flood damage simulation model of Taggart and van de Lindt ( 2009) and van de Lindt and Taggart (2009). The damage and loss model relates correlated probabilistic resistances of building components to wind speeds and inundation depths, considering the effects of wind pressure and missiles, building orientation relative to the wind, and the increase in internal pressure that results when the building envelope is breached.
Table 4 summarizes, for key features of the damage and loss modeling, the settings currently available in STARR. In general, any available damage and loss model that can provide similar output can be used, such as, HAZUS (FEMA 2022), IN-CORE (van de Lindt et al. 2023), and R2D (Deierlein et al. 2020;McKenna et al. 2024), as long as it produces the output required by the other modules and the final STARR results.
In particular, the Damage and Loss Module should be able to represent retrofits, preferably in a way that corresponds to a physical change (rather than an arbitrary fragility curve shift) so that the associated retrofit cost can be estimated and household decisions about implementing them can be modeled. In general, higher precision is required for loss model results in STARR than when estimating only total regional loss because the decision modules require information about the individual household economic losses.
The purpose of this module is to recommend a decision among specified government policy interventions so as to best meet specified objectives subject to specified constraints. Its inputs include (1) loss L imcvh associated with each building of type m, resistance c, and value v in location i in hurricane scenario h; (2) the set of long-term Y-year scenarios s with their associated occurrence probabilities P s ; and (3) an understanding of how households J and insurers N of each type would respond to each possible policy. It outputs two final framework outputs: (1) recommended government policy G, and (2) outcomes for government, O G .
The government is the leader in the outer Stackelberg leaderfollower game. The Government Decisions Module, therefore, determines the recommended government decisions based on the hurricane risk and an understanding of how the insurers and households are likely to act in response. Importantly, government decisions are forward-looking, based on knowledge of past hurricane events and the expectation of what hurricanes are possible in the future, with their relative likelihoods and losses. That is, as in reallife, the Government Decisions Module assumes that the past is known and future risk is known, but there is no crystal ball about what will actually happen in the future. This is a normative, not descriptive, decision module, and it is stochastic to consider the uncertainty associated with hurricane occurrence.
The policies considered in the Government Decisions Module are evaluated based on a defined government objective (e.g., minimize expected total economic losses) under given constraints (e.g., government budget). A range of policy alternatives can be evaluated through the STARR Government Decisions Module, including grant assistance for structural retrofits or property acquisition (buy-out) offers.
Table 5 summarizes, for key features of the government decisionmaking, the settings used to date in STARR. When designing the Government Decisions Module, for each policy alternative considered, the Household Decisions Module must indicate how households would respond to it, the Insurer Decisions Module must indicate how insurers would respond to it, and the Damage and Loss Module should be able to represent any effects of its implementation. For example, including a retrofit grant as a government policy choice requires that the household decision model is able to predict the likelihood that households will retrofit with and without the grant and that the loss model can predict losses with and without implementation of the retrofit.
This module determines the decisions that insurers are likely to make based on specified government decisions and with knowledge of likely household responses. Its inputs include (1) loss L imcvh associated with each building of type m, resistance c, and value v in location i in hurricane scenario h; (2) the set of long-term Y-year scenarios s with their associated occurrence probabilities P s ; (3) government decisions G; and (4) an understanding of how households J of each type would respond to each insurer decisions. The outputs are (1) predicted insurer decisions N to input into the Household Decision Module and as final framework output, and (2) outcomes for insurers O N as a final framework output.
The insurers interact with households and possibly other insurers within the insurance market, which is modeled as a Cournot-Nash inner game of the STARR framework. The Insurer Decisions Module determines the decisions the insurers are likely to make based on the hurricane risk and an understanding of how households will respond to risk-based insurance policy prices. It is a descriptive decision model. That is, rather than offer recommendations for what insurers should do, the model aims to describe how insurers will act based on an understanding of their objectives, available alternatives, constraints, and decision processes. The insurer decision model is a stochastic decision model to consider the uncertainty associated with hurricane occurrence. It also captures the cumulative effects of hurricane losses over time, as that determines the insurers' solvency and profitability. Table 6 summarizes, for key features of the insurer decisionmaking, the settings used to date in STARR. The Insurer Decisions Module can assume that there is either a single insurer or multiple carriers interacting with household demand within an insurance market. The former is simpler and forms a benchmark of the highest profitability that can accrue to the insurance industry; the latter is more typical as market concentration in the primary insurance market can lead to significant differences in insurers' operational decisions. Both are possible in STARR. As the number of insurers within the market increases, insurance premium prices converge to competitive prices with profitability typical of a normal rate-of-return that we would expect as a benchmark for regulated prices.
In the US, flood insurance is offered by the National Flood Insurance Program (NFIP); some standard homeowners' policies cover wind damage, some do not; and insurance is regulated by state agencies. The STARR Insurer Decisions Module considers flood and wind insurance separately.
This module determines decisions households are likely to make based on specified government and insurer decisions and with knowledge of the hurricane risk. Its inputs include (1) loss L imcvh associated with each building of type m, resistance c, and value v in location i in hurricane scenario h; (2) the set of long-term Y-year scenarios s with their associated occurrence probabilities P s ;
(3) household attributes X imcvjt ; (4) government decisions G; and (5) insurer decisions N. It outputs (1) predicted household decisions J to input into the Insurer Decisions Module and Government Decisions Modules and as a final framework output, and (2) outcomes for households O J as a final framework output.
More specifically, the Household Decisions Module determines the decisions households are likely to make based on the hurricane risk, government policy choices, and attributes of the household and/or house. It is a descriptive decision model that aims to capture how households will act based on an understanding of their objectives, available alternatives, constraints, and decision processes.
Rather than assume household compliance with a specified policy, this module represents decisions that involve household choice and preference, and how they can be influenced through the terms of the decision, such as cost of insurance net of subsidies, affordability, deductible offered, terms of retrofit grants, and acquisition offers.
The Household Decisions Module allows different households to make different decisions, and implicitly or explicitly considers the uncertainty associated with hurricane occurrence. Table 7 summarizes, for key features of the household decisionmaking, the settings currently available in STARR. The STARR framework allows a choice of two possible household decision models: expected utility models or discrete choice models (DCM). Utilitybased models, driven by rational choice theory, tend to be more sensitive to the terms of the decisions than the discrete choice models.
For the DCMs, we have used survey data to fit and implement (1) mixed logit models to represent insurance purchase as a function of premium, deductible, and attributes of the household and house (Wang et al. 2017); (2) mixed logit models of retrofit implementation as a function of grant terms (amount, percentage of price offered) and attributes of the household and house (Chiew et al. 2020); and (3) a pooled probit model of property acquisition offer acceptance as a function of price offered, whether the offer is made before or after damage has occurred, and attributes of the household and house (Frimpong et al. 2019). As is the case with all predictive human decision models, the utility models and DCMs are theory-driven and empirically specified. They represent a couple of the myriad ways of capturing these decisions. Both assume that each household acts in its own best interest and both types include an affordability constraint expressed as a percentage of house value.
In addition to development of the discrete choice models implemented directly in STARR, insights from related studies of household decision-making have influenced the assumptions in and interpretations of results from STARR. For example, about one-third of people have never engaged in making a protective action decision (Stock et al. 2021). It is not that they weighed pros and cons and decided not to do something; rather, they just never engaged in the decision at all. Perceived attributes of retrofit (e.g., effort required, whether it adds to the resale value, and attractiveness) influence the decision as well as cost (Zou et al. 2020). Low-interest loans and insurance premium reduction incentives for retrofit are not as influential as grants (Jasour et al. 2018).
The purpose of this module is to determine the impact of hurricane loss on the regional economy. Its inputs include (1) initial state capital stock in regional economy (i.e., value of buildings); (2) loss L imcvh associated with each building of type m, resistance c, and value v in location i in hurricane scenario h; (3) the set of long-term Y-year scenarios s with their associated occurrence probabilities P s ; (4) stakeholder decisions (households, insurers, and government); and (5) household attributes (income). The output is the (1) predicted metrics of economic output E as a final framework output.
To assess the economy-wide impacts of hurricanes, STARR uses a static computable general equilibrium (CGE) simulation model of the regional economy, which includes multiple counties networked within a county management association (CMA). The current application uses 1-year time steps for the multiyear scenarios.
The CGE model is a stylized computational representation of the circular flow of the economy. The output specifications, factor prices, firms' productivity, and household income are all estimated at a general systemwide equilibrium (Dixon and Jorgenson 2013;Wing 2011;Wing and Balistreri 2018). The CGE model in STARR divides the regional economy into county-level microregions with 28 constituent industry sectors, each of which is modeled as a representative firm characterized by a nested constant elasticity of substitution (CES) technology to produce a single good or service.
In each microregion, households are grouped into nine income classes, each of which is modeled by a representative agent with nested CES preferences and constant marginal propensity to save out of income. Federal, state, and local governments in each microregion are also represented in a simplified fashion. Their role in the circular flow of the economy involves collecting taxes from industries and passing some of the resulting revenue to the households as a lump-sum transfer, in addition to purchasing commodities to create a composite government good that is consumed by the households. Two factors of production are represented within the model, labor and capital, which is sector-specific. Both factors are owned by the representative agent and rented out to the firms in exchange for factor income. Each microregion is modeled as an open economy that engages in trade with other CMA counties, the rest of the US and the rest of the world using the Armington (1969) specification (imports from other states and the rest of the world are imperfect substitutes for goods produced in each state).
The model is specified as a square system of nonlinear equations that are numerically calibrated using region-specific social accounting matrices from IMPLAN Model (2020). These data record the flows of commodities and factors among households and sectors in the study area's CMA counties for a specified year. The model computes the prices and quantities of goods and factors of production that equalize supply and demand in all markets in the economy, subject to constraints on the external balance of payments. Direct impacts of hurricanes are modeled as sector-based reductions to the residential and commercial capital stock. This shock increases the marginal cost of capital in the affected counties and sectors, which induces producers and consumers to substitute other inputs for capital, and in turn triggers price and quantity adjustments that ripple across all markets in the study area's microregions. Table 8 summarizes the settings currently available in STARR for key features of the regional economy modeling. In general, other specifications of CGE models and parameterizations can be used, as long as they produce the output required by the other modules and the final STARR results.
The STARR dynamic modeling framework is a simulationoptimization. It is solved by, for each long-term scenario s, stepping through the timeline year by year (Fig. 2). First, solve the government optimization based on full knowledge of the annual probability that a hurricane or multiple hurricanes will occur in the future and their expected associated losses. It determines the optimal government policies (e.g., property acquisition and retrofit programs) for y ¼ 1. Second, perform the insurer optimization to calculate the equilibrium risk-based insurance prices based on homeowner demand for insurance as of year y. Third, simulate household decisions based on offered government programs and insurance prices, given the current inventory of households and buildings. Fourth, distribute the loss associated with the hurricanes in the current year of the scenario s among the stakeholders based on the up-to-date building inventory and household insurance purchase decisions.
Fifth, update the households and/or buildings in the inventory to reflect structural retrofits, hurricane damage, and any other nonhurricane-related changes (e.g., typical population growth). Sixth, update the state of the economy (e.g., county-level GDP). Seventh, step forward to the next year y þ 1 and repeat Steps 1-6 until the end of the timeline. Finally, compile the results. Guo et al. (2022) described the models and solution algorithm in greater detail.
STARR may be implemented for a region for practical purposes, e.g., to identify recommended government policies, or for research purposes, e.g., to better understand the interactions among the many stakeholders and context factors within the regional disaster risk management system it represents. Although the specific outputs depend on the particular settings chosen for each module (Tables 2345678), the outputs generally include, for each stakeholder, the (1) decisions, and (2) outcomes (Table 9) each expressed over the Y-year duration of the analysis and including uncertainty in hurricane occurrence.
Results can be disaggregated spatially, temporally, by household type, and in other ways. Examining the many disaggregated results can help address questions like: How would each stakeholder fare under each policy and event? Based on that, what policies are actually likely to be implemented? How equitable are different policy solutions? How do different policies and stakeholder actions interact? Are there win-win policies?
As in all modeling, judgments are required to determine what to capture explicitly and how. The aim is to keep the models as simple as possible while ensuring they are meaningful and can answer the questions posed. Thus, to determine which settings to use for a STARR application (i.e., choices in Tables 2345678), the following questions should first be answered based on the intended uses of the analysis:
• Which hazards should be considered (e.g., hurricane wind, coastal flooding, or inland flooding)? • Which interventions should be considered (e.g., property acquisition, retrofit, or insurance)? • What outputs are desired (e.g., household expenditures, probability of insurer insolvency, and disaggregated or not)? • What data and models are available (e.g., What is the geographic and temporal resolution of inventory data? Is there a discrete choice model that captures household decisions related to the intervention that is being studied)? • What key assumptions should be made (e.g., insurer surplus is capped or not)? • What computational capacity is available?
These answers will guide the setting decisions. It is important to realize, however, that although each module is described independently herein, they work together to serve the STARR framework's overall goals and therefore are intertwined in a number of obvious and more subtle ways. In particular, there are inputs and outputs For each policy intervention included in the Government Decisions Module (e.g., offering a retrofit grant), the Insurer Decisions Module and Household Decisions Module must be able to capture how the insurers and households, respectively, will respond to them (i.e., how their decision will differ depending on what the government does), and the Damage and Loss Module must represent its effect on the magnitude, nature, and/or distribution of losses. Finally, the combined numbers of modeled decisions, hurricane scenarios, modeled years, and inventory resolution (e.g., definitions of locations i and building types m) drive the computational demand of the overall framework. Therefore, each decision of the STARR architecture should not only consider the effects on all modules in STARR but should also consider the computing resources required.
STARR was developed for hurricane risk in the US, considering government, households, and primary insurers as the key stakeholders. However, like with the original regional loss estimation models, there are many opportunities for extension and improvement. Some are relatively easy to implement, some are not. Improvements that could be made within the current version of STARR (with the currently available settings) include refitting household decision discrete choice models with additional data to improve their generalizability, enhancing the description of the building inventory, and improving the predictive power of the damage and loss model. There are also many more variations that could be developed and tested within the seven modules depicted in Fig. 1. The Appendix lists some example possibilities.
More demanding extensions to STARR include expanding the framework to consider additional hazards (e.g., climate change effects and earthquakes), policies (e.g., land-use policies that apply to new rather than existing construction), stakeholder types (e.g., mortgage lenders and commercial building owners), loss metrics (e.g., time a house is uninhabitable), and/or sources of uncertainty (e.g., trajectories of evolution of building/economy). Additional applications of STARR to more geographic regions would also be valuable to examine the extent to which findings from STARR results are generalizable across regions. Like the regional loss estimation models, STARR could provide a structure for decades of future research.
Community resilience has become a guiding ideal for addressing escalating impacts of natural hazard events in the US. Nevertheless, it remains challenging to achieve in practice. We hypothesize that increasing practical execution of risk management efforts requires not only better estimates of risk and the effectiveness of new interventions, but also a broader framing of the problem. With that in mind, we introduce STARR, which aims to help practitioners design and evaluate policies that align with stakeholders' natural decisionmaking processes, result in improved positions for all stakeholders, and thus are more likely to be implemented. STARR provides an improved understanding of the complex, dynamic system by which natural disaster risk is managed, including the many interactions among various stakeholders and multiple types of public-policy and private-sector interventions (e.g., insurance, retrofit grants, and property acquisition) over many years.
For one example of an area in which STARR can provide useful insights, consider the current crisis in homeowners insurance (Flitter and Flavelle 2024;Flitter 2023). A viable insurance market-a critical determinant of postdisaster recovery-requires both homeowner demand for insurance products at an affordable price and an insurance industry that can profitably supply those products and remain solvent. The STARR framework allows us to evaluate the impact of policy on both the homeowners and the insurers to find sustainable solutions.
As a framework rather than a single model, STARR offers multiple modeling options depending on the specific questions being examined, desired assumptions, and available data and computational capacity. Like the early regional loss estimation (cat) models, it also offers opportunities for improvement of existing modules and expansion of the scope. By defining the framework, however, our hope is that future work related to the different modules (e.g., household mitigation decision-making research) can be seamlessly integrated within the larger system and thus supports exploration of the broad, system-level disaster risk management questions.
Nat. Hazards Rev., 2025, 26(3): 04025031 ${copyrightStatement}