The Effect of Natural/Human-Made Hazards on Business Establishments and their Supply Chains

Abstract

This paper examines the impact of natural and human-made hazards on payroll, GDP, employment, and establishment survival/creation in the year of hazard occurrence in the U.S. economy and more specifically in the U.S. manufacturing/goods producing industry. Many of the papers that examine economic impacts of hazards consider upstream impacts of supply chain disruption. Measures of downstream impacts are often limited, particularly in measuring the short-term impacts. This paper examines how manufacturers and other establishments are impacted, at the industry and total economy level, by a disruption in supplies of goods with low substitutability, which is often referred to as the ripple effect. In this paper, eight models are developed to explore supply chain vulnerability, at the industry-level, to hazard events across geographic areas of the U.S. during the 2005 to 2016 time period. The most severe impacts are due to hazards in the manufacturing/goods industry supply chain, where payroll, GDP, and employment declined 2.9 %, 3.9 %, and 8.6 %, respectively. For all establishments, payroll and employment declined 5.3 % and 3.0 %, respectively. The results further suggest that the compound effect of hazards through the supply chain possibly exceeds that of the local hazard (i.e., direct impact). This can create an incentive misalignment. The establishment that invests in mitigation efforts and experiences the hazard locally does not directly experience the majority of the net benefit. The findings also suggest there is a need to better understand the short-term downstream impacts from all hazards, especially at the aggregated national level.

1. Introduction

In 2020, there were 22 natural disasters with losses exceeding $1 billion each in the U.S. (NOAA NCEI, 2021). Economic effects on the areas impacted directly were significant; Swiss Re estimated insured losses to be $83 billion, making 2020 the fifth costliest year on record since 1970 (Swiss Re, 2020). An increase in direct capital losses from such events is documented; however, specific effects on supply chains, especially indirect losses from hazards outside of the location directly physically impacted by a natural disaster, are less well developed in the literature.

At the macroeconomic level, disaster impacts rarely have a meaningful impact on a national economy; as noted by Mileti (1999), capital markets are simply too large to be disturbed beyond a short period of time by natural disasters. However, the propagation of impact throughout the supply chain and the connection between the local disaster impact and impacts felt elsewhere are not often part of the disaster loss calculation.

There have been notable changes over the last few decades to supply chain structures and management through globalization and innovation rates. This also indicates a growing interdependency between firms and increasing supply chain complexity (e.g., Kamalahmadi and Parast, 2016). Industry is increasingly aware of the growing volatility across a range of business parameters from energy cost, to raw material availability, and currency exchange rates (e.g., Neiger, Rotaru, and Churilov, 2009; Christopher and Holweg, 2011; Vlajic, van Lokven, Haijema, and van der Vorst, 2013). There is evidence that increased extreme weather and associated hazard events will make these vulnerabilities more pronounced in the future (e.g., Karl, 2009; Allison, et al. 2009, Bouwer, 2019; Willner et al. 2018; Koks et al. 2019) and businesses appear to be aware of potential effects. For example, the Supply Chain Management’s (SCM) “World’s 2017 Future of Supply Chain” survey found a pronounced shift in responses in the level of perceived risk in areas outside of our traditional span of control; the greatest jumps are in the percentage of respondents who report being “very concerned” about the effects of natural disasters on supply chain viability (Forbes, 2018).

There has been research on potential indirect losses, including how production losses due to disasters propagate to regions not directly hit through firms’ supply chains. For example, Park, Hong, and Roh (2013) analyze the downstream manufacturing supply chain effects of the 8.9-magnitude earthquake on the northeast coast of Japan and the resulting tsunami through a series of case studies of top production companies in Japan. They report that because Japan produces about 60 % of the world’s silicon for semiconductor chips, global prices for computer memory components spiked by 20 % immediately after the disaster. Meanwhile, several U.S. automobile plants were forced to halt production until specialized paints and computer chip shipments could resume. Inoue and Todo (2019) use an agent-based model with supply chain data (in the form of supplier-customer relations) for nearly one million firms in Japan to estimate the direct and total economic effects of the 2011 Great East Japan earthquake. As insightful as this study is, it does not provide insight into the total effect or risk of disruptions due to hazards, as it studies individual events.

There are several data analysis techniques used to assess the indirect and income effects of natural disasters. These techniques include surveys, econometric models, input-output models, and computable general equilibrium (CGE) models, among others (e.g., Cochrane, 2004; Chang, 2003; Hallegatte, 2015; Hosseini et al., 2019; Ribeiro and Barbosa-Povoa, 2018). The relative strengths and limitations of these approaches are noted in the following section.

Our research aims to improve understanding of business and supply-chain disruptions, and the associated losses, using a methodology based on empirical data. This is contrast to the many papers that use input-output analysis and other model-based approaches in order to estimate supply-chain losses, which rely on assumptions on the propagation of disaster losses and adaptations within the macroeconomic system. We examine the impact of hazards on payroll, GDP, employment, and the survival/creation of establishments post-disaster event. Eight models were developed to test eight hypotheses focusing on the manufacturing/goods producing industries, and all industries combined. Annual observations of disaster-related incidents and damage at the county level between 2006 and 2016 are used in the model to measure their impact on the U.S. economy and U.S. manufacturing in particular. Data on the geographic supply network combined with disaster data at these locations is used to examine the impact of disasters within the supply chain (i.e., the ripple effect). A simulated impact of damage both locally and through the supply chain was conducted for each model over the entire study period. This approach fills an important research gap and provides evidence on the presence, and magnitude, of supply-chain losses, and reveals aggregated impacts of disaster events both 1) directly and 2) through the supply chain, a topic that is crucial to understanding risk and providing policy-relevant insights on mitigation of disaster impacts.

This paper is organized as follows. Section 2 provides additional research context and a review of relevant background literature. Our methodological approach and data sources are then introduced in Section 3 and Section 4. The results are presented in Section 5 and then summarized in Section 6. Conclusions are presented and finally future research in this space is discussed in Section 7.

2. Literature Review and Context

In their study of supply chain characteristics relevant to a firm’s exposure to supply chain risk, Wagner and Bode (2006) provide a thorough discussion of the concepts of supply chain risk, supply chain disruption, and supply chain vulnerability. In this article, we follow the definition of supply chain disruption as “an unintended, untoward situation, which leads to supply chain risk” (Wagner and Bode, 2006) and acknowledge that supply chain disruptions can materialize from supply-side or demand-side risks and both these types of risk can be exacerbated by natural disasters.

A supply chain’s resiliency lies on a continuum and thus a supply network can be classified as being more or less resilient. The operation of a less vulnerable (i.e., less resilient) supply chain is not capable of operating without interruption when a disaster occurs (Sheffi and Rice, 2005); this may be the outcome of physical infrastructure failure upon which the supply chain depends or the human choice and interactions. Conversely, supply chain resilience is defined as “the capability of supply chains to respond quickly to unexpected events so as to restore operations to the previous performance level or even to a new and better one” (Pereira, Christopher, and Da Silva, 2014).

The ripple effect occurs when a disruption in the supply chain cascades downstream to impact supply chain performance, as opposed to remaining localized (Dolgui, Ivanov, and Sokolov 2018). The annual impacts of the ripple effect (industry- and economy-wide levels) due to natural hazards has received limited research to date. Many studies on the impact of natural hazards examine the upstream impact. Additionally, few papers estimate the national impacts that result from the ripple effect or estimate the average impact experienced by establishments downstream. A great deal of research on the ripple effect focuses on supply chain and inventory management from the perspective of the individual establishment/enterprise.

This focus leaves at least three non-trivial issues to be considered. The first is that many data observations are needed to accurately reveal the risk of infrequent hazards and many enterprises do not, by themselves, have sufficient data to estimate it. These enterprises will, potentially, make less than optimal risk mitigation decisions. The second issue is that without an understanding of the costs/losses resulting from natural hazards, there may be under investment in research efforts on risk mitigation. The third issue is that each establishment within a supply chain makes decisions based on their stakeholder perspective and incentives for the good along the rest of the supply chain may likely be unaligned. A competitive economy is, likely, one that considers system level issues that cross between enterprises along a supply chain. To any individual entity along the supply chain, a loss induced by natural disaster impacts may not be large enough to warrant risk mitigation actions (e.g., seek a new supplier or invest in a generator); however, summed together the losses can become a significant impact on the economy, especially when you consider the potential delays further downstream.

Much of the literature to date that is specific to supply chain analysis and improved resilience is focused at the enterprise-level and stresses preparedness through individual business continuity plans. There are two major pieces that are not often covered effectively in such business continuity planning. Firstly, there are several risks outside the control of an individual business (often on the community scale) that may instigate closures and prevent recovery (e.g., transportation closures and delayed utility recovery) (e.g., Tierney, 2014). Secondly, business continuity plans may stress supplier diversity, but understanding and planning for the potential effects on a given business by disruptive events that occur non-locally (e.g., out-of-state or out-of-country) remains a challenge.

Literature examining the macroeconomic impacts of natural disasters often measure impacts in terms of production outputs and inputs (e.g., Belasen and Polachek, 2008; Spencer and Polachek, 2015; Mohan, 2016; Okuyama and Rose, 2019). Studies at the macroeconomic level also tend to either identify setbacks for economic growth (e.g., Noy, 2009; Cavallo and Noy, 2011) or find that disasters have no lasting perverse effects on economic growth (e.g., Hochrainer, 2009; Strobl, 2011). Loayza, Olaberria, Rigolini, and Christiaensen (2012) find that disaster events do not always affect economic growth negatively and that moderate disasters show positive (localized) growth effects in some sectors, while more severe disasters do not.

There are several data analysis techniques used to assess the indirect and income effects of disasters including surveys, econometric models, I/O models, computable general equilibrium (CGE) models, and economic accounting models (Cochrane, 2004; Chang, 2003; Zimmerman et. al., 2005). There are relative strengths and limitations associated with each of these approaches, as noted in Table 1.

Table 1.

Analysis methods for examining indirect effects of disasters.

	Strengths	Limitations
Surveys	Provides direct information from those impacted or in close association with those directly affected by disasters.	Non-response bias and limitations on representative of the broader population.
Econometric models	Fully-partialed effects of a disaster event can be modeled as an intrusion on a series of data.	Data availability limitations and data often published with a significant time-lag post disaster event.
I/O models	Models produce three types of impact assessments: direct, indirect, and induced.	Static (measuring economic relationships at a particular point in time), linear relationships.
Computable General Equilibrium Models	Incorporate a range of input substitutions and different elasticities of supply and demand can be applied across different tiers of economic activity (e.g., Rose and Liao, 2005).	Often cover only a limited set of industrial sector and emphasize equilibrium states, which is unlikely post-disaster event.

Many of the studies examining the economic impact of disasters at a macroeconomic-level use input-output (I/O) analysis, social accounting matrices (SAM), and computable general equilibrium (CGE) modeling techniques. Recent advances in these models for disaster impact analysis are compiled and reviewed in Okuyama and Rose (2019) and Botzen, Deschenes, and Sanders (2019). Social accounting matrices use a similar dataset as input-output analysis in that it is a matrix of economic transactions, but includes other economic data, such as capital and labor. CGE models use both SAM and IO data along with additional datasets. Unlike IO analysis, CGE models incorporate price changes and each CGE model is unique, having its own variant, which can reflect downstream impacts to some extent; for example, Prager, Chen, and Rose (2018) use a CGE model and generate GDP impacts using Monte Carlo simulations of key drivers of disaster losses. CGE models are not, generally, considered forecasts or predictions, but rather they present a theoretical alternative future state. There are emerging models that include both demand and supply side effects; for example, Koks and Thissen (2016) present a supply-use model which considers production technologies and allows for supply side constraints; this model is demonstrated on three floods in Rotterdam, Netherlands.

Generally, no data or evidence from the post-hazard supply chain is used to estimate losses or impacts in I/O, SAM, or CGE models. When downstream impacts are represented, it is through long-run mechanisms such as price changes and substitution effects, which take time to be realized in an economy. Subsequently, a CGE model is not likely to capture the economic impact from intermediate goods failing to arrive - or failing to arrive on time - at the factory floor immediately following a hazard/disaster, as this is not realized through price changes or other long-run mechanisms. CGE models might capture a portion of the downstream impact, but it tends to be estimated future impacts at an economic equilibrium. The results of a CGE model are indispensable for community resilience planning; however, individual manufacturers and businesses are undoubtedly concerned with the short-term impacts and risks that result from hazards and in understanding these risks for supply chain management.

Although I/O and CGE models provide significant insights about natural hazard disruption, like all models, they have limitations; neither CGE nor IO models completely measure the immediate short-term impact in the downstream supply chain and, currently, they are not readily applied in a way that shows the total aggregated downstream supply-chain loss that results from all hazards, large and small. Two alternatives for capturing the short-term impact is to collect data on the supply chain post hazard or use correlation studies. This paper employs the latter. This method is not without its limitations either. Correlation studies do not prove causation and our method does not capture the total up and down stream impacts, but it can capture short-term impacts albeit along with some longer-term impacts. We do not propose using correlation studies in place of IO or CGE models, but rather be used to supplement the current existing approaches, especially where there are limitations.

In terms of the findings from disaster impact studies, Mohan, Spencer and Strobl (2019) find that hurricanes have a positive national impact on production efficiency using a panel data set across hurricanes and production in the Caribbean. Benson and Clay (2004) credited GDP increases occurring post-natural hazard to a catch-up effect and investment in reconstruction activities, as opposed to productivity increases. It is worth noting that the literature also provides evidence that productivity improvements are possible when natural hazard events are relatively minor (e.g., the recovery process is more efficiently managed) (Halkos, Managi, and Tzeremes, 2015). Sarmiento (2007) showed that on average, as workers left areas, aggregate local employment in the U.S. decreased by 3.4% following floods. In their study of hurricane impacts in Florida, Belasen and Polachek (2008) identified a decrease in the labor supply and an increase in post-hurricane labor demand. The result was a 4.35 % increase in local income.

Findings at the firm level are mixed. De Mel, McKenzie and Woodruff (2011) find that the firms that suffered more asset damages from the 2004 Sri Lankan tsunami exhibited smaller profits, sales, and capital stock. In addition, Tanaka (2015) found that following the 1995 Kobe Earthquake, plants located in the most devastated districts had lesser employment and value-added growth. Alternatively, Leiter, Oberhofer, and Raschky (2009) found that European firms located in regions affected by a major flood in 2000 had higher asset and employment growth compared with non-affected firms. Finally, Cole, Elliott, Okubo, and Strobl (2013) highlight a short-run productivity increase for damaged plants after the Kobe Earthquake.

As supply chains continue to grow in spatial complexity, they face challenges such as high demand variability, short life of products, and different expectations and requirements of customers (e.g., Infor, 2016). In turn, adapting to these challenges has largely increased supply chain complexity and resulted in more instability and unpredictability (e.g., Ghadge, Dani, and Kalawsky, 2012, Pereira, Christopher, and Da Silva, 2014).

In this paper we demonstrate a means to identify and assess general trends across industries as disaster events occur in a given location. Our research aims to improve understanding of business and supply-chain disruptions, and the associated losses, using a methodology based on empirical data. This is contrast to the many papers that use input-output analysis and other model-based approaches in order to estimate supply-chain losses, which rely on assumptions on the propagation of disaster losses and adaptations within the macroeconomic system.

3. Data

A number of datasets were used for this analysis. The dependent variables, which include annual payroll, gross domestic product (GDP), employment, and count of establishments at the county level, were taken from the County Business Patterns, an annual series from the US Census Bureau (2018). Payroll, employment, and the count of establishments from the CBP for the manufacturing industry¹ (i.e., North American Industry Classification System (NAICS) code 31–33) and for the total economy were used in the models (U.S. Census Bureau 2017). We put additional focus on the manufacturing industry, as we suspect that it is more vulnerable to supply disruption than other industries. Many other industries can continue to operate without a constant flow of supplies; however, manufacturing can be immediately stopped as a result of disruption. This paper also utilizes annual data on total US GDP and goods producing industry GDP at the county and national level from the Bureau of Economic Analysis (2018a, 2018b). All dollar figures in the model were adjusted to 2016 using the Consumer Price Index from the Bureau of Labor Statistics (2018) except the county level GDP which is in chained 2012 dollars.

Another dataset used was the domestic flow data from the Freight Analysis Framework (FAF) accessed through the US Department of Transportation (2018). The FAF data provides shipments of goods by origin and destination for 122 zones (see Figure 1) covering the entire US for 2012 through 2016. It is not believed that the misalignment of the FAF zones and the counties causes an issue, as the method (discussed below) uses a fixed effects model, which would address a potential issue. For this paper, the dollar value of a selection of goods was used for 13 categories of commodities classified by the Standard Classification of Transported Goods systems (see Table 2). These industries were selected to represent intermediate goods and goods that might have low levels of substitutability.

Figure 1:

Shipments by FAF Origin, Lower 48 States (2016)

Table 2:

Commodities used in Analysis

20 Basic Chemicals

23 Chemical Products

24 Plastics and Rubber

25 Logs and other Wood in the Rough

26 Wood Products

30 Textiles, Leather, and Articles of Textiles or Leather

31 Nonmetallic Mineral Products

32 Base Metal in Primary or Semi-Finished Forms and in Finished Basic Shapes

33 Articles of Base Metal

34 Machinery

35 Electronic and Other Electrical Equipment and Components, and Office Equipment

36 Motorized and Other Vehicles (including parts)

37 Transport Equipment (other)

The annual county data on count and damage caused by hazards/perils was taken from the Spatial Hazard Events and Losses Database for the United States (SHELDUS) accessed through Arizona State University (2018). SHELDUS^™ covers natural hazards such as thunderstorms, hurricanes, floods, wildfires, and tornados; as well as perils such as flash floods and heavy rainfall that may be considered “stressors.” Although these are primarily natural hazards, some (e.g., wildfires) can be human caused. In this paper, hazards and perils are treated in the same category and are universally referred to as hazards. The data is for all total hazards and perils, as defined in the database; the resulting models are disaster agnostic in nature. It appears that there may be some challenges in estimating accurate and precise damage levels in the SHELDUS data. For instance, damages are often reported in round numbers such as $50 000 or $75 000, suggesting there is some level of detail that is absent. These errors are likely to diminish statistical significance and our estimated impacts, making our estimates more conservative. Figure 2 provides a map of the total dollar value of hazard damage by county between 2005 and 2016. As can be seen, there are a few areas of concentrated hazard damage; however, a large proportion of the counties have less than $5 million in damage over the entirety of the study period.

Figure 2:

Hazards by County, Lower 48 States (2005–2016)

A summary of selected data used in the model is provided in Table 3, including the mean number of establishments, employees, and payroll for manufacturing industries and total industry for each year. On average, a given county contains 95–108 manufacturing establishments and 2388–2518 total establishments. An establishment is a physical location of economic activity where a company or enterprise could own multiple establishments. The average property damage, as listed in the SHELDUS database, was between $1.4 and $31.4 million. Property damage includes damage to establishments as well as other types of damage such as that of residential property. There are only between 4.1 and 7.3 hazards per county per year and there are many counties that have zero for the year (not shown). The average annual GDP for a county is between $4.9 billion and $5.7 billion with manufacturing accounting for between $0.9 and $1.1 billion. Manufacturing seems to account for a larger percent of GDP than it accounts for that of the number of establishments.

Table 3:

Data Summary

	Manufacturing				Total Industry
Year	County Mean Number of establishments	County Mean Number of employees	County Mean of Annual payroll ($million)	County Mean of Real GDP ($million)*	County Mean Number of establishments	County Mean Number of employees	County Mean of Annual payroll ($million)	County Mean of Real GDP ($million)	Mean of Property Damage by County ($million)	Mean Number of Hazards by County
2005	108	5 259	231	1 005	2 433	37 150	1 435	4 897	31.4	4.4
2006	108	5 247	239	1 051	2 466	38 163	1 530	5 038	2.3	5.2
2007	108	5 450	246	1 051	2 498	38 356	1 603	5 116	1.8	4.7
2008	106	5 411	247	1 025	2 467	38 440	1 635	5 104	6.7	6.7
2009	101	4 863	218	952	2 418	36 533	1 551	4 984	2.0	5.5
2010	98	4 515	218	949	2 402	35 707	1 575	5 103	3.3	5.9
2011	96	4 664	230	967	2 388	36 141	1 642	5 174	6.7	7.3
2012	97	4 760	241	986	2 414	36 835	1 717	5 262	10.7	5.6
2013	95	4 819	246	1 013	2 436	37 559	1 780	5 353	3.1	5.2
2014	95	4 900	260	1 042	2 454	38 226	1 871	5 478	1.9	5.0
2015	95	4 196	238	1 082	2 488	39 023	1 965	5 643	1.4	4.7
2016	95	4 192	240	1 070	2 518	39 608	2 010	5 729	5.9	4.1

^*Goods producing GDP

4. Methods

This paper tests a series of hypotheses regarding the impact of natural and human-made hazards on manufacturing and associated supply chains:

1. Natural/human-made hazards have a negative effect on payroll locally.

2. Natural/human-made hazards have a negative effect on output locally.

3. Natural/human-made hazards have a negative effect on employment locally.

4. Natural/human-made hazards drive some establishments out of business and/or prevent establishment formation.

5. Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on payroll locally.

6. Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on output locally.

7. Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on employment locally.

8. Natural/human-made hazards in the upstream supply chain (i.e., suppliers) drive some establishments out of business and/or prevent establishment formation locally.

The testing of multiple hypotheses and models allows us to publish results that do not support a hypothesis, which often do not get published. To test the hypotheses, two sets of four models, for a total of eight, were developed. Each set of models examines: 1) payroll, 2) GDP, 3) employment, and 4) the number of establishments. The first set of models addresses the manufacturing/goods industry, while the second set considers the total economy. Each model is in the form of a Cobb-Douglas production function and utilizes annual county data from 2005 through 2016. The number of observations used for estimating each model ranged between 8669 and 30 131. The natural log of each structural equation was used and then parameters were estimated using linear regression. Zero values in the data were replaced with 1.0 since the natural log of zero is undefined. Dummy variables were used in the model to indicate when zero values occurred.

There have been a number of studies examining the effects of research/development and productivity increases on production (Ugur et al., 2016). There are two approaches that are commonly used: the primal approach (production function) and the dual approach (cost function) with the primal approach being far more prevalent (Ugur et al., 2016). Given its prominence, this paper adopts the primal approach, which uses a Cobb-Douglas production function. It tends to model real output on research and development capital, capital stock, labor (number of employees or hours worked), and technological progress:

$$Q={Ae}^{\lambda }{C}^{{\beta }_{x1}}{L}^{{\beta }_{x2}}{K}^{{\beta }_{x3}}{𝓔}^{{\beta }_{x4}}$$

where

$Q$ = Real output

$C$ = Real capital stock

$K$ = Real research and development capital

$L$ = Labor (number of employees or labor hours worked)

${Ae}^{\lambda }$ = is technological progress with a rate of disembodied technological change $\lambda$

${\beta }_{xn}$ = Estimated parameters

We utilize the primal approach, but, in this instance, we are using lagged dependent variables to control for capital, labor, and other factors. The use of a Cobb-Douglas production function has also been used to examine natural hazard impacts (Mohan et al., 2019).

There are at least four types of damage that result from hazard events: 1) damage that does not disrupt the measured economy (e.g., GDP), 2) direct damage of intermediate goods that are part of the year’s economic activity, 2) damage that directly disrupts the production of goods and services, and 3) damage that has cascading disruptions to the production of goods and services. Some losses do not disrupt the measured economy (i.e., GDP). The damaging of some other types of physical capital, however, can have a linear relationship with economic losses. For instance, the loss of goods inventory can reduce GDP. Yet other damage can render some capital as temporarily useless, resulting in a multiplicative effect. For instance, the loss of utilities or the interruption in the delivery of supplies can leave machines sitting idle. This damage can be multiplicative in that the larger the economy, the more impact that results. For instance, the loss of a transformer in Frederick, Md probably results in less economic loss than that of a similar transformer in New York City. As losses increase, however, there are often cascading effects where, for instance, the loss of one utility results in losing another utility. These impacts increase both the physical and temporal breadth of the impact on the economy and can grow exponentially. Additionally, research from Shughrue (2020) and Koks (2016) confirm that damages grow non-linearly. For these reasons, this paper uses a Cobb-Douglas production function, which captures both multiplicative and exponential effects.

Hazard events can have both a positive and negative impact on GDP, payroll, employment, and the number of establishments (i.e., establishment survival and creation). For example, a positive economic impact is the result of consumers replacing damaged goods, while a negative impact results from damage to infrastructure needed to facilitate production activities. One- and two-year lagged values of the dependent variable were used in each model to predict the current year’s value. The 1-year lag is interacted with both the hazard count and hazard damage to account for the scale of damage relative to the size of the county’s economy. Each model controls for the number of local hazards and hazard damage in dollars. Hazards comprise of all hazards and perils in SHELDUS, including but not limited to: earthquakes, flooding, fog, hail, heat events, hurricanes, tropical storms, landslides, lightning, thunderstorms, tornados, tsunami, volcanos, wildfires, wind events, and winter weather. The number of hazards is intended to capture the positive impact that might result from hazards, while the damage captures the negative impact. A higher number of hazards would increase the probability of more goods being damaged and more consumers, businesses, and governments replacing or repairing those goods. It also increases the probability of damaged infrastructure and construction spending.

There are two variables that account for the top 24 suppliers to each county; the top 24 was selected since it covers approximately 20 % of the 122 FAF locations. The first variable is an interaction variable (${SUPCHN}_{TOTAL24}$) between the proportion of the supply chain represented and the damage occurring within that supply chain zone. The total of the top 24 supply chain zones is, then, summed together. For instance, consider Kalamazoo County in Michigan. Kalamazoo falls within the “rest of Michigan” category, which is the yellow portion of Michigan shown in Figure 1. The variable ${SUPCHN}_{TOTAL24}$ for this county is the amount supplied to that FAF zone from the largest supplier (excluding self-supply) divided by the total supplied to the region from all U.S. locations (including FAF region self-supply). This ratio is then multiplied by the total hazard damage that occurred in the FAF supplier zone. This is calculated for the top 24 locations and summed together. This interaction variable weights the damage occurring at that supply chain zone (i.e., FAF location) by its importance to the destination county. A supplier is defined as one of the 122 FAF zones seen in Figure 1. Note that the supply chain variables are only domestic suppliers. The supply chain data is available from 2012 through 2016; thus, 2012 is used for earlier years and an indicator variable for 2012 and earlier is included in the model. The second supply chain variable is the total count of hazards occurring at the top 24 supply chain zones.

It is important to note that the models in this paper do not take into account the effect of varying hazard types or business types. Future research might explore these combinations to further refine economic loss estimates. It is also important to note that the models examine the losses experienced in the year of hazard occurrence; thus, there are losses that are not measured, including those that occur into the next calendar year. The models are suited for estimating the losses that occur in a calendar year as a result of that year’s hazards. Hazards occurring near the beginning of the year will have more representation than those at the end of the year. This is a limitation of the data available. Future research might focus on additional losses in subsequent years.

Although this paper uses linear regression, the relationship between hazards and the dependent variable (i.e., GDP, employment, payroll, and establishments) being examined is not linear. The relationship is in the form of a Cobb-Douglas production function (see discussion above). The parameters being estimated are the exponents while the independent variables (e.g., hazard damage) are the base. Each of the bases and their exponents are multiplied together. Thus, despite using linear regression, the relationship being tested is not linear, but exponential and multiplicative. Taking the natural log of each side of an equation puts this relationship in a linear format in order to estimate the parameters, but the relationship is still mathematically equivalent to being exponential and multiplicative (Kennedy 2003; Greene 2008). The linear format for the equations is provided below.

Model 1

The first model examines the effect that hazard damage has on production, using manufacturing payroll to capture decreases in labor that do not necessarily result in reductions in employment. Note that in 2016, the top 24 domestic suppliers represented, on average, 74 % of the total domestic supplies from outside FAF zones. The structural equation for Model 1 in log terms is represented as:

$$\ln \left({PR}_{MAN,x}\right)={\beta }_{1a}\ln \left({PR}_{lag1,MAN,x}\right)+{\beta }_{2a}\ln \left({PR}_{lag2,MAN,x}\right)+{\beta }_{3a}\ln \left({HZRD}_{DMG,x}\right)$$

$$+{\beta }_{4a}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{5a}{INTRCT}_{1a}+{\beta }_{6a}{INTRCT}_{2a}+{\beta }_{7a}\ln \left({SUPCHN}_{TOTAL24,x}\right)$$

$$+{\beta }_{8a}\ln \left({HZRD}_{TOTAL24,x}\right)+{\beta }_{9a}\ln \left(YR\right)+{\beta }_{10a}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{11a}\ln \left({ZERO}_{CNT,LOC,x}\right)$$

$${\beta }_{12a}+{\sum }_{s\in S}{\beta }_{sa}{C}_s+𝓔$$

where

$${INTRCT}_{1a}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({PR}_{lag1,MAN,x}\right)$$

$${INTRCT}_{2a}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({PR}_{lag1,MAN,x}\right)$$

$${SUPCHN}_z=\sum _{z=1}^{24}\frac{{SC}_{Top-z,x}}{\sum _{\mathrm{i}=1}^{122}{SC}_i}{HZRD}_{DMG,Top-z}$$

and where

${PR}_{MAN,x}$ = Manufacturing industry payroll in county x. Note that lag1 and lag2 indicate 1-year and 2-year lags.

${HZRD}_{DMG,X}$ = The total damage in county x caused by all hazards and perils listed in the SHELDUS database

${HZRD}_{CNT,X}$ = The total number of hazards and perils from SHELDUS in county x listed in the SHELDUS database

${HZRD}_{DMG,Top-z}$ = The total damage from SHELDUS for all counties in supply chain zone ranked z for county x where z equals ranks 1 through 24. Supply chain zones are those used by the Freight Analysis Framework.

${HZRD}_{TOTAL24,x}$= The total number of hazards and perils listed in SHELDUS for all counties in the top 24 supply chain zones for county x

$YR$ = Indicator variable for 2013 and later where $\ln \left(YR\right)$ equals 1 when the observation year is greater than 2012

${SC}_{Top-z,x}$ = The value supplied to county x from the z^th largest supplier where z is between 1 and 24.

${ZERO}_{CNT,LOC,x}$ = Indicator variable for zero hazard incidents locally for county x where $\ln \left({ZERO}_{CNT,LOC,x}\right)$ equals 1 when there are zero hazard incidents

${ZERO}_{DMG,LOC},x$ = Indicator variable for zero hazard damage locally for county x where $\ln \left({ZERO}_{DMG,LOC,x}\right)$ equals 1 when there are zero hazard incidents

$𝓔$ = Error term

${\beta }_{ya}$ = Parameter set a to be estimated where y is parameter 1 through 12

$C_s$ = An indicator variable for county s, where S is the set of counties

Note that the proposed relationship between GDP and hazard damage can be obtained by taking the inverse of the natural log of both sides of the equation. This renders an equation similar to the Cobb-Douglas production function. The first few variables for the relationship (i.e., taking the inverse natural log of both sides of the equation above) would appear as the following:

$${PR}_{MAN,x}={{PR}_{lag1,MAN,x}}^{{\beta }_{1a}}\times {{PR}_{lag2,MAN,x}}^{{\beta }_{2a}}\times {{HZRD}_{DMG,x}}^{{\beta }_{3a}}\times {{HZRD}_{CNT,x}}^{{\beta }_{4a}}\times \dots$$

Only the linear formats of the equations are presented below.

Model 2

Model 2 examines GDP from goods producing industries. The structural equation in log terms for Model 2 is represented as:

$$\ln \left({GDP}_{GOODS,x}\right)={\beta }_{1c}\ln \left({GDP}_{lag1,GOODS,x}\right)+{\beta }_{2c}\ln \left({GDP}_{lag2,GOODS,x}\right)$$

$$+{\beta }_{3c}\ln \left({HZRD}_{DMG,x}\right)+{\beta }_{4c}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{5c}{INTRCT}_{1c}+{\beta }_{6c}{INTRCT}_{6c}+$$

$${\beta }_{7c}\ln \left({SUPCHN}_{TOTAL24,x}\right)+{\beta }_{8c}\ln \left({HZRD}_{TOTAL24,x}\right)+{\beta }_{9c}\ln \left({ZERO}_{DMG,LOC},x\right)$$

$$+{\beta }_{10c}\ln \left({ZERO}_{CNT,LOC,x}\right)+{\beta }_{11c}+{\sum }_{s\in S}{\beta }_{sc}{C}_s+𝓔$$

where

$${INTRCT}_{1c}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({GDP}_{lag1,MAN,x}\right)$$

$${INTRCT}_{2c}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({GDP}_{lag1,MAN,x}\right)$$

and where

${GDP}_{GOODS,x}$ = GDP for goods producing industries in county x. Note that lag1 and lag2 indicate 1- and 2-year lags

${\beta }_{yc}$ = Parameter set c to be estimated where y is 1 through 11

Model 3

Model 3 examines the effect that hazard damage has on manufacturing employment. The structural equation in log terms for Model 3 is represented as:

$$\ln \left({EMP}_{MAN,x}\right)={\beta }_{1d}\ln \left({EMP}_{MAN,lag1,x}\right)+{\beta }_{2d}\ln \left({EMP}_{MAN,lag2,x}\right)+{\beta }_{3d}\ln \left(UR\right)+$$

$${\beta }_{4d}\ln \left({HZRD}_{DMG,x}\right)+{\beta }_{5d}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{6d}{INTRCT}_{1d}+{\beta }_{7d}{INTRCT}_{2d}+$$

$${\beta }_{8d}\ln \left({SUPCHN}_{TOTAL24}\right)+{\beta }_{9d}\ln \left({HZRD}_{TOTAL24}\right)+{\beta }_{10d}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{11d}YR$$

$$+{\beta }_{12d}\ln \left({ZERO}_{CNT,LOC,x}\right)+{\beta }_{13d}+{\sum }_{s\in S}{\beta }_{sd}{C}_s+𝓔$$

where

$${INTRCT}_{1d}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EMP}_{lag1,MAN,x}\right)$$

$${INTRCT}_{2d}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EMP}_{lag1,MAN,x}\right)$$

and where

${EMP}_{lag1,MAN,x}$ = Manufacturing employment in county x. Note that lag1 and lag2 indicate 1-year and 2-year lags.

$UR$ = National unemployment rate

${\beta }_{yd}$ = Parameter set d to be estimated where y is 1 through 13

Model 4

Model 4 examines the effect that hazard damage has on the survival and creation of manufacturing establishments. The structural equation in log terms for Model 4 is represented as:

$$\ln \left({EST}_{MAN,x}\right)={\beta }_{1e}\ln \left({EST}_{lag1,MAN,SMLL,x}\right)+{\beta }_{2e}\ln \left({EST}_{lag1,MAN,MED,x}\right)+{\beta }_{3e}\ln \left({EST}_{lag1,MAN,LG,x}\right)+{\beta }_{4e}\ln \left({HZRD}_{DMG,x}\right)+{\beta }_{5e}\ln \left({HZRD}_{CNT,x}\right)$$

$$+{\beta }_{6e}\ln \left({INTRCT}_{1e,SMLL}\right)+{\beta }_{7e}\ln \left({INTRCT}_{1e,MED}\right)+$$

$${\beta }_{8e}\ln \left({INTRCT}_{1e,LG}\right)+{\beta }_{9e}\ln \left({INTRCT}_{2e,SMLL}\right)+{\beta }_{10e}\ln \left({INTRCT}_{2e,MED}\right)$$

$$+{\beta }_{11e}\ln \left({INTRCT}_{2e,LG}\right)+{\beta }_{12e}\ln \left({SUPCHN}_{TOTAL24}\right)+{\beta }_{13e}\ln \left({HZRD}_{TOTAL24}\right)+{\beta }_{14e}\ln \left(YR\right)$$

$$+{\beta }_{15e}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{16e}\ln \left({ZERO}_{CNT,LOC,x}\right)+{\beta }_{17e}+{\sum }_{s\in S}{\beta }_{se}{C}_s+𝓔$$

where

$${INTRCT}_{1e,SMLL}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EST}_{lag1,MAN,SMLL,x}\right)$$

$${INTRCT}_{1e,MED}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EST}_{lag1,MAN,MED,x}\right)$$

$${INTRCT}_{1e,LG}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EST}_{lag1,MAN,LG,x}\right)$$

$${INTRCT}_{2e,SMLL}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EST}_{lag1,MAN,SMLL,x}\right)$$

$${INTRCT}_{2e,MED}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EST}_{lag1,MAN,MED,x}\right)$$

$${INTRCT}_{2e,LG}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EST}_{lag1,MAN,LG,x}\right)$$

and where

${EST}_{MAN,z,x}$ = Total number of establishments of establishments with size z for manufacturing in county x where size z is small (i.e., establishments with 1 to 19 employees), medium (i.e., establishments with 20 to 500 employees), or large (i.e., establishments with 500 or more employees);

${\beta }_{ye}$ = Parameter set e to be estimated where y is 1 through 17.

Model 5

Model 5 examines the effect that hazard damage has on all industry output, using payroll for all industries as a proxy for output. The structural equation in log terms for Model 5 is represented as:

$$\ln \left({PR}_{ALL,x}\right)={\beta }_{1f}\ln \left({PR}_{lag1,ALL,x}\right)+{\beta }_{2f}\ln \left({PR}_{lag2,ALL,x}\right)+$$

$${\beta }_{3f}\ln \left({HZRD}_{DMG,x}\right)+{\beta }_{4f}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{5f}{INTRCT}_{1f}+{\beta }_{6f}{INTRCT}_{2f}+$$

$${\beta }_{7f}\ln \left({SUPCHN}_{TOTAL24}\right)+{\beta }_{8f}\ln \left({HZRD}_{TOTAL24}\right)+{\beta }_{9f}\ln \left(YR\right)+$$

$${\beta }_{10f}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{11f}\ln \left({ZERO}_{CNT,LOC,x}\right)+{\beta }_{12f}+{\sum }_{s\in S}{\beta }_{sf}{C}_s+𝓔$$

where

$${INTRCT}_{1f}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({PR}_{lag1,ALL,x}\right)$$

$${INTRCT}_{2f}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({PR}_{lag1,ALL,x}\right)$$

and where

${PR}_{x,ALL}$ = Total payroll for all industries in county x; note that lag1 and lag2 indicate 1-year and 2-year lags for this variable;

${\beta }_{yf}$ = Parameter set f to be estimated where y is parameter 1 through 12.

Model 6

Model 6 examines total county GDP. The structural equation in log terms for Model 6 is represented as:

$$\ln \left({GDP}_{ALL,x}\right)={\beta }_{1g}\ln \left({GDP}_{lag1,ALL,x}\right)+{\beta }_{2g}\ln \left({GDP}_{lag2,ALL,x}\right)+{\beta }_{3g}\ln \left({HZRD}_{DMG,x}\right)$$

$$+{\beta }_{4g}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{5g}{INTRCT}_{1g}+{\beta }_{6g}{INTRCT}_{2g}+{\beta }_{7g}{SUPCHN}_{TOTAL24}$$

$$+{\beta }_{8g}{HZRD}_{TOTAL24}+{\beta }_{9g}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{10g}\ln \left({ZERO}_{CNT,LOC,x}\right)+{\beta }_{11g}$$

$$+{\sum }_{s\in S}{\beta }_{sg}{C}_s+𝓔$$

where

$${INTRCT}_{1g}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({GDP}_{lag1,MAN,x}\right)$$

$${INTRCT}_{2g}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({GDP}_{lag1,MAN,x}\right)$$

${GDP}_{ALL,x}$ = GDP for all industries in county x. lag1 and lag2 indicate 1- and 2-year lags.

${\beta }_{yg}$ = Parameter set g to be estimated where y is 1 through 11

Model 7

Model 7 examines the effect that hazard damage has on total employment. The structural equation in log terms for Model 7 is represented as:

$$\ln \left({EMP}_{ALL,x}\right)={\beta }_{1h}\ln \left({EMP}_{lag1,ALL,x}\right)+{\beta }_{2h}\ln \left({EMP}_{lag2,ALL,x}\right)+{\beta }_{3h}\ln \left({HZRD}_{DMG,x}\right)+$$

$${\beta }_{4h}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{5h}{INTRCT}_{1h}+{\beta }_{6h}{INTRCT}_{2h}+{\beta }_{7h}\ln \left({SUPCHN}_{TOTAL24}\right)$$

$$+{\beta }_{8h}\ln \left({HZRD}_{TOTAL24}\right)+{\beta }_{9h}\ln \left(YR\right)+{\beta }_{10h}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{11h}\ln \left({ZERO}_{CNT,LOC,x}\right)$$

$${\beta }_{12h}\ln \left(UNEMP\right)+{\beta }_{13h}+{\sum }_{s\in S}{\beta }_{sh}{C}_s+𝓔$$

where

$${INTRCT}_{1h}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EMP}_{lag1,ALL,x}\right)$$

$${INTRCT}_{2h}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EMP}_{lag1,ALL,x}\right)$$

and where

${\beta }_{yh}$ = Parameter set h to be estimated where y is 1 through 13;

${EMP}_{lag1,ALL,x}$ = Total employment for all industries in county x. Note that lag1 and lag2 indicate 1-year and 2-year lags.

Model 8

Model 8 examines the effect that hazard damage has on the survival and creation of all establishments. The structural equation in log terms for Model 8 is represented as:

$$\ln \left({EST}_{ALL,x}\right)={\beta }_{1i}\ln \left({EST}_{lag1,ALL,SMLL,x}\right)+{\beta }_{2i}\ln \left({EST}_{lag1,ALL,MED,x}\right)+{\beta }_{3i}\ln \left({EST}_{lag1,ALL,LG,x}\right)$$

$${\beta }_{4i}\ln \left({HZRD}_{DMG,x}\right)+{\beta }_{5i}\ln \left({HZRD}_{CNT,x}\right)+{\beta }_{6i}\ln \left({INTRCT}_{1i,SMLL}\right)+{\beta }_{7i}\ln \left({INTRCT}_{1i,MED}\right)+$$

$${\beta }_{8i}\ln \left({INTRCT}_{1i,LG}\right)+{\beta }_{9i}\ln \left({INTRCT}_{2i,SMLL}\right)+{\beta }_{10i}\ln \left({INTRCT}_{2i,MED}\right)$$

$${+{\beta }_{11i}\ln \left({INTRCT}_{2i,LG}\right)+\beta }_{12i}\ln \left({SUPCHN}_{TOTAL24}\right)+{\beta }_{13i}\ln \left({HZRD}_{TOTAL24}\right)+{\beta }_{14i}\ln \left(YR\right)$$

$$+{\beta }_{15i}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{16i}\ln \left({ZERO}_{DMG,LOC,x}\right)+{\beta }_{17i}+{\sum }_{s\in S}{\beta }_{si}{C}_s+𝓔$$

where

$${INTRCT}_{1i,SMLL}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EST}_{lag1,ALL,SMLL,x}\right)$$

$${INTRCT}_{1i,MED}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EST}_{lag1,ALL,MED,x}\right)$$

$${INTRCT}_{1i,LG}=\ln \left({HZRD}_{DMG,x}\right)\mathrm{*}\ln \left({EST}_{lag1,ALL,LG,x}\right)$$

$${INTRCT}_{2i,SMLL}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EST}_{lag1,ALL,SMLL,x}\right)$$

$${INTRCT}_{2i,MED}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EST}_{lag1,ALL,MED,x}\right)$$

$${INTRCT}_{2i,LG}=\ln \left({HZRD}_{CNT,x}\right)\mathrm{*}\ln \left({EST}_{lag1,ALL,LG,x}\right)$$

and where

${EST}_{MAN,z,x}$ = Total number of establishments of establishments with size z for all industries in county x where size z is small (i.e., establishments with 1 to 19 employees), medium (i.e., establishments with 20 to 500 employees), or large (i.e., establishments with 500 or more employees);

${\beta }_{yi}$ = Parameter set i to be estimated where y is 1 through 17.

Three versions of the Breusch-Pagan and Cook-Weisberg test for heteroskedasticity (Stata 2013a) were run. The results indicated that heteroskedasticity was present in the data. To address this issue, we fit a fixed-effects model using a “GLS estimator (producing a matrix-weighted average of the between and within results)” (Stata 2013b). This approach has been identified as providing robust estimates for data with this issue (Hoechle 2007).

Simulation

Simulations were conducted using each model to estimate the impact of damage caused by hazards. First, a simulation was run to estimate the impact of hazard damage over the study period by using the estimated parameters. The estimate of the dependent variable (i.e., payroll, GDP, employment, and the number of establishments) was then compared to two additional simulations where: 1) no damage occurred locally and 2) no damage occurred in the supply chain. The first examined the impact of hazards locally by setting hazard damage (${HZRD}_{DMG,x}$) to equal 1 and the indicator variable for zero damage (${ZERO}_{DMG,LOC}$) set so that $\ln \left({ZERO}_{DMG,LOC,x}\right)$ equals 1. This simulates a world where the hazards occurred (i.e., the count still captures the hazards) but did not cause damage. The percent change was then calculated:

$${PC}_{LOC,d}=\frac{\sum _{x=1}^n\left({DEP}_{LOC,DMG,d,x}-{DEP}_{LOC,NO-DMG,d,x}\right)}{\sum _{x=1}^n{DEP}_{LOC,NO-DMG,d,x}}$$

where

${PC}_{LOC,d}$ = Percent change in dependent variable d due to local damage, where d is either payroll, GDP, employment, or the number of establishments;

${DEP}_{LOC,DMG,d,x}$ = Estimate of the dependent variable d in county x estimated with local hazard damage, where d is either payroll, GDP, employment, and the number of establishments;

${DEP}_{LOC,NO-DMG,d,x}$ = Estimate of the dependent variable d in county x estimated with no local hazard damage, where d is either payroll, GDP, employment, and the number of establishments.

The second comparison is between the simulations with and without damage in the supply chain, where the dependent variable is estimated with ${SUPCHN}_{TOTAL24}$ equaling 1 or $1 in damage. The percent change was then calculated:

$${PC}_{SUPCHN,d}=\frac{\sum _{x=1}^n\left({DEP}_{SUPCHN,DMG,d,x}-{DEP}_{SUPCHN,NO-DMG,d,x}\right)}{\sum _{x=1}^n{DEP}_{SUPCHN,NO-DMG,d,x}}$$

where

${PC}_{SUPCHN,d}$ = Percent change in dependent variable d due to hazard damage in the supply chain, where d is either payroll, GDP, employment, or the number of establishments;

${DEP}_{SUPCHN,DMG,d,x}$ = Estimate of the dependent variable d in county x estimated with hazard damage in the supply chain, where d is either payroll, GDP, employment, and the number of establishments;

${DEP}_{SUPCHN,NO-DMG,d,x}$ = Estimate of the dependent variable d in county x estimated with no hazard damage in the supply chain, where d is either payroll, GDP, employment, and the number of establishments.

The 95 % confidence intervals were calculated using a bootstrapping procedure where the total impact is estimated for a random selection of observations. This process was iterated 5000 times.

5. Results

A guide to the models and hypotheses is provided in Table 4, which identifies the variables in each model and the associated hypotheses. Results are reported in Table 5 with simulation results found in Table 6. As seen in Table 6, the adjusted R² is between 0.9749 and 0.9997. Each model includes two lags of the dependent variable, accounting for the high R² value. Results are reported in Table 5. Hazard damage and the zero indicator (${HZRD}_{DMG}$ and/or ${ZERO}_{DMG,LOC}$) were not statistically significant in any of the models except for the model of total payroll. The impact in total payroll, however, is minimal as seen in the simulation results shown in Table 6. No decrease in the number of establishments was detected. Moreover, the results suggest there are no discernable local impacts due to local hazards.

Table 4:

Guide to Models

		Model
	Model Characteristics	1	2	3	4	5	6	7	8
Hypothesis	1. Natural/man-made hazards have a negative effect on payroll locally.	X				X
	2. Natural/man-made hazards have a negative effect on output locally.		X				X
	3. Natural/man-made hazards have a negative effect on employment locally.			X				X
	4. Natural/man-made hazards drive some establishments out of business and/or prevent establishment formation				X				X
	5. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on payroll locally.	X				X
	6. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on output locally.		X				X
	7. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on employment locally.			X				X
	8. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) drive some establishments out of business locally and/or prevent establishment formation				X				X
Dependent Variables	Payroll (PR)	X1				X2
	GDP		X1				X2
	Employment (EMP)			X1				X2
	Establishments (EST)				X1				X2
Independent Variables	Payroll lagged 1 year (PRlag1)	X1				X2
	Payroll lagged 2 year (PRlag2)	X1				X2
	Interaction of hazard count and payroll (INTRCT2a or 2b or 2f)	X1				X2
	Interaction of hazard damage and payroll (INTRCT1a or 1b or 1f)	X1				X2
	GDP lagged 1 year (GDPlag1)		X1				X2
	GDP lagged 2 years (GDPlag2)		X1				X2
	Interaction of hazard count and GDP (INTRCT2c or 2g)		X1				X2
	Interaction of hazard damage and GDP (INTRCT1c or 1g)		X1				X2
	Employment lagged 1 year (EMPlag1)			X1				X2
	Employment lagged 2 years (EMPlag1)			X1				X2
	Interaction of hazard count and employment (INTRCT2d or 2h)			X1				X2
	Interaction of hazard damage and employment (INTRCT1d or 1h)			X1				X2
	Establishments lagged (small, medium, and large) 1 year (ESTlag1)				X1				X2
	Interaction of hazard count and establishments (small, medium, and large) (INTRCT2e or 2i)				X1				X2
	Interaction of hazard damage and establishments (Small, medium, and large) (INTRCT1e or 1i)				X1				X2
	Local hazard count (HZRDCNT)	X	X	X	X	X	X	X	X
	Local hazard damage (HZRDDMG)	X	X	X	X	X	X	X	X
	Hazard damage at top 24 supply locations (SUPCHNTOTAL25)	X	X	X	X	X	X	X	X
	Hazard count at top 24 supply locations (HZRDTOTAL25)	X	X	X	X	X	X	X	X
	Indicator for 2012 and Earlier	X		X	X	X		X	X
	National Unemployment Rate (UR)			X				X
	Indicator for zero damage (ZERODMG,LOC)	X	X	X	X	X	X	X	X
	Indicator for zero damage (ZEROCNT,LOC)	X	X	X	X	X	X	X	x
	Indicator for county s (Cs)	X	X	X	X	X	X	X	X

¹For manufacturing or goods producing only

²For all industries

Table 5A:

Results for Models 1 through 8

	Model Number and Dependent Variable
	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7	Model 8
	PRMAN	GDPGOODS	EMPMAN	ESTMAN	PRALL	GDPALL	EMPALL	ESTALL
EMPlag1,MAN			0.5627***
EMPlag2,MAN			−0.0205*
EMPlag1,ALL							0.5973***
EMPlag2,ALL							−0.0043
ESTlag1,LG				0.0276***				0.0024*
ESTlag1,MED				0.1648***				0.0917***
ESTlag1,SMLL				0.1997***				0.6478***
GDPlag1,MAN		0.5922***
GDPlag2,MAN		0.0252*
GDPlag1,ALL						0.779***
GDPlag2,ALL						−0.0134
HZRDCNT	−0.0222	−0.0068	0.004	0.0003	0.0028	−0.0005	0.0013	−0.0029
HZRDDMG	−0.0012	−0.0044	−0.0007	−0.0001	−0.0031**	−0.0027	0.0002	0.0019**
HZRDTOTAL24	−0.0019	−0.0075	0.0724***	0.0124***	−0.0086*	−0.0216***	0.0309***	−0.0057***
SUPCHNTOTAl24	−0.0022***	−0.0029**	−0.0068***	0.0002	−0.004***	−0.0006	−0.0023***	0.0003**
ZERODMG,LOC	−0.0058	−0.0065	0.0045	−0.0006	−0.005	−0.0016	0.0006	0.0017
ZEROCNT,LOC	−0.0006	0.0062	−0.0082	0.0011	−0.0049	0.0024	−0.0039*	−0.0017
INTRCT1a	0.0001
INTRCT1b		0.0003
INTRCT1c			0.0001
INTRCT1d,LG				−0.0007*
INTRCT1d,MED				−0.0011**
INTRCT1d,SMLL				0.0011**

^***Significant at the 0.1 level,

^**Significant at the 0.05 level,

^*Significant at the 0.01 level

Table 6:

Simulation and Results Summary

Model	Dependent Variable	Adjusted R2	Observations	Simulated Impact of Local Damage (2006–2016)			Simulated Impact of Supply Chain Damage (2006–2016)
				Est.	95 % Confidence Interval		Est.	95 % Confidence Interval
1	PRMAN	0.9919	22 518	-	-	-	−2.9%	−5.2%	−0.6%
2	GDPGOODS	0.9749	23 878	-	-	-	−3.9%	−6.8%	−0.1%
3	EMPMAN	0.9893	20 981	-	-	-	−8.6%	−10.6%	−6.7%
4	ESTMAN	0.9914	8 669	-	-	-	-	-	-
5	PRALL	0.9973	27 183	0.0%	−0.7%	0.6%	−5.3%	−5.4%	−3.2%
6	GDPALL	0.9966	29 985	-	-	-	-	-	-
7	EMPALL	0.9980	30 131	-	-	-	−3.0%	−3.9%	−2.2%
8	ESTALL	0.9997	13 487	-	-	-	-	-	-

⁻Hyphen indicates that either the local or supply chain damage variables were not statistically significant at the 0.1 level

The variable for hazard damage within the supply chain (i.e., ${SUPCHN}_{TOTAL24}$) was significant in five of the eight models, including one model of GDP (i.e., Models 2) and both for employment (i.e., Models 3 and 7). It was also significant for both payroll models (i.e., Models 1 and 5). Note that payroll includes all forms of compensation measured in dollars; however, temporary employees (e.g., paid through a third party) are not included. These results suggest that hazards in the supply chain have a statistically significant negative effect on economic activity in the downstream supply chain. The magnitude of this impact can be seen in the results of the simulation (see Table 6), where a world without hazard damage is simulated. The impact of hazards in the supply chain decreased goods GDP by 3.9 %. Payroll is decreased for manufacturing by 2.9 % and for total payroll the decrease is 5.3 %. Employment for manufacturing and the total economy decreased 2.9 % and 5.3 %, respectively. For manufacturing/goods producing industries, employment decreased more than GDP. This type of effect can also be seen in some recessions, such as the recession of the late 2000s where, in percent terms, employment decreased more than GDP (Thomas, 2020)

6. Summary and Discussion

This study examines the impact of hazards on payroll, GDP, employment, and the survival/creation of establishments post-hazard. Eight hypotheses were presented in this paper that were tested using eight models. Presented below is a discussion on how the evidence supports or does not support the hypotheses. Table 7 summarizes which hypotheses were supported by the models.

Table 7:

Do the Models Support the Hypotheses

	Model Characteristics	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7	Model 8
Hypothesis	1. Natural/man-made hazards have a negative effect on payroll locally.	No				No
	2. Natural/man-made hazards have a negative effect on payroll output.		No				No
	3. Natural/man-made hazards have a negative effect on employment locally.			No				No
	4. Natural/man-made hazards drive some establishments out of business and/or prevent establishment formation				No				No
	5. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on payroll locally.	Yes				Yes
	6. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on output locally.		Yes				No
	7. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on employment locally.			Yes				Yes
	8. Natural/man-made hazards in the upstream supply chain (i.e., suppliers) drive some establishments out of business locally and/or prevent establishment formation				No				No

Hypothesis 1:

The first hypothesis is that “Natural/human-made hazards have a negative effect on payroll locally.” This hypothesis was somewhat supported by the statistical significance of the hazard damage variable (i.e, ${HZRD}_{DMG}$) and the interaction variable with hazard damage in Model 5. As is seen in the simulation, however, the impact is relatively small and 95 % confidence interval ranged from −0.7 % to 0.6 % impact as seen in Table 6. This hypothesis was not supported in Model 1 where the zero-damage variable (i.e., ${ZERO}_{DMG,LOC}$), hazard damage variable (i.e, ${HZRD}_{DMG}$), and interaction variable with hazard damage were not statistically significant. A decrease in payroll is intended to represent an overall decrease in labor. Theoretically, employment could be unchanged while hours are decreased. This type of change would likely be reflected in payroll changes.

Hypothesis 2:

The second hypothesis, “Natural/human-made hazards have a negative effect on output locally,” is not supported by the models. The zero-damage variable (i.e., ${ZERO}_{DMG,LOC}$) in Model 2 was not statistically significant nor was the hazard damage variable (i.e, ${HZRD}_{DMG}$) or interaction variable with hazard damage. The lack of significance could result from challenges to estimate accurate and precise damage levels. For instance, damages are often reported in round numbers such as $50 000 or $75 000, suggesting there is some level of detail that is absent.

Hypothesis 3:

The third hypothesis, “Natural/human-made hazards have a negative effect on employment locally,” was not supported, as the zero-damage variable (i.e., ${ZERO}_{DMG,LOC}$) along with the hazard damage variable (i.e., ${HZRD}_{DMG}$) and the interaction variable with hazard damage were not statistically significant in Model 3 or Model 7, where employment is the dependent variable. It is possible that imprecision in damage estimates result in a lack of significance in these variables.

Hypothesis 4:

The fourth hypothesis, “Natural/human-made hazards drive some establishments out of business and/or prevent establishment formation,” was not supported due to the lack of statistical significance in the zero-damage variable (i.e., ${ZERO}_{DMG,LOC}$) and hazard damage variable (i.e., ${HZRD}_{DMG}$) in Model 4 and Model 8, where the number of establishments is the dependent variable. Again, it is possible that imprecision in damage estimates result in a lack of significance in these variables.

Hypothesis 5:

The fifth hypothesis, “Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on payroll locally,” was supported by the models. The statistical significance of the supply chain hazard damage variable (i.e., ${SUPCHN}_{TOTAL24}$) in Model 1 and Model 5, where manufacturing payroll and total payroll is the dependent variable, supports this hypothesis. The 95 % confidence interval for manufacturing and the total economy overlap, suggesting there may be little difference in the impact on these two. The significance of this hypothesis suggests that labor goes down either through job loss or decreased hours per employee.

Hypothesis 6:

The sixth hypothesis, “Natural/man-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on output locally,” was supported by the statistical significance of the supply chain hazard damage variable (i.e., ${SUPCHN}_{TOTAL24}$) in Model 2, where the dependent variable is goods GDP. This variable was not significant in Model 6 where total GDP was the dependent variable. The simulations indicated that these hazards had a substantial impact, reducing goods GDP by 3.9 %. The statistical significance in goods producing industries and the lack thereof for the total economy could be due to the supply chain for physical goods facing more risk due to the many potential points where disruption can occur (i.e., higher level of exposure). It is also possible that critical items with limited value can stop large value items from being produced. For instance, in the mid-2000’s the Boeing 787 Dreamliner experienced significant delays in production due to a shortage of fasteners (i.e., bolts). Although these bolts/fasteners are unique in that they are aluminum and titanium, their value is smaller when compared to that of the 787 production (Reuters 2007). The implication of this larger impact is that mitigation investments can be targeted toward those areas that have higher risk, resulting in a higher return on investment. It is also possible that the supply chain disruption for non-goods producing industries is not as critical. For instance, a retailer can experience a supply chain disruption for one item but continue to sell other items while a manufacturer may have to stop production due to a disruption in supplies.

Hypothesis 7:

The seventh hypothesis, “Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on employment locally” is supported in the statistical significance of the supply chain hazard damage variable (i.e., ${SUPCHN}_{TOTAL24}$) in Model 3, where manufacturing employment is the dependent variable, and Model 7, where total employment is the dependent variable. The simulations indicate that supply chain hazards decrease employment by 8.6 % in the manufacturing industry and 3.0 % in the total economy. It is interesting to note that the elasticity for the supply chain variable in the GDP model correlates with that of the employment model and payroll model (see Figure 3).² For manufacturing/goods, the correlation coefficient between goods GDP and manufacturing employment is 0.999. The correlation between goods GDP and manufacturing payroll is also 0.999. A similar situation occurs for the total economy. The correlation coefficient between the elasticities for total payroll and total employment is 0.963. These correlations provide some evidence to confirm the effect of hazards on the supply chain, as payroll, GDP, and employment are expected to move together similarly.

Figure 3:

Elasticity for Supply Chain Damage (SUPCHN_TOTAL24)

Note: Other variables are held at the mean.

Note: Elasticities were calculated at percentile 1 and then at 2, 4, 6, …, 98, 100 of the SUPCHN_TOTAl24 variable.

It is interesting that downstream effects have a statistically significant impact on employment and payroll in Models 5 and 7 but not on GDP in Model 6. This could be due to inaccuracies in tracking hazard damage. It is important to note that an impact is detected in goods producing industries, which is part of the total economy being measured in model 6.

Hypothesis 8:

The last hypothesis, “Natural/human-made hazards in the upstream supply chain (i.e., suppliers) drive some establishments out of business and/or prevent establishment formation locally” is not supported by either the model of manufacturing establishments (i.e., Model 4) or total establishments (i.e., Model 8). This suggests that, although hazard damage in the supply chain diminishes economic activity, it does not drive establishments out of business or prevent their creation. This could be due to establishments’ ability to acquire new sources of materials or it could be that the effect takes more than one year to realize and is not fully captured. This study examined the short-term impact of natural hazards where the impact measured occurs in the same year as the hazard. Thus, the impact on subsequent years is not considered. It is also possible that imprecision in damage estimates result in a lack of significance in the variables.

Discussion:

The literature tends to show that there is a mix of responses to natural hazards with there being increases in economic growth in some sectors and decreasing growth in other sectors (Loayza, Olaberria, Rigolini, and Christiaensen 2012; Mohan, Spencer and Strobl 2019; Benson and Clay 2004; Koks and Thissen 2016). Additionally, studies often find that there are either setbacks for economic growth or that disasters have no lasting perverse effects on economic growth (e.g., Hochrainer, 2009; Strobl, 2011). In the short run, there are a mix of results. Albala-Bertrand (1993) for instance, show a neutral or positive effect on economic growth (0.4 % effect) while Raddatz (2007) shows a negative effect from climatic events (−2 % effect on GDP per capita) and humanitarian events (−4 % effect on GDP per capita), but geological events were not statistically significant. Strobl (2008) shows an immediate −0.8 % impact from hurricanes. Noy (2009) shows a positive 1.33 % impact on GDP in the short run while Hochrainer (2009) shows a −0.5 % impact on GDP after the first year of an event.

The results from the local impacts on payroll, GDP, employment, and establishments in our study was, largely, not statistically significant and would be consistent with the mixed responses found in other literature on short-run impacts. The lack of significance could be that local businesses are not experiencing losses, or it could be the result of some businesses benefiting from increased demand while others experience losses. It is also possible that imprecision in the data or data variation results in a lack of statistical significance. Finally, the presence of hazards that have zero damage could make it difficult to detect economic losses. Moreover, there are a number of possible explanations for the lack of statistical significance for the local impacts.

Our study suggests that there are substantial annual losses to economic growth that result from hazards occurring in that year, particularly in the downstream supply chain for manufacturing and goods producing industries. The supply chain variables were statistically significant for manufacturing payroll, manufacturing GDP, manufacturing employment, total payroll, and total employment with declines ranging from 2.9 % to 8.6 % in the simulations. These losses are not the result of any single event, but rather a result of the aggregate annual occurrence of hazards. These results would be more consistent with research that suggests a negative impact from disasters such as Raddatz (2007) and Hochrainer (2009),

The losses shown in our study, particularly the supply chain losses, are not well represented in the existing literature because many studies that examine the economic impact of disasters at the macroeconomic-level have fundamental assumptions about the response to supply chain disruption and do not fully include the losses resulting from the immediate disorganization following a hazard event. Thus, some analyses tend to exclude the element that this paper aims to measure while this paper tends not to measure the elements measured in other papers. Many studies also tend to focus on single events whereas ours focuses on the repeating occurrence of events. These differences make it difficult to make comparisons. Other types of models (e.g., CGE and IO models) are indispensable for community resilience planning; however, our results suggest they should not be solely relied on for understanding total national macroeconomic losses due to hazard events. The model presented in this paper also suggests that economic losses grow in a non-linear fashion, which concurs with existing literature (Shughrue 2020; Koks 2016).

The effect of hazards along the supply chain appears to exceed that of the local impact, at least during the year of hazard occurrence. This can create a misalignment of incentives where the establishment that invests in mitigation efforts and experiences the hazard does not necessarily experience the majority of the impact. This could result in a less than optimal level of investment in hazard mitigation, as the stakeholder making the mitigation investment does not receive all the benefits. One might surmise that the risk of losing customers might drive an establishment to invest; however, the establishment would, likely, need some means for showing that they are a more reliable supplier because of their investments in mitigation in order to see substantial return on investment.

7. Conclusions

This study provides evidence supporting three of the eight proposed hypotheses:

• Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on payroll locally.

• Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on output locally.

• Natural/human-made hazards in the upstream supply chain (i.e., suppliers) have a negative effect on employment locally.

The hypotheses that natural/man-made hazards locally affect payroll, output, employment, or establishment formation/survival was not supported by the models. It is possible that either there is little effect on businesses or that there are both winners and losers that result from hazards. The latter seems the more likely given the results from examining downstream supply chain effects along with the results in the literature. The hypothesis that natural/man-made hazards in the supply chain, measured using supplies of goods with low substitutability, affect payroll, output (measured in GDP per capita), and employment was supported by the models; however, it was not supported in the model of total output, suggesting that it has a greater impact on certain industries. The models did not support the hypothesis that natural/man-made hazards in the upstream supply chain drive some establishments out of business locally and/or prevent establishment formation. It is important to note that there appears to be some imprecision in tracking hazard incidents and damage. This imprecision could result in some parameters being found statistically insignificant or underestimate the measured impact. For instance, hypothesis 1 through 4 or hypothesis 6 could be falsely rejected due to data imprecision. Additionally, the presence of hazards that have zero damage could make it difficult to detect economic losses.

The more severe impacts seem to be due to hazards in the manufacturing/goods industry supply chain, where GDP and employment declined 3.9 % and 8.6 %, respectively. Payroll for all industries was estimated to be higher than that for manufacturing; however, the 95 % confidence intervals have significant overlap suggesting that they may not be statistically different. The next largest effects were on the total economy where employment and payroll declined 3.0 % and 5.3 %, respectively, due to hazards in the supply chain. However, no discernable loss was observed in total GDP despite losses detected in goods GDP. Local damage due to local hazards was not significant in any of the models except for the payroll model for the total economy. The impact was minimal, however, and the 95 % confidence interval ranged from −0.7 % to 0.6 %.

The lack of statistical significance in local damage while supply chain damage is significant may seem perplexing at first; however, there are a number of possible reasons for this. For instance, at the local level, there may be winners and losers. Moreover, only limited conclusions can be made from the lack of statistical significance. Some studies may not capture the effects presented in this paper, as many papers focus on regional effects while this study is at the national level. Other approaches do not always capture the short-term effects (e.g., CGE and IO models) while this study does tend to capture them. Finally, this study aims to measure the impact of all hazards small and large across the entire United States while many other studies focus on single incidents. The implications of the results suggest that current methods (e.g., IO, SAM, and CGE models) do not alone represent total losses, particularly at the aggregated national level. We do not propose using correlation studies in place of these methods, but rather be used to supplement the current existing approaches, especially where there are limitations. Moreover, we propose that a full accounting of economic impacts will require multiple methods to estimate the different types of impacts. For instance, IO and CGE models can be used to estimate upstream impacts and long-term downstream impacts while correlation studies can be used to estimate short-term local impacts and short-term downstream impacts. Estimates of national aggregated losses, such as these, can be used to make investment decisions regarding hazard resilience, particularly investments by local, state, and national governments. It can also provide businesses insight into the threat that hazards pose to businesses and their supply chains for investment decisions. The findings from this paper also suggest there is a need to better understand the short-term downstream impacts from all hazards, especially at the aggregated level. In terms of research investments, the findings show that the ripple effect from natural hazards has a significant economic impact and that there is a potential market failure regarding investments in risk mitigation due to a misalignment of incentives. In terms of investments in natural hazard mitigation, the findings provide an estimate of the magnitude of impact from natural hazards and the ripple effect.

This paper has limitations in that it is not measuring total impacts of natural hazards but rather it focuses on short-term impacts locally and in the downstream supply chain. Although this paper presents evidence to support or refute a series of hypotheses, it is examining correlation, which does not prove causation. Therefore, additional research is needed to further substantiate findings.

Since the data for this analysis is from the US, it is difficult to generalize the findings to other locations. It is likely that other industrialized nations will have similar impacts from natural hazards where the impact from the ripple effect exceeds that of the direct impact; however, less developed nations are likely to have varying supply chain relationships. Also, the US is a large nation, both in terms of population and geography. The importance of imported supply chains is, likely, more critical for smaller nations. Also, the frequency and severity of natural hazards vary significantly by geographic location; thus, the magnitude of the impact of these incidents may vary significantly by nation. Additional research might reveal some of these differences in impact.

Future research might focus on identifying the impacts of hazards on individual industries, as this paper has identified that the risk varies by industry (i.e., between manufacturing/goods and the total economy). Some industries may be more vulnerable than others to supply disruption. If so, identifying the at-risk industries can facilitate targeted hazard mitigation efforts. In addition to vulnerability of the destination industry, there may be variations in vulnerability of the goods being supplied. Again, identifying the at-risk areas can facilitate targeted mitigation.

Additional research might also examine the impact of disasters outside of US national borders. The risks posed to domestic manufacturing by imported supply chains were not examined and may represent a greater/lower risk than that of domestic supply chains. Mitigating this risk may also require a different approach than that needed to mitigate domestic supply chain risk.

Another potential avenue of future research is in examining beyond the tier 1 supply chain. This paper focused on the impact of local hazards and hazards occurring in the suppliers to a geographic location. The tier 2 and beyond suppliers (i.e., the suppliers to the suppliers) are likely to add to this risk/impact.

Finally, future research might examine the effects of hazards beyond the initial year of impact. This paper focused on examining the impact of natural hazards in the year of hazard occurrence. Understanding the total impact of hazards on the economy requires further examination of hazard impacts in subsequent years.

Table 5B:

Results for Models 1 through 8

	Model Number and Dependent Variable
	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7	Model 8
	PRMAN	GDPGOODS	EMPMAN	ESTMAN	PRALL	GDPALL	EMPALL	ESTALL
INTRCT1e					0.0002**
INTRCT1f						0.0002
INTRCT1g							0.0000
INTRCT1h,LG								0.0000
INTRCT1h,MED								0.0005*
INTRCT1h,SMLL								−0.0006**
INTRCT2a	0.0015
INTRCT2b		0.0005
INTRCT2c			−0.0007
INTRCT2d,LG				0.0004
INTRCT2d,MED				0.0068**
INTRCT2d,SMLL				−0.006**
INTRCT2e					−0.0002
INTRCT2f						0.0000
INTRCT2g							−0.0001
INTRCT2h,LG								0.0001
INTRCT2h,MED								−0.0017
INTRCT2h,SMLL								0.0017
PRlag1,ALL					0.6654***
PRlag2,ALL					−0.0504***
PRlag1,MAN	0.645***
PRlag2,MAN	−0.0605***
UR			−0.1586***				−0.0637***
YR	0.0363***		−0.0109***	−0.0116***	0.0328***		0.004***	0.0087***
Constant	4.6407***	4.8598***	3.4381***	3.1044***	4.9185***	3.3967***	3.6417***	2.2103***

^***Significant at the 0.1 level,

^**Significant at the 0.05 level,

^*Significant at the 0.01 level