Hurricane fatalities and hurricane damages: are safer hurricanes more damaging?

Sutter, Daniel

1. Introduction

Hurricanes have long threatened the coastal areas of the United States. The nation has invested millions of dollars to understand and forecast hurricanes. Research efforts led by the National Hurricane Center (Simpson 1998) have succeeded in making land-falling hurricanes less deadly. In the 1990s, the modernization of the National Weather Service, featuring the installation of the Advanced Weather Interactive Processing System to process data from radar, satellites, and surface observations at high speeds and a nationwide network of Doppler weather radars, contributed to improved forecasts of weather hazards (Friday 1994). Annual hurricane fatalities have fallen from 0.5 per million residents nationally during the 1950s to 0.05 per million residents during the 1980s and 1990s. Kunkel, Pielke, and Changnon (1999) attribute the decline to improved hurricane forecasts. (1)

Although hurricanes have become less deadly over time, hurricane damages have increased, particularly in recent years. By 1995, hurricane damage in the 1990s had already exceeded total damage in the 1970s and 1980s combined. This escalation has led to interest among policy makers and researchers regarding the causes of increasing hurricane damages. Some observers attribute rising damages to an increase in the number and severity of hurricanes; for instance, a 1995 congressional report asserts that hurricanes "have become increasingly frequent and severe over the last four decades as climatic conditions have changed in the tropics" (cited in Pielke and Landsea 1998, p. 623). This explanation, however, is simply false. Katz (2002) for instance finds no statistically significant increase in the number of land-falling hurricanes over time. (2) And the period from 1991 to 1994 had the fewest tropical storms of any four-year period in the last 50 years.

Increasing societal vulnerability, that is, more people and wealth along hurricane-prone coasts, seems to explain increasing hurricane damages. Figure 1 illustrates the increase in coastal county populations during the 20th century. The figure graphs the population growth rates by decade for 130 Atlantic and Gulf coast counties and for the United States overall. As illustrated, the coastal counties grew faster than the nation in each decade. A wealthier population will also have more property vulnerable to destruction by a hurricane. Pielke and Landsea (1998), Changnon and Hewings (2001), and Katz (2002) find no time trend for hurricane damages after normalizing for changes in population and wealth in addition to inflation.

An understanding of increasing hurricane losses requires an explanation for the increase in coastal county populations, and several have been advanced. One is the rising standard of living in the United States: wealthier people will spend more on luxuries, such as living near the ocean. Another possibility involves low-probability event bias. Considerable evidence suggests that people do not behave according to expected utility theory with respect to low-probability, high-consequence events like hurricanes. Instead of considering the expected cost of these events, which is considerable, people act as if such events "couldn't happen to me" and treat the low probability as a zero probability (Kunreuther et al. 1978; Camerer and Kunreuther 1989). Finally, a number of government policies, including subsidized insurance, disaster assistance, and structural restoration measures (e.g., rebuilding roads and restoring beaches after storms) contribute to overbuilding on hurricane-prone coasts (Platt 1999). (3)

We consider an alternative explanation, one which, to our knowledge, has not been widely discussed, namely the very reduction in hurricane lethality. Through improved hurricane warnings, better evacuation, and engineering advances, the probability of fatalities has been reduced, thereby decreasing the expected cost of living along hurricane-exposed coasts. The law of increasing demand consequently explains at least a part of the increase in coastal populations. (4) We provide evidence of the impact of reduced hurricane fatalities on damages using a database of land-falling hurricanes in the United States between 1940 and 1999. We do not argue that reduced lethality is the exclusive cause of increasing hurricane damages, only that it is a contributing and overlooked factor. Our explanation utilizes the concept of offsetting behavior in response to an exogenous change in the riskiness of an activity, first proposed by Peltzmaal (1975) for automobile safety.

The remainder of this paper is organized as follows: Section 2 presents an expected utility model of a household's location choice and shows how a reduction in the probability of deaths from a hurricane makes a household more likely to live along a hurricane-prone coast. In particular, the effect of reduced fatalities will be greatest when the probability of a hurricane is highest. Section 3 explains our econometric model. We first estimate a time-varying measure of hurricane lethality in a Poisson model of hurricane fatalities. We then use this measure of lethality with a lag to explain hurricane damages. We also interact this measure with the probability of a hurricane. Section 4 presents the empirical results, and Section 5 offers a brief conclusion.

2. Hurricane Forecasts and Locational Choice

In this section we examine a simple model of household location choice to derive testable predictions concerning hurricane lethality and damages. Consider a representative household's choice to live on a hurricane-exposed coast. Let [pi] be the probability of a hurricane and let [sigma] be the probability that the household suffers a casualty given that a hurricane strikes the household's residence on the coast. Let I be the household's income, which we assume does not depend on location decision, and let L be the dollar value of property losses that occur if the household lives on the coast and their residence is struck by a hurricane. The household can purchase insurance against property damage. Let x be the dollar value of coverage purchased and let p be the price per dollar of coverage. The household's total premium is [p.sup.*]x, and the household receives a payment of x if a hurricane loss occurs. Let y denote the disposable income spent on consumption goods.

Household utility is a function of disposable income y, the household's location, and the household's state of health. Let [theta] denote the household's state of health, with [[theta].sup.h] indicating full health and [[theta].sup.i] indicating that the household has suffered a hurricane casualty. (5) We assume that utility is lower (and the marginal utility of income higher) when the household suffers a hurricane casualty. Let a superscript on the utility function designate the household's location choice, with c representing the hurricane-vulnerable coast and o the location away from the coast. Let [U.sup.c] (y,[theta]) be the household's expected utility if they choose to live on the coast, which can be written

(1) [U.sup.c](y, [theta]) = (1 - [pi]) x [U.sup.c](I - px, [[theta].sup.h]) + [pi] x (1 - [sigma] x [U.sup.c](I - L - px + x, [[theta].sup.h]) + [pi] x [sigma] x [U.sup.c](I - L - px + x, [[theta].sup.i])

We assume that x is the household's expected utility-maximizing insurance purchase. Utility if the household chooses to live inland is [U.sup.o](y,[[theta].sup.h]), which is the household's reservation utility level. The household will live on the coast if [U.sup.c](y,[theta]) [greater than or equal to] [U.sup.o](y,[[theta].sup.h]).

We examine the comparative statics of the household's location decision. Consider first the effect of a change in the probability of a casualty, [sigma]. Forecasts allow residents to evacuate in advance of an approaching hurricane, so improved warnings will reduce [sigma], but not the probability of a hurricane, [pi]. A change in [sigma] does not affect the reservation level of utility, [U.sup.o](y,[[theta].sup.h]). Thus, the effect on [U.sup.c](y,[theta]) is

(2) [differential][U.sup.c]/[differential][sigma] = [pi] x [[U.sup.c](I - L - px + x, [[theta].sup.i] - [U.sup.c](I - L - px + x, [[theta].sup.h]),

which is negative given that the marginal utility of income is higher when the household suffers an injury, [U.sup.c](y,[[theta].sup.i]) > [U.sup.c](y,[[theta].sup.h]) a typical assumption. A reduction in the probability of injury from a hurricane raises expected utility from living on the coast and will, ceteris paribus, increase the population on the vulnerable coast. If all households, including the new residents, suffer similar losses, L, the increase in population will increase the property damage from a hurricane. From Equation 2, we also see that the effect on utility of a reduction in o depends on the probability of a hurricane. Thus, a reduction in hurricane fatalities will have a greater impact in coastal areas facing a greater risk of hurricane landfall. (6) This leads to our main testable prediction.

An increase in income also affects the household's location choice. An increase in income increases the household's reservation level of utility, [differential][U.sup.o]/[differential]I > 0. The effect of an increase in income on the utility of living on the coast (ignoring the effect of the change in 1 on losses from a hurricane or insurance purchase) can be written

(3) [differential][U.sup.c]/[differential]I = (1 - [pi])[differential][U.sup.c](I - px, [[theta].sup.h])/ [differential]y + [pi](1 - [sigma])[differential][U.sup.c](I - L - px + x, [[theta].sup.h])/[differential]y + [pi][sigma][differential][U.sup.c](I - L - px + x, [[theta].sup.i])/[differential]y.

An increase in income raises the utility of living on the coast. With the standard assumptions of diminishing marginal utility of income and higher marginal utility of income given a lower state of health, then it follows that [differential][U.sup.c]/[differential]I, and an increase in income will increase coastal populations and hurricane property damage.

Finally, the effect of a change in the price of insurance, ignoring the effect on the quantity of insurance purchased, is

(4) [differential][U.sup.c]/[differential]p = -(1 [pi])[differential][U.sup.c](I - px, [[theta].sup.h]/[differential]y - [pi](1 - [sigma])[differential][U.sup.c](I - L - px + x, [[theta].sup.h])/ [differential]y - [pi] [sigma][differential][U.sup.c](I - L - px + x, [[theta].sup.i])/[differential]y.

An increase in the price of insurance lowers the utility of living on the coast, and the impact of the price change on the quantity of insurance purchased does not alter this result. Consequently, a reduction in the price of insurance because of a public subsidy or cross-subsidization in regulated insurance rates also increases coastal populations and hurricane damages. We lack a direct measure of insurance subsidy over time in different coastal areas. States regulate insurance companies, which suggests the value of including state fixed effects in our analysis of hurricane damage.

We noted earlier the reduction in hurricane lethality apparent in the raw time series of hurricane fatalities. We presume that improved forecasts and better evacuations are responsible for declining fatalities, but an improvement in construction techniques that allows buildings to better withstand hurricanes could also produce lower fatalities. Improved construction techniques would reduce both [sigma] and L; more households would locate on hurricane exposed coasts but lower losses per household imply that damages may not increase. Fronstin and Holtman (1994), however, found that newer subdivisions suffered greater damage in Hurricane Andrew, which indicates that construction techniques have not improved significantly.

3. Econometric Specification and Data

We employ a two-stage estimation procedure. We first estimate fatalities directly caused by a hurricane as a function of storm strength and other control variables. We also include decade dummy variables to capture changes in the lethality of hurricanes over time. We then estimate the determinants of hurricane damages to find whether a change in hurricane lethality affects damages.

Our data set is taken from the National Hurricane Center's archive of land-falling hurricanes in the United States. (7) Damage estimates are missing for a number of hurricanes prior to 1950, so we use hurricanes during 1940-1999 in our fatalities regression and 1950-1999 in our damage regression. Table 1 reports the breakdown of land-falling hurricanes by category on the Saffir-Simpson scale and by decade. The Saffir-Simpson scale measures the intensity of the hurricane and its destructive potential. Ratings on the scale are integer values from 1 to 5, with a category 5 hurricane the most intense, and are based on wind speed, storm surge, and potential damage. A category 1 storm is a minimal hurricane and has sustained wind speeds of 74-95 miles per hour and a 4-5 foot storm surge, while a category 5 hurricane has sustained winds in excess of 155 miles per hour and a storm surge in excess of 18 feet. Note that the damages corresponding to the five categories do not increase in linear fashion; a category 4 hurricane would be expected to cause 100 times the damage of a category 1 hurricane. (8) A total of 94 hurricanes made landfall between 1940 and 1999, with 73 striking between 1950 and 1999. Category 1 hurricanes (at landfall) were most common (32 of 94); only 7 storms reached Category 4 and one was rated Category 5. Mean fatalities were 24, with a median of 3 and range of 0 to 394. Mean damages were $1.54 billion, with a median of $242 million and range of $1.14 million to $28.8 billion (Hurricane Andrew in 1992).

Our first-stage regression estimates the determinants of the number of persons killed by a hurricane. We model the number of persons killed by hurricane i as follows:

(5) [Fatalities.sub.i] = f([Category.sub.i], [Density.sub.i], [D40.sub.i], [D50.sub.i], [D60.sub.i], [D70.sub.i], [D80.sub.i]).

Fatalities is the number of persons directly killed by hurricane i and does not include deaths from inland flooding. Category is the rating of the hurricane on the Saffir-Simpson Hurricane scale at the time of landfall. Density is the average population density in persons per square mile of the counties struck by the hurricane, as listed in the National Hurricane Center's hurricane archive. The population for a county in a given year was estimated using linear interpolation from the decennial censuses. A higher population density of the storm path should increase the number of fatalities. D40, D50, D60, D70 and D80 are dummy variables that equal one if the hurricane occurred in the decades 1940s, 1950s, 1960s, 1970s or 1980s respectively, or zero otherwise, with the 1990s the omitted decade. Thus, we allow the lethality of hurricanes to vary over the decades, with the decade dummies capturing the effects of improved hurricane warnings and public dissemination of these warnings. We expect that hurricanes have become less lethal over time, so we expect positive coefficients on the decade dummy variables, with the magnitude of the coefficients becoming smaller.

The number of fatalities produced by a hurricane is a count variable, taking on integer values with a high proportion of zeros. Of the 94 hurricanes in our sample, 23 produced no direct fatalities, and the median number of fatalities is 3 compared with a mean of 24.3. Thus we estimate the fatalities function using Poisson regression (Greene 2000, pp. 880-886). The Poisson model assumes that the number of persons killed by hurricane i, [Y.sub.i], is distributed as a Poisson random variable. The probability of a given number of fatalities is

(6) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII.]

The parameter [[lambda].sub.i] depends on the vector of independent variables [x.sub.i] described above.

Our second stage estimates the determinants of property damage caused by a hurricane. We model damages as follows:

(7) [Damage.sub.i] = [[beta].sub.0] + [[beta].sub.1]x[Category.sub.i] + [[beta].sub.1]x[Density.sub.i] + [[beta].sub.3]x[Income.sub.i] + [[beta].sub.4]xYear + [[beta].sub.5]xRF[R.sub.i] + [[beta].sub.6]xP[H.sub.i] + [[beta].sub.7]RF[R.sub.i]*P[H.sub.i] + [[epsilon].sub.i].

Damage is the value of property damage caused by the hurricane in millions of dollars, adjusted for inflation using the GDP deflator. Category is the rating of the hurricane on the Saffir-Simpson scale; we expect that stronger hurricanes will produce more damage, [[beta].sub.1] > 0. Density is the population density of the counties affected by the hurricane and is expected to increase damages, [[beta].sub.2] > 0. Income is the per capita income of the counties struck by the hurricane. Because the value of real and personal property on a high-income coastal area is higher, the dollar value of damage should be higher, [[beta].sub.3] > 0. But higher-income individuals will also spend more to protect themselves and their property against hazards, which could reduce total damage. Thus, either a positive or a negative value for [[beta].sub.3] could be observed. Year is a time trend included to capture any effects of improved construction techniques or changes in building codes over time that might affect property damage. RFR is our time-varying measure of the deadliness of hurricanes, based on the coefficient point estimate of the decade dummy variable from our first stage estimation. A decline in hurricane fatalities reduces the cost of living on a hurricane-prone coast, so we expect that this will increase coastal population and damages. Clearly a lag is required for people to recognize that hurricanes have become less deadly and move into hurricane-exposed coasts. Consequently, we use the coefficient from the previous decade's dummy variable as the RFR for a hurricane in year t. Thus, the coefficient on D70 in the fatalities regression is the value of RFR for any hurricane occurring during the decade of the 1980s. We follow Sobel and Nesbit's (2002) investigation of offsetting behavior in NASCAR racing. They use the number of fatalities divided by the number of accidents for the previous 110 races as a measure of the recent fatality rate. We must control for the strength of the hurricane and set some time limit for recent hurricanes because of the randomness in the occurrence of land-falling hurricanes. PH is an estimate of the annual probability of a major hurricane at different points along the coastline. This variable was taken from estimates for various cities along the Atlantic and Gulf coasts contained in Sheets and Williams (2001). In the expected utility model, an increase in [pi] ceteris paribus reduces the utility of living on the coast, but we observe different [pi]'s at different locations, so the utility of living on these different stretches of coast may differ, rendering a prediction for PH difficult. The expected present value of hurricane loss reduction mechanisms, for instance, will depend on the annual probability of a hurricane. If more hurricane-prone areas employ better building techniques or other loss-reduction mechanisms, PH will have a negative value. Alternatively, if hurricane-prone states subsidize or cross-subsidize hurricane insurance, PH could have a positive value. RFR*PH is an interaction term capturing the combined effect of the recent fatality rate and probability of a hurricane. (9) A decrease in hurricane lethality will have its greatest impact on damages in the most hurricane-prone coastal areas. A negative value on this interaction term, [[beta].sub.7] < 0, provides the sharpest test of the damage augmenting effect of hurricane forecasts and warnings.

4. Results

Table 2 presents our first-stage Poisson estimates of hurricane fatalities. Not surprisingly, Category is a positive and highly significant determinant of fatalities; a one-category increase in the strength of a hurricane almost triples expected casualties. Density is also positive and significant at better than the 1% level. As expected, hurricanes that strike more highly populated coastal areas are more deadly. The decade dummy variables are all statistically significant at better than the 1% level, except D70, which is significant at only the 10% level. All of the decade dummies are positive except D80, which is negative and significant. Roughly speaking, a downward trend in hurricane lethality is evident, because the coefficients on D40 and D50 are the largest, whereas the 1980s and 1990s are the least lethal decades. The differences between the decade dummy variables are significant at the 5% level as well, so from the 1950s through 1980s we see consistent and statistically significant reductions in lethality each decade.

Table 3 presents our second-stage ordinary least squares estimation of hurricane damages. (10) The first column displays estimates using the point estimates of the dummy variables from Table 2 as the RFR variable. All of the control variables are significant at the 10% level or better. Category and Density have positive values, so a stronger hurricane striking a more densely populated coast will cause greater damage, as expected. A one-category increase in the strength of a land-falling hurricane increases expected damages by about $1.4 billion, which is just less than the mean damage of all hurricanes in the sample of $1.54 billion. Income has a negative association with damages. Although the value of real and personal property is higher in higher income areas, wealthier residents seem to take more precautions to mitigate hurricane losses. Because windborne debris is a major contributor to structural damage, destruction of poorly constructed homes can damage other structures in the neighborhood. The negative sign on Income is actually consistent with Fronstin and Holtman's (1994) result that subdivisions with higher average home prices suffered less damage in Hurricane Andrew. Year has a positive coefficient, so, ceteris paribus, more recent hurricanes have been causing greater damage, which is also consistent with Fronstin and Holtman's (1994) finding that newer subdivisions suffered greater damage in Hurricane Andrew. Year may be capturing the effect of increasing wealth over time, with our Income variable capturing the cross-sectional impact of wealth on losses. The coefficient on PH, the probability of a major hurricane, is positive and significant. After controlling for category, population density, and income, regions with a higher probability of a hurricane still suffer greater damages. (11) This is a surprising result because durable loss-reduction measures such as strengthened building techniques and hurricane shutters have higher expected benefits in more hurricane-prone regions and thus should be more likely to be installed (or to have their installation mandated). Our result is consistent with possible insurance cross-subsidization or weak enforcement of building codes in hurricane-prone regions.

Our measure of recent hurricane lethality provides evidence on offsetting behavior. RFR has a positive and significant (at the 5% level) direct effect on damages and a negative and significant (at the 1% level) effect when interacted with the probability of a major hurricane. The interaction coefficient provides the strongest test of the role of reducing the lethality of hurricanes or hurricane damages, and we see that the reduction in the lethality of hurricanes does increase damages in the following decade in more hurricane-prone regions. (12) The marginal effect of a decrease in RFR becomes positive when the annual probability of a major hurricane exceeds about 3.9%, a threshold exceeded in most counties of south Florida and along the Texas gulf coast. The magnitude of the impact of the declining fatality rate on damages is quantitatively quite significant. The increase in expected damages because of the observed decline in the fatality rate is $5.1 billion when the probability of a major hurricane is 7% and $10.9 billion when the probability of a major hurricane is at its maximum of 10.5%. (13,14)

We also estimated the damage model using the lower bounds and upper bounds of the 95% confidence intervals for the estimates of the coefficients of the decade dummy variables to determine if our results were robust to plausible changes in the estimated lethality of hurricanes. The second and third columns of Table 3 present the results. Our results are not affected in any substantial way. The estimated impact of the observed decrease in hurricane lethality with a 7% probability of a major hurricane is $4.8 billion with the lower bounds estimate and $5.5 billion with the upper bounds.

The potential for state policies, particularly regulation of the insurance industry, to create subsidies for living on hurricane-exposed coasts was noted in Section 2. To explore this possibility, we created state effect variables. Because some hurricanes struck more than one state, the state variables were defined to equal the fraction of the population of the area struck by the hurricane living in that state, based on the counties listed for each storm. The fourth column of Table 3 presents this estimation, which uses the point estimates of the decade dummy variables for the RFR variable, with the state variables omitted to conserve space. Inclusion of state effects does not affect the estimates very much at all, and the state variables are both individually and jointly insignificant. (15) The state effects model does produce a slightly higher estimate of the impact of the observed reduction in hurricane lethality on damages of $5.6 billion (with a 7% probability of a hurricane), compared to $5.1 in the model in column 1.

5. Conclusion

Economists since Peltzman (1975) have identified a number of offsetting behaviors, such that as technology or regulation reduce the full cost of risky behavior, people will engage in more of the risky behavior. We have considered an application of offsetting behavior to natural hazards, and specifically hurricanes. Advances in meteorology, engineering, and emergency management have combined to make hurricanes less deadly over time. Yet if hurricanes are less likely to produce fatalities and injuries, living along an exposed coast becomes more inviting and coastal populations will increase. Therefore, hurricanes will kill fewer people but will produce more property damage. We offer evidence for this proposition through an analysis of land-falling hurricanes in the United States between 1940 and 1999. Our results suggest that the reduction in hurricane lethality has a statistically significant and quantitatively large effect on damages on the portions of the coast most prone to hurricanes.

Scientific or engineering approaches to natural hazards can sometimes exacerbate hazards (Mileti 1999). Improved weather forecasts and other measures that reduce hazard deaths provide obvious benefits to society. But offsetting behavior will increase societal vulnerability, leading perhaps to an increase in damages. We have examined only the case of hurricanes here, but offsetting behavior should lead to a lethality/damage tradeoff for other hazards.

Increasing populations along exposed coasts provide a potential new hurricane hazard. As Dow and Cutter (2002) stress, the growth of coastal populations threaten to exceed the capacity of the highway infrastructure to allow timely evacuation. Indeed, the prospect of massive traffic jams affected residents' evacuation decision in advance of Hurricane Floyd in 1999. Traffic congestion, the impact of a household's decision to live along the coast on others' ability to evacuate, is a negative externality that households are unlikely to take into account. Thus, even if residents bear the full expected cost of hurricane damage, an evacuation externality might result in greater than optimal coastal populations, and be exacerbated as hurricanes become less deadly.

A reduction in the lethality of hurricanes may increase expected hurricane damages but still raise social welfare. If the risk to life and limb deterred some prospective residents from living along a hurricane-exposed coast, this is also a social cost of hurricanes in addition to property damage. But the risk to life and limb is one borne by residents, whereas other costs of hurricanes can be externalized. If the regulation of insurance or disaster relief subsidizes coastal living, however, making hurricanes less deadly can lower social welfare. As hurricanes become less deadly, the cost to society of socializing property losses increases.

References

Camerer, Colin F., and Howard Kunreuther. 1989. Decision processes for low probability events: Policy implications. Journal of Policy Analysis and Management 8:565-92.

Changnon, Stanley A., and Geoffrey J. D. Hewings. 200l. Losses from weather extremes in the United States. Natural Hazards Review 2: 113-23.

Dow, Kirstin, and Susan L. Cutter. 2002. Emerging hurricane evacuation issues: Hurricane Floyd and South Carolina. Natural Hazards Review 3:12-8.

Friday, Elbert W. Jr. 1994. The modernization and associated restructuring of the National Weather Service: An overview. Bulletin of the American Meteorological Society 75:43-52,

Fronstin, Paul, and Alphonse G. Holtman. 1994. The determinants of residential property damage caused by Hurricane Andrew. Southern Economic Journal 61:387-97.

Garrett, Thomas A., and Russell S. Sobel. 2003. The political economy of FEMA disaster payments. Economic Inquiry 41: 496-509.

Greene, William H. 2000. Econometric analysis. 4th edition. Upper Saddle River, NJ: Prentice Hall, pp. 880-6.

Katz, Richard W. 2002. Stochastic modeling of hurricane damage. Journal of Applied Meteorology 41:754-62.

Kunkel, Kenneth E., Roger A. Pielke Jr., and Stanley A. Changnon. 1999. Temporal fluctuations in weather and climate extremes that cause human health impacts: A review. Bulletin of the American Meteorological Society 80:1077-98.

Kunreuther, Howard, R. Ginsberg, L. Miller, P. Sagi, P. Slovic, B. Borkan, and N. Katz. 1978. Disaster insurance protection: Public policy lessons. New York: Wiley & Sons.

Mileti, Dennis S. 1999. Disasters by design. Washington, DC: Joseph Henry Press, pp. 17-40.

Pagan, Adrian. 1984. Econometric issues in the analysis of regressions with generated regressors. International Economic Review 25:221-47.

Peltzman, Sam. 1975. The effects of automobile safety regulation. Journal of Political Economy 83:677-725.

Pielke Jr., Roger A., and Christopher W. Landsea. 1998. Normalized hurricane damages in the United States, 1925-1995. Weather and Forecasting 13:621-31.

Platt, Rutherford H. 1999. Disasters and democracy: The politics of extreme natural events. Washington, DC: Island Press, pp. 11-41.

Sheets, Bob, and Jack Williams. 2001. Hurricane watch: Forecasting the deadliest storms on earth. New York: Vintage Books, pp. 292-4.

Simmons, Kevin M., Jamie B. Kruse, and Douglas A. Smith. 2002. Valuing mitigation: Real estate market response to hurricane loss measures. Southern Economic Journal 68:660-71.

Simpson, R. H. 1998. Stepping stones in the evolution of a national hurricane policy. Weather and Forecasting 13:617-20.

Sobel, Russell S., and Todd M. Nesbit. 2002. Automobile safety and the incentive to drive recklessly: Evidence from NASCAR. Unpublished paper, West Virginia University.

(1) The National Hurricane Center maintains a continuous watch for tropical cyclones throughout hurricane season, May 15 through November 30. The Center issues watches and warnings for hurricanes threatening landfall, and orders evacuations based on the warnings. Throughout the remainder of the year, the Center provides training for emergency managers from the United States and other countries affected by tropical storms and conducts research on hurricanes and forecasts.

(2) See also our Table 1 reporting land-falling hurricanes in the U.S. by decade.

(3) Garrett and Sobel (2003) document political influence on presidential disaster declarations and the dollar value of disaster assistance provided under the Stafford Act.

(4) Imagine an island exposed to frequent hurricanes. Absent any type of hurricane forecast, residents of the island could be surprised any time during the hurricane season. Under such circumstances, the island may remain uninhabited, but it may well become inhabited once residents can be warned in time to evacuate from an approaching hurricane.

(5) In this simple formulation, we consider all casualties equivalent. Gradations of casualties could be introduced but would not affect the testable hypotheses derived here.

(6) Fronstin and Holtman (1994) argue that an ability to evacuate from an approaching hurricane encourages residents to substitute lower quality construction, which would provide an additional method by which improved forecasts can increase damages. Note that the effect of a decrease in the probability of hurricane casualties for a household on the overall number of casualties is theoretically ambiguous because of the Peltzman (1975) offsetting behavior effect.

(7) The hurricane archive was accessed at http://www.nhc.noaa.gov/pastall.shtml.

(8) For details on the Saffir-Simpson scale, see www.nhc.noaa.gov/aboutsshs.shtml.

(9) On market incentives for the installation of loss-reduction measures like hurricane shutters, see Simmons, Kruse, and Smith (2002).

(10) A Breusch-Pagan heteroscedasticity test failed to reject the null hypothesis of homoscedasticity at even the 10% significance level. The test statistic was 44.44, with a p-value of 0.1085.

(11) Note that, because of the interaction term, the partial effect of hurricane probability on damages becomes negative if RFR is greater than 1.21, which it is with the 1950s value. We also estimated the damages model using an estimate of the probability of any hurricane also reported in Sheets and Williams (2001), because we do not know a priori what measure of hurricane risk people might use in estimating B. The signs of the estimated coefficients were the same as reported in Table 3, but the model overall did not perform as well, with an adjusted [R.sup.2] of only 0.199. Consequently, we conclude that the probability of a major hurricane seems to approximate the public's subjective measure of hurricane risk.

(12) We also estimated the fatalities model using a linear time trend and constructed an RFR variable in this fashion. The time trend variable had a negative and significant sign in the fatalities equation, and the interaction term in the damages regression was again negative and significant.

(13) The observed reduction in the hurricane fatality rate is assumed to equal the difference between the mean of the point estimates of D40 and D50 and the point estimate of D80 and the omitted decade, the 1990s, so [DELTA]RFR = -1.38.

(14) Our use of an estimated parameter from the first stage as our RFR variable creates the potential for a generated regressor bias as noted by Pagan (1984), which could bias the estimate of the standard errors downward. Unfortunately, there is no widely accepted correction for this type of bias in our type of model. To examine the robustness of our results, we estimated our models using Newey-West and White's standard error. The interaction term remained significant in both cases, at the 10% level using Newey-West standard errors and at the 10% level in a one-tailed test with White's standard errors.

(15) Both Wald and F-tests failed to reject the null hypothesis of joint insignificance of the state variables at even the 10% level. The test statistic for the Wald test was 14.82 with 13 degrees of freedom and a p-value of 0.3185, and the test statistic for the F-test was 1.140 with a p-value of 0.3488.

Nicole Cornell Sadowski * and Daniel Sutter ([dagger])

* Department of Economics, University of Oklahoma, Norman, OK 73019-2103, USA; E-mail nicole.L.cornell-1@ou.edu; Present address: Department of Business Administration, York College, York, PA 17405, USA.

([dagger]) Department of Economics, University of Oklahoma, Norman, OK 73019-2103, USA; E-mail dsutter@ou.edu; corresponding author.

We would like to thank Robin Grier, Cindy Rogers, Aaron Smallwood, two referees, and session participants at the 2003 SEA meetings for useful comments on an earlier draft.

Received February 2004; accepted February 2005.

Table 1. Land-Falling U.S. Hurricanes using Saffir-Simpson Scale,
by Decade

 Number of Storms, by Category

Decade 1 2 3 4 5 Total

1940s 5 8 7 1 0 21
1950s 4 1 8 2 0 15
1960s 4 5 3 2 1 15
1970s 6 2 4 0 0 12
1980s 8 2 4 1 0 15
1990s 5 6 4 1 0 16

Total 32 24 30 7 1 94

Table 2. Poisson Regression of Hurricane Fatalities

 Standard 95%
Independent Variable Estimate Error Confidence Interval

Category of hurricane 1.081 ** 0.0255 1.031 1.131
Population density 0.0007180 ** 0.0000 0.0006 0.0008
1940s dummy (D40) 0.9937 ** 0.0912 0.8149 1.173
1950s dummy (D50) 1.354 ** 0.0884 1.181 1.528
1960s dummy (D60) 0.4865 ** 0.0965 0.2974 0.6757
1970s dummy (D70) 0.2145 * 0.1270 -0.0344 0.4634
1980s dummy (D80) -0.4082 ** 0.1286 -0.6602 -0.1562
Intercept -0.6580 ** 0.1130 -0.8794 -0.4366

Number of observations = 94. Dependent variable is the natural
logarithm of expected fatalities.

* Significant at the 10% level.

** Significant at the 1% level.

Table 3. Analysis of Hurricane Damages

Independent Variable RFR Point Estimates RFR Lower Bounds

Category of hurricane 1427 ** (3.98) 1430 ** (3.99)
Population density 1.322 * (2.09) 1.312 (2.07)
Income -0.3177 * (2.15) -0.3189 * (2.16)
Year 141.4 * (1.95) 143.4 * (1.95)
Recent fatality rate (RFR) 4675 * (2.28) 4474 * (2.27)
Probability of major
 hurricane (PH) 1454 ** (3.90) 1181 ** (3.78)
RFR*PH -1199 ** (3.39) -1138 ** (3.38)
Intercept -6038 * (2.30) -5032 * (2.13)
Adjusted [R.sup.2] 0.3141 0.3134

Independent Variable RFR Upper Bounds State Fixed Effects

Category of hurricane 1422 ** (3.97) 1386 ** (3.64)
Population density 1.333 * (2.10) 6.639 ** (2.89)
Income -0.3161 * (2.15) -0.3737 * (2.05)
Year 139.0 (1.96) 160.2 (1.66)
Recent fatality rate (RFR) 4883 * (2.30) 5610 * (2.15)
Probability of major
 hurricane (PH) 1755 ** (3.93) 1767 ** (4.15)
RFR*PH -1266 ** (3.40) -1385 ** (3.62)
Intercept -7125 * (2.42) -1234 (0.06)
Adjusted [R.sup.2] 0.3149 0.3421

The first column presents estimates using the point estimates of the
Recent Fatality Rate variable from Table 2, and the second and third
columns use the lower and upper bounds of the 95% confidence interval
of the estimates from Table 2. The fourth column includes state fixed
effects, which are not presented here to conserve space. Number of
observations = 73. t-statistics are in parentheses. Source: Author
estimation using U.S. Census Data.

* Significant at the 10% level.

** Significant at the 1% level.