Environmental & Resource Economics (2006) 33: 273–297 DOI 10.1007/s10640-005-3610-4
Ó Springer 2006
Can Natural Experiments Measure Behavioral Responses to Environmental Risks? JARED C. CARBONE, DANIEL G. HALLSTROM and V. KERRY SMITH* Department of Agricult. and Resource Econ., North Carolina State University, Box 8109, Raleigh, North Carolina, 27695-8109, USA; *Author for correspondence (e-mail: kerry_
[email protected]) Accepted 15 May 2005 Abstract. Efforts to measure people’s responses to spatially delineated risks confront the potential for correlation between these risks and other, unobserved characteristics of these locations. The possibility of correlation arises in part because individuals observe other locational attributes that can be expected to influence the hedonic equilibrium. One response to this problem is to use events from nature to exploit both temporal and spatial variation in the behavioral responses of interest. This paper evaluates the use of hurricanes as a source of new risk information to households in coastal counties potentially subject to the effects of these storms. We study the extent to which housing prices before and after hurricane Andrew, a hurricane with unprecedented property loss, reveal how Floridians responded to the risk information provided by the storm. Two counties are selected – one without and another with damage from the hurricane. To evaluate the plausibility of using quasi-random experiments for locations not directly affected by natural events, we compare Lee County’s results to those of Dade County, where the majority of the damage occurred. Our findings suggest, after controlling for ex post storm damage and changes in insurance markets, there is a reasonably high level of consistency in a repeat sales model’s ability to estimate the effects of the risk information conveyed by the storm for both counties. Key words: hurricane risk, repeat sales, hedonic models JEL classification: Q51, Q54
* Department of Economics, Williams College, Affiliated Economist, CEnREP, North Carolina State University and University Distinguished Professor, North Carolina State University, and Resources for the Future University Fellow, respectively. Senior authorship is not assigned. Thanks are due to Shelby Gerking and two anonymous reviewers for careful and constructive comments that substantially improved the paper. Michael Darden and Jaren Pope provided excellent research assistance and Alex Boutaud and Susan Hinton helped to make sense out of numerous drafts of this work. Smith’s contribution was partially supported by the United States Department of Homeland Security through the Center for Risk and Economic Analysis of Terrorism Events (CREATE), grant number EMW-2004-GR-0112. However, any opinion, findings, and conclusions or recommendations in this document are those of the author(s) and do not necessarily reflect views of the U.S. Department of Homeland Security.
274
JARED C. CARBONE ET AL.
1. Introduction An ideal test of any factor hypothesized to influence behavior would randomly vary it among a sample of the people who are expected to respond. Economic tests rarely meet this standard. As a rule, we must be satisfied with quasi-random experiments, where control is not guaranteed to meet the standards of a completely random assignment. Often, these experiments arise from natural events or exogenous changes that provide clear-cut sources of variation in the effect of interest. As Meyer (1995) observes, the careful logic developed as part of these quasi-experiments usually documents why the variation in a treatment can be considered to be exogenous. It emphasizes the importance of understanding how measured effects are identified. Recent uses of this logic in environmental economics have relied on differences due to policy decisions (Chay and Greenstone 2005), variations in aggregate economic conditions (Chay and Greenstone 2003), and natural events (see Beron et al. 1997 for the case of earthquakes). In each case the unanticipated change serves to designate who receives a variation in the treatment. As a result, it is argued that the analyst can be reasonably confident assuming the effect experienced was not the result of a separate behavioral choice that is related to the tradeoff estimates (or the tests) derived from the analysis. When these analyses involve economic outcomes such as housing values and treatments that arise from differences in non-market environmental amenities, two assumptions are important to the chain of logic that assures the experiment isolates the effect being measured. The first concerns the hypothesis that is usually of direct interest. People are assumed to recognize the ‘‘amount’’ of the amenity and respond to it. Second, the analysis requires some phenomenon to provide a consistent and recognizable variation in the amenity (or disamenity) of interest. This paper summarizes the finding of an ongoing research program on hurricanes as a source of risk information to coastal residents.1 Our objective here is to describe how this second assumption can be important. In our case the exogenous source of variation in information is a hurricane. Before it happens, people who live or wish to live along the coast have one set of risk beliefs. After, they have another. Our question is simple – what is the spatial extent of the new information provided by a large storm? Do we assume it is limited to the area directly affected by the hurricane or does it extend beyond these locations? To investigate these questions we use the largest hurricane (in terms of insurance industry losses) in U.S. history, Hurricane Andrew. We consider two counties in Florida: the one that was hit (Dade County) and experienced about 20 billion 2004 dollars in insured losses, and a second where the county was close to the hurricane’s landfall path in Florida (Lee County) but far enough that residents did not experience structural damage.2 We use a repeat
CAN NATURAL EXPERIMENTS MEASURE BEHAVIORAL RESPONSES
275
sales model for each county separately as well as the spatial location of private homes that sold twice between 1983 and 2000. Our test considers whether the new risk information we attribute to Andrew influenced the change in housing prices for those sales that bracket Andrew and were in flood hazard zones versus those that did not satisfy both conditions. A comparison of separate models for each county provides a plausibility check on our research strategy. While each analysis faces specialized considerations, a comparison of our estimates of the effects of the hurricane, as new information, after these added controls are taken into account, provides one means to confirm the general logic. In contrast to many other sources of environmental risk, insurance is available for the property losses due to a number of natural hazards, including flood and hurricane damage. Our analysis of the effects of risk information controls for the effects of changes in insurance terms. Finally, in the case of Dade County, to assure we can separate the effects of damage from new information, a unique database is used to control for the effects of storm damage. That is, homes damaged by the storm may have included modernization and structural upgrades in the repairs. Thus, price differences measured for sales that bracket the storm reflect both the effects of risk information and any structural changes arising from how repairs were undertaken. The analyses in both counties confirm that Andrew appears to have altered households’ perceptions of the risks of living in coastal locations. The estimated effect of Andrew is remarkably consistent for the two counties, ranging from about 20 to 30 percent reduction in the pace of increase of housing prices for the highest flood risk locations. Efforts to control for the effects of insurance have a large impact on the estimated effect of the risk information in Dade County, but not in Lee County. Section 2 outlines a simple expected utility model for describing how risk preferences can be deduced from a hedonic price function. Section 3 reviews the basic framework for our natural experiment and describes how it is implemented in a repeat sales hedonic model. Section 4 discusses our data and Section 5 summarizes the findings. We conclude with some discussion of the implications of our overall research program. 2. Behavioral Responses to Risk Economic models of behavioral risk adjustments hypothesize that people have incentives to respond to risks they know about. The nature of their response will depend on the costs and benefits of averting the source of the risk or mitigating its potential impact. If these behaviors influence housing prices, then the analyst may have sufficient information to measure how people tradeoff money for risk.
276
JARED C. CARBONE ET AL.
The hedonic property value model is one of the earliest economic frameworks to offer a way to include location with its summary of behavioral choices among different types of houses. Early discussions of the model sought to describe the price function as a reduced form relationship for a matching equilibrium. A key concern has been the ability to recover measures for individual tradeoffs among attributes from the derivatives of hedonic price equations (see Rosen (1974) and Ekeland et al. (2004)). When the attributes associated with a house are ‘‘attached’’ because of its location, there may be other reasons for questioning our ability to estimate their effects on price without bias. Rosenzweig and Wolpin’s (2000) review makes the problem’s source clear. Suppose property values, R, are related to the risk, p, of an undesirable outcome. This risk is conveyed by the location of each home. The hedonic price function assumes the relationship between prices and spatially differentiated risks is the result of an equilibrium matching of buyers and sellers. The risk is likely to be correlated with important, but unknown, sources of variation in the attributes of the property as well as the characteristics of both the buyers and the sellers. Attitudes towards risk, differences in households’ ability to respond to risk, and variation in their knowledge are all good candidates for the source of this unobserved heterogeneity.3 As a result, ordinary least squares (OLS) estimates of this equilibrium relationship may not provide unbiased measures of the marginal value of a risk reduction to each participant in a matched exchange.4 It is also reasonable to expect in some cases that the locations which are associated with higher risk may have other desirable attributes. For example, homes on or near the ocean have improved access for recreation, can have better views of ocean vistas, and may even have better air quality. Risks of storm damage and flooding are higher for these locations as well. When there are opposite effects correlated with distance to the coast, additional information is required to recover an estimate of the effect of risk. Of course, it is important to recognize that the interpretation of the estimated marginal effect depends in part on the identifying information. To help fix these ideas we use a simple expected utility model. Households have state-dependent utility functions over two outcomes. In one state, a serious event takes place, while in the other there is no such event. In our case, the event is a severe hurricane. A homeowner’s expected utility is given in Equation (1).5 V ¼ pðr; IÞ UH ðr; h; m Rðr; h; i0 ; pðr; IÞÞ Lðr; h; i0 ÞÞ þ ð1 pðr; IÞÞ UNH ðr; h; m Rðr; h; i0 ; pðr; IÞÞÞ where
ð1Þ
CAN NATURAL EXPERIMENTS MEASURE BEHAVIORAL RESPONSES
277
m=income (or wealth) less any hazard insurance expectations h=housing characteristics r=site attributes that can relate to both the risk of storm hazards and coastal amenities (e.g. on the coast or a canal) R(.)=hedonic price function (measured in annual terms if m is annual income) i0=insurance rate per dollar of coverage L(.)=monetary loss due to storm p(r,I)=household’s subjective probability of hurricanes at a given location and with specific information set (I) Uj (.) = utility for state j, j = H and NH H labels the utility realized with a case of a hurricane and NH no storm. Both the level of utility and the marginal utility of income may change with the state. We assume that households maximize expected utility by selecting r and h conditional on the information (I), insurance rates (i0), and their income. Assuming an equilibrium that matches buyers and sellers with each fully informed, the marginal price of the attribute r is: Rr ¼
pUHr þ ð1 pÞUNHr pUHm Lr pUHm þ ð1 pÞUNHm pUHm þ ð1 pÞUNHm pr ðUH UNH Þ þ pUHm þ ð1 pÞUNHm
ð2Þ
where UHi indicates partial derivative of UH with respect to i. This equation illustrates part of the difficulty with interpreting marginal prices (Rr) with respect to locational influences (r) on risk. If we interpret r as a locational attribute, such as distance to the coast, then we would expect that changes in r would lead to a change in hurricane risk. However, the behavioral interpretation for this partial derivative is more complex. It can reflect a composite of any risk changes along with any other contributions that coastal amenities make to individual well-being aside from risk (the first term). There may also be aspects of a location that contribute to losses from a storm (the second term). Moreover, in this simple interpretation we are implicitly assuming that the ‘‘correct’’ equilibrium hedonic price function is known. To meet our overall objectives, we would like to recover an estimate of Rp, the ex ante incremental option price for a risk change. As noted, spatial differences in prices in relation to a measure of distance from the coast are unlikely to provide a reliable basis for gauging behavioral response to risk. Hurricanes as a source of new risk information offer an opportunity to use the logic of natural experiments to measure RI = RpÆ pI. Combining temporal variation in risk perceptions with spatial variation in risk characteristics can also help in avoiding the endogeneity problems discussed earlier. Suppose an exogenous event occurs that provides new
278
JARED C. CARBONE ET AL.
information about a location’s risk. In equilibrium, the marginal impact of that new information on the hedonic price is: RI ¼
pI ðUH UNH Þ pUHm þ ð1 pÞUNHm
ð3Þ
Dividing both sides of Equation (3) by pI provides RI/pI. This relationship measures the incremental option price an individual would be willing to pay as an ex ante premium for a home that is located so as to reduce the risk of damage and disruption from severe coastal storms. This incremental value is revealed through the response of housing values to information. It also requires an external estimate for how p changes with new information. Hence, a composite of spatial attributes and changes in information about risk (using the framework in Equation (3)), along with a measure of pI offer the elements needed to measure the ex ante incremental value of risk reductions. In the context of severe hurricanes it is important to acknowledge that insurance terms may also change.6 Thus, the effect we describe in Equation (3) as exclusively associated with information may also lead to changes in i0. In this case, the measurement of the incremental option price for a risk change would require taking into account these insurance related effects as separate influences on price. In simple terms, if i0 also responds to I and we ignore changes in insurance coverage, then Equation (4) would replace (3) as the description of the information effect. This process assumes that the appropriate interpretation of the housing price change is a response to both the storm as information and as a stimulus for adjustment in the insurance market. RI ¼
pI ðUH UNH Þ pUHm Li0 i0I Ri0 i0I pUHm þ ð1 pÞUNHm pUHm þ ð1 pÞUNHm
ð4Þ
In the absence of considerable background on the insurance market, the best we can hope for is to ‘‘take account’’ of the last two terms on the right side of Equation (4) as a composite but not necessarily to measure them separately.7 We hypothesize that residents of a coastal community will learn and update their risk perceptions from hurricanes as ‘‘new’’ events. This prospect is especially relevant to situations when there has been a recent history of few storms.8 Nonetheless, this proposal has several important qualifications. The first requires that we use the storm to isolate new information about risk of these extreme weather events. This treatment cannot be confounded with the damage due to the actual storm. This concern is one of the issues motivating the comparison in this paper. In the first of our efforts in this area, Hallstrom and Smith (2005), we selected a ‘‘near miss’’ – a housing market with preexisting, known risk of hurricanes that was close to, but not hit by, a severe storm. This strategy seeks to avoid the effects of storm related damages on the models for property values. It implicitly maintains that the selection of a
CAN NATURAL EXPERIMENTS MEASURE BEHAVIORAL RESPONSES
279
location generally at risk of storm damage, but that has not been ‘‘hit’’, serves to identify a group of individuals who should be aware of the possibility of hurricanes. They may not, depending on past history in relation to their location choices, have accurate perceptions of the probabilities of storms. Under these circumstances, a storm that is ‘‘noticed’’ but that does not directly affect them provides new information. To play this role, Hallstrom and Smith select hurricane Andrew, the global insurance industry’s natural event of record, and the responses of housing prices in the Special Flood Hazard Areas (SFHA) for Lee County, Florida. In the 20 years prior to August 1992 (when Andrew made landfall) no major hurricanes had passed within 150 miles of Lee County. Andrew made landfall in Dade County and progressed westward toward the Gulf Coast. At its closest point, the hurricane was 75 miles south of Fort Myers. This proximity (along with the assurance that no physical damage occurred in the area) is what Hallstrom and Smith use to characterize the county as a ‘‘near miss’’. This strategy, of course, is simply one ‘‘informational story’’. There are other possibilities, each with important implications for the design of public disaster policy. For example, an alternative hypothesis is that markets have already perfectly capitalized risk characteristics into property prices. In this case, the occurrence of a near miss storm is simply a realization from a known probability distribution with no actual effects. The storm conveys no new information for the residential housing market. Under these conditions there would be no change in the equilibrium prices. In addition, location specific characteristics such as past hurricane history and extent of the coastal hazard area might be expected to affect how much new information a serious storm conveys. It seems reasonable to ask how we might test the ‘‘near miss’’ strategy. There is no decisive test. As noted at the outset, we propose to evaluate it by considering what happened to housing prices in the county that was directly hit by Andrew. This alternative analysis of the storm’s information adds a new consideration. It must control for the effects of renovations that might take place as part of repairing the hurricane’s damage. It does have some advantages. The spatial risk designations available to homeowners through the FEMA flood maps in Dade County use the same basic scale as in Lee County. Both areas experienced a relative lull in activity prior to Andrew. Thus, households in each area might have fairly comparable storm risk beliefs. 3. Using Natural Quasi-Experiment for Environmental Applications To introduce the logic for our model for uncovering the effect of information on risk preferences, consider a simple regression version of the model used in difference-in-differences studies.9 Assume there are two dimensions
280
JARED C. CARBONE ET AL.
distinguishing the structure of the quasi-experiment – the group assignment (j) for each entity in the study and the timing (t) of the outcome that is observed for each entity. Translating the vocabulary used in the quasirandom experiment to our hedonic model, the price (in annualized terms) R designates the outcome we observe for each entity. An entity in our application will be a house. A group will be a location and the timing will be the date of the sale in relation to an external natural event – a hurricane. The simplest form of a difference-in-differences framework assumes a linear specification (in parameters) with effects due to the time (a1), the group (a1), the treatment response (e.g. requiring the simultaneous presence of both time and group effects) (b), and the observable and unobservable sources of heterogeneity (z jit for the observable features and ejit for the unobservable). ln R jit ¼ a0 þ a1 dt þ a1 d j þ bd jt þ cz jit þ e jit
ð5Þ
j
In Equation (5) dt, d are dummy variables equal to one if the group or time designation is satisfied and zero otherwise. d jt is the interaction effect of the two conditions (i.e. dtÆ d j). If we assume, as is often the case in hedonic models, that we do not completely observe all the attributes of a home and its location that contribute to its market price, then we will have incomplete knowledge of zjit. A common response (see Palmquist (1982) for the first application), and one that is consistent with the basic logic of quasi-experiments, is to use a repeat sales model and focus the analysis on the change in prices for the same houses over time. To illustrate, suppose there are two sales periods, t and s. If we use the difference in sales prices for the same homes (ln Rt)ln Rs), then this transformation is written in Equation (6). ðlnR jit lnR jis Þ ¼ a1 ðdt ds Þ þ bðd jt d js Þ þ cðz jit z jis Þ þ ðe jit e jis Þ
ð6Þ
^ is the estimate of the incremental price paid to acquire the condition b represented by the group and time designations. We see from Equation (7) that it assumes the difference in the logs of the annualized prices before (s=0) and after (t=1) the hypothesized outcome causing the new information for the group of interest (j=1) compared to the baseline group is constant. ^ ¼ ln R1 ln R1 ln R0 ln R0 ð7Þ b 1 0 1 0 For the repeat sales to capture this difference in mean effects, there must be no other change in observable variables (z jit = z jis), that contribute to housing price differences and the unobservables represented by the difference in the errors, (e jit e jis ), should not be correlated with the effect being measured. Linking estimates of b to individual risk preferences, as envisioned in Equation (3), requires that we consider the connections between the price
CAN NATURAL EXPERIMENTS MEASURE BEHAVIORAL RESPONSES
281
equation and our model. This connection will be somewhat different for our two counties. In the case of Lee County (i.e. the near miss) there was no damage from the storm, but the information is nonetheless hypothesized to affect properties in the FEMA flood hazard zones differently from those outside them. In addition, Andrew led to substantial turmoil throughout the state in the insurance market. Florida’s overall response to the insurance crisis Andrew created involved the development of two state-run property and casualty underwriting associations. However, popular descriptions of the program (see Longman (1994)) indicate that property insurance remained underpriced and did not adequately signal the risks posed to structures in coastal areas. Others note that after Andrew, some companies left the market for coastal properties completely.10 These events imply that the storm/market insurance response described in general terms in Equation (4) needs to be considered in modeling price changes for both counties. In the case of Dade County there is the added concern raised by the requirement to control for whether a property was damaged by the hurricane. Commercial data sets available for estimating hedonic models usually do not include this type of information. We used two strategies for controlling for damages. The first relies on a unique database published by the Miami Herald in December 1992. These data correspond to a damage assessment by housing subdivision conducted by the National Oceanic and Atmospheric Administration (NOAA) after the storm. This survey reports the percentage of homes in a neighborhood judged to be uninhabitable after the storm. The second uses a geo-coded map of the storm’s path and the wind maps prepared by Wakimoto and Black (1994) for Andrew to estimate a band where storm damage was most likely.11 Each of these effects is represented by a series of interaction terms linking an attribute (i.e., in the FEMA flood zone or in a zone with hurricane damage if the analysis is for Dade County) with the timing of the housing sales. To illustrate this point consider the risk and information relationship which is hypothesized to be relevant for both counties. Re-writing Equation (5) we identify a different perceived risk of hurricane damage based on location in the FEMA Special Flood Hazard Zones (SFHA). We begin by assuming that only the designation for being in the zone versus outside it is relevant. In our final model we investigate whether a more detailed resolution of sub-zones is possible. Equation (8) describes this single destination – perceived risk based on being inside or outside the SFHA. X ck zk þ Fi ðbt þ bpt þ gi þ eit Þ þ ð1 Fi Þðct þ b/t þ gi þ eit Þ: ð8Þ lnRit ¼ k
Notice we replace ejit with gi+eit. The subscript i identifies the property and t the date of the sale. gi is idiosyncratic unobserved heterogeneity. bt and ct are
282
JARED C. CARBONE ET AL.
time effects for properties inside and outside an SFHA, respectively, Fi is a qualitative variable identifying the location of properties inside (=1) and outside (=0) an SFHA and eit is an iid error term. Homeowners’ subjective risk assessments of hazards at time t are pt for properties within an SFHA and /t for properties outside. Our model assumes that information from a major hurricane causes households to update risk assessments from their baseline levels by zone. Thus, if pt=p0 is the baseline risk assessment inside an SFHA before the information and pt=p1 after, and if we use the same convention for /t, we can augment the model to reflect this change. Define At=1 if Andrew occurred and At=0 otherwise, recognizing our hypothesized discrete change in risks for the two locations using Equations (9a) and (9b), respectively, and substituting, we have our revised model in (10). pt ¼ At p1 þ ð1 At Þp0
ð9aÞ
/t ¼ At /1 þ ð1 At Þ/0
ð9bÞ
ln Rit ¼
X
ck zki þ Fi ðbt þ bðAt p1 þ ð1 At Þp0 Þ þ gi þ eit Þ
k
þ ð1 Fi Þðct þ bðAt /1 þ ð1 At Þ/0 Þ þ gi þ eit Þ:
ð10Þ
Differencing the models for RiT when T=t and T=s for the same property i, we have Equation (11). X ck ðzkit zkis Þ þ ðct cs Þ þ Fi ððbt bs Þ ðct cs ÞÞ lnðRit =Ris Þ ¼ k
þ bð/1 /0 ÞðAt As Þ þ b ððp1 p0 Þ ð/1 /0 ÞÞFi ðAt As Þ þ ðeit eis Þ
ð11Þ
Equation (11) allows for a change in housing attributes between sales that span Andrew. The identifying restrictions required to estimate a pure information effect in this simple case are that: (a) there are no significant changes in attributes between the two time periods (e.g. the zj’s remain the same); (b) the partial effects of structural attributes on the log of the sale prices are constant (i.e. the cj’s do not change); and (c) the unobserved heterogeneity is not differentially influenced by the event or the group. Thus, our re-formulation simply adds more context to Meyer’s simplified expression. The interaction term indicating the sales bracketed Andrew and that a property is in a SFHA measures b[(p1 ) p0))(/1 ) /0)], the incremental option price scaled by the differential risk, for the areas with significant hazard compared to those without. Notice our model assumes that the effect of subjective risks is the same in both areas as in the Meyer basic formulation.
CAN NATURAL EXPERIMENTS MEASURE BEHAVIORAL RESPONSES
283
To introduce insurance effects we must consider whether they are differentially relevant to homes in the SFHA zones. Our analysis assumes they are. We do not have property specific insurance rates. Our primary basis for accounting for the effects of insurance is through the requirement, in 1994 legislation, that mortgage lenders assure flood insurance policies are in place for all home sales requiring financing. Thus, sales that bracket Andrew and have their most recent sale after the implementation period (1995–1996) are assumed to be differentially impacted by the hurricane. To consider a simple treatment of this effect we write bt and ct as follows, bt ¼ At B1 Dt þ ð1 At ÞB0 Dt
ð12aÞ
ct ¼ At c1 Dt þ ð1 At Þc0 Dt
ð12bÞ
with B1 and B0 designating the effect for the new insurance inside the SFHA zone before and after Andrew and the same for c1 and c0 for homes outside an SFHA zone. Dt=1 if the most recent sale was after 1996 and 0 otherwise. Considering the difference ln Rit)ln Ris we add another interaction term. ððB1 B0 Þ ðc1 c0 ÞÞ Fi ðAt As Þ Dt
ð13Þ
This term measures the differential effect on insurance on the SFHA properties if they bracket Andrew and were sold after the flood insurance changes. Finally when we introduce damage effects for Dade County these add a different spatial unit. In addition to the Special Flood Hazard Area we must take account of the prospects for storm damage. Two strategies are used to take account of whether a property is located where hurricane damage took place – the Miami Herald database versus a band using the estimates for Andrew’s wind speeds. The basic logic is comparable to what is used for the flood hazard – suppose di measures either the proportion uninhabitable based on the Miami Herald data or a fixed effect if a property is in the zone with high winds likely to cause damage and zero otherwise.12 We hypothesize that those within the damage zone that have sales bracketing Andrew are more likely to have price adjustments due to repairs of that damage. Those with either both sales before or both sales after Andrew are assumed to have characteristics that are not changed by the hurricane. Thus, this alteration is not a factor in sales price changes. 4. Data We purchased information on all the sales of residential homes in both Lee County and Dade County between 1980 and 2000 from a commercial vendor (First American Real Estate Solutions, FARES). These data include detailed records on the characteristics of properties at the time of sale, the date of each sale (year, month, and day), the sales price, the latitude and longitude
284
JARED C. CARBONE ET AL.
coordinates, and a variety of other variables describing the properties.13 The spatial coordinates for each property allowed them to be merged with the Federal Emergency Management Agency’s G3 flood maps. The housing sales data for our analysis were also cleaned to remove several types of transactions, including: properties that sold for less than $100; properties that were bought and sold within a period of several months and had a price difference exceeding $500,000; and properties where the first sale was for land only and the second sale included land and a structure. Finally, the National Flood Insurance Program includes several provisions – two were important to limiting our sample. The first of these relates to properties built before 1974, making them eligible for subsidized flood insurance. The second is due to the Coastal Barrier Resources Act of 1982. It provided the first specific requirement that lenders notify borrowers that a home was in a Special Flood Hazard Area. At the same time, there was a sharp change in flood insurance rates. These two factors together imply large differences (for distinct reasons) in the homes built prior to 1982 from those built later. As a result, we limited our attention to properties built after 1982. Using the transactions with current sales records from 1993 to 2000 and the immediate past sales for these properties, we reconstruct a set of price differences. We can also identify the timing of sales in relation to the hurricane in August 1992. When the model is applied to repeat sales for Lee County, there is no need to consider the prospect for storm related damage. Properties with two sales isolating changes in risk perceptions about their locations satisfy two conditions. They must be located in a SFHA and have the two sales bracket Hurricane Andrew. This interaction of the two qualitative variables corresponds to the interaction term Fi(At ) As) in Equation (11). The parameter for this variable bÆ((p1 ) p0))(/1 ) /0)) is our differencein-differences estimate of the effects of information about the risks of storms for these locations. The term for Andrew without the SFHA designation isolates b(/1 ) /0). The sum of these terms then yields b(p1 ) p0). The model for Dade County takes several forms depending on how we attempt to account for the effects of storm damage. Our preferred specification includes the Miami Herald summary of the NOAA survey. Table I decomposes the sample for each county based on whether a repeat sale is within the SFHA and whether it brackets the date of Hurricane Andrew. For Lee County there is a fairly even split in sales by area and time. The decomposition for Dade County is not as even, partially due to the spatial distribution of homes in this area. To provide some gauge of the sample coverage due to hurricane damage, in the table we use a band approximately 18 miles wide centered at the path of the eye of the storm. Most of the damage to residential properties was north of the path of the eye. Our wind damage zone includes properties with a distance from the eye of the storm that is within 9 miles. The resulting distribution of our sample suggests
CAN NATURAL EXPERIMENTS MEASURE BEHAVIORAL RESPONSES
285
Table I. Spatial and temporal dimensions of the natural experimentsa Spatial dimensions of natural experiments
Temporal dimensions of natural experiments Total
Lee County Bracket Andrew
Special Flood Hazard Area Out (sum of 2,805 1,136 sub categories) d
