Risk Analysis, Vol. 35, No. 5, 2015

DOI: 10.1111/risa.12307

Seeing is Believing? Evidence from Property Prices in Inundated Areas Ajita Atreya1,∗ and Susana Ferreira2

We use hedonic property models to estimate the changes in implicit flood risk premium following a large flood event. Previous studies have used flood hazard maps to proxy flood risk. In addition to knowing whether a property lies in the floodplain, we use a unique data set with the flood inundation map. We find that the price discount for properties in the inundated area is substantially larger than in comparable properties in the floodplain that did not get inundated. This suggests that, in addition to capturing an information effect, the larger discount in inundated properties reflects potential uninsurable flood damages, and supports a hypothesis that homeowners respond better to what they have visualized (“seeing is believing”).

KEY WORDS: Difference-in-differences; flood risk; hedonic model; inundation

1. INTRODUCTION

Regarding floods, individuals in a community may perceive the risk of being flooded differently due to discrepancies in the assessed probability of risk (e.g., as homes differ in terms of their location with respect to the floodplain). However, a different, perhaps more realistic perception can be expected of the risks with which people have some direct or indirect experience.(2) Intuitively, we would therefore expect the flood risk perception of individuals to be more pronounced in those areas directly hit by a flood. In addition, a heightened risk perception after a flood event is expected to result to a large extent from actually being inundated (direct experience) rather than just from the assessed probabilities of the risk. This article considers the 1994 “flood of the century” in the city of Albany, Georgia, as a source of flood risk information to homeowners in Albany, and determines the changes in implicit flood risk premium following the flood. In particular, we compare the price discount for properties in the inundated area to that for properties in the floodplain but outside the inundated area. In the United States, the Federal Emergency Management Agency (FEMA) uses a system of flood zones to convey information about the severity or type of flooding. For example, the 100-year flood

A key element in hazard and disaster management is the understanding of how individuals perceive risk. Inadequate distinction between assessed risk and perceived risk is a major difficulty in managing flood risk. Risk assessment estimates the chances of a specific set of events occurring and/or their potential consequences. However, scientists and engineers need to provide the users of these data with a picture of what is known regarding the nature of a particular risk and the degree of uncertainty surrounding these estimates.(1) Risk perception, on the other hand, is concerned with psychological and emotional factors that have been shown to have an enormous impact on behavior.(1) For example, those hazards for which the person has little knowledge and most dread are perceived as being the most risky. 1 The

Wharton School, University of Pennsylvania, Philadelphia, PA, USA. 2 Department of Agriculture & Applied Economics, University of Georgia, Athens, GA, USA. ∗ Address correspondence to Ajita Atreya, The Wharton School, University of Pennsylvania, 3730 Walnut Street, Jon M. Huntsman Hall #500, Philadelphia, PA 19104, USA; [email protected].

828

C 2014 Society for Risk Analysis 0272-4332/15/0100-0828$22.00/1 

Seeing is Believing? zone or Special Flood Hazard Area (SFHA) represents a 1% chance of flooding per year, whereas the 500-year flood zone represents a 0.2% chance per year. Previous studies have used FEMA designated flood hazard maps and the location of properties with respect to the floodplain as delineated in these maps as a proxy for flood risk,(3–7) and specific flood events as information shocks on perceived flood risk.(8–10) The typical finding in these studies is that large floods increase perceived risk. This can be explained by the availability heuristic,(11) through which individuals assess the probability of an event’s occurrence by how easily examples of such events come to mind. Reinforcing this intuition, recent studies from Atreya et al.(12) and Bin and Landry(13) find that the risk premium for 100-year floodplain properties in Dougherty County, Georgia, and Pitt County, North Carolina, increased in the wake of the 1994 “flood of the century” and Hurricanes Fran and Floyd, respectively, but that the effect was transitory, disappearing a few years after the flood events. Previous papers, however, do not control for the actual inundation or the damages caused by flooding as these data are not readily available. In this article, we analyze the same large flood as Atreya et al.,(12) but we limit our analysis to a smaller area in Dougherty County encompassing the city of Albany, for which damage data were available and are able to control for the damages caused by the 1994 flood. Moreover, in addition to FEMA hazard maps classifying floodplain areas, we use a map of the area that was not just at risk but inundated by the 1994 flood in Albany to tease out a “pure information effect” associated with just being located in the floodplain from an “inundation effect.” To the extent of our knowledge, this is the first study that uses flood inundation maps to determine the effects of flood events on property prices. We hypothesize that, following a large flood, the price of properties in the inundated area will be discounted for two reasons: first, because homeowners in inundated areas are more likely to have experienced physical and other damages from the flood; and second, in terms of heightened flood risk perception because people respond better to what they have experienced directly (“seeing is believing”). While, in principle, under conditions of perfect information and full insurance, flood insurance could cover reconstruction and cleaning costs, an increase in the discount for properties in the

829 inundated area would reflect the incremental insurance costs and the residual risk for noninsurable losses (personal items with sentimental value, risk of injury and death, hassle of being displaced, damage to community infrastructure, etc.), or the difference between the loss from flooding and the payment from the insurance company. We use a hedonic property model in a differencein-differences (DD) framework to determine the impact of the flood event on the estimated flood risk discount and to tease out the inundation effect from the information effect of the flood. Some studies have shown that flood risk is capitalized into property prices and the price difference between houses inside and outside of the floodplain is often consistent with the discounted sum of future flood insurance payments.(6,8) Like Bin and Landry,(13) we do not find a significant discount for floodplain properties before the flood, but in the wake of the 1994 flooding event we find a significant, large discount for floodplain properties. Consistent with previous studies,(12,13) the flood discount vanishes over time. However, it is not possible to tell whether the estimated discount for floodplain properties is the result of a pure information effect or an inundation effect. Many floodplain properties were in fact inundated and we suspect that a big part of the discount arises from the inundation and not from just being located in the floodplain. To tease out a pure information effect, if any, from an inundation effect we divided our study area into four mutually exclusive groups: inundated and in the floodplain, inundated outside the floodplain, noninundated and in the floodplain, and noninundated outside the floodplain. If there is an information effect associated with being located in the floodplain, one might expect to see a discount for noninundated floodplain properties as well, but our results suggest that this is not the case. We did not find any discount for the floodplain properties when we dropped the inundated properties from the model. The noninundated properties, including those in the floodplain, were not discounted significantly following the flood. Our results suggest that it is an inundation effect rather than a pure information effect that is capitalized into property prices. Despite federal requirement for flood insurance for floodplain properties, the lack of evidence of a pure information effect and of a persistent risk premium suggest limited awareness about flood hazards among buyers of floodplain properties.

830

Atreya and Ferreira

Fig. 1. Flood inundation study area: Albany, Georgia.

2. STUDY AREA The city of Albany is located along the banks of the Flint River in southwest Georgia (Fig. 1). It comprises a total area of 55.8 square miles and is home to over 77,000 people.(14) In 1994, a

severe flood caused by tropical storm Alberto destroyed parts of downtown and south Albany, causing 15 deaths and displacing almost 22,000 people. Peak discharges greater than the 100-year flood discharge were recorded at all U.S. Geological Survey (USGS) Flint River gauging stations.(15) At Albany,

Seeing is Believing? the Flint River peaked at a stage about 5 ft higher than the 1925 flood, which was the maximum flood ever recorded at that gauging station. Albany is one of the nearly 20,000 communities across the United States and its territories voluntarily participating in the National Flood Insurance Program (NFIP). The NFIP makes federally backed flood insurance available to homeowners, renters, and business owners in participating communities that in exchange adopt and enforce floodplain management ordinances to reduce future flood damage. In order to actuarially rate new constructions for flood insurance and create broad-based awareness of the flood hazards, FEMA maps 100- and 500-year floodplains in participating communities.3 Homes and buildings in high-risk flood areas, those with 1% or greater chance of flooding in any given year, and with mortgages from federally regulated or insured lenders are required to have flood insurance. In addition to FEMA hazard maps, typically used in the literature to capture flood risk, in this article we use a map of the area that was inundated by the 1994 flood. The USGS, along with partners at the National Weather Service (NWS), the U.S. Army Corps of Engineers (USACE), FEMA, state agencies, local agencies, and universities, has developed a web-based tool for flood response and mitigation with a major goal of reducing vulnerability of people and areas most at risk from natural hazards. It provides digital geospatial flood inundation maps that show flood water extent and depth on the land surface. USGS has modeled flow characteristics of flooding along a 4.8-mile reach of the Flint River in Albany, Georgia, using recent digital-elevationmodel data and the USGS finite-element surfacewater modeling system for two-dimensional flow in the horizontal plane. Fig. 1 show the study area and the inundated area for a water surface altitude of 192.5 ft at Flint River, which corresponds to the 1994 flood caused by tropical storm Alberto. 3. METHODS Following the standard hedonic approach,(16,17) we model the price of a property, P, as a function of its structural characteristics, S (e.g., number of rooms, size of the house), neighborhood and location characteristics, L (e.g., distance to river, distance to parks), and its location with respect to the inundated

831 area (IND) in addition to its location with respect to the floodplain (FP). We use a quasi-experimental DD approach to measure the effect of a large flood event on floodprone property prices. In the conventional application of the DD model to measure flood risk,(10,12,13) properties that fall within the floodplain are the treatment group and properties outside the floodplain are the control group. The DD design allows to isolate the effect attributable to the flood from other contemporaneous variables (e.g., macroeconomic shocks, changes in local housing or labor markets), since the control group experiences some or all of the contemporaneous influences that affect property values in the treatment group but offers lower flood risk. Much of the debate around the validity of a DD identification strategy revolves around the possible endogeneity of the intervention.(18) The use of a DD identification strategy in our study seems appropriate as the occurrence of the flood was exogenous. Moreover, we do not have a reason to believe that preexisting trends in home prices in the treatment and control groups would have diverged in the absence of the flood. As robustness checks, we explore potential changes in the composition of houses that sold after the flood that could affect the interpretation of our estimates. We also conduct falsification tests by shifting the date of the flood, both to before and after its true occurrence. In the models we also include spatial fixed effects to difference away the potential bias from time-invariant omitted variables. However, endogeneity of treatment is only one of the many threats to validity in DD analysis.(19) Particularly relevant for this article is the possibility of omitted interaction variables. Houses that are inundated can be repaired and mitigated (perhaps even upgraded) relative to the rest of the housing stock; this is, in essence, an omitted interaction variable and may cast some doubt on the interpretation of the decay effect.4 To address this concern, as an additional robustness check we ran a separate regression removing the properties that were substantially damaged during the 1994 flood from the sample. Excluding those properties ensures that the coefficients in the DD model do not pick up a damage effect that has potential to contaminate our interpretation of shadow values reflecting the flood risk discount after a flood event. We estimated three different models.

3 100- and

500-year floodplain refer to the flooding probability, 1% and 0.2% in a given year, respectively.

4 We

thank an anonymous referee for pointing this out.

832

Atreya and Ferreira

3.1 Model 1: Hedonic Model with Floodplain Dummies As in recent papers studying the capitalization of flood risks into property prices,(10,12,13) our simplest model uses FEMA-designated floodplain maps to measure flood risk in the following specification: ln(Pit ) = β0 + β  1 ln Li + β  2 S it + β3 F Pi +β4 Floodit + β5 FPi ∗ Floodit +β6 f (year s) + β7 f (year s)∗ FPi +β8 Damagei + γi + δt + εit .

(1)

In the DD design, FP is a dummy variable equal to 1 if the property falls in the floodplain and 0 otherwise, and the control group is properties outside the floodplain. The variable Flood is a dummy variable equal to 1 if the sale happened after the flood (July 1994). The interaction term between the floodplain variable (FP) and Flood tells us how the 1994 flood might have affected the prices of floodplain properties sold after the 1994 flood. This is our variable of interest. The specification in Equation (1) includes an interaction term between the floodplain variable and f(years), where years is the number of years after the 1994 flood. This allows us to examine the persistence of the risk premium over time after the 1994 flood. Previous papers(12,13) find this temporal decay effect to be important. In addition to using a linear time trend (f(years) = years), we explored an alternative, nonlinear functional form, the natural logarithm, f(years) = ln(years). With the variable Damage we seek to control for the physical damages to the properties affected by the flood. Although we do not have information on flood damages suffered by the properties in the inundated area at the structure level, we do have information on two alternative indicators. First, we consider the average dollar losses from the 1994 flood at census-block level, generated using HAZUS, a software developed by FEMA to estimate flood damages.5 We controlled for damages in the 114 census blocks in the inundated study area. These census blocks are small, with just 12 houses on average. Second, assuming that the depth of the flood water acts as a proxy for the degree of structural damages to the properties in the inundated 5 We

thank Paul P. Hearn, Chief, Research Station, Eastern Geographic Science Center, U.S. Geological Survey, and Eastern Region for kindly providing the damage data: http://wim.usgs.gov/FIMI/FloodInundationMapper.html.

area, we extracted the flood depth in the inundated study area at a gauge height of 43 ft and an altitude of 192.5 ft (corresponding to the 1994 flood) using a raster map developed by USGS. Because inundation depth reflects both flood severity and the ground elevation of a building,(20) we subtracted the land surface elevation from the model elevation. We expect structural damages to be greater for properties located at higher flood depths, and thus expect to find a larger price discount for those properties. Regarding the functional form, we performed a Box-Cox transformation of the dependent variable and compared the residual sum of squares to conclude that the natural log of price as the dependent variable was the best specification for our model. After testing several transformations of the independent variables, the location variables L were best fitted in their log form while the other attributes S were fitted best in their linear specification. Census-block fixed effects (γ i ) and year fixed effects (δ t ) were included to capture time-invariant characteristics at the census-block level and yearly shocks that may affect all the properties. Subscripts i and t represent property and time, respectively. 3.2 Model 2: Hedonic Model with Inundation Dummies In order to explore the inundation effect and to determine the changes in property prices in the inundated area after the 1994 flood, we used the USGS inundation map of the study area. In the specification below, the treatment group is composed of properties in the inundated area, while the control group is composed of properties that are outside the inundated area: ln(Pit ) = β0 + β  1 ln Li + β  2 S it + β3 INDi +β4 Floodit + β5 INDi ∗ Floodit +β6 f (year s) + β7 f (year s) ∗ INDi (2) +β8 Damagei + γi + δt + εit , where the variable IND (inundation) is a dummy variable equal to 1 if the property was inundated by the 1994 flood and 0 otherwise. That is, Equation (2) replaces the floodplain dummy variable in Equation (1) with an inundation dummy variable. 3.3 Model 3: Hedonic Model with Floodplain and Inundation Dummies Finally, we overlay the inundation and flood hazard maps to divide the properties within the study

Seeing is Believing? area into four mutually exclusive groups: inundated and in the floodplain (IN ̲ FP), inundated outside the floodplain (IN ̲ OFP), noninundated and in the floodplain (NIN ̲ FP), and noninundated outside the floodplain (NIN ̲ OFP). Model 3 explores the effect of the 1994 flood in each of these mutually exclusive groups using the properties outside both the floodplain and inundated areas, NIN ̲ OFP, as the control group (and thus the dummy variable excluded in the regression below): ln(Pit ) = β0 + β  1 ln Li + β  2 S it +β3 I N F Pi + β4 I N OF Pi + β5 NI N F Pi +β6 Floodit + β7 I N F Pi ∗ Floodit +β8 I N OF Pi ∗ Floodit +β9 NI N F Pi ∗ Floodit + β10 f (year s) (3) +β11 f (year s) ∗ I N F Pi ∗ +β12 f (year s) I N OF Pi +β13 f (year s) ∗ NI N F Pi +β14 Damagei + γi + δt + εit . An important econometric issue in hedonic models concerns the potential spatial dependence of the observations. Neighboring properties are likely to share common unobserved location features, similar structural characteristics due to contemporaneous construction, neighborhood effects, and other causes of spatial dependence. Ignoring the problem could result in inefficient or inconsistent parameter estimates.(21) The baseline results reported in this article are estimated using ordinary least squares (OLS), but as evidenced in Equations (1)–(3), all the specifications include spatial and neighborhood controls (L), and census-block level fixed effects. An alternative is to estimate the models explicitly accounting for the potential spatial dependency, for example, by fitting a spatially lagged and autoregressive disturbance model, frequently referred to as a SARAR model.(22) As a robustness check, we estimate such a model that allows for spatial interactions in the dependent variable, the exogenous variables, and the disturbances. Spatial interactions in the dependent variable are modeled through a spatial lag structure that assumes an indirect effect based on proximity (i.e., the weighted average of other housing prices affects the price of each house). The error term incorporates spatial considerations through a spatially weighted error structure. As noted in recent papers,(23–26) spatial econometric modeling per se is no panacea. It is valid as a descriptive approach, but does not provide a valid approach to causal identification. In this article, we use it as a robustness check

833 to our traditional hedonic model that uses the large 1994 flood in a DD setting for identification. 4. DATA We used four data sources to construct our data set: Dougherty County’s Tax Assessor’s Office for individual property sales of residential homes in the city of Albany; Georgia’s Geographic Information System (GIS) clearinghouse for parcel level GIS data; FEMA for floodplain maps of the Flint River at Albany; and USGS for flood inundation maps and to estimate losses from the 1994 flood at census-block level. Each property is a single-family residence sold between 1985 and 2007. Individual property sales data contain information on housing structural characteristics S, such as number of bedrooms, number of bathrooms, heated square feet, and presence of garage, in addition to sale date, and sale price P. Property sale prices were adjusted to 2007 constant dollars, using the housing price index for the Albany metropolitan area from the Office of Federal Housing Enterprise Oversight (OFHEO). The GIS database was utilized to determine the location attributes of the properties, L, such as proximity to rivers, railroad, major roads, and parks. Other neighborhood characteristics (median household income and percent of nonwhite residents) were determined at the block-group level using 2000 census data.6 The floodplain map published as Q3 data by FEMA was used to determine if the parcel was within or outside the 100-year floodplain.7 Similarly, the USGS flood inundation map for a water surface altitude of 192.5 ft at Albany’s stream gauge corresponding to the 1994 flood was used to determine if the parcel was within or outside the inundated area. We confined our study area to the flood inundation study area at Flint River, Albany, prepared by USGS (Fig. 1), which includes 2,685 single-family residences. We contacted several personnel at USGS to inquire about the validity of the flood inundation map. Because the flood inundation map is modeled, 6 Block

groups generally contain between 600 and 3,000 people, with a typical size of 1,500 people. 7 This is a 1996 map. We contacted a FEMA representative to see if there had been any updates to the flood maps between 1996 and 2007 and discovered that although the FEMA map was updated in 2001, those maps became effective only in 2009. Because the large flood event in our study occurred in 1994 and all of our sales transactions occurred before 2009, in our analysis we use the 1996 map.

834

Atreya and Ferreira

Table I. Number of Properties in Different Flood Hazard Categories, Albany Category

Number of Properties

Inundated (IN) Noninundated (NIN) Inside floodplain (FP) Outside floodplain (OFP) Inundated and in floodplain (IN ̲ FP) Inundated but outside floodplain (IN ̲ OFP) Noninundated and in floodplain (NIN ̲ FP) Non inundated outside floodplain (NIN ̲ OFP)

793 1,892 615 2,070 551 242 64 1,828

“[e]ven under ‘perfect’ conditions the inundated area is going to be somewhat different than what the model predicts because there are so many factors involved in a dynamic flood which a static model cannot take into account” (Jonathan W. Musser, Hydrologist, USGS, personal communication, January 31, 2014). However, we are confident that the flood inundation map is valid for making inferences regarding perceptions of flood risk, especially when compared to the alternative, widely used, FEMA floodplain maps. First, “[t]he Digital Elevation Model (DEM) that was used in computing the inundation polygons was created after the 1994 flood, so it should account for any changes in channel hydraulics due to sediment movement or other factors” (Paul P. Hearn, USGS, personal communication, October 9, 2012). Second, the flood inundation maps are based on the USGS guage on the Flint River in Albany specific to the 1994 flood, and were calibrated to the high water marks available at the time. In conjunction, the 100-year floodplain map and the inundation map define the four mutually exclusive groups employed in the estimation of model 3: inundated and in the floodplain (IN ̲ FP), inundated outside the floodplain (IN ̲ OFP), noninundated and in the floodplain (NIN ̲ FP), and noninundated outside the floodplain (NIN ̲ OFP, the reference category). Table I shows the number of properties in each group. Given the few number of observations in the NIN ̲ FP category, we did not consider a separate set of variables for properties in the 500-year floodplain, as there were only 183 properties in total in the 500-year floodplain. In addition, homeowners are not required to buy flood insurance if they are located in the 500-year floodplain and therefore might be unaware of the flood hazard associated with the location of their properties. However, we did try a broader definition of the floodplain variable by merging both the 100- and 500-year floodplain properties. This increased the number of observations in the

NIN ̲ FP category to 133. The results discussed in the next section were robust to this change. The NIN ̲ FP variable is critical for identifying a difference between the inundation and information effects, but the low proportion of observations for this variable casts doubts on whether we are identifying a “true” effect. Therefore, we conducted a post hoc power analysis to test whether the data have enough power to detect an effect.(26) The power is the probability of detecting a true effect when it exists. We used the software package G* Power.(27) We set the sample size at 2,685 and a 32 predictor variable equation as a baseline. The effect size, which is the magnitude of the treatment effect, was calculated to be 0.005 based on the variance explained by the predictor of interest (NIN ̲ FP). The α level used was 0.05. Alternatively, we ran a Monte Carlo simulation for power analysis using Stata software (using the sample sizes, means, and standard deviations of all the four groups). In both cases, the estimated statistical power was 0.97 for detecting an effect size of 0.005, which is more than 0.80, a convention proposed for general use. Table II reports the descriptive statistics of all the variables considered in the analysis. The mean property price was $80,221 in 2007 constant dollars. The oldest property was built in 1883. The average property had 0.26 acres. The maximum elevation was 216 ft and the minimum elevation was 175 ft. Mean distance to the nearest river was 671 m. Of all the sales between 1985 and 2007, 23% of the properties were in the 100-year floodplain and 29% of the properties in the sample were inundated during the 1994 flood. In Table III, we present the summary statistics separately for the inundated and the noninundated groups in the pre- and postflood periods. In both the pre- and the postflood time periods, we find some differences between properties in areas that inundated relative to those that did not. Houses in the low-lying areas were found to be less pricey even before the flood occurred. The mean property price for a property in the noninundated area (pre-1994), with mean elevation of 196 ft, was $52,688, whereas the mean property price for a property in the inundated area, with mean elevation of 183 ft, was $43,499. We believe that these differences arise due to differences in elevation and other housing and neighborhood socioeconomic characteristics for which we control in the regressions. In fact, we ran a regression of the sale price on a dummy variable indicating the treated houses for the years prior to the 1994 flood. The dummy, whether the treatment was defined as

Seeing is Believing?

835 Table II. Variable Definition and Descriptive Statistics, Full Sample

Variable Price Depth Avg˙Dam Years Elevation River Lake Railroad Roads Utilities Park School Year built Acres Bedrooms Fullbths Halfbths Htdsqft Fireplace AC Garage Brick Flood FP IND IN FP IN OFP NIN FP NIN OFP Income PcBlk

Description

Mean

SD

Min

Max

Sale price of property (2007 constant dollars) Flood depth (for inundated properties) (ft) Average damage by census block (thousand $) Number of years after 1994 flood Elevation of property (ft) Distance to nearest river (m) Distance to nearest lake (m) Distance to nearest railroad (m) Distance to nearest road (m) Distance to nearest utility lines (m) Distance to nearest park (m) Distance to nearest school (m) Year the property was built Total acreage of the property Number of bedrooms Number of full baths Number of half baths Heated square feet Number of fireplaces 1 if central AC present, else 0 1 if garage present, else 0 1 if brick exterior, else 0 1 if sold after July 1994, else 0 1 if in 100-year floodplain, else 0 1 if in inundated area during 1994 flood, else 0 1 if inundated in FP, else 0 1 if inundated outside FP, else 0 1 if noninundated in FP, else 0 1 if noninundated outside FP, else 0 Median household income by census block group Percent of nonwhites by census block group (%)

80,221 5.04 0.37 4.88 191.82 671 727 1,045 30 3,462 1,771 854 1,962 0.26 2.81 1.31 0.10 1,195 0.14 0.67 0.03 0.03 0.70 0.23 0.29 0.21 0.09 0.02 0.68 20,545 85

153,157 2.20 2.32 4.56 9.32 466 351 639 23 1,446 737 435 22 0.20 0.58 0.51 0.30 427 0.35 0.47 0.16 0.17 0.46 0.42 0.45 0.40 0.29 0.15 0.47 6,110 20

1,854 0.09 0 0 175 6 10 21 0.02 734 47 44 1,883 0.01 0 1 0 480 0 0 0 0 0 0 0 0 0 0 0 6,907 15.8

1,400,000 14.23 27.5 13 216 2,346 1,681 2,750 154 6,268 3,137 2,036 2,008 3.73 8 7 2 4,714 1 1 1 1 1 1 1 1 1 1 1 42,964 100

being in the floodplain or being inundated, was statistically insignificant, suggesting that the sales price in the preflood time period for the treated group is comparable to the control group, ceteris paribus (Table IV). We also note that the validity of the DD approach is based on the assumption that the underlying trends in the sale prices are similar for treatment and control group before being treated. Fig. 2 shows the trends in sales price during the preflood time period for the four mutually exclusive groups that we used in our analysis. As shown in the figure, the trend in sales price in treatment and control groups is not systematically different in the preflood period, suggesting that the assumption of common trends is a reasonable one. Economic theory does not provide definite guidance on the data range to be used in hedonic models, except that the contribution of the various characteristics to the value of the house should have been relatively stable over that time period.(28) Our sample period (1985–2007) was chosen to ensure that there

were sufficient transactions in the noninundated but in the floodplain (NIN ̲ FP) category, which is key to our identification strategy, and at the same time to capture the time trend in the flood risk discount following the 1994 flood while excluding more recent, postrecession observations. In order to check the stability of the housing attributes during this time period, we performed a series of paired t-tests on the characteristics of the average property before and after the 1994 flood and we failed to reject the null of equal means for the two time periods for most of the attributes. Notably, the proportion of houses in the floodplain, and most of the structural variables, including important attributes such as lot size (acres) and the heated squared footage, were not significantly different across the two subsamples.8

8 Results

of these tests are available in Table S1 in the online appendix.

836

Atreya and Ferreira Table III. Summary Statistics by Inundation and Pre- or Postflood Period Status Pre-1994 Noninundated

Pre-1994 Inundated

Post-1994 Noninundated

Post-1994 Inundated

Variable

Mean

SD

Mean

SD

Mean

SD

Mean

SD

Price Elevation River Lake Railroad Roads Utilities Park School Acres Year built Bedrooms Full baths Half baths Htdsqft Fireplace Brick AC Garage Income PcBlk

52, 688 196 869 903 904 33 3,900 2,293 957 0.27 1,957 2.85 1.33 0.12 1,227 0.18 0.06 0.63 0.04 21,545 80.08

40,381 7.90 504 394 562 26 1,834 722 490 0.20 21.12 0.56 0.55 0.33 457 0.38 0.24 0.48 0.19 7,153 21.80

43,499 183 382 651 1,876 33 4,081 1,462 1,072 0.23 1,967 2.79 1.23 0.13 1,111 0.08 0.05 0.77 0.02 20,184 93.07

39,526 6.13 273 277 765 23 1,221 777 452 0.15 18 0.51 0.42 0.34 334 0.27 0.23 0.42 0.13 6,044 13.37

104,388 195 948 935 962 34 3,999 2,150 927 0.26 1,957 2.78 1.31 0.08 1,201 0.16 0.02 0.65 0.03 20,426 82.17

198,013 8.17 545 406 560 26 1,755 786 508 0.22 21.97 0.61 0.53 0.27 456 0.37 0.12 0.48 0.16 6,398 21.36

69,358 184 388 586 1745 36 3,819 1,697 1,080 0.22 1,974 2.85 1.35 0.12 1,185 0.08 0.02 0.75 0.03 20,185 92.21

125,952 7.34 269 286 742 26 1,324 907 468 0.14 20.31 0.52 0.49 0.32 333 0.28 0.14 0.43 0.16 4,249 13.04

Table IV. Regression of Sale Price on Treatment Dummy for Preflood Period (1985–1993) Model 1 Treatment = FP FP

Model 2 Treatment = IND

−0.346 (0.314)

IND

0.0125 (0.178)

IN ̲ FP

−0.0101 (0.0156) 25.53 (18.69) Included

−0.00624 (0.0173) 19.79 (18.66) Included

−0.451 (0.302) 0.173 (0.205) −0.0360 (0.493) −0.00907 (0.0155) 25.43 (19.00) Included

671 0.551

671 0.548

671 0.555

IN ̲ OFP NIN ̲ FP ELEVATION Constant Structural attributes/location attributes/year and census block FEs Observations R2

Model 3 Treatment = IN ̲ FP, IN ̲ OFP, NIN ̲ FP

Notes: OLS estimates; dependent variable is ln(Price); robust standard errors in parentheses; ***p < 0.01, **p < 0.05, *p < 0.1.

In addition, because we do not want to attribute the evolution of the flood risk premium immediately following the flood to changes in the number of homes sold in the different areas, or to differences in the composition of the housing stock, we performed

two additional tests. As a first test of market stability, we analyzed the number of houses that transacted five years before and after the 1994 flood to see if there was an abrupt change in the number of homes sold, which would signal a potential instability

Seeing is Believing?

837

Fig. 2. Trends in sales price prior to the 1994 flooding for the four mutually exclusive groups.

Fig. 3. Log count of housing transactions (1989–1999).

in the housing market associated with the flood. Fig. 3 shows the log of number of properties sold between year 1989 and 1999, with the properties classified in two groups corresponding to their location with respect to the floodplain and the inundated area. We observe no dramatic percentage change in the

number of properties sold in any of the groups, suggesting that the housing market was relatively stable over this time period according to this metric. Second, we ran a series of linear regression with our key housing attributes as the dependent variable and the flood risk variables and their interactions

838

Atreya and Ferreira Table V. Housing Composition Effects Before and After the 1994 Flood

Variables FP IND Flood FP* Flood IND* Flood Observations R2

(1) Acres

(2) Age

(3) Bedrooms

(4) Fullbths

(5) Halfbths

(6) Htdsqft

(7) Fireplace

0.0737*** (0.0253) −0.0387** (0.0193) −0.0286 (0.0302) −0.0139 (0.0181) 0.00559 (0.0189) 2685 0.473

6.287** (3.004) −3.434 (2.521) 12.94*** (2.293) −3.533 (2.441) −0.387 (2.253) 2685 0.764

0.00261 (0.0877) −0.0825 (0.0566) −0.250 (0.184) −0.0113 (0.0669) 0.0891 (0.0629) 2685 0.308

−0.132* (0.0759) 0.00578 (0.0615) −0.259 (0.201) 0.0389 (0.0642) 0.0439 (0.0618) 2685 0.445

−0.0572 (0.0561) −0.0200 (0.0439) 0.104* (0.0610) 0.0366 (0.0500) 0.0111 (0.0470) 2685 0.243

−47.63 (51.23) −47.33 (37.73) −147.1 (99.07) −12.88 (40.70) 81.72** (38.45) 2685 0.358

0.0823 (0.0600) −0.0742 (0.0542) −0.0305 (0.0861) −0.0564 (0.0565) 0.0840 (0.0569) 2685 0.204

Notes: Robust standard errors in parentheses; ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

with the 1994 flood variable as our independent variables. Significant interaction terms of the flood risk variables with the flood variable would signal that there are changes in the housing characteristics after the flood. Table V shows the result of these regressions; all the regressions include census-block and year fixed effects. We find that out of the 14 interactions only one interaction coefficient is statistically significant (at a 5% level, to explain heated area square footage in properties in the inundated area), suggesting that there are no strong housing composition changes before and after the 1994 flood.

5. RESULTS 5.1. Model 1: Hedonic Model with Floodplain Dummies In the first column of Table VI, we present the estimates of model 1, which, following previous studies, simply estimates the effect of the flood on the prices of floodplain properties irrespective of whether the property was inundated or not. Flood risks, as measured by being located in the floodplain (FP), are capitalized into property prices only after the flood. We do not find a significant discount for floodplain properties before the 1994 flood. We find that properties in the floodplain were sold for 37% less than an equivalent property outside the floodplain immediately following the 1994 flood, which is equivalent to about $29,477 when evaluated at the

average-price home.9 However, consistent with previous studies that find a temporal decay effect,(12,13) we find that the large discount is short lived, as suggested by a positive and significant coefficient in FP* f(years) variable.10 5.2. Model 2: Hedonic Model with Inundation Dummies In the second column, we present the estimates of the effect of the 1994 flood on flood-prone properties in the study area as determined by whether they fall inside or outside the inundated area (model 2). We find that immediately after the flood, properties in the inundated area sold for a substantial 46% less than equivalent properties outside the inundated area. Even controlling for flood damages using the census-block average damage, this is a 9% larger discount than for floodplain properties in model 1.11 Again, we find that the discount for the properties in the inundated area diminished over time. 5.3 Model 3: Hedonic Model with Floodplain and Inundation Dummies To tease out the effect of being inundated from the information effect of being in the floodplain, we 9 In

semi-logarithmic equation such as Equation (1), the marginal effect of the dummy FP*Flood in the first column is given by exp(−0.458) − 1 = −0.37.(32) 10 The results are robust to using an alternative, logarithmic, decay function. The coefficients are available in Table S2 in the online appendix. 11 The 46% discount is calculated as exp(−0.609) − 1 = −0.46.

Seeing is Believing?

839 Table VI. OLS Estimates of Difference-in-Difference (DD) Model

Variables FP FP* Flood FP* f(years)

Model 1

Model 2

0.0659 (0.165) −0.458*** (0.150) 0.0471*** (0.0155)

IND

0.0671 (0.109) −0.609*** (0.145) 0.0718*** (0.0150)

IND* Flood IND* f(years) IN FP IN OFP NIN FP FLOOD 1994

0.687*** (0.187)

0.744*** (0.189)

0.0127 (0.0226)

0.0138 (0.0228)

−0.0251*** (0.00803) −0.00126 (0.00608) −0.0889*** (0.331) 0.0684 (0.131) 0.120 (0.150) −0.0145 (0.0249) 0.475 (0.453) 0.0292 (0.233) 0.211 (0.146) 0.187 (0.213)

−0.0228*** (0.00793) −0.00138 (0.00599) −0.0874*** (0.323) 0.0746 (0.131) 0.147 (0.149) −0.0151 (0.0248) 0.327 (0.459) 0.0263 (0.227) 0.168 (0.143) 0.249 (0.209)

IN FP* Flood IN OFP* Flood NIN FP* Flood f(years) IN FP* f(years) IN OFP* f(years) NIN FP* f(years) Damage (= census block average damage) Elevation Ln(River) Ln(Lake) Ln(Railroad) Ln(Road) Ln(Utilities) Ln(Park) Ln(School) Acres

Model 3

0.104 (0.179) 0.0798 (0.136) 0.211 (0.243) 0.762*** (0.190) −0.648*** (0.161) −0.451** (0.202) 0.155 (0.499) 0.0130 (0.0228) 0.0762*** (0.0169) 0.0434** (0.0216) −0.0697 (0.0458) −0.0208*** (0.00800) −0.00137 (0.00600) −0.0910*** (0.334) 0.0818 (0.134) 0.136 (0.149) −0.0128 (0.0249) 0.397 (0.455) 0.0424 (0.234) 0.187 (0.145) 0.198 (0.213) (Continued)

840

Atreya and Ferreira Table VI. Continued

Variables Age Bedrooms Full baths Half baths Heated sqft Fireplace Brick (=1 for brick exterior) AC (= 1 if central AC) Garage (= 1 if garage present) Ln(Income) % Black Year FEs Census block FEs Observations R2

Model 1

Model 2

Model 3

−0.0113*** (0.00268) 0.0792 (0.0591) −0.0848 (0.0716) −0.0747 (0.0818) 0.000400*** (0.000107) 0.0734 (0.0791) −0.148 (0.129) −0.0312 (0.0643) 0.124 (0.163) 0.313 (0.780) 0.0393 (0.750) Yes Yes 2,685 0.320

−0.0114*** (0.00259) 0.0824 (0.0601) −0.0916 (0.0720) −0.0739 (0.0817) 0.000385*** (0.000111) 0.0860 (0.0806) −0.125 (0.128) −0.0200 (0.0636) 0.133 (0.163) 0.287 (0.767) 0.0416 (0.728) Yes Yes 2,685 0.327

−0.0114*** (0.00259) 0.0802 (0.0598) −0.0910 (0.0724) −0.0793 (0.0821) 0.000392*** (0.000109) 0.0834 (0.0800) −0.119 (0.129) −0.0195 (0.0636) 0.145 (0.161) 0.289 (0.777) 0.0638 (0.744) Yes Yes 2,685 0.331

Notes: Dependent variable is ln(Price). Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

estimated model 3 wherein we divided the study area into four mutually exclusive groups. Compared to noninundated properties outside the floodplain (NIN ̲ OFP), we find that inundated floodplain properties (IN ̲ FP) were discounted by 48% immediately after the flood. Similarly, we find a significant discount of 36% for the properties that were inundated but were located outside the floodplain (IN ̲ OFP). Interestingly, the noninundated properties were not discounted significantly even if they were located in the floodplain (NIN ̲ FP). This suggests that it is in fact the inundation effect that is capitalized into property prices. We suspect that the significant discount found for floodplain properties in studies that do not control for whether the properties were inundated or not might be due to the inundation effect rather than an information effect associated with being located in the floodplain. Across all the specifications, we find that the price discount is temporary, decaying over time. The parameter estimates of the interaction term between the flood risk variables

(FP, IND, IN ̲ FP, and IN ̲ OFP) and the years variable are always positive and statistically significant. Regarding other controls, across all the specifications, flood damages have an expected negative sign and are significant (at a 1% level), indicating an almost one-to-one correspondence between damages and a drop in property prices; on average a $1,000 increase in damages reduces the sale price by $975– $979. Controlling for flood risk, proximity to the river is an amenity. It increases the property prices and so does the heated square footage. We find that a 1 ft increase in the heated square footage is expected to increase the property price by $32. As expected, age, on the other hand, depresses property prices. As mentioned in the previous section, the interpretation of our results is not confounded with a change in the composition of the houses that transacted before and after the flood (see Table V). A large difference in the housing composition would mean that the housing price effects that we observe are partially driven by supply rather than demand.(29)

Seeing is Believing?

841

Finally, we note that from the use of a flood inundation map that was recreated to represent the past flood, we can expect two types of measurement errors, arising from (1) the omission of properties that fall outside the inundation area but that were actually inundated and from (2) the inclusion of properties that fall within the inundation area but that were not inundated. Both will tend to bias our estimates upwards (i.e., to reduce the estimated flood risk discount). Our results can thus be considered conservative estimates.

5.4 Robustness Checks We implement a number of procedures to check the robustness of the results. First, in addition to including spatial census-block level FE and other location controls in the regressions, to explicitly account for spatial dependence we use a spatially lagged and autoregressive disturbance model frequently referred to as a SARAR model.(21)12 In the first column of Table VII, we present the estimates of the effect of the flood on the prices of floodplain properties irrespective of whether the property was inundated or not as in the first column of Table VI. Consistent with the OLS results, we do not find a significant discount for floodplain properties before the 1994 flood. We find that properties in the floodplain were sold for 33% less than an equivalent property outside the floodplain immediately following the 1994 flood, which is equivalent to about $26,500 when evaluated against the average-price home.13 Additionally, consistent with the OLS results, we find a statistically sig-

12 In

the SARAR specification, spatial dependence in the dependent variable is modeled by incorporating an additional term in the right-hand side of Equations (1)–(3): λ wi j In( pi t ) to alj

low for the weighted average of the other housing prices affecting the price of each house. The error term also incorporates spatial considerations through a spatially weighted error structure:  εit = ρ mi j ε jt + μit , where the disturbance μit is assumed to be identically distributed. wij and mij are the elements of W and M, two n × n spatial weights matrices, and λ and ρ are the spatial autocorrelation parameter and spatial autoregressive coefficient, respectively. Concerning the spatial weights matrices, W and M, we used a contiguity matrix, where adjacent properties get a weight of one and zero otherwise. Another popular choice is an inverse distance weights matrix (IDW), but because IDW requires the observations to be unique, we would lose half of our observations, leading to a substantially smaller sample size. 13 This is calculated as (exp(−0.395) − 1) × (1/(1 − 0.00557)) = −0.33. Note that, we also take into account the spatial multiplier 1/(1 − lambda) when determining the marginal effect.

nificant temporal decay effect; the large discount is short-lived. In the second column, we estimated the effect of the 1994 flood on flood-prone properties in the study area as determined by whether they fall in the inundated area (as in the second column of Table VI). We find that immediately after the flood, properties in the inundated area sold for 39% less than equivalent properties outside the inundated area, which is a 6% larger discount than for floodplain properties in model 1.14 Again, we find that the discount for the properties in the inundated area diminished over time. Finally, we reestimated the richer model 3. Compared to noninundated properties outside the floodplain, we find that inundated floodplain properties (IN ̲ FP) and properties that were inundated but were located outside the floodplain (IN ̲ OFP) were discounted by 41% and 33%, respectively. Consistent with the OLS results, the noninundated properties were not discounted significantly even if they were located in the floodplain (NIN ̲ FP). The significant spatial autocorrelation parameter (ρ) and spatial autoregressive coefficient (λ) toward the bottom of Table VII suggest that for all the specifications there is, in fact, spatial dependence among the properties in our data set in the expected direction: a positive adjacency effect. However, the impact of the dependency is very small, 0.0052–0.0057 across specifications, indicating that if the weighted average of neighboring houses’ sale price increases by 1%, the sale price of an individual house increases by approximately 0.0055%. Regarding the interpretation of the regression coefficients, in the spatial lag model, marginal effects are calculated by multiplying the estimates times a spatial multiplier, 1/(1 − λ).(30) A larger λ means a larger spatial dependence and thus, a larger spatial multiplier. As a second robustness check, we used an alternative proxy for flood damages: flood depth. In Table VIII, we report the OLS estimates of the DD models for this alternative proxy for flood damages. This variable also has the expected negative sign, although it is not statistically significant at the conventional levels. However, the results for the variables of interest are virtually identical: a 38% discount for floodplain properties (FP) in model 1, a 46% discount for inundated properties (IND) in model 2, and 39% discount is calculated as (exp(−0.496)−1) × (1/(1 − 0.00552)) = −0.39.

14 The

842

Atreya and Ferreira Table VII. SARAR Estimates of Difference-in-Difference (DD) Model

Variables FP FP* Flood FP* Years

Model 1

Model 2

−0.0664 (0.100) −0.395*** (0.122) 0.0503*** (0.0126) −0.0445 (0.0889) −0.496*** (0.115) 0.0649*** (0.0115)

IND IND* Flood IND* Years IN ̲ FP IN ̲ OFP NIN ̲ FP Flood

0.149 (0.196)

0.188 (0.196)

IN ̲ FP* Flood IN ̲ OFP* Flood NIN ̲ FP* Flood f(years)

−0.000836 (0.00838)

−0.00842 (0.00858)

−0.0270*** (0.00870) Yes Yes Yes 0.00557*** (0.000890) 0.0585*** (0.00501) 2,685

−0.0254*** (0.00863) Yes Yes Yes 0.00552*** (0.000894) 0.0591*** (0.00500) 2,685

IN ̲ FP* f(years) IN ̲ OFP* f(years) NIN ̲ FP* f(years) Damage (= census block average damage) Structural attributes Location attributes Census-block and year FEs λ ρ Observations

Model 3

−0.0999 (0.111) 0.0164 (0.138) 0.0269 (0.198) 0.183 (0.196) −0.526*** (0.129) −0.397** (0.200) 0.387 (0.354) −0.00555 (0.00868) 0.0709*** (0.0134) 0.0458** (0.0179) −0.0604* (0.0342) −0.0236*** (0.00873) Yes Yes Yes 0.0056*** (0.000894) 0.0590*** (0.00501) 2,685

Notes: Dependent variable is ln(Price). Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

a 48% discount for inundated floodplain properties (IN ̲ FP), 36% discount for inundated and outside the floodplain properties (IN ̲ OFP), and an insignificant discount for noninundated floodplain properties in model 3. To check for whether or not the decrease in housing prices after the flood in flood-prone

properties is due to a differential trend in housing prices in these areas rather than to the 1994 flood, we conducted several falsification tests. We estimated models 1–3, but this time using false flooding dates. In Table IX, we present the results of using 1991 as a false flooding date (three years before the actual flooding). The results in Table IX show no evidence

Seeing is Believing?

843

Table VIII. OLS Estimates of Difference-in-Difference (DD) Model with Alternative Definition of Damages Variables FP FP* Flood FP* f(years)

Model 1

Model 2

0.0558 (0.165) −0.491*** (0.150) 0.0548*** (0.0151)

IND

0.123 (0.123) −0.631*** (0.144) 0.0771*** (0.0145)

IND* Flood IND* f(years) IN ̲ FP IN ̲ OFP NIN ̲ FP Flood

0.830*** (0.186)

0.874*** (0.187)

0.0271 (0.0226)

0.0260 (0.0227)

−0.0166 (0.0163) Yes Yes Yes 2,685 0.318

−0.0219 (0.0201) Yes Yes Yes 2,685 0.325

IN ̲ FP* Flood IN ̲ OFP* Flood NIN ̲ FP* Flood f(years) IN ̲ FP* f(years) IN ̲ OF* f(years) NIN ̲ FP* f(years) Damages (= flood depth) Structural attributes Location attributes Census-block and year FEs Observations R2

Model 3

0.150 (0.183) 0.158 (0.145) 0.197 (0.242) 0.880*** (0.189) −0.677*** (0.161) −0.453** (0.202) 0.150 (0.492) 0.0235 (0.0226) 0.0833*** (0.0163) 0.0432** (0.0214) −0.0689 (0.0452) −0.0233 (0.0200) Yes Yes Yes 2,685 0.330

Notes: Dependent variable is ln(Price). Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

of a decrease in property prices after the false flooding dates, suggesting that there was no existing trend in the housing prices (the interaction between flood risk variables and the false flood date is insignificant) that could influence our results. This falsification test supports the causal interpretation of our DD model estimates of the impact of flood on housing prices. We also conducted two falsification tests using postflooding dates. With a falsification test using a date that is postflood (1997, for instance), we

suspected to see some remaining existing trend on the discount and decay effect that are spill-over effects of the 1994 flood. These effects, however, should be small. We present the results of the test in Table X. Conforming to our intuition, the 1997 false flooding date has an estimated negative impact on property prices that, compared to Table VI (with results for the true flood date), is much smaller and statistically insignificant in all the cases but for properties that were inundated and outside the floodplain

844

Atreya and Ferreira Table IX. Falsification Test: 1991 as Hypothetical Flood Year

Variables

Model 1

FP

−0.384 (0.285) 0.356 (0.225) 0.148** (0.0630)

FP* Flood FP* f(years)

Model 2

−0.348 (0.211) 0.333 (0.201) 0.167*** (0.0552)

INUNDATION IND* Flood IND* f(years) IN ̲ FP IN ̲ OFP NIN ̲ FP Flood

−0.226 (0.168)

−0.252 (0.178)

−0.0162 (0.0476)

−0.0463 (0.0575)

Yes Yes Yes 2,685 0.317

Yes Yes Yes 2,685 0.317

IN ̲ FP* Flood IN ̲ OFP* Flood NIN ̲ FP* Flood f(years) IN FP* f(years) IN OFP* f(years) NIN FP* f(years) Structural attributes Location attributes Census-block and year FEs Observations R2

Model 3

−0.498 (0.334) 0.0958 (0.233) 0.323 (0.411) −0.227 (0.177) 0.457* (0.239) −0.159 (0.239) −0.538 (0.410) −0.0403 (0.0626) 0.199*** (0.0691) 0.0679 (0.0725) −0.0586 (0.180) Yes Yes Yes 2,685 0.320

Notes: Dependent variable is ln(Price). Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

(and in this case both the magnitude and the level of significance are reduced). Similarly, both the economic and statistical significance of the decay effects in Table X are greatly reduced. We also conducted a falsification test using the false flood date of 1998, where, as expected, these effects (not reported here) are even smaller and always statistically insignificant. Finally, to address the possibility of omitted interaction variables associated with the repairs to severely damaged houses and to check for the validity of the decay effect, we removed the properties that were substantially damaged dur-

ing the 1994 flood from our sample. That is, we were able to parse out the census blocks where there were substantially damaged buildings due to the 1994 flooding. After removing these properties from the sample we retained 2,443 observations. As shown in Table XI, we find that the results are robust: the inundated properties were discounted irrespective of whether they fell in or outside the floodplain. The results also show that there was no significant effect on the properties that were not inundated even if they were in the floodplain. Regarding the decay effect, as in the

Seeing is Believing?

845 Table X. Falsification Test: 1997 as Hypothetical Flood Year

Variables

Model 1

FP

−0.0954 (0.160) −0.0383 (0.206) 0.0294 (0.0284)

FP* Flood FP* f(years)

Model 2

−0.118 (0.0991) −0.176 (0.182) 0.0650** (0.0253)

Inundation IND* Flood IND* f(years) IN ̲ FP IN ̲ OFP NIN ̲ FP Flood

0.507* (0.278)

0.558* (0.289)

−0.0339 (0.0225)

−0.0551** (0.0237)

Y Y Y 2,685 0.316

Y Y Y 2,685 0.320

IN ̲ FP* Flood IN ̲ OFP* Flood NIN ̲ FP* Flood f(years) IN FP* f(years) IN OFP* f(years) NIN FP* f(years) Structural attributes Location attributes Census-block and year FEs Observations R2

Model 3

−0.114 (0.167) −0.0158 (0.113) 0.320 (0.279) 0.612** (0.289) −0.117 (0.231) −0.378* (0.201) −0.583 (0.461) −0.0486** (0.0244) 0.0594* (0.0321) 0.0619** (0.0301) −0.0212 (0.0567) Y Y 2,685 0.326

Notes: Dependent variable is ln(Price). Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

benchmark specification, the discount for the properties in the inundated area diminished over time. 6. CONCLUSIONS Natural hazards provide exogenous risk information. Previous studies have found that the information provided by large floods is capitalized into property prices. These studies use floodplain maps to measure flood risk. Our study utilizes a flood inundation map in addition to the traditional FEMA floodplain map to determine the effect of the 1994

“flood of the century” caused by tropical storm Alberto in Albany, Georgia. This allows us to explore whether a potential effect of the flood on property prices was due to a pure information effect or due to an inundation effect. The overall pattern of the findings suggests that there was a discount for floodplain properties only after the flooding, particularly in the floodplain properties that were inundated as well. Our results accord with Tversky and Kahneman’s (11) availability heuristic, which postulates judgments of probability to be reflective of the availability of information.

846

Atreya and Ferreira Table XI. OLS Regression Removing Buildings with Substantial Damage

Variables FP FP* Flood FP* Years

Model 1

Model 2

Model 3

−0.0219 (0.106) −0.285** (0.143) 0.0473*** (0.0157) −0.0627 (0.0860) −0.454*** (0.126) 0.0673*** (0.0139)

IND IND* Flood IND* Years IN ̲ FP IN ̲ OFP NIN ̲ FP Flood

0.174 (0.110)

0.175 (0.110)

0.0120 (0.0152)

0.0135 (0.0151)

3.894 (2.923) Y 2,443 0.195

3.695 (2.925) Y 2,443 0.200

IN ̲ FP* Flood IN ̲ OFP* Flood NIN ̲ FP* Flood f(years) IN ̲ FP* f(years) IN ̲ OFP* f(years) NIN ̲ FP* f(years) Constant Structural/location/year FE/census-block FE Observations R2

−0.0451 (0.126) −0.0339 (0.123) −0.0479 (0.155) 0.171 (0.110) −0.450*** (0.149) −0.428** (0.196) 0.581 (0.481) 0.0134 (0.0152) 0.0705*** (0.0169) 0.0563*** (0.0200) −0.0560 (0.0429) 3.446 (2.953) Y 2,443 0.201

Notes: Dependent variable is ln(Price). Robust standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. All the regressions include census-block and year fixed effects.

When the impacts of floods are visualized and remembered, subjective risk probabilities are high, but as the effects of flooding fade over time (e.g., due to reconstruction), subjective probabilities decay and eventually vanish. In the absence of additional flooding, the price differential between the floodplain properties and the nonfloodplain properties did indeed decay and vanish. However, there was no significant discount associated with properties in the floodplain if they were not in the inundated area. These results suggest that most of the discount in

property prices in the area affected by a large flood event comes from being in the actually inundated area, and that not accounting for whether properties in the floodplains are also in the inundated area may overestimate the informational effect of large flood events and underestimate the value loss of inundated properties. The discount in inundated properties captures potential uninsurable reconstruction and psychological costs, and supports a hypothesis that homeowners respond better to what they have visualized (“seeing is believing”).

Seeing is Believing? The hypothesis of “seeing is believing” may have applicability beyond that of flooding, as evidenced in previous studies.15 For instance, Meyer(31) discusses the case where the residents of South Florida were underprepared during Hurricane Wilma in 2005 despite early hurricane warnings and ample time to prepare. One explanation for this outcome as presented by the author is that while residents had extensive experience in preparing for storms, far fewer had direct experience recovering from them, which led to underpreparation. These results offer important lesson for public risk communication. Flood risk, for instance, is often communicated using the probabilities of occurrence, but this might not be sufficient for the individuals never directly affected by flooding to prompt them to behave proactively to protect against flood risk. Future public risk communication strategies might consider explicitly addressing those individuals with little or no direct experience dealing with the hazard.

ACKNOWLEDGMENTS The authors thank Howard Kunreuther, Nicholas P. Magnan, Warren P. Kriesel, Elizabeth Kramer, Jeffrey H. Dorfman, two anonymous reviewers, the area editor, and participants at the annual meetings of the European Association of Environmental and Natural Resource Economists and the Agricultural and Applied Economics Association for useful comments. Financial support of the USGS/Georgia Water Resources Institute Grant #2521RC282387 is gratefully acknowledged. The usual disclaimer applies.

REFERENCES 1. Kunreuther H. Risk analysis and risk management in an uncertain world. Risk Analysis, 2002; 22(4): 655–664. 2. Thompson CP, Mingay D. Estimating the frequency of everyday events. Applied Cognitive Psychology, 1991; 5:497– 510. 3. Shilling J, Benjamin J, Sirmans C. Adjusting comparable sales for floodplain location. Appraisal Journal, 1985; 53(3):429– 436. 4. MacDonald D, Murdoch J, White H. Uncertain hazards, insurance, and consumer choice: Evidence from housing markets. Land Economics, 1987; 63(4):361–371. 5. Speyrer JF, Ragas WR. Housing prices and flood risk: An examination using spline regression. Journal of Real Estate Finance and Economics, 1991; 4(4):395–407.

15 We

thank the area editor for making this observation.

847 6. Harrison DM, Smersh GT, Schwartz AL. Environmental determinants of housing prices: The impact of flood zone status. Journal of Real Estate Research, 2001; 21(1):3–20. 7. Beatley T, Brower D, Schwab A. An Introduction to Coastal Zone Management, 2nd ed. Washington, DC:Island Press, 2002. 8. Bin O, Polasky S. Effects of flood hazards on property values: Evidence before and after Hurricane Floyd. Land Economics, 2004; 80(4):490–500. 9. Carbone J, Hallstrom D, Smith V. Can natural experiments measure behavioral responses to environmental risks? Environmental and Resource Economics, 2006; 33(3):273– 297. 10. Kousky C. Learning from extreme events: Risk perceptions after the flood. Land Economics, 2010; 86(3):395–422. 11. Tversky A, Kahneman D. Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 1973; 5(2):207–232. 12. Atreya A, Ferreira S, Kriesel W. Forgetting the flood: Changes in flood risk discount over time. Land Economics, 2013; 89(4):577–596. 13. Bin O, Landry C. Changes in implicit flood risk premiums: Empirical evidence from the housing market. Journal of Environmental Economics and Management, 2013; 65(3):361– 376. 14. U.S. Census Bureau. State and County Quick Facts: Dougherty County, Georgia. Available at: http://quickfacts. census.gov/qfd/states/13/13095.html, Accessed February 20, 2012. 15. Stamey TC. Summary of data-collection activities and effects of flooding from tropical storm Alberto in parts of Georgia, Alabama, and Florida. U.S. Geological Survey Open-File Report 96-228. Atlanta, GA: U.S. Geological Survey 1996. 16. Rosen S. Hedonic prices and implicit markets: Product differentiation in pure competition. Journal of Political Economy, 1974; 82(1):34–55. 17. Freeman A. The Measurement of Environmental and Resource Values: Theory and Methods. Washington, DC:RFF Press, 2003. 18. Bertrand M, Duflo E, Mullainathan S. How much should we trust differences-in-differences estimates? Quarterly Journal of Economics, 2004; 119(1):249–275. 19. Meyer BD. Natural and quasi-experiments in economics. Journal of Business and Economic Statistics, 1995; 13(2):151– 161. 20. Zhai G, Fukuzono T, Ikeda S. Effect of flooding on megalopolitan land prices: A case study of the 2000 Tokai flood in Japan. Journal of Natural Disaster Science, 2003; 25(1):23–36. 21. Anselin L, Bera AK. Spatial dependence in linear regression models with an introduction to spatial econometrics. Pp. 237– 290 in Ullah A, Giles D (eds). Handbook of Applied Economic Statistics, Vol. 155. New York: Marcel Dekker, 1998. 22. Anselin L, Florax R. Small sample properties of tests for spatial dependence in regression models: Some further results. Pp. 21–74 in Anselin L, Florax RJGM (eds). New Directions in Spatial Econometrics. Berlin: Springer, 1995. 23. McMillen DP. Issues in spatial data analysis. Journal of Regional Science, 2010; 50(1):119–141. 24. Pinske J, Slade ME. The future of spatial econometrics. Journal of Regional Science, 2010; 50(1):103–117. 25. Gibbons S, Overman HG. Mostly pointless spatial econometrics? Journal of Regional Science, 2012; 52(2):172–191. 26. Cohen J. Statistical Power Analysis for the Behavioral Sciences, 2nd ed. Hillsdale, NJ: Lawrence Erlbaum Associates Publishers, 1988. 27. Faul F, Erdfelder E, Lang AG, Buchner A. G*Power 3: A flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 2007; 39:175–191.

848 ¨ 28. Palmquist RB. Property value models. Pp. 763–819 in Maler KG and Vincent J (eds). Handbook of Environmental Economics: Valuing Environmental Changes, Vol. 2. Amsterdam, The Netherlands: Elsevier, 2005. 29. Pope D, Pope J. When Walmart comes to town: Always low housing prices? Always? NBER Working Paper No. 18111. Cambridge, MA: National Bureau of Economic Research Available at: http://www.nber.org/papers/w18111. 2013. 30. Kim CW, Phipps TT, Anselin L. Measuring the benefits of air quality improvement: A spatial hedonic approach. Journal of Environmental Economics and Management, 2003; 45(1):24– 39. 31. Meyer RJ. Why we under prepare for hazards. Pp. 153– 173 in On Risk and Disaster: Lessons from Hurricane Katrina.

Atreya and Ferreira SUPPORTING INFORMATION Additional Supporting Information may be found in the online version of this article at the publisher’s website:

Table S1: t-Tests on the Characteristics of the Average Property Before and After the 1994 Flood. Table S2: OLS estimates of Difference-inDifferences (DD) model using logarithmic decay function.