Abstract
We propose a theoretical framework characterizing (1) the attenuation of flood risks revealed by the flood zone designation in flood insurance rate maps and (2) the asymmetric impacts of adding versus removing flood zone status on property values. We apply spatial fixed effects models to empirically investigate the impacts of flood zone status and test the proposed theory. The results indicate that housing values decrease by more than 11% when a property is assigned into a flood zone. However, property values do not rebound when flood zone status is removed. (JEL Q51)
1. Introduction
Accurate flood risk maps are essential for homeowners and potential homeowners to make housing decisions, for insurance firms1 to set actuarially fair insurance rates, and for communities to set appropriate land use regulations and flood preparation plans. However, the accuracy of flood risk maps will change over time with shoreline erosion, the loss of wetlands, development patterns, climate change, changing hydrology, and advances in data and modeling accuracy (Association of State Flood plain Managers 2013).2
Given that flood risks change over time, the National Flood Insurance Program (NFIP) Reform Act of 1994 mandated the Federal Emergency Management Agency (FEMA) to review the flood insurance rate maps (FIRMs) every five years. However, in 2008, 70% of NFIP’s approximately 100,000 FIRMs were more than 5 years old and 50% were more than 15 years old (U.S. Government Accountability Office 2008). Furthermore, in 2009, only 21% of the continental U.S. population lived in areas that had flood insurance maps meeting the data quality standards of FEMA (National Research Council 2009). Making huge strides, by November 2015, 94% of the U.S. population lived in areas with updated flood maps (U.S. Government Accountability Office 2016). Since the creation of NFIP in 1968, $6.2 billion has been spent on flood mapping efforts (Association of State Floodplain Managers 2013).3
This paper uses the remapping of flood plains as a natural experiment to investigate how housing prices reflect flood risks revealed from the information provided in FIRMs. In performing cost-benefit analyses for improvements in flood map accuracy, FEMA classifies changes to property values as a benefit of remapping efforts because it reduces housing market inefficiencies due to misinformation about flooding risks (Association of State Floodplain Managers 2013). In order for this benefit to be realized, housing prices must ac-curately reflect the information provided by remapping, in other words, houses mapped into flood plains must reflect the new increased risk of flooding and houses mapped out of flood plains must reflect the decreased risk of flooding.
This research makes two contributions to the literature. First, this is the first paper, to our knowledge, that uses changes in flood risk mapping to isolate the effect of flood risk on housing values from other (dis-)amenities associated with location in a flood plain, such as adjacency to open spaces and water front as well as (in)accessibility to roads and other facilities. Second, we estimate how housing prices capitalize increases and decreases in flood risks, because we observe houses mapped into and out of flood plains. The premise is that homeowners are unlikely to have perfect information about flood risks, so they will largely depend on FIRMs to update their information about flood risk. The information will then be capitalized into the housing prices.
Studies have found evidence that flood risk discounts increase after a major storm event (Bin and Polasky 2004; Carbone, Hall-strom, and Smith 2006; Hallstrom and Smith 2005), and the risk discounts decay over time (Atreya, Ferreira, and Kriesel 2013; Bin and Landry 2013; Kousky 2010).4 Scholars have suggested several explanations for such behavior. Prospect theory argues that humans often poorly weight small-probability events (Kahneman and Tversky 1979) such as flooding. Availability heuristics suggests that individuals can overestimate the probability of an event that just happened because of its high saliency (Tversky and Kahneman 1973).
These mental shortcuts are exhibited in flood insurance market failure: low take-up rates and huge government spending on postdisaster aid.5
In light of these findings, we provide a theoretical framework to characterize the impacts of newly assigned flood zone status and potential decay and asymmetry of the impacts of flood zone status change on housing prices, building on the theory of incremental option value by MacDonald, Murdoch, and White (1987). Flood maps provide information about flood risks with areas designated as having a 1% annual chance of flooding (known as the Special Flood Hazard Area, SFHA), a 0.5% annual chance of flooding, and a less than 0.5% annual chance of flooding. Properties with a federally backed mortgage in the SFHA are required by the 1973 Flood Disaster Protection Act to purchase flood insurance. Therefore, the effect of being mapped into or out of the SFHA (or the flood plain) on housing prices will depend on the extent to which the new designation changes homeowners’ subjective beliefs about flooding risk and their annual flood insurance premiums.
When properties are mapped into flood zones, their sale prices should decrease due to updated beliefs of higher flood risks and the mandatory requirement to purchase flood insurance.6 This negative impact could attenuate over time without the occurrence of major flooding events as homeowners update their subjective beliefs about flood risks and as homeowners let their flood insurance policies lapse.7 When properties are mapped out of flood zones, their sales prices should increase due to updated beliefs of lower flood risk and the relaxation of the requirement to purchase flood insurance. However, it is possible that being mapped out of a flood plain does not change homeowners’ beliefs about flood risk, due to either a decay in risk assessment without the occurrence of floods or a failure to realize the lowered flood risk.
We test the theory with spatial fixed effects models that allow for asymmetrical responses mapped into versus mapped out of a flood zone. The models also allow the flood risk discount to decay over time. We use property transaction data with repeat sales in Centre County, Pennsylvania, from 1996 to 2012, and two FIRMs published in 1996 and 2004.8 We are able to focus on the impacts of flood map change in our empirical analysis because there were no major flood events recorded in Centre County during the study period.9
The results confirm the asymmetric impact of flood status change. Being mapped into a flood plain decreases property prices by more than 11%, but, surprisingly, there is no positive impact on property prices when the flood zone status is removed. We do not find strong evidence suggesting that the flood risk discount attenuates over time in our study region.
2. Theory of Impacts of Flood Risk on Housing Prices
We provide an extension of the theory of MacDonald, Murdoch, and White (1987) and Bin and Landry (2013) by accounting for the information of flood maps. Homeowners maximize their utility by choosing a location to purchase a home, and their state-dependent expected utility, E(U), from buying a house in a given location under two states—the states of flooding (F) and no flooding (NF)—can be written as [1] where q(i,SFHA) is the subjective probability of flooding given the information set, i, which can change due to the occurrence of flood events, and whether a property is in or out of a flood zone (SFHA ∈ (in, out)); the indirect utility, V, in each state is a function of a vector of housing attributes (a) and income (y), the hedonic price function of the property (H), the losses during a flood event (L), the insurance coverage (C), and the insurance premium, I(q,C).
We define option values (OV) as the maximum marginal willingness to pay to decrease the likelihood of the unfavorable state, F, by σ. With ŷSFHA = y - H(a, q(i, SFHA)), equation [1] can be rewritten as [2]
When a property is fully insured, option values can be interpreted as the expected differences between the losses and the insurance claim payments. In the case of no insurance, option values can simply be the expected sum of the losses. For owners with flood insurance, the option values can be reflected by their insurance costs. For owners without insurance, the option values would be realized by investing in any flood mitigation, such as elevating a house or waterproofing important areas.
The marginal change in OV with respect to the change in σ can be derived using the implicit function theorem:10 [3]
Maximizing the expected utilities in equation [1] with respect to the probability of flooding (q) yields the marginal implicit price of flood risks: [4]
Similarly, the special case of no insurance can be written as
[5]Based on equations [4] and [5], the impacts of a property’s flood zone status change on the marginal implicit price of flood risks can therefore be discussed in four cases.
Case 1.
Properties mapped out of floodplains that never had flood insurance.
The change in the implicit price of flood risk when properties that never had flood insurance are mapped out of floodplains can be written as [6]
In this case, we focus on the change in the implicit price of flood risk due to the difference in OV when mapped in and out of the flood zones. If the flood zone status provides information to homeowners about the probability of flooding such that homeowners in (or outside of) the flood zones believe that they have ≥ 1% (or ≤ 1%) chance of being flooded annually, q(i,in) (or q(i,out)), then Δ(∂H/ ∂q) >0.
Even if homeowners are located in floodplains, they could underestimate their risk of flooding because of several reasons, such as underweighting low-probability events (McClelland, Schulze, and Coursey 1993). It is also possible that homeowners do not believe that their probability of flooding has truly decreased even when mapped out of the flood plain. Therefore, if the flood zone status change does not change homeowners’ subjective flood risks, then Δ(∂H/∂q)=0 since q(i,in) = q(i,out).
It is also possible that the flood zone status provides information about flood risks, but the impacts of the information decay over time as homeowners update their subjective flood risks due to experience with floods. When properties are first mapped into a floodplain, homeowners may believe that they have a 1% (or even higher) chance of being flooded. Over time, if they do not experience a flood, they may reduce their subjective flood risk until there is no difference between the subjective flood risk for homes inside and outside of floodplains. As a result, q(i,in) will decrease over time and eventually reach some number close to q(i,out), and Δ(∂H / ∂q) → 0. The evolution of flood risk discount in this case is illustrated by the dashed line in Figure 1.
Case 2.
Properties with mandatory flood insurance mapped out of floodplains.
These properties can now be sold without the mandatory flood insurance premiums being capitalized into the value of the house.
[7]A priori it is unclear how the change in implicit price of flood risk is related to the reduction in mandatory flood insurance premiums,
In equation [3], if the losses from flooding (L) are perceived to be equal to the payment from flood insurance (C), then the indirect utility of the two states will be equal so that
Note that it is possible that which could occur if the losses from flooding are perceived to be greater than the payments
Thus, the change in the implicit price of flood risk is from flood insurance, that is, L > C.11 Conversely, if L < C, then
The capitalization of decreased flood risks and the removal of mandatory flood insurance purchases into housing prices requires that this information be disclosed to home buyers. The Real Estate Seller Disclosure Act in Pennsylvania, passed in July 1996, required the disclosure of flood damages to a home and also mandatory flood requirements. However, the decreases in flood risk or mandatory flood insurance payments may not be fully disclosed and believed by home buyers. Therefore, this information might not be capitalized into housing prices when properties are mapped out of floodplains.
Case 3.
Properties mapped into floodplains that do not purchase flood insurance.
Similar to case 1, the only change in the implicit price of flood risk is due to the difference in option value when mapped in and out of the flood zones. [8]
If the change in flood zone status makes homeowners update their subjective probability of flooding upward, then Δ(∂HI=0
The change in the marginal implicit price of flood risk will be less than the change in insurance premiums:
that is,
if being in the flood zones increases homeowner’s subjective flood risk or if the losses from flooding are perceived to be greater than the payments from flood insurance, so that L > C. The risk discount could, similar to case 3, change over time as subjective flood risk decays or as homeowners update their perceptions of how payments from flood insurance ∂qout→in) < 0. However, homeowners may not update their beliefs about the probability of flooding with new information provided in the maps, or the impacts of the information provided by the flood zone status may decay over time as homeowners update their subjective probability of flooding, q(i,in), due to experience with (no) floods. Again, when no flood event occurs, it is reasonable to assume that q(i,in) will approach q(i,out) over time, so equation [8] approaches zero. The evolution of flood risk discount is illustrated by the solid line in Figure 1, which is similar to that proposed by Tobin and Newton (1986) under the impact of a rare flood event on property values.
Case 4.
Properties mapped into floodplains that purchase flood insurance.
Properties cannot be sold without mandatory flood insurance after the status change,12 thus the net present value of these premiums should be capitalized into the housing value. [9] compare to losses. If homes in floodplains are deteriorating over time as flood damages exceed insurance claims, then Δ(∂H / ∂q) will decrease more over time than -[∂I(q,C)]/∂q. However, if homes in floodplains are improving over time because damaged structures or contents are being upgraded after floods, then Δ(∂H/∂q) will decrease less over time than -[∂I(q,C)] / ∂q.
These four cases provide the following testable hypotheses on the effect of flood insurance mapping on housing prices.
Hypothesis 1A. When homes are mapped out of floodplains, there are positive changes in housing prices, namely, Δ(∂HI∂0 / ∂qin→out) > 0.
Hypothesis 1B. When homes are mapped into floodplains there are negative changes in housing prices, namely, Δ(∂HI∂0 / ∂qout→in) < 0.
Hypothesis 2. The change in housing prices after being mapped into a floodplain decays over time, Δ(∂HI∂0 / ∂qout→in) → 0. Therefore, Δ(∂HI∂0 / ∂qin→out) < 0 but approaches zero over time as homeowners reduce their subjective flood risk until there is no difference between the subjective risk for homes inside and outside of floodplains: Δ(∂HI∂0 / ∂qout→in)=0.
3. Data
We collected two major datasets for our analysis: (1) housing characteristics and transaction data, and (2) FIRMs. First, we have historical transaction prices, the sale date, and current detailed housing characteristics data provided by the Tax Assessment Office of Centre County, Pennsylvania. Second, we collected two digitized FIRMs, which were released in 1996 and 2004, generating variations in whether a parcel is inside a flood zone when the transaction happens. The present analysis uses the transactions between 1996 and 2012.13 The first release of the FIRMs also coincided with the Real Estate Seller Disclosure Act in Pennsylvania, passed in July 1996. Therefore, a widespread unawareness of properties mapped into flood plains among buyers was unlikely.14 The sale prices are indexed to the fourth quarter of 2012, using the quarterly all-transaction house price index of State College metropolitan statistical area, provided by Federal Housing Financial Agency. A property is assigned inside a flood zone if any portion of the parcel is inside a flood zone. The 2004 flood zone map and the township boundary of Centre County are shown in Figure 2.
The sample is further limited to single-family residential houses and excludes seasonal properties, mobile homes, and properties with some commercial function. We dropped the transactions with sale prices smaller than the lowest adjusted price of a valid sale ($5,450). The selection criteria leave the data with 32,874 transactions from 20,363 parcels. In order to exploit the parcel-level fixed effect, parcels with only one transaction during the study period are dropped. This reduces the sample to 21,137 transactions from 8,628 parcels. The summary statistics of these observations are presented in Table 1. Note that the housing attributes observed in the data are those in the most recent transaction of each parcel, so all the housing attributes except whether located in flood zones and house age are assumed to be time invariant. In addition, minutes to the campus of Pennsylvania State University is calculated using ArcGIS with the GIS street layer.15
Summary Statistics of Housing Transaction Data in Centre County, Pennsylvania, 1996–2012
Except for the most recent transaction of a parcel, the housing transaction data do not specify whether a transaction is a valid arm’s-length transaction. Therefore, among all the repeat sale transactions, we applied a “normal sale screening” procedure to trim the transactions that are unlikely to be arm’s-length ones. The ith transaction of a parcel with price pi is dropped if [10] where Ti,i-1 and Ti,i+1 are the time (in years) between the ith transaction and its previous (i – 1) and next (i + 1) transaction, respectively.16 This screening procedure essentially drops those transactions with prices much lower than the previous or following transactions of the same parcel. Transactions with (abnormally) low prices are likely to be those between families or foreclosures. Moreover, if both transactions were officially recorded as valid sales but one was with a price much lower than the other, this could indicate that the property was significantly deteriorated, improved, or even have changed some of the housing attributes.
This procedure reduces the sample to 14,926 transactions on 6,342 parcels. The summary statistics of this sample are reported in Table 1. Among these transactions, those in a flood zone account for about 4%, which is comparable to many studies on more flood prone areas. The summary statistics of the remaining transactions by their flood zone status are reported in Appendix Table A2.
4. Empirical Model
Because flood zone status is often correlated with other environmental (dis)amenities, such as adjacency to open spaces and waterfront and (in)accessibility to roads and other facilities, the estimate of flood risk discount can be biased if those related attributes are not controlled. Unfortunately, those potentially related attributes are not directly observable in the present dataset (and in many other hedonic studies). Spatial fixed effects are frequently used to overcome omitted variable bias in hedonic price analysis (Parmeter and Pope 2011).17 Many of the reviewed studies on the impacts of flood risks or flood zone status apply a spatial error/lag model with a spatial weights matrix to correct spatial dependence and autocorrelation, such as those of Bin and Landry (2013) and Atreya and Czajkowski (2016). Fixed effects with clustered errors can also treat the problems of spatial autocorrelation and dependence, where the fixed effects correct the former and error clustering corrects the latter (Heintzelman and Tuttle 2012). The present study therefore employs the same strategy by including fixed effects at two different spatial scales—the township level and the parcel level—with clustered errors.18
The hedonic regression with township fixed effects can be written as [11] where pijt is the index-adjusted sale price of property i in township j at time t; Fijt is a vector of flood status–related variables, and β is a vector of the marginal implicit prices of the flood-related variables; xijt is a set of housing attributes, with γ representing the vector of marginal implicit prices of the corresponding attributes; δj is a set of dummy variables of township characteristics, representing the fixed effects, while τt consists of the time (year and quarter) dummies; lastly, νjt and εit are the error terms at the township and the individual level, respectively. Note that, although housing characteristics can certainly be time variant, the housing characteristics included in our analysis, other than the flood-related variables and the age of house, are assumed to be time invariant given the data availability.
Similarly, the model with parcel fixed effects is [12] where κi, replacing δj, represents the fixed effects at parcel level; Ait is the age of the house, and α corresponds to the marginal price. As noted above, we specify the errors to be clustered at each corresponding spatial level to treat the issue of spatial dependence.
In a parcel fixed effects model, the implicit prices of the variables of interest are identified through the temporal variations of the attributes. In the current context, the implicit price of flood risk can be identified only when there are temporal variations in flood zone status, before and after the remapping in 2004, for at least some parcels. The asymmetry when estimating the implicit prices of flood risks can be illustrated by rewriting equation [12] in the following equivalent form: [13] where FIN _ it =1 if a transaction happened in t ∂ 2004 and the parcel was outside of a flood zone before 2004 but was in a flood zone in t ∂ 2004, and FIN_it = 0 otherwise.19 βIN is therefore expected to be negative. Conversely, FOUT_it = 1 if a transaction happened in t ∂ 2004 and the parcel was in a flood zone before 2004 but was not in t ∂ 2004, and FOUT_it =0 otherwise. βOUT is then expected to greater than or equal to zero, as shown in equation [6]. If there is no such proposed asymmetry, then we have βIN = -βOUT =β<0, as β= (βIN - βOUT)/ . That is, there is a risk discount for properties located in flood zones regardless of the direction of flood zone statusβOUT-A)UT (or |βIN > ∂Æ)UT), so βIN <β<-βOUT. Therefore, without distinguishing this two-way effect of flood risks but simply looking at the coefficient estimate of β, we would obtain a confounded effect of flood risks.
Furthermore, we use the following specification to examine the evolution of the flood risk discount: [14] where g(FIN.TIMEit) is a function of the time (in years) since being mapped into a flood zone for property i at time t.
5. Results
Flood Risk Discounts
We use three samples—most recent valid sales, repeat sales, and screened repeat sales— and two fixed effect scales—township and parcel level—to estimate five models.20 The model results are presented in Table 2, and we first discuss the models using our preferred parcel fixed effects.
The flood risk discount estimates from the parcel fixed effects models (models 4 and 5) are negative but statistically insignificant. These results highlight the argument in our empirical section regarding the confounded effect of flood risk when the effects of mapping in and out are not distinguished, which we will explore in our models on the asymmetry of the flood risk discount. Comparing models 4 and 5, where the latter uses the sample with the normal sale screening procedure, model 5 appears to have a dominant model fit (with adjusted R2 = 0.8865 in model 5 vs. 0.4966 in model 4). This is indicative that the screening procedure effectively filtered some of the non-arm’s-length transactions with unreasonable transaction prices. In addition, the estimated flood risk discount in model 5 (4.94%) is lower than that in model 4 (8.20%). This suggests that there are more transactions with low prices being dropped for properties in flood zones. We argue that this is a logical inference because properties in flood zones are more likely to have improvements, such as those for flood mitigation or flood aftermath cleanup, and many of the improvements with lower transaction prices recorded are trimmed by our screening procedure.
Regardless of the set of samples used, all township fixed effects models indicate a negative and significant risk discount ranging from 6.16% to 13.40%.21 Comparing models 2 and 3, the latter of which uses the normal sale screening sample, model 3 appears to have a better fit. The adjusted R2 and the estimate of flood risk discount in model 3 are comparable to those in model 1, which uses only the most recent valid sale samples and the same township fixed effects. This is another indication that the normal sale screening procedure effectively mimicked a repeat sale sample set with valid transactions.
Comparing the models using parcel and township fixed effects (i.e., model 2 vs. 4 and model 3 vs. 5), we find that using parcel fixed effects not only offers better model fits but also controls for, at least partly, the omitted variable biases stemming from using coarser fixed effects. In particular, the comparison indicates that the township fixed effects models could be suffering from an overestimated negative impact of flood risks. The downward-biased estimates in the township fixed effects models suggest that there are disamenities of properties in flood zones being omitted. We consider, in our case, that some of the omitted attributes are associated with less ideal locations in a given township, such as (longer) distances to shopping, schools, and hospitals. The estimates for the marginal implicit price of house age are significantly negative only in the parcel fixed effects models. This also indicates possible omission biases in the township fixed effects models.
Most of the coefficient estimates of other housing characteristics are significant and of the expected signs. People are willing to pay more for a larger parcel, more living space, more bedrooms, a larger garage for more cars, a basement, a finished basement, a fireplace, a pool, and being closer to the city center (i.e., Penn State campus). The negative estimates of full baths could be explained by smaller living area or bedrooms, given that the size of living area and number of bedroom have been controlled for.
Asymmetry in Flood Risk Discounts
To examine the asymmetric impacts of flood zone status change, the model presented in equation [13] is estimated. The results are shown in Table 3, and we first examine the results from our preferred model, based on the previous discussion, with the screened repeat sale sample and parcel fixed effects, which is model 5A. Model 5A shows that being mapped out of flood zones has no statistically significant impact on property prices, but being mapped into flood zones decreases prop-erty prices by 11.64%. This is equivalent to an average decrease of about $19,981 in adjusted sale prices. The number is nearly identical to the net present value of the average insurance payment in Centre County in perpetuity: $20,000.22 Similar results are found in model 4A, where we, once again, observe the more negative impact of flood risk when the sample is not treated with our normal sale screening procedure. These results largely support our argument regarding the confounded effect of flood risk when the effects of being mapped in and out are not explicitly modeled.
The results of the township fixed effects models (models 2A and 3A) indicate similar impact of being mapped into flood zones: a decrease in property prices of around 15%. However, there are serious biases in the impact of being mapped out of flood zones. The estimates show that being mapped out of a flood zone will decrease property prices by more than 7%, regardless which set of samples is used. This again highlights the need to control for the effects from any unobserved and time-invariant property characteristics.
Attenuation of Flood Risk Discount and Explanations for the Asymmetry
Table 4 presents models that characterize the attenuation of the flood risk discount over time. The five models (1D–5D) use a chisquare distribution with different degrees of freedom (3 to 7). This specification implies that there is a large initial (negative) change in risk discount and recovery at a decreasing rate. The model estimates are translated into flood risk discount by year and illustrated in Figure 3. The figure shows that the discounts can reach more than 15% in the first few years after a property has been mapped into a flood zone and attenuate over time. After 14 years, the discount is less than 2% regardless of specifications.
22 The average flood insurance premium in Centre County, Pennsylvania, is around $1,000. The net present value is calculated assuming a 5% discount rate.
Imposing the functional form of the decay could potentially force the decay to appear. We tested several other specifications for the patterns of risk discount over time, such as adding a quadratic term or a natural logarithm of years after change, and using a Box-Cox transformation or a ratio form, which impose less extreme curvature compared to the natural logarithm form. Some of the specifications were used by Bin and Landry (2013) to model the change of the flood risk discount after a major flood event. The tested specifications are reported in Appendix Table A3, with the results following in Table A4.23 The initial decrease in property values after the assignment of flood zone status is identified in most of the models, but the decay pattern is not robust. Therefore, we have only some weak evidence to support our second hypothesis: the flood risk discount decays over time without the reoccurrence of major flood events.24
We acknowledge that there are potential biases in our parcel fixed effects models because of the constraints of our data. First, we assume buyers are fully aware of whether a property is located in a flood zone, given the disclosure act in Pennsylvania. In reality, this is still not the case for all buyers. Failure to control for the impact of flood risk information disclosure would underestimate the flood risk discount. Pope (2008) exploited a mandate disclosure act in North Carolina and found no such discount without the disclosure act. Therefore, the flood risk discount of 11.64% we found for properties being mapped into flood zones could be only a lower bound. Conversely, in the case when a property is mapped out of a flood zone and the previous owner was not aware of the flood risk of the property, we would observe a smaller premium for removing the flood risk. This can also be another potential cause for the asymmetry. After all, we consider the impact of information disclosure to be minimal in both the cases because there was no evidence that the extent of disclosure had changed dramatically during our study period.25
Second, we observe the characteristics of a property only in the most recent transaction and assume the housing characteristics were the same in all previous transactions. In our parcel-fixed effect models, this assumption would result in the changes in flood zone status capture the effects of housing attribute and condition changes. Specifically, the flood risk discount for properties being mapped in could again be underestimated because of not considering the housing attribute changes (particularly improvements). For those properties being mapped out, if housing characteristics changes result in lifting the house out of flood zone, then there will be no bias but an incorrect interpretation for the price premium. If the change in characteristics does not relate to updated flood risks, the estimate for being mapped out would be biased upward. We consider this possibility as less likely to be relevant because we do not find positive impacts of being mapped out across all of our models. In summary, these potential omissions and corresponding biases may exist but do not change our findings on the asymmetry impact of flood zone status change.
6. Conclusions
Since the FIRMs have been produced, under the National Flood Insurance Program (NFIP), how flood risks revealed by flood zone status are capitalized into housing prices has increasingly drawn the attention of researchers and policy makers. This study, exploiting the revisions of FIRMs, investigates how housing prices reflect flood risks as revealed from the information provided by FIRMs. We propose an extension of the theory on the marginal implicit prices and option value of avoiding flood risks. The theory characterizes the decay of the impacts of being mapped into flood zones on housing prices and the asymmetric impact of flood zone status change.
We applied fixed effects models with property transaction data from Centre County, Pennsylvania, to estimate the marginal implicit price of flood risks. The results of the parcel fixed effects models show a negative but insignificant flood risk discount when the effects of being mapped in and out are not distinguished. However, the township fixed effects models suggest a more than 7% flood risk discount. This finding indicates omitted variable bias in the township fixed effects models. There are housing or neighborhood attributes that can negatively affect property prices that are correlated with flood zone status.
The analysis then explicitly models the impacts of being mapped into and out of a flood zone. We find that, when previously flood-free properties are assigned into flood zones, their sale prices, on average, decrease by 11.64% based on our preferred model (screening repeat sales sample and parcel fixed effects). The price difference before and after being mapped in is comparable to flood insurance premiums in the county. Moreover, we model the temporal change of the flood risk discount, although we do not find strong evidence showing the risk discount attenuates over time. For the other direction of the change, properties previously located in flood zones becoming flood-free has no significant impact on sale prices. These asymmetric impacts of being mapped into and out of flood zones also lead to the insignificant estimate of flood risk discount when the asymmetry is not accounted for.
In summary, the findings suggest that keeping flood risk maps updated is important so that homeowners can accurately assess their flood risks (or, at least, be aware of the risks), particularly given that this flood risk information from the map updates can decay. Without exposure to floods, homeowners can do a poor job of assessing flood risks. Given the path of the risk discount decay over time, it seems that the mandated five-year review cycle for FIRMs in the NFIP Reform Act of 1994 could keep citizens aware of their flood risks.
We note that it is possible that previous flood risk will still make these properties undesirable even after the risk is alleviated. Identifying such a stigma effect is of interest for future studies. One other interesting potential investigation would involve the intensity of the flood map change. That is, the effects of being mapped in or out could be associated with how close a property was to the old or new flood zones before the map revision.26 Lastly, the framework proposed in our study can be utilized to study how the flood risk discount evolves in response to flood zone mapping and actual storm/flood events simultaneously, possibly when a sufficiently large sample is available.
Acknowledgments
The authors greatly appreciate the help and data support from the following people in the Centre County government: Nick Barger (GIS Director), Joseph Davidson (Recorder of Deeds), and Jennifer Pettina (Assistant Chief Assessor). We would also like to thank Scott Colby, Minghao Li, Richard Ready, James Shortle, and two anonymous reviewers for helpful comments.
Footnotes
↵1 In the case of flood insurance, currently the Federal Emergency Management Agency sets insurance rates (Federal Emergency Management Agency 2002).
↵2 For example, the Federal Emergency Management Agency studied the effect of planned development on the increase in expected flood damages to current buildings and found that the effect ranged from a 20% increase in Fort Collins, Colorado, to more than a 1,200% increase in Harris County, Texas (Congressional Budget Office 2009). Furthermore, the area with a 1% annual chance of flooding is expected to increase by 45% in riverine areas and 55% in coastal areas across the United States due to climate change (AECOM 2013).
↵3 This amount includes funds generated from flood insurance fees and appropriations from Congress. For example, Congress appropriated $1.2 billion between 2003 and 2008 to update flood maps under the 1997 Flood Map Modernization Initiative (U.S. Government Accountability Office 2010). To give a sense of the magnitude of spending on mapping flood risk, the NFIP has collected approximately $56 billion in earned premiums since its inception.
↵4 In addition, many studies have investigated the impact of flood hazard on housing values, for example, Bin, Kruse, and Landry (2008), Daniel, Florax, and Rietveld (2009), de Koning, Filatova, and Bin (2017), Donnelly (1989), Harrison, Smersh, and Schwartz (2001), MacDonald et al. (1990), Rambaldi et al. (2013), Samarasinghe and Sharp (2010), and Troy and Romm (2004).
↵5 For example, in 2005, four hurricanes, including Katrina, incurred FEMA nearly $24 billion in debt (American Institutes for Research 2005).
↵6 A summary of the national estimates of market penetration rates for flood insurance is presented in Appendix Table A1.
↵7 Bin and Landry (2013) and Gallagher (2014) find that the number of flood insurance policies in force increase after major storm events and then decrease over time. In addition, Michel-Kerjan, Lemoyne de Forges, and Kunreuther (2012) find that only 20% of the flood insurance policies purchased in 2001 were still in force in 2009.
↵8 These two historical FIRMs were provided through personal contact and discussion with Nick Barger, GIS Direc-tor, at the Centre County government. The current FIRM, published in 2012 and not used in our analysis, is available from Pennsylvania Spatial Data Access (http://www.pasda.psu.edu.).
↵9 According to the NOAA storm event database, during 1996–2017, the estimated damages due to floods and flash floods are $6.72 million in Centre County; only $0.22 million damages were incurred between 1996 and 2012. Note that the damage estimates are not based on insurance claims, and the database FAQ page states that “the National Weather Service makes a best guess using all available data at the time of the publication. The damage amounts are received from a variety of sources, including those listed above in the Data Sources section. Property and Crop damage should be considered as a broad estimate” (NOAA National Centers for Environmental Information 2018).
↵10 If L includes noninsurable losses, such as sentimental ones, even if a homeowner is fully insured, the losses will be greater than coverage, L-C>0, so VNF(⋅)> VF(⋅) and dOVSIFHA /dσ> 0. It is also possible that the losses are perceived to be less than the payment received from flood insurance. In this case, VNF(⋅)< VF(⋅)and . This could be because the replacement value of items damaged in floods is worth more to people than the items themselves. Therefore, when insurance is presented in the state-dependent expected utility, the sign of the incremental option value is indeterminate. In the case of no insurance, C and I(q,C) both equal zero; and since the no-flooding state is preferred to the flooding state, namely, VNF(⋅)> VF(⋅, we have
↵11 Homes in floodplains may be underinsured because the NFIP caps the coverage at $250,000 for primary residences.
↵12 Grandfathering allows properties that were in compliance with their previous FIRM to maintain their (lower) flood insurance premiums when FIRMs are adjusted to reflect higher flood risks, but still requires properties mapped into the SFHA to purchase flood insurance, although at a lower rate. The Biggert-Waters Flood Insurance Reform Act of 2012 removed grandfathering; however, the Homeowner Flood Insurance Affordability Act of 2014 reinstated grandfathering (National Research Council 2015).
↵13 Although available, we do not include the transactions after 2012 because the grandfather clause was removed by the Biggert-Waters Flood Insurance Reform Act of 2012 but was later reinstated with the Homeowner Insurance Affordability Act of 2014. These legislation changes also make the identification difficult without data on flood insurance purchases at the household level.
↵14 The Real Estate Seller Disclosure Act in Pennsylvania (P.L. 500, No. 84) was first passed on July 2, 1996. This act included a Seller’s Property Disclosure Statement and questions regarding property’s flooding problems.
↵15 Centre County is tested to be a very monocentric area (Coulson 1991), where many amenities, such as shopping, hospitals, and schools, surround the campus of Penn State. Therefore, we consider adding one such proximity variable to be sufficient to control for a majority of other amenities.
↵16 As one reviewer suggested, we performed robustness checks by trimming the top 1%, the bottom 1%, and both ends of the distribution, based on the screened repeat sales sample. The results do not suggest any qualitative differences. In addition, we applied other screening rules by dropping parcel i if pi<(d+r×Ti,i-1)pi-1 or pi<(d+r×Ti,i+1)pi+1, with d = 1.5, 1.4, 1.3, or 1.2 and r = 0.05, 0.04, 0.03, 0.02, or 0.01, but the results do not suggest any qualitative differences. We therefore picked the midpoints of d and r.
↵17 A parcel fixed effects model is arguably the best strategy to tackle the issue of omitted variables, because any time-invariant attributes of a parcel will be completely controlled for in the included fixed effects. In addition, a spatial fixed effects model is able to correct the problem of endogeneity, although such a problem is much less likely to exist in our context since the price of a house is unlikely to affect the chance of a house being located in a flood zone.
↵18 Using such an approach, particularly with parcel fixed effects, in hedonic price analysis within a quasi-experiment setting is increasingly common. See, for example, Heintzelman and Tuttle (2012) on wind power facilities and Buck, Auffhammer, and Sunding (2014) on irrigation water. Among the studies on floods, Kousky’s (2010) also applied this approach to study the impacts of a major flood event on the housing market.
↵19 We also used different cutoffs—the ends of each quarter in 2004—and the results are not qualitatively different.
↵20 We cannot estimate the most recent valid sales with parcel-level fixed effects because we need repeated sales for such estimation.
↵21 Previous U.S. studies find the marginal (negative) impacts of flood risk on housing prices to range from 2.8% to 12%.
↵23 We also investigated a semiparametric form that interacts being mapped out of a flood zone with year dummies, but most of the estimates turned out to be insignificant.
↵24 With the inclusion of extra samples from 2013 to 2015, we tested whether risk discount increased again due to Hurricane Sandy and a flash flood in 2013 but did not detect any significant patterns. We suspect the data does not have enough observations/variations after 2012 for us to identify any potential patterns. Hurricane Sandy forced Penn State to cancel classes on October 29, 2012, for the first time since at least 1994 for other than a snowstorm, although there was no significant damage or loss reported in Centre County. Sandy also brought one of the most serious and widespread flooding events for the northeastern coast. In addition, a flash flood on June 27, 2013, hit multiple locations throughout the county. The flood resulted in property damage estimated at $1.5 million, and it was the costliest flood event in the county since the NOAA storm event database started collecting flood event data in 1996.
↵25 Bin and Landry (2013) point out that recent migrants are likely to be less informed about local flood risks, so their purchases can mitigate the negative impacts of flood risk on property prices. We consider this as less likely to be the case in our study area because local residents have little experience with flood events in the region.
↵26 We thank an anonymous reviewer for pointing out this issue. We did attempt to include the distances to old/new flood zone boundaries, but, possibly due to our small sample size, we did not find significant impacts of the distances on the effects.