Housing market dynamics of the post-Sandy Hudson estuary, Long Island Sound, and New Jersey coastline are explained by NFIP participation

This study centered on 496 coastal census tracts across New York, New Jersey, and Connecticut (figure 1). The average number of housing units per census tract was relatively stable with a mean of 1890 units in the year 2000 reaching a peak of 2031 units in 2010 and then dropping to 1970 in 2020 (Steven et al 2022, U.S. Census Bureau 2022a). Median HPs increased from $274 500 in 2000 to $504 100 in 2010 and $581 700 in 2020 (Steven et al 2022, U.S. Census Bureau 2022b).

Zoom In Zoom Out Reset image size

This region has a documented history of flooding from storm surge, riverine, and pluvial flooding (Depietri et al 2018). The spring season is characterized by moderately intense precipitation on snow, potentially inducing riverine or pluvial flooding. Late summer and fall are characterized by heavy tropical storms that bring both intense precipitation and wind-driven storm surge. Nor'easter cyclones are most common during the winter season and can cause elevated storm surge and intense precipitation. The average annual maximum daily precipitation was 7.4 cm¹ day⁻¹ with nine precipitation events exceeding 10 cm¹ day⁻¹ from 1970–2021 (NOAA 2022a). Mean high tide at the battery, NY is 0.76 m MSL. From 1970 to 2021, the region experienced an average annual maximum storm surge of 1.57 m MSL, with three events exceeding 2 m MSL (figure S1) (1992, 2011, 2012) (NOAA 2022b).

From 1979 to 2021, the average annual number of flood insurance claims submitted through NFIP in this region was 3,476 claims¹year⁻¹ with an average annual claim total of $112 million. Historical NFIP flood losses were highest with 75 247 claims in 2012, due to rainfall and storm surge from Hurricane Sandy (peak water elevation of 3.5 m MSL at Battery, NY). Reported NFIP claims in this year were approximately $4.86 billion. Following Hurricane Sandy the rate of new NFIP flood insurance policies decreased from a maximum of 182 530 insurance policies held in 2013–144 954 in 2021 (US FEMA 2022).

We studied relationships between HP dynamics and community awareness of flooding with a SE model (Baldassarre et al 2015, Barendrecht et al 2019). This model simulates annual economic losses (L), population density (D), community flood awareness (A), and flood preparedness (P) based on a time series of peak annual flood hazards (figure 2). Flood hazards (i.e., high surge elevations, extreme rainfall, (W)) above some threshold (T) result in a simulated loss (L), based on current housing density (D) and a parameter controlling the hazard to loss ratio (β_R ). Losses greater than 0 cause A (and preparedness, P) to increase and D to decrease. In the absence of floods, A and P gradually decay and D increases. This model structure has been shown to accurately simulate human-flood dynamics across a variety of case studies (Barendrecht et al 2019, Knighton et al 2021). Datasets used in model development are presented in section S1.

Zoom In Zoom Out Reset image size

For application to the study region, we modified this model to separate the effects of surge- and rainfall-driven flooding events, which yielded a significant improvement in simulations across the study region (figure S3). A full description of this modification is presented in section S2. We hereafter refer to the SE model with this modification as the base model. The SE model was also adapted to include three separate approaches to simulating temporal changes in HPs due to flooding impacts: a sigmoid recovery (Pryce et al 2011), an interrupted time series (ITS) recovery, and a K-shaped recovery (Beltrán et al 2019) (figure 2). A full description of time varying HP models is presented in section S3. We calibrated the SE model for each census tract individually with data spanning 1970–2021. The difference between normalized total annual NFIP claims and simulated L were summarized with the Kling-Gupta Efficiency (KGE_L) (equation 1), where r is the linear correlation coefficient between simulated and observed values, σ_sim is the standard deviation of simulated values, σ_obs the standard deviation of observed values, μ_sim the average of simulated values, and μ_obs the average of observed values

$$\begin{equation}{\text{KGE}} = 1 - {\text{ }}\sqrt {{{\left( {r - 1} \right)}^2} + {{\left( {\frac{{{\sigma _{{\text{sim}}}}}}{{{\sigma _{{\text{obs}}}}}} - 1} \right)}^2} + {{\left( {\frac{{{\mu _{{\text{sim}}}}}}{{{\mu _{{\text{obs}}}}}} - 1} \right)}^2}} .\end{equation} \tag{ 1 }$$

KGE_A was computed between annual simulated awareness (A) and the normalized value of active NFIP policies. KGE_D was computed between simulated population density (D) and normalized observed population density. KGE_HP was computed between simulated HPs and normalized observed HPs. All metrics were first compared to the KGE mean observational benchmark, above which models have predictive skill (Knoben et al 2019). We then evaluated the relative performance of each HP model through Mann Whitney U-tests comparing all KGE scores across all census tracts. This calibration approach was used to determine which HP model was most appropriate and to estimate relationships between HP dynamics and community perceptions of flooding.

Integration of housing models into an SE model supports an exploration of relationships between housing market values, external flooding hazards, and perceptions of future risk. Through model calibration, we estimated parameters related to community-scale flood behaviors. Anxiousness (ɑ_A ) and forgetfulness (μ_A ) control how population density changes immediately following a flood-related loss and through periods of no loss, respectively. The decay of precautionary measures (μ_P ) controls how floods impact simulated awareness (A) and preparedness (P) (figure 2) which in turn impacts the loss from future surges. We also estimated parameters related to HP change. Parameters b2 and b3 which control initial drop in HPs due to flooding and the rate of HP recovery, respectively. We examined relationships between these calibrated model parameters to those parameters describing temporal changes in the housing market using Spearman's ranked correlation.

Calibration of the models to observed NFIP claims, policies, population density, and mean HPs demonstrated that the ITS model provided the most accurate representation of flooding dynamics across the study region (see section S3 for further details). Each model was capable of adequately simulating the period of rising HPs from 1970 through 2011, with some disagreement on the impact of minor flooding events (figures 3(a)–(d)). Each model simulated different HP responses to flood events in 2011 (tropical storm Lee) and 2012 (hurricane Sandy). The Sigmoid and K-shaped models each imposed immediate drops in HPs and monotonic recoveries starting in 2013 (figures 3(b) and (d)). As this pattern was not evident in the observed data, the calibration algorithm minimized model residuals by estimating parameters for each model that predicted small losses and limited recovery (figures 3(b) and (d)). Calibration of the ITS model similarly reproduced the case of a limited initial impact on HPs followed by a period in which HPs steadily declined (figure 3(c)).

Zoom In Zoom Out Reset image size

We compared correlations between SE model parameters describing community perceptions and responses to floods, and parameters describing changes in housing values (figure 4). The analysis was performed with calibrated ITS model parameter values across the 496 census tracts within the study region (figure S6 and S7). Across all tracts, the b2 parameter describing the initial drop in HPs immediately following a flood loss was significantly positively correlated with anxiousness (ɑ_A ), effectiveness of preparedness (ɑ_R ), forgetfulness (μ_A ), and decay of precautionary measures (μ_P ) (figure 4(a)). The b3 parameter describing the gradual decrease in HPs after flooding was significantly positively correlated with ɑ_A, μ_A , and μ_P (figure 4(a)). State-level variations in the relationships between model parameters describing flood risk perceptions and HP dynamics are evident. The regional patterns are driven largely by NY (figure 4(c)). In NJ and CT, there are similar significant positive correlations between ɑ_A and b2 (figures 4(b) and (d)). In CT, μ_P and b3 are significantly positively correlated (figure 4(d)).

Zoom In Zoom Out Reset image size

Understanding HP dynamics is critical for regional flood mitigation planning. From 1970 to 2012, the study region experienced a steady increase in HPs, which were potentially temporarily slowed by several minor flood events occurring in the 1980s and 1990s (figures 3(a) and (c)). In the absence of substantial flooding losses between 1970 and 2012, regional awareness of flooding risk may have gradually decreased (Meyer et al 2014), resulting in an inaccurate discounting of the negative impacts of flood risk on property values (Lamond and Proverbs 2006).

After Hurricane Sandy in 2012, HPs continued to rise, but more gradually than in the preceding 30 years (figure S2) (U.S. Census Bureau 2022b), which has been proposed as evidence of the limited effect of flooding on coastal property values (Murfin and Spiegel 2020). Our findings generally support the idea that the region exhibited a lower immediate impact of flooding on housing values; however, we observed a clear drop in relative HPs (figures 3 and S2). This result is similar to a study in Brisbane which found increasing home values after a flood, but also a sustained differential between the prices of flood-impacted and non-impacted properties (Eves and Wilkinson 2014). The observed trend of home values after Hurricane Sandy is counter to the frequent observation of immediate decreases in home values followed by relatively quick recoveries in other case studies (Atreya and Ferreira 2015, Beltrán et al 2019, Chandra‐Putra and Andrews 2020). The slow decline in HPs after Hurricane Sandy relative to the national mean HP may be related to the capacity for NFIP to limit the exposure of residents to economic losses related to flooding. Surveys conducted 5 years after Hurricane Sandy indicated that coastal residents in the region slightly underestimated flooding risks and that coastal prices exceeded expected prices by 6%–13% (Bakkensen and Barrage 2022), suggesting that HPs have not been corrected to the true flood risk profile as the memory of Hurricane Sandy fades.

The coastal communities of the study region remain highly vulnerable to flooding under climate change (Kirshen et al 2008, Sweet et al 2022). Future regional precipitation will likely be characterized by more intense rainfall interspersed with extended droughts (Hayhoe et al 2008, Xue and Ullrich 2022). Predictions of coastal storm surge are less certain with some forecasts suggesting increased (Michaelis et al 2017) or stable (Lin et al 2019) wave and surge inundation regimes. In aggregate, the region is likely to experience more frequent flooding and economic loss. The SE model developed in this research suggests that the immediate negative consequences of an isolated intense flooding event may continue to be absorbed by NFIP; however, predicted changes to the flooding regime may result in a sustained decreases in property values.

Flood risk is the aggregation of compound flooding events (Ghanbari et al 2021), variations in community exposure and resilience (Knighton et al 2021, Tate et al 2021, Wing et al 2022), perceptions of flooding (Merz et al 2015), and governance (Tullos 2018, Snel et al 2020). Risk perception itself is intricate, with the characteristics that shape it being both unclear and specific to research problem and applied methods (Lechowska 2018). These complexities necessitate that we study SE systems with tools that consider how these processes interact.

Significant positive correlations between the SE model parameter describing changes community risk awareness in response to flooding losses (ɑ_A ) and the initial drop in HPs post-flood (b2) suggested that census tracts in which individuals were most flood-aware (and increased NFIP policy uptake) experienced the largest immediate changes in HP. Similarly, positive correlations between b2, b3, forgetfulness (μ_A ) and the decay of precautionary measures (μ_P ) possibly suggested that awareness of flooding reflected by NFIP participation is correlated with the memory of past floods that is encoded in housing market values. Both μ_A and μ_P are related to the lapse rate of NFIP policies in the years after floods. Where these parameters indicated a greater lapse in NFIP policies, HPs were more likely to decrease than those with less variability in the number of active policies. Together, these correlations possibly suggest that increases in the participation in NFIP across the study region was driving short-term perceptions of increased risk which then steadily drove down housing values over the next decade (Colby and Zipp 2021). Positive correlations between effectiveness of preparation (ɑ_R ) and HP changes (b2) further reinforced that housing values in the region were driven more by perceptions and NFIP participation than actual changes to flood hazards. Variations in correlations across the states may be due to differences in social vulnerability, which in turn drives variations in NFIP participation (Dixon et al 2017). Alternately, the variations may be related to the number of tracts (n) considered from each state, where more tracts in NY could support more robust statistical inferences (figure 4).

While NFIP appears to have shielded the housing market from a substantial immediate economic loss (figure 3) it possibly had the unintended consequence of conditioning housing values on perceived risks (figure 4). There is empirical evidence that regional perceptions of risk across the study region are decoupled from actual risk (Bakkensen and Barrage 2022). Though they are often treated as objective information about the world, property values are unavoidably and fundamentally subjective determinations. They can reflect all manner of human biases–from cognitive heuristics that shape natural hazard risk perception (Kahneman et al 1982, Slovic 1987, Kunreuther et al 2001), to prejudices that associate the presence of racial/ethnic minorities with 'riskiness,' as the history of redlining has shown (Rothstein 2017). As Freddie Mac noted in its 2016 report, it is therefore hard to predict, in the context of climate change, the time path of house prices (Freddie Mac 2016). Perhaps house prices will decline gradually as the expected life of the property becomes shorter as we observed post-Sandy (figure 3). Or perhaps they will reach a tipping point and plummet when a cautious mortgage lender or insurer refuses to underwrite a property, or when homeowners begin to sell defensively.

The development of an SE model to predict time-varying HPs as a function of both external floods and community perceptions could substantially contribute to our understanding of socio-economic disparities in flood risk related to home value (Siders et al 2019) and the capacity for communities to recover (Ranasinghe 2020). The ITS model which best reproduced the combined NFIP and housing market system throughout the study period (figure 3(c) and S4) has the flexibility to simulate a sudden drop in HPs, but also a lagged decrease in value (figures 2 and 3). In contrast, the Sigmoid (Pryce et al 2011) and K-shaped (Beltrán et al 2019) models each imposed an initial sudden drop in value followed by a monotonically increasing recovery (figures 2, 3(b) and 4(d)). The housing value dynamics described by the Sigmoid and K-shaped models may prove to be more appropriate in other settings where housing markets have incurred sudden losses and then steadily recovered (Pryce et al 2011, Atreya et al 2013, Beltrán et al 2019, Chandra‐Putra and Andrews 2020). The sigmoid and K-shaped recovery models may be most appropriate for systems not modulated by insurance, or for modeling systemic losses incurred across both the housing and insurance markets (Thomson et al 2023).

Studies have demonstrated that complex feedbacks can emerge between flood insurance uptake and home values (Petrolia et al 2013, Gallagher 2014, de Koning et al 2019, Indaco et al 2019). We therefore anticipated that simulating time-varying house prices would improve the representation of NFIP losses and policy uptake in subsequent years over that provided by the base model. Simulating flood impacts on HPs significantly improved the simulations of both HPs (KGE_HP) and community awareness (KGE_A) (figure S5), supporting this concept. This result was also in agreement with observations of flood-induced insurance uptake (Gallagher 2014). We observed trade-offs between KGE_HP and loss (KGE_L) (figure S8), indicating the models evaluated did not support simultaneously reproducing all aspects of the housing and NFIP system (figure S4). This may be caused by imperfect forcing data (i.e., overly generalized surge elevations), imperfect census and NFIP data (i.e., missing or incomplete records), or possibly an overly-simplistic model formulation. Alternatively, this result may be due to the lack of large flooding events in sequence within the historical record and not necessarily a deficiency of the model or input data.