We employ a flexible analytical framework that can be applied at international, national, regional, and local scales, and can accommodate varying levels of data availability and resolution. Our approach involves three main steps: (1) assessing the social vulnerability and demographic characteristics of populations in the coastal zone; (2) estimating SLR impacts and responses in the study area; and (3) comparing these results spatially. We identified disadvantaged communities along the contiguous U.S. coastline (step 1) by calculating Social Vulnerability Index (SoVI) scores according to the methods outlined in Cutter et al. (2003) for each coastal census tract. As described in Sections 2.1 and 2.2 below, we used the SoVI index to analyze a large set of socio-economic and demographic variables; however, demographic indicators such as wealth and race can be used in the absence of detailed data. We then applied an existing analytical model (step 2) previously developed by EPA and collaborators called the SLR national coastal property model (NCPM), for which Neumann et al. (2010a, b) have documented the full methodological details. We assessed low, mid, and high SLR scenarios to identify areas at risk and the economically efficient adaptation response (abandon, protect, or nourish) for each of these areas. For developing nation coastlines, many of which lack detailed elevation, property value, and protection cost estimates, global coastal risk models, such as DIVA (Vafeidis et al. 2008), could be employed to assess the potential impacts of SLR. Lastly, we spatially compared the social vulnerability and SLR risk information (step 3) in a Geographic Information System (GIS) to determine if socially vulnerable communities are likely to experience disproportionate impacts of SLR and/or bear a disproportionate level of the costs in order to effectively respond to the threat. While using a GIS offers substantial advantages, particularly over larger geographic areas, simpler methods can be employed to compare SLR risk and social vulnerability information. For example, over smaller areas, spreadsheet based evaluations are possible, using data compiled for gridded areas or for political or census divisions.
Existing literature on estimating social vulnerability in the U.S.
A number of indices or models exist that seek to identify disadvantaged communities. We review several here that are most relevant to our objective, but this paper does not attempt to review, as others have, the available literature on general social vulnerability (Rygel et al. 2006; Cutter 1996; Dow 1992). In most cases, EJ indices incorporate environmental risk factors (e.g., health risks), as well as social vulnerability factors (e.g., income, age).
Moss et al. (2001) developed the Vulnerability-Resilience Indicator Model which aggregates 17 social and environmental variables into sectors, then into sensitivity and adaptive capacity values, and finally into a resilience index. The Environmental Justice Screening Method was developed by Poastor et al. (2009) for the California Air Resources Board. This method is designed to identify and map cumulative impacts and vulnerability at the neighborhood-level using a comprehensive set of 24 indicator metrics covering hazard proximity and land use, health risk and exposure, and social and health vulnerability. A recent paper by Wilson et al. (2010) calculates vulnerability scores at the county level using data on race/ethnicity and socio-economic status, pollution sources and levels, and health. Similarly, the Environmental Justice Strategic Enforcement Assessment Tool (EJSEAT) was designed to enable the U.S. Environmental Protection Agency (EPA) to identify communities or areas that are susceptible to experiencing disproportionate environmental and public health burdens and is based on 18 individual variables collected at the census tract level (USEPA 2010).
Cutter et al. (2003) developed the SoVI as a way to quantify social vulnerability using county-level socio-economic and demographic data. The index comprises 42 variables which are reduced to 11 independent factors that are then used to compute a summary score. Oxfam (2009) applied SoVI in an analysis to identify counties in the Southeastern U.S. that are hotspots of high social vulnerability, then overlaid these data in a mapping tool with projections of climate change, including drought, floods, hurricane-force winds, and SLR. However, this study did not include consideration of the costs of adaptation to climate change risks, as we do here for SLR risks. Schmidtlein et al. (2008) published a paper testing the sensitivity of SoVI to different aggregation levels of the census data, including tract-level data. Given that coastal counties in the U.S. can contain highly variable levels of vulnerability to SLR inundation, an application at a finer level of census resolution (i.e., tract level instead of county) is more appropriate for the objectives of our analysis.Footnote 2
Description of approach used for estimating social vulnerability
After reviewing the available literature on EJ indices, we determined that SoVI is best suited for this analysis. Our reasoning is that the index has been applied extensively by others (Borden et al. 2007; Schmidtlein et al. 2008; Burton and Cutter 2008; Wood et al. 2010), and because it is a pure social vulnerability index that does not include any environmental risk factors, including the coastal risks which we model explicitly using the NCPM. For example, an index that includes climate change risks in the vulnerability score (Moss et al. 2001) might present issues of effectively double counting climate risk when this information is combined with detailed data on SLR vulnerability. In our work, we also find that a rigorous separation of climate and socio-economic vulnerability allows us to make a clearer distinction between those factors which can be affected by climate policy (the degree of greenhouse gas mitigation, which in turn influences the rate of SLR, and the adaptation response, which involves decisions of policy-makers), and those policies that make up a “baseline” social vulnerability (mainly, indicators of socio-economic status). Without that clear separation of climate and social vulnerability, the issue of causality can be muddied – in other words, is the region vulnerable because of climate risks, or because of pre-existing socio-economic status? We wish to identify coastal areas inhabited by socially vulnerable populations because we believe they may have less capacity and resources over time to make necessary investments to protect development from coastal hazards.
As described in the next section, our approach seeks to reduce this uncertainty and improve modeling accuracy by separating the calculation of socio-economic risk and vulnerability to climate change, in this case SLR. As described in further detail below, we do this in three steps: calculating SoVI scores for each coastal census tract, estimating SLR risks and likely adaptation responses across a range of scenarios, and spatially comparing these two sets of information.
We chose to construct SoVI at the census tract level closely following the work of Schmidtlein et al. (2008). SoVI construction is a four-tiered process: (1) collect and clean input data, (2) standardize input data; (3) conduct principal components analysis (PCA); and (4) place the components in an additive model.
In the first step, we collected census tract level data for 26 demographic variables from the U.S. Census Bureau’s American Fact Finder website. For purposes of applying SoVI and categorizing data in this analysis, we assign coastal counties in the contiguous U.S. to one of four regions: North Atlantic (Maine through Virginia), South Atlantic (North Carolina through Monroe County, Florida), Gulf (Collier County, Florida through Texas), and Pacific (California through Washington). Census tracts with populations reported as zero are omitted from the analysis. Where census tracts are missing data for a particular variable, we replace the missing value with the mean value for the region.Footnote 3 Next, all input variables were standardized to z-scores with zero means and unit variances.Footnote 4 The input variables are standardized to avoid potential problems that may arise from using variables of different magnitudes in PCA.
The z-scores for each of our 26 variables are then used in PCA to reduce the variables to a smaller set of components. The resulting components are linear combinations of correlated variables that represent a broader measure of how certain components contribute to vulnerability.Footnote 5 Component values are calculated for each census tract by multiplying a variable’s value by the estimated weight for the component.Footnote 6 The PCA for the North Atlantic region produces a total of seven components, which explain 71% of the variance among the census tracts in that region. The PCA for the Gulf region also produces a total of seven components explaining 73% of the variance. The PCA for the South Atlantic and Pacific regions each produce six components explaining 71 and 68% of the variance, respectively.Footnote 7 Table 1 presents the components and their associated variance explained and dominant variable for each regional PCA.
Table 1 Components for each regional analysis in the contiguous United States. The table displays the Social Vulnerability Index components, including the dominant variables and their associated variance explained for each regional principle components analysis
Once the components were derived, we made adjustments to their directionality based on their known influences on vulnerability. A positive directionality was assigned to components believed to increase vulnerability (e.g., poverty), while a negative directionality was assigned to components believed to decrease vulnerability (e.g., wealth). The components were then placed in an additive model where each component is assigned an equal weight.Footnote 8 This additive model generates the overall SoVI score for each census tract. Following Schmidtlein et al. (2008), we standardized the resulting SoVI values to z-scores. This standardization is an important step that allows for comparative interpretations of the final SoVI scores for each area. For this analysis, we preformed the standardization at the U.S. regional level: North Atlantic, South Atlantic, Gulf, and Pacific.
An alternative to using an index to identify disadvantaged communities would be to simply analyze various demographic indicators using census data, such as wealth and race. We conducted a preliminary analysis using census block group level data, however we determined that this approach is less useful for spatially-comprehensive climate change impact assessment models, like the NCPM, because SoVI provides an analytically rigorous tool that combines multiple measures of social vulnerability in a single index value.
Estimating vulnerability to SLR
To estimate human response to the threat of SLR and the economic impacts on coastal property, we apply the NCPM. This framework employs GIS to structure and overlay available data (including coastal elevations and parcel-level property value data) using a 150-meter grid cell network covering the contiguous U.S. coastline. The model combines elevation dataFootnote 9 with SLR and sub-regional land subsidence rates to identify inundation risk information along the coast. The tool then models a response to the SLR threat over time, and reports estimates of response mode, property at-risk, property damages, and costs of adaptation in both graphical (map and chart) and tabular form. Protection decisions are determined on a grid cell basis according to calculations which compare the value of the land to the cost of various protection measures to determine an economically efficient response for each cell. Property value is assigned to grid-cells based on parcel level assessed value data, which is available in electronic form in many counties and regions of the U.S. Where the cost of protective measures is less than the benefit of avoided property value loss, the impact of SLR is estimated as the capital cost of seawall construction plus ongoing maintenance costs – in areas with beach frontage, maintenance of the existing beach profile is achieved through periodic beach nourishment, the cost of which is evaluated through estimates of sand volume required and a fixed cost per unit of sand. All cost estimates are consistent with those estimated from an analysis of U.S. Army Corps of Engineers beach protection projects. Where the cost of protection exceeds the benefit, retreat/abandonment is the estimated response, and the impact of SLR is lost structure and land value. In general, the model estimates that high-value land will be protected with hard structures or nourished with sand, while low-value land will be abandoned. The model estimates a trajectory of response to SLR through the 21st century, for each potentially threatened grid cell, based on scenarios of SLR over the same period. Thus while the algorithm for estimating response is relatively straightforward, the spatial dataset is very large, and the result is a highly-spatially resolved and differentiated estimate of areas threatened by SLR and the likely mode and cost of the adaptation response, which can be regenerated for a wide range of scenarios and input data to support uncertainty and sensitivity analyses. See Neumann et al. (2010a, b) for more detailed discussions of the model.
We acknowledge that many decisions of this type in the coastal zone are not made with strict cost-benefit decision rules, particularly at the local level. Other factors may include local zoning bylaws, future land use plans, the presence of development-supporting infrastructure, or proximity to sites with high cultural value. However, the analytical framework of the NCPM provides a good indication of areas which, for example, might be supported in adaptive measures by the U.S. Army Corps of Engineers, which requires benefit-cost analyses to support, and in many cases justify, coastal infrastructure and beach nourishment projects.Footnote 10
The NCPM is capable of simulating a wide range of SLR trajectories through 2100. For this application, we applied three scenarios, with rates of SLR derived from MAGICCFootnote 11 modeling (Wigley 2008) and Rahmstorf (2007):
-
1.
Low scenario, which is based on MAGICC processing of an IPCC B1Footnote 12 scenario implying 28.5 cm SLR by 2100 compared to 1990 levels.Footnote 13
-
2.
Mid scenario, also based on MAGICC processing, but of the IPCC A1B scenario, implying 66.9 cm SLR by 2100 compared to 1990 levels.Footnote 14
-
3.
High scenario, based on Rahmstorf (2007) maximum, and implying 126.3 cm SLR by 2100 compared to 1990 levels.
We apply our analytical framework to estimate adaptation costs for areas in the U.S. affected by Atlantic, Gulf, and Pacific coastal risks, but excluding Alaska, Hawaii, and island territories. The model forecasts which grid cells are likely to be abandoned, protected, and nourished in response to the threat of inundation from SLR. Where a cell straddles two or more block groups, that cell is assigned to the cell’s predominant block group (and by extension, that cell’s predominant tract). This information is then used in conjunction with SoVI to analyze the overlay of social and environmental vulnerability in coastal areas.
We note that the results presented in this analysis are calculated using estimates of the timing of inundation that are based on gradual changes in sea level. A key limitation at this stage of the research is that we do not assess how storm surge would affect the timing of inundation, and therefore, the decision of whether protection is warranted. However, the framework of the coastal property model is suitable to such an analysis, and efforts are underway to build capacity in the model to estimate damages from storms, and joint impacts of storms and SLR. Integration of risks from storm surge flooding will address many of the important impacts described in Kirshen et al. (2008), which indicate that the projected decrease in return frequency of coastal storms will have large effects on flood damages in the Northeastern U.S. Another important limitation is that we rely on current data for the socio-economic status and location of disadvantaged populations. Over the 21st century, as climate change unfolds, it is reasonable to expect that populations will migrate, and incomes to grow or, in some cases, fall. Unfortunately, we have not identified a reliable approach to estimating migrations and income changes over the long time scales in which climate change and SLR will manifest.