LETTERS
PUBLISHED ONLINE: 14 MARCH 2016   DOI: 10.1038/NCLIMATE2961

Millions projected to be at risk from sea-level rise
in the continental United States
Mathew E. Hauer1*, Jason M. Evans2 and Deepak R. Mishra3
Sea-level rise (SLR) is one of the most apparent climate change
stressors facing human society1 . Although it is known that
many people at present inhabit areas vulnerable to SLR2,3 , few
studies have accounted for ongoing population growth when
assessing the potential magnitude of future impacts4 . Here
we address this issue by coupling a small-area population
projection with a SLR vulnerability assessment across all
United States coastal counties. We find that a 2100 SLR
of 0.9 m places a land area projected to house 4.2 million
people at risk of inundation, whereas 1.8 m affects 13.1 million
people—approximately three times larger than indicated by
current populations. These results suggest that the absence of
protective measures could lead to US population movements of
a magnitude similar to the twentieth century Great Migration
of southern African-Americans5 . Furthermore, our population
projection approach can be readily adapted to assess other
hazards or to model future per capita economic impacts.
Sea-level rise is widely recognized as one of the most likely
and socially disruptive consequences of future climate change2 .
Scenarios of future SLR at the year 2100 range from a low of 0.3 m to
a high scenario of 2.0 m associated with collapse of polar ice sheets3 .
Understanding the specific locations at risk of SLR impacts is a high
priority in climate change research6 and adaptation planning7,8 .
Although there is growing worry and debate that climate change
could cause widespread human migration over the next century2,9,10 ,
relatively few studies have attempted to merge climate change
scenarios with population growth trends and projections in highrisk areas (however, see ref. 11). Notably, several previous studies
have estimated the populations at risk of future SLR inundation
through the use of current population data12 . Given the rapid
growth of population in coastal areas13 , such temporal mismatch
of data sets (that is, present population and future SLR) seems
likely to underestimate the impacts SLR will have on future
populations. Other research has tied small-area flood inundation
risk to populations at a county scale14 . Such spatial mismatch is
likely to overestimate the future populations at risk of SLR, as
populations located on higher ground within a coastal county may
be erroneously assumed to flood.
The mutability of many sub-county geographic units (for
example, Census Tracts and Census Block Groups) at each decennial
Census cycle is a classic example of the modifiable areal unit
problem15 , and generally limits the development of long-range
projections to areas in which geographic boundaries remain stable16 .
Using a novel approach, we overcome the methodological issues
related to spatial and temporal mismatch and the mutability of
sub-county units17 by synthesizing spatially explicit environmental

data (that is, elevation and associated flood risk) with small-area
population projections developed with a modified version of the
Hammer method17,18 in a dynamic flood hazard model. By spatially
and temporally aligning small-area population projections from
coastal states in the continental United States (US) to 2100, we are
able to assess who could be at risk from future SLR.
This approach addresses two fundamental questions concerning
the vulnerability of future coastal populations in the United States:
How many people are potentially at risk of impact from SLR? and
What areas in the US are likely to experience the greatest population
exposure to SLR? Accordingly, our results can be used to inform
local adaptation infrastructure and growth management strategies,
alerting officials to the areas where interventions and policies are
most needed.
We assess the populations at risk of SLR by using the National
Oceanic and Atmospheric Administration’s (NOAA) 0 m through
1.8 m (6 feet) SLR data sets for twenty-two coastal states and the
District of Columbia19 . These data sets simulate expected changes
in the mean higher high water (MHHW) mark on areas that
are hydrologically connected to coastal areas, without taking into
account additional land loss caused by other natural factors such as
erosion. Notably, the state of Louisiana was not included in the data
set at the time of analysis owing to local hydrologic complexities
associated with coastal levees and accelerated land subsidence;
however, we have recreated NOAA’s hydrologic connectedness
approach for Louisiana using USGS’s National Elevation Dataset
(NED) (Methods).
We used a linear/exponential extrapolation approach for
projecting Census Block Groups (CBGs) from 2010 to 2100. We
included only CBGs (n = 72,664) located in counties (n = 319)
expected to experience impact under the 1.8 m scenario. A
detailed technical description is available in Methods. Detailed
projections of exposure for all 319 coastal counties are also found
in Supplementary Fig. 1 and Supplementary Tables 1 and 2. The
population at risk of SLR is dynamically assessed as the proportion
of the CBG underwater when SLR is expected to exceed 0.3 m
intervals under the 0.9 m and 1.8 m scenarios. With a recreation
of NOAA’s hydrologic connectedness approach for Louisiana
at 0 m, 0.9 m, and 1.8 m, we assessed Louisiana’s population at
0.9 m intervals rather than 0.3 m intervals. As populations become
exposed under each SLR scenario in each block group, projected
populations are dynamically adjusted to account for this exposure
to ensure no persons are double counted.
We find that in the continental US approximately 13.1 million
people are at risk under the 1.8 m scenario (Fig. 1). The projected
number for the US is nearly triple the current population estimates

1 Carl Vinson Institute of Government, University of Georgia, 201 N. Milledge Avenue, Athens, Georgia 30602, USA. 2 Department of Environmental Science
and Studies, Stetson University, 421 N. Woodland Boulevard, DeLand, Florida 32723, USA. 3 Center for Geospatial Research, Department of Geography,
University of Georgia, 210 Field Street, Athens, Georgia 30602, USA. *e-mail: hauer@uga.edu

NATURE CLIMATE CHANGE   ADVANCE ONLINE PUBLICATION   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

1

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961

Table 1   Projected populations at risk of sea-level rise by 2100.
State

Current populations
0.9 m SLR in 2100
14,648
94,217
17,249
545
19,782
385,436
25,061
412,648
38,232
30,300
6,849
12,379
59,884
3,299
117,553
48,933
4,374
2,537
5,188
52,443
52,600
45,521
11,178
1,461,854

AL
CA
CT
DC
DE
FL
GA
LA
MA
MD
ME
MS
NC
NH
NJ
NY
OR
PA
RI
SC
TX
VA
WA
Tot

Projected populations

1.8 m SLR in 2100
25,326
216,174
39,482
1,257
35,811
1,499,509
48,426
441,011
155,335
68,667
13,233
20,075
109,756
6,211
300,923
221,056
8,985
7,288
13,150
126,498
114,797
109,507
26,597
3,609,955

0.9 m SLR in 2100
38,238
472,248
53,566
2,005
44,597
1,221,837
93,036
758,584
103,552
92,584
15,230
50,385
163,260
8,670
308,662
198,257
12,754
9,939
14,875
163,492
173,025
181,130
43,436
4,310,983

±
7,801
98,343
7,189
410
7,708
236,103
18,683
263,827
13,329
14,730
1,848
10,254
27,210
1,131
47,436
32,543
1,903
1,858
1,646
38,527
45,306
38,072
7,229
923,086

1.8 m SLR in 2100
57,303
1,046,757
128,048
4,629
76,836
6,057,419
178,787
1,286,877
427,549
188,624
29,028
76,901
297,917
15,432
827,449
901,366
25,614
27,427
36,546
374,395
405,423
475,871
94,139
13,115,250

±
11,584
208,343
17,947
948
14,061
1,216,806
37,263
292,676
57,669
31,624
3,574
16,721
52,013
2,024
137,272
159,124
4,163
5,659
3,977
86,058
106,301
102,952
16,040
2,584,797

7

16
14

1.8 m SLR in 2100
0.9 m SLR in 2100

12
10
1.5 m

8
1.2 m

6
0.9 m

4

2020

2030

2040

0.6 m

2

0

0.3 m
2070

2080

2090

2100

Figure 1   Cumulative projected at-risk populations for the continental
United States, 2010–2100. Projections reflect assumed growth/decline
rates for 72,664 census block groups in 319 coastal counties. The shading
indicates the 90% confidence interval of the projection models.

for 2010 in these areas (Fig. 2 and Table 1), suggesting an
underestimation of risk when using current population estimates.
Florida accounts for nearly half of the total at-risk population.
Whereas other southeastern states have substantially fewer people
at risk, states such as Georgia, South Carolina, and Louisiana have
over 10% of future coastal populations at risk under the 1.8 m
scenario. The southeastern US alone represents nearly 70% of the
entire projected populations at risk, suggesting the impacts of SLR
will be highly regionalized in nature.
Our results also suggest a hyperlocalized impact from SLR (Fig. 3
and Supplementary Table 2). Although the median percentage
2

3

1

0.6 m

2050 2060
Year

4

NJ
VA
M
A
TX
SC
NC
M
D
GA
CT
W
A
M
S
DE
AL
RI
M
E
PA
O
R
NH
DC

0
2010

0.3 m

5

FL
LA
CA
NY

2

Projected population at risk
Current population at risk

6
Population (millions)

Cumulative potentially inundated populations (millions)

We considered only census block groups and counties expected to experience any inundation under 1.8 m of sea-level rise in 2100. ± values are the 90th percentile from the projection values.

Figure 2   Projected cumulative populations at risk of sea-level rise in
2100 under the 1.8 m scenario. We considered 22 states and the District of
Columbia. Black bars are the projected population at risk and the grey line is
the current population at risk based on Census 2010.

of the population subject to SLR impact across all 319 coastal
counties is just 3.5% under the 1.8 m scenario, several lowlying counties would be likely to experience extreme exposure.
Three counties in particular—Tyrrell, North Carolina (94% of the
projected population located in land area at risk of inundation),
Monroe, Florida (88%), and Hyde, North Carolina (82%)—could
see catastrophic impacts with 1.8 m SLR. Broward, Miami-Dade
and Pinellas, Florida; San Mateo, California; and Jefferson and
Orleans, Louisiana are projected to see more than 100,000 residents
potentially impacted with a 0.9 m SLR. An additional 25 counties
would have more than 100,000 impacted persons with a 1.8 m
SLR. Miami-Dade and Broward counties in Florida alone account

NATURE CLIMATE CHANGE   ADVANCE ONLINE PUBLICATION   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961

WA
ME

VT NH

OR

NY

MA

ID
CT
PA

RI

NJ

100 miles

DC MD DE

NV

WV

CA

VA
Population at risk
0.9 m in 2100

NC

0−8,017
8,018−23,795

AZ

23,796−57,070

SC

57,071−102,992
102,993−231,336

TX

LA

MS

AL

GA

FL

Figure 3   Cumulative projected populations at risk of SLR under the 0.9 m scenario by 2100 for US counties. Counties not included in the study are
coloured in grey.

for more than a quarter of the people impacted under the 1.8 m
scenario. Expanded results for all 319 counties can be found in
Supplementary Table 2.
Cities such as Tampa–St Petersburg, Florida; Charleston, South
Carolina; Poquoson, Virginia; and Cape May, New Jersey may
experience serious levels of population impact under the 1.8 m
SLR scenario. Other areas such as Hartford, Connecticut; Fairfax,
Virginia; and San Diego California, by contrast, may expect to see
very little impact from SLR. Owing to geographic variability, onesize-fits-all national approaches for tackling SLR, such as recent
changes to the US federal government’s National Flood Insurance
Program20 , could prove problematic or inadequate as adaptation
strategies alone.
Adaptation strategies for SLR rely on accurate information about
the geographies, timescales, economies and populations at risk.
Recent estimates of annual global costs for adapting coastal flood
protection infrastructure to a 2100 SLR of 2.0 m are approximately
US$421 billion (2014 values) per year21 . Although such cost
inventories22 are helpful, they do not take into account expansions
in population and infrastructure that are likely to take place before

inundation occurs. Our work indicates that existing estimates of
future adaptation cost may, in fact, be deceptively low if future
population growth is not taken into account.
Similarly, proposed managed retreat solutions could also prove
troublesome if population projections are left out of the equation.
So far, managed retreats have tended to involve small populations
and areas23,24 , but future action could be needed in areas with
areas with much larger and growing populations. Not only could
the costs of relocating a community be greatly underestimated if
that population is growing, but the challenge of finding suitable
areas for relocation could be problematic as well. With current
estimates as high as US$1 million per resident in some small
Alaskan villages25 , each decade both increases that population’s
exposure to SLR and increases their vulnerability to the economic
costs of inaction. Potential growth management strategies in
high-risk areas experiencing rapid population growth could also
prove more effective than relocation. Population projections are
not a panacea for these problems, but they move us towards
evaluating the potential SLR impacts on future, rather than
current, populations.

NATURE CLIMATE CHANGE   ADVANCE ONLINE PUBLICATION   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

3

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961

Research indicating how populations might adapt to SLR is
still in its infancy, thus limiting our ability to model how future
populations might organically adapt to rising seas and the loss
of both current and future coastal human habitat. For instance,
Venice, Italy has seen its population remain stable over the past
decade26 in spite of widely documented tidal flooding from both
land subsidence and SLR, suggesting a complicated relationship
between population dynamics and SLR. Furthermore, adaptation
and mitigation strategies are likely to be employed, shaping future
population scenarios through unknown future public policies. Our
projections of inundated populations could be biased upwards by
the limited interaction between SLR and population growth.
Uncertainty in our projections result from the sensitivity of
long-term population to both the selection of base period length
and projection horizon length27 . By using the longest possible
base period, we do find acceptable accuracy for these projections,
with approximately half of the coastal states exceeding accuracy
expectations. There were notable exceptions, however, as four
states fell far below accuracy expectations—Massachusetts, Maine,
Mississippi, and Rhode Island—with another eight states falling
just below expectations (Supplementary Table 1). In spite of these
issues, our out-of-sample validation found the projections to be
reasonably calibrated, as four of the top six most affected states
exceed expectations.
Past trends do not guarantee future trends. Local growth
ordinances and population saturation points could improve future
population projections. Furthermore, vertical land movement
will exacerbate the impacts of SLR, specifically in southeast
Louisiana and the Chesapeake Bay areas12,28 . Although we do not
model vertical land movement, our results could be considered
conservative in the aforementioned areas expected to see the
greatest land subsidence, as it is the combination of SLR and vertical
land movement that can prove the most destructive.
The approach demonstrated in this paper allows for spatially
and temporally aligning population data with any type of hazard
modelling requiring small-area spatio-temporal population
projections that can be readily used by decision makers and
researchers. For example, other by-products of SLR, such as loss of
coastal wetlands, saltwater intrusion, and higher storm surges from
tropical cyclones29–31 could also be modelled, as well as economic
impacts from these hazards. For instance, using the example of
the cost for relocating some Alaskan coastal villages25 of US$1
million per resident, the cost of relocation could exceed US$14.0
trillion (2014 values). More precise cost estimates could incorporate
our approach. There is high potential for coupling population
projections in dynamic systems simulations that incorporate
such stressors into multivariate scenario modelling. We note,
however, that our small-area projection method requires detailed
demographic information on the age of housing stock, thus limiting
the applicability of the approach to nations and jurisdictions where
such data are regularly collected and available.

3. Vermeer, M. & Rahmstorf, S. Global sea level linked to global temperature.
Proc. Natl Acad. Sci. USA 106, 21527–21532 (2009).
4. Parris, A. et al. Global Sea Level Rise Scenarios for the United States National
Climate Assessment (US Department of Commerce, National Oceanic and
Atmospheric Administration, Oceanic and Atmospheric Research, Climate
Program Office, 2012).
5. Gregory, J. N. The Southern Diaspora: How the Great Migrations of Black and
White Southerners Transformed America (Univ. North Carolina, 2005).
6. Wu, S.-Y., Yarnal, B. & Fisher, A. Vulnerability of coastal communities to
sealevel rise: a case study of Cape May county, New Jersey, USA. Clim. Res. 22,
255–270 (2002).
7. Lutsey, N. & Sperling, D. America’s bottom-up climate change mitigation
policy. Energy Policy 36, 673–685 (2008).
8. Titus, J. et al. State and Local government plant for development of most land
vulnerable to rising sea level along the US Atlantic Coast. Environ. Res. Lett.
4, 044008 (2009).
9. Black, R., Bennett, S. R. G., Thomas, S. M. & Beddington, J. R. Migration as
adaptation. Nature 478, 447–449 (2011).
10. Gray, C. & Bilsborrow, R. Environmental influences on human migration in
rural Ecuador. Demography 50, 1217–1241 (2013).
11. Hugo, G. Future demographic change and its interactions with migration and
climate change. Glob. Environ. Change 215, 521–533 (2011).
12. Haer, T., Kalnay, E., Kearney, M. & Moll, H. Relative sea-level rise and the
conterminous United States: consequences of potential land inundation in
terms of population at risk and GDP loss. Glob. Environ. Change 23,
1627–1636 (2013).
13. Crossett, K., Ache, B., Pacheco, P. & Haber, K. National Coastal Population
Report, Population Trends from 1970 to 2020 (National Oceanic and
Atmospheric Administration, Department of Commerce/US Census Bureau,
2014); http://oceanservice.noaa.gov/facts/coastal-population-report.pdf
14. Curtis, K. & Schneider, A. Understanding the demographic implications of
climate change: estimates of localized population predictions under future
scenarios of sea-level rise. Popul. Environ. 33, 28–54 (2011).
15. Cromley, R. G., Ebenstein, A. Y. & Hanink, D. M. Estimating components of
population change from census data for incongruent spatial/temporal units and
attributes. J. Spat. Sci. 54, 89–99 (2009).
16. Swanson, D. A., Schlottman, A. & Schmidt, B. Forecasting the population of
census tracts by age and sex: an example of the Hamilton–Perry method in
action. Popul. Res. Policy Rev. 29, 47–63 (2010).
17. Hauer, M., Evans, J. & Alexander, C. Sea-level rise and sub-county population
projections in coastal Georgia. Popul. Environ. 37, 44–62 (2015).
18. Hammer, R. B., Stewart, S. I., Winkler, R. L., Radeloff, V. C. & Voss, P. R.
Characterizing dynamic spatial and temporal residential density patterns from
1940–1990 across the North Central United States. Landscape Urban Plan. 69,
183–199 (2004).
19. Sea Level Rise and Coastal Flooding Impacts (NOAA, 2014);
https://coast.noaa.gov/slrdata
20. Fox, S. This is adaptation: the elimination of subsidies under the National
Flood Insurance Program. Columbia J. Environ. Law 39, 205–249 (2014).
21. Nicholls, R. J. et al. Sea-level rise and its possible impacts given a ‘beyond
4 ◦ C world’ in the twenty-first century. Phil. Trans. R. Soc. A 369,
161–181 (2011).
22. Gornitz, V., Couch, S. & Hartig, E. K. Impacts of sea level rise in the New York
City metropolitan area. Glob. Planet. Change 32, 61–88 (2001).
23. Arenstam Gibbons, S. J. & Nicholls, R. J. Island abandonment and sea-level rise:
an historical analog from the Chesapeake Bay, USA. Glob. Environ. Change 16,
40–47 (2006).
24. Abel, N. et al. Sea level rise, coastal development and planned retreat: analytical
framework, governance principles and an Australian case study. Environ. Sci.
Policy 14, 279–288 (2011).
25. Huntington, H. P., Goodstein, E. & Euskirchen, E. Towards a tipping point in
responding to change: rising costs, fewer options for Arctic and global societies.
Ambio 41, 66–74 (2012).
26. UN Statistics Division Demographic Statistics (UNdata, 2015);
http://data.un.org/Data.aspx?d=POP&f=tableCode:240
27. Tayman, J., Smith, S. & Lin, J. Precision, bias, and uncertainty for state
population forecasts: an exploratory analysis of time series models. Popul. Res.
Policy Rev. 26, 347–369 (2007).
28. Nicholls, R. J. & Leatherman, S. P. Adapting to sea-level rise: relative sea-level
trends to 2100 for the United States. Coast. Manage. 24, 301–324 (1996).
29. Nicholls, R. J. & Cazenave, A. Sea-level rise and its impact on coastal zones.
Science 328, 1517–1520 (2010).
30. Nicholls, R. J. Planning for the impacts of sea level rise. Oceanography 24,
144–157 (2011).
31. Burton, D. A. Comments on ‘‘Assessing future risk: quantifying the effects of
sea level rise on storm surge risk for the southern shores of Long Island,
New York’’. Nat. Hazards 63, 1219–1221 (2012).

Methods
Methods and any associated references are available in the online
version of the paper.
Received 4 April 2015; accepted 15 February 2016;
published online 14 March 2016

References
1. Sweet, W. P. J., Marra, J., Zervas, C. & Gill, S. Sea Level Rise and Nuisance Flood
Frequency Changes Around the United States NOAA Technical Report NOS
CO-OPS 073 (NOAA, 2014); http://tidesandcurrents.noaa.gov/publications/
NOAA_Technical_Report_NOS_COOPS_073.pdf
2. IPCC Climate Change 2014: Impacts, Adaptation, and Vulnerability. Part A:
Global and Sectoral Aspects (eds Field, C. B. et al.) (Cambridge Univ.
Press, 2014).
4

NATURE CLIMATE CHANGE   ADVANCE ONLINE PUBLICATION   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961
Acknowledgements
Publication supported in part by an Institutional Grant (NA10OAR4170098) to the
Georgia Sea Grant College Program from the National Sea Grant Office, National
Oceanic and Atmospheric Administration, US Department of Commerce. Data reported
in the paper are available in the Supplementary Methods. The authors are grateful for the
assistance and constructive comments from K. Devivo, C. Hopkinson, J. M. Shepherd,
S. Holloway, T. Mote, J. Baker and W. Anderson.

Author contributions
M.E.H. produced the small-area population projections and the projections of
inundation, contributed to the methodological design, wrote the paper, and is the
corresponding author to whom requests for materials should be addressed. J.M.E.

contributed significantly to the methodological design, conceptual framing, and editing
of the paper. D.R.M. produced the inundation modelling for Louisiana and contributed
to the editing of the paper.

Additional information
Supplementary information is available in the online version of the paper. Reprints and
permissions information is available online at www.nature.com/reprints.
Correspondence and requests for materials should be addressed to M.E.H.

Competing financial interests
The authors declare no competing financial interests.

NATURE CLIMATE CHANGE   ADVANCE ONLINE PUBLICATION   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

5

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961

Methods
The methodology for overlaying projected small-area population with sea-level rise
(SLR) inundation layers is outlined in this section. First, we describe the data sets
and basic methodology behind SLR inundation layers. Second, the methodology to
historically estimate housing units is introduced. Third, the methodology to
convert housing units to population is reviewed. Fourth, the extrapolation
approach undertaken to produce population projections is reviewed. Next, the
determination of at-risk populations through intersection with SLR curves and
inundation models is described. Last, we evaluate the accuracy of our
population projections.
Data. Many assessments of the populations at risk from SLR have used an elevation
based ‘bath-tub’ approach for inundation modelling12,14,32 , whereby all areas under
a given threshold (usually 1 m, 2 m, 3 m, or 6 m) are flooded without explicit
consideration of hydrological connectedness. A known limitation of the simple
bath-tub approach is that areas protected from inundation by dykes or levees will
be shown as inundated. For example, much of New Orleans, Louisiana is located at
an elevation well below local mean sea level. Under a simple bath-tub approach, all
areas of New Orleans located below sea level, including those protected from
floodwaters by dykes and levees, would be shown as inundated under even an
initial condition (0 m) SLR model.
For this research, we used SLR inundation data sets developed by the National
Oceanographic and Atmospheric Administration (NOAA) as the basis for
simulating future SLR impacts on human populations in the coastal United States
(US)19,33 . The NOAA data sets are based on a 1/3 arcsecond (10m) resolution digital
elevation model (DEM), which is then used to simulate expected 0.3 m increment
(1 foot) changes in the mean higher high water (MHHW) mark, up to a maximum
scenario of 1.8 m (6 feet), on areas of the continental US that are hydrologically
connected to the coastal zone. The low SLR value of 0.3 m and the high SLR value
of 1.8 m in the NOAA data sets generally represent the range of low to high SLR
scenarios defined by the most current US National Climate Assessment3 . The
underlying 1/3 arcsecond DEM is used by the US federal government for
development of floodplain contours, rated as ±0.3 m at the 85% confidence
interval34 . Because such floodplain contour maps are used to set flood insurance
rates at the parcel-scale, there is confidence in applying a similar 0.3 m interval
assessment of SLR as generalized across much larger geographic areas (for example,
counties and Census boundaries). However, we also note that the NOAA SLR data
set does not take into account additional land loss caused by other natural factors
such as erosion, subsidence, or future construction, and are provided ‘as is’ without
warranty to their performance.
We used the 1/9 arcsecond (3 m) NED data to develop the SLR projection
model for Louisiana. A 3 m MHHW surface was created using NOAA’s vertical
datum conversion software, VDatum (http://vdatum.noaa.gov) and a triangulated
irregular network (TIN) was created and used for hydrologic connectivity mapping
for the 0 m depth grid (current condition). A linear superposition method was used
by adding 0.9 m (3 ft) and 1.8 m (6 ft) to the 0 m depth grid to map SLR scenarios.
A small-area housing unit projection method was used to produce sub-county
population projections for all US coastal counties expected to have direct impacts
from the 1.8 m SLR scenario (n = 319). The sub-county unit for these projections
was Census Block Groups (CBGs), with geographies defined by the 2010 US
Census. Data for conducting the population projections come from three main
sources. The first source of data comes from the American Community Survey
(ACS) 2008–2012 estimates. The ACS provides the ‘year structure built’ data, and
the 2010 boundaries for CBGs. The second piece of data is the actual historic count
of housing units (HU) and population for each county. This data is available as
digitized records from the Census Bureau’s website. For 1940 to 1990, data can be
found at http://www.census.gov/prod/cen1990/cph2/cph-2-1-1.pdf. Census 2000
data can be downloaded through American FactFinder. Finally, our Group
Quarters (GQ) population data come from the 2010 Census. It should be noted
that the ACS data, although similar to decennial data, is subject to sampling error,
but all released ACS data have confidence limits above 90% (ref. 35). Furthermore,
GQ tends to be the most volatile aspect of the Census Bureau’s Estimates Program
and ACS (ref. 36), but is an important aspect of the HU method.
Estimates of historic housing units. Demographic projections of small-areal units
(that is, sub-county units) tend to be less robust than projection methodologies at
larger scales16,37 . The changeability of many sub-county boundaries (for example,
Census Tracts and CBGs) at each decennial Census cycle provides a classic example
of the modifiable areal unit problem (MAUP), thus effectively limiting the
development of more long-range projections to areas in which geographic
boundaries remain stable16 . In the US, counties are the smallest geographies with
boundaries that tend to remain stable over time.
We use a modified version of the Hammer method17,18 based on a proportional
fitting algorithm to project sub-county populations38 . Hammer’s method is
essentially a combination of a growth-allocation and proportional fitting approach,
where the growth between time periods is allocated to each block group and

proportionally fitted to the marginals. Equation (1) demonstrates this proportional
fitting approach.

Ñ
Ĥijt =

é

Cjt
t−1
P

t
j

∗

t−1
X

H

Hijt

(1)

i=1939

i=1939

The number of housing units in county j as counted in the census taken in time
t is denoted as Cjt and the number of housing units in block group i in county j
based on the ‘year structure built’ question in the ACS is denoted as Hijt . Thus, any
estimate of housing units in any given block group in county j is given as a
proportionally adjusted estimate based on the ratio of the total number of housing
units as counted in the Census to a county’s estimated housing units from the ACS
for t − 1. For instance, an estimate of the number of housing units for block group i
in county j for the year 1980 would be equal to the number counted at the county
level according to the 1980 census, Cj1980 , divided by theP
number of housing units at
1979
the county level in the ACS for the period 1939–1979,
Hj1980 , multiplied by
i=1939
the number of housing units
observed in the ACS for the period 1939–1979 for
P1979
block group i in county j,
Hij1980 . This process is iterated for each decade
i=1939
until the most recent time period, that is, the 2010 census. These estimates of
housing units for each block group in each county provide the key input needed to
convert an estimate of housing units into an estimate of total population.
Housing units to population. Equation (2) demonstrates the approach employed
here to make use of the Housing Unit (HU) method to convert an estimate of
Housing Units to an estimate of population.
Pt = H ∗ PPHU + GQ

(2)

Where H is the number of housing units, PPHU is the persons per household, and
GQ is the group quarters population. Any error associated with the HU method is
attributable to the quality of the inputs39 , as the HU method is considered a
demographic identity. The Hammer method, outlined above, can provide a
long-range back cast of housing units for normalized boundaries in any given
census geography (whether its 1990, 2000, or 2010 geographies). Whereas
Census-designated boundaries may change, housing units typically do not move18 .
Based on the ‘year structure built’ question in Census data, the method produces
proportionally adjusted housing unit estimates at the sub-county CBG, which is the
smallest geography possible for such projections using US Census data.
Equation (3) demonstrates the approach employed here to use the HU method
Ûijt+1 can be
to project a population. While PPHU and GQ are held constant, H
projected though any set of extrapolation methods40–43 .

Ûijt+1 ∗ PPHUtij + GQtij
Pt+1 = H

(3)

Projection approach. We employed a linear/exponential (LIN/EXP),
regression-based extrapolation based on the past 70 years of population change for
1940–2010. Geographies that have experienced growth used a linear regression
whereas geographies that have experienced decline use an exponential regression.
A LIN/EXP model is used to ensure that long-range linear projections of decline do
not project negative populations, and that long-range exponential projections of
growth do not produce extreme values of runaway growth. Recent research
suggests that a LIN/EXP model outperforms both a linear and an exponential
model, respectively44 . Included within the regression formulae is an adjustment
factor allowing for the projected and observed populations at launch year to be
identical. This is computed by adding the residual of the estimate at time t back
into the regressed estimate of time t. This allows the projection to go through the
launch year population. The small data requirements make these extrapolation
methods ideal for small-area projections, and the use of a regression-based
extrapolation allows for estimates of projection intervals.
If the base housing stock is growing:

Ûijt+z = (α + βz) + H t − (α + βt)
H
 

 

(4)

If the base housing stock is declining:

Ûijt+z = eβ ∗ z α + H t − (eβ ∗ t α )
H
 

 

(5)

The use of a regression-based extrapolation allows for the creation of projection
intervals. We follow a long line of inquiry in determining the credibility of
population projections using projection intervals45–50 . These projection intervals
use the standard error of the estimate for the models and their sample sizes.
Intervals were generated using equations 4.1 and 4.2 from Hyndman &
Athanasopoulos’ Forecasting: Principles and Practice51 . We have chosen to produce
a set of three population projections for each block group, an upper, middle and

NATURE CLIMATE CHANGE   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961
lower bound based on the 90% projection interval. Thus we produce a set of
210,942 projections—one for every block group in the study area (n = 72,664) as
well as for the upper and lower bound.
Assessing at-risk populations. At-risk projected populations of sea-level rise
under prescribed SLR scenarios were calculated using equation (6).

t
ij

PR =

X
PRt−1
ij

+

Pijt −

X

!

!
∗A

t
ij

(6)

PRt−1
ij

where the population at risk of sea-level rise (PRt ) is equal to the population
projected at time t (P t ) minus the sum of the previously impacted populations
(PRt−1 ) multiplied by the land lost due to SLR (At ). We subtract out previously
impacted populations to ensure populations are not double counted. We consider
this approach a first-order, one-way interaction between population dynamics and
inundation modelling. Supplementary Fig. 1 demonstrates this first-order, one-way
interaction between population dynamics and SLR in four select counties.
Land lost due to SLR is calculated with a spatial overlay workflow in ArcGIS
10.1 as one minus the percentage of land lost under the preceding amount of SLR,
that is, 0.3 m divided by 0 m, 0.6 m divided by 0.3 m, and so on. The first step in the
analysis was to use a base, 0 m MHHW layer, which was derived from NOAA’s 0 m
scenario, and used as the initial condition to calculate a base of dry land area
contained within the geographies of 2010 CBGs. The resulting calculation is
therefore a total area of dry land, without any distinction between habitable and
uninhabitable dry land, available at present for human habitation within each CBG
geography. Each subsequent scenario is expressed as the ratio of each scenario to
the previous scenario.
Next, we used the method developed for the US National Climate Assessment4
to determine the years SLR could be expected to exceed 0.3 m intervals. The
following quadratic equation was used as the basis for calculating deterministic
curves for high (1.8 m) and medium (0.9 m) SLR scenarios at 2100:
E(t) = at + bt 2

(7)

where E(t) = eustatic SLR, in metres, at time t; a = global linear trend SLR
constant of 0.0033 m yr−1 ; t = years since 2010; b = SLR acceleration coefficient
(units of m yr−2 ), with bhigh = 1.86 × 10−4 ; bmedium = 7.44 E × 10−5 .
These curves were then used to find the years when SLR would exceed 0.3 m
increments under the high (1.8 m) and medium (0.9 m) curves. These correspond
to 2058, 2082 and 2100 for the medium curve (0.9 m) and 2045, 2061, 2073, 2083,
2092 and 2100 for the high curve (1.8 m). With a recreation of NOAA’s hydrologic
connectedness approach for Louisiana at 0 m, 0.9 m and 1.8 m, we assessed
Louisiana’s population at 0.9 m intervals rather than 0.3 m intervals. This
corresponds to the years 2100 for the 0.9 m curve and 2073 and 2100 for the
1.8 m curve.
We explicitly do not migrate those who are projected to be at risk from SLR.
Our current understanding of the human migratory response to environmental
events is not robust enough to model where these inundated persons will
potentially move, or if they will move at all. There are several hypotheses on human
migration and climate change, mostly drawing from environmental events in the
twentieth century14,52–55 . These hypotheses, however, result in empirical migration
effects that are highly dependent on the type of environmental pressure. Drought,
flooding, tropical cyclones, and tsunamis all exhibit differing migration
patterns56–58 , with very little research suggesting the effect of SLR on human
migration systems14 . Furthermore, very little research has been undertaken that
would be the bedrock of modelling who moves, where, and in what proportion55 .
Will impacted populations migrate landwards? Could future coastal cities resemble
Venice, Italy, complete with populations still adapting to rising sea levels? Or will
populations move to more land-locked cities for protection? These questions still
remain unanswered. For these reasons, our approach is strictly a model of the
confluence between two processes, SLR and population growth. Although this
confluence implies a high level of societal impact (for example, coastal flood
protection, architectural adaptation, migration, and so on) in the most general
sense, our approach here makes no prediction as to what the specific impacts will
be in any particular location.
Evaluation of projections. Projection intervals, produced through the use
of a regression-based projection, allow us to determine the degree of feasibility in a
projection. Previous analyses have used the 2/3 (or 66%) projection interval to assess
the degree of accuracy in a population projection27,46 representing empirical ‘low’
and ‘high’ scenarios from cohort-component projections59 . The use of a 2/3 interval
is ‘‘neither so wide as to be meaningless nor too narrow to be overly restrictive’’50 .
To assess the degree of feasibility, we assess all intervals on the 2008–2012 ACS
estimate of HU for each CBG in the study area. We produce projections based on

the equations in the preceding section with base period 1940–2000. If less than 2/3
of the ACS estimates of HU in 2010 falls within the 2/3 projection interval, then the
results would suggest less than ideal accuracy in terms of long-range projections.
Alternatively, if greater than 2/3 of the ACS estimates of HU falls within the 2/3
projection interval, then the results would suggest an ideal amount of accuracy in
terms of long-range projections. It should be noted in the consideration of these
inputs that the ACS data, although similar to decennial data, is subject to many
types of error. Although all released ACS data have confidence limits above 90%
(ref. 60), the ‘true’ estimate from the ‘year structure built’ question cannot be
known. Our evaluation should be considered in lieu of the limitations of
ACS accuracy.
Supplementary Table 1 shows the number of ACS housing unit projections that
fall within the 2/3 projection interval. Overall, 68.1% of the 2010 estimates fell
within the projection interval, suggesting an adequate degree of feasibility
associated with these projections in the aggregate. Seven states greatly exceed the
target 2/3 projection interval. Four states, however, fell far below the target 2/3
projection interval—Massachusetts, Maine, Mississippi, and Rhode Island—with
another eight states falling just below the target.
Projections inherently rely on historic trend data, and therefore performance
tends to suffer when growth deviates the greatest from historical patterns.
State-level aggregation might hide underlying geographic variability, and the
variation in the projected exposure to SLR is heavily influenced by areas with the
greatest deviation in past population growth. To assess these patterns, we
considered the block-group-specific coefficient of variation. Panel A in
Supplementary Fig. 2 demonstrates the coefficient of variation for each block
group’s population projection model. We find that overall variation in projected
populations is generally relatively low, with the greatest variation occurring in parts
of Louisiana, southern Texas, and inland North Carolina and Virginia. The Pacific
Coast also tends to have lower overall variation compared to the Gulf and Atlantic
coasts. By comparison, if we assess the overall contribution to uncertainty in
projected populations in panel B of Supplementary Fig. 2 (the standard error), we
find most uncertainty in the Gulf Coast region, specifically from Mississippi
through South Florida. Three of the four states with greatest observed downward
deviation in accuracy from the 66% interval show some of the lowest standard
errors, with Mississippi being the exception. The northeast states, including those
that fall under the 66% threshold, nevertheless show low coefficients of variation.
These results provide confidence that our overall small-area projections meet or
exceed accepted feasibility standards for more standard projection geographies, and
thus are well-suited for finer-grain assessments of future human hazard exposure.

References
32. Lam, N. S.-N., Arenas, H., Li, Z. & Liu, K.-B. An estimate of population
impacted by climate change along the U. S. Coast. J. Coast. Res. 56,
1522–1526 (2009).
33. Marcy, D. et al. in Proc. 2011 Solutions to Coastal Disasters Conference,
Anchorage, Alaska (eds Wallendorf, L. A., Jones, C., Ewing, L. & Battalio, B.)
474–490 (American Society of Civil Engineers, 2011).
34. FEMA Revised Procedure Memorandum No. 38—Implementation of
Floodplain Boundary Standard (Section 7 of MHIP V1.0) (2007);
http://www.fema.gov/media-library-data/1437593713238cadcd346d3c4b9739304a26be5c12af7/Revised_PM_38_10_2007.pdf
35. Swanson, D. A. & Tayman, J. Sub-national Population Estimates
(Springer, 2012).
36. Beaghen, M. & Stern, S. in Joint Statistical Meetings: Proc. Survey Research
Methods 2123–2137 (American Statistical Association, 2009).
37. Baker, J., Alcantara, A., Ruan, X. M., Watkins, K. & Vasan, S. A comparative
evaluation of error and bias in census tract-level age/sex-specific population
estimates: component I (net-migration) vs component III (Hamilton–Perry).
Popul. Res. Policy Rev. 32, 919–942 (2013).
38. Deming, W. E. & Stephan, F. F. On a least squares adjustment of a sampled
frequency table when the expected marginal totals are known. Ann. Math. Stat.
11, 427–444 (1940).
39. Siegel, J. & Swanson, D. A. Methods and Materials of Demography 2nd edn
(Emerald Group Publishing, 2008).
40. Smith, S. K. & Cody, S. Evaluating the housing unit method: a case
study of 1990 population estimates in Florida. J. Am Plann. Assoc. 60,
209–221 (1994).
41. Bogue, D. J. A technique for making extensive population estimates. J. Am. Stat.
Assoc. 45, 149–163 (1950).
42. Starsinic, D. E. & Zitter, M. Accuracy of the housing unit method in preparing
population estimates for cities. Demography 5, 475–484 (1968).
43. Armstrong, J. S. Principles of Forecasting: A Handbook for Researchers and
Practitioners (Springer, 2001).
44. Wilson, T. New evaluations of simple models for small area population
forecasts. Popul. Space Place 21, 335–353 (2014).

NATURE CLIMATE CHANGE   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved

 LETTERS

NATURE CLIMATE CHANGE DOI: 10.1038/NCLIMATE2961

45. Cohen, J. E. Population forecasts and confidence intervals for Sweden: a
comparison of model-based and empirical approaches. Demography 23,
105–126 (1986).
46. Swanson, D. A. & Beck, D. M. A new short-term county population projection
method. J. Econ. Soc. Meas. 20, 25–50 (1994).
47. Swanson, D. A., Tayman, J. & Barr, C. F. A note on the measurement of accuracy
for subnational demographic estimates. Demography 37, 193–201 (2000).
48. Smith, S. K., Tayman, J. & Swanson, D. A. State and Local Population
Projections: Methodology and Analysis (Plenum, 2001).
49. Swanson, D. A., Tayman, J. & Bryan, T. MAPE-R: a rescaled measure of
accuracy for cross-sectional subnational population forecasts. J. Popul. Res. 28,
225–243 (2011).
50. Swanson, D. A. & Tayman, J. Emerging Techniques in Applied Demography
93–117 (Springer, 2015).
51. Hyndman, R. J. & Athanasopoulos, G. Forecasting: Principles and Practice
(On Demand Publishing, LLC-Create Space, 2014).
52. McLeman, R. A. & Hunter, L. M. Migration in the context of vulnerability and
adaptation to climate change: insights from analogues. Wires Clim. Change 1,
450–461 (2010).

53. McLeman, R. A. Climate and Human Migration: Past Experiences, Future
Challenges (Cambridge Univ. Press, 2013).
54. Gutmann, M. P. & Field, V. Katrina in historical context: environment and
migration in the US. Popul. Environ. 31, 3–19 (2010).
55. Fussell, E., Curtis, K. J. & DeWaard, J. Recovery migration to the City of New
Orleans after Hurricane Katrina: a migration systems approach. Popul. Environ.
35, 305–322 (2014).
56. Hunter, L. M., Murray, S. & Riosmena, F. Rainfall patterns and U. S. migration
from rural Mexico. Int. Migr. Rev. 47, 874–909 (2013).
57. Thiede, B. & Brown, D. Hurricane Katrina: who stayed and why? Popul. Res.
Policy Rev. 32, 803–824 (2013).
58. Kayastha, S. L. & Yadava, R. P. in Population Redistribution and
Development in South Asia (eds Kosiński, L. A. & Elahi, K. M.) 79–88
(Springer, 1985).
59. Stoto, M. A. The accuracy of population projections. J. Am. Stat. Assoc. 78,
13–20 (1983).
60. Swanson, D. A. & Tayman, J. in Subnational Population Estimates Ch. 6
Vol. 31, Ch. 6 (Springer Series on Demographic Methods and Population
Analysis Vol. 31, 2012).

NATURE CLIMATE CHANGE   www.nature.com/natureclimatechange

© 2016 Macmillan Publishers Limited. All rights reserved