Data from: Modeling population vulnerabilities to climate change-driven hurricanes and tropical storms in the North Atlantic Basin

Published Sep 22, 2026 on Dryad. https://doi.org/10.5061/dryad.qv9s4mwwq

Abstract

Hurricanes and tropical storms are among the deadliest and costliest disasters in the United States, and climate change is expected to make the strongest storms more frequent. Knowing which communities are most likely to suffer deaths during a storm helps officials plan evacuations, strengthen buildings, and direct help where it is needed most. We combined records from twenty-four of the most damaging storms to strike the United States, Puerto Rico, and the U.S. Virgin Islands since 1992 with detailed information about the people living within 25 miles of the coast. For each storm, we measured rainfall, wind, and flooding, described the makeup of every neighborhood, and used a computer-learning method to estimate how at risk each small area would be. We then grouped storms by strength to map where danger is greatest for each storm category. Risk was highest in densely populated coastal cities and in low-lying inlets, peninsulas, and sounds, and it grew as storms strengthened. These maps are available in a public, interactive online dashboard so that residents, planners, and emergency managers can explore local risk and prepare before the next storm.

Access this dataset on Dryad

This dataset contains the derived, block-level mortality risk surfaces that underlie The Hurricane Dashboard (https://www.hurricanedashboard.org) and the associated manuscript, "Modeling Population Vulnerabilities to Climate Change-Driven Hurricanes and Tropical Storms in the North Atlantic Basin."

Storm-specific boosted-tree models were trained on 23 of the deadliest and costliest North Atlantic tropical cyclones to make U.S. landfall between 1992 and 2024, using NOAA hazard surfaces (wind swath, rainfall, storm-surge inundation) together with 2020 Decennial Census demographic, American Community Survey socioeconomic, and CDC/ATSDR Social Vulnerability Index covariates. Each model predicts population-standardized mortality for every U.S. census block within 25 miles of the Atlantic or Gulf coast, including Puerto Rico and the U.S. Virgin Islands. Block-level predictions were then averaged within Saffir–Simpson tiers to produce six category-level risk surfaces (tropical storm and Categories 1–5).

What is in this deposit: six category-level surfaces, each published in five or six files — one non-spatial attribute table (CSV), two summary tables (CSV), and two or three GeoPackages carrying the same attributes joined to census block or census tract geometry. The GeoPackages are the map layers served by the dashboard and are ready to load in ArcGIS Pro, ArcGIS Online, QGIS, R (sf), or Python (geopandas).

Important framing: this is an ecological (population-level) study. Values are modeled mortality risk for a hypothetical storm of a given Saffir–Simpson category, not observed deaths, and they support no individual-level causal inference. See "Known issues and caveats" before reuse.

Description of the data and file structure

Directory layout

hurricane/
├── tropical_storm/
├── category_1/
├── category_2/
├── category_3/
├── category_4/
└── category_5/

Each folder holds the files for one Saffir–Simpson tier. File names follow the pattern:

hurricane_<TIER>_<CONTENT>_per_100k.<ext>

where <TIER> is tropical_storm or category_1 … category_5. The per_100k suffix indicates that the primary display field is expressed per 100,000 population; raw and per-10,000 versions of the same quantity are also carried in every file.

Files in each tier folder

File	Format	Approx. size	Contents
`hurricane_<TIER>_block_attributes_no_geometry_per_100k.csv`	CSV	~1.0 GB	One row per census block. All 89 attribute fields, no geometry. Join to Census TIGER/Line 2020 blocks on `GEOID20`.
`hurricane_<TIER>_blocks_FULL_for_AGOL_per_100k.gpkg`	GeoPackage	~2.2 GB	One feature per census block, full-resolution TIGER/Line geometry, same attributes as the CSV.
`hurricane_<TIER>_blocks_SIMPLIFIED_for_AGOL_per_100k.gpkg`	GeoPackage	~2.2 GB	One feature per census block, generalized geometry for web rendering, same attributes.
`hurricane_<TIER>_tract_OVERVIEW_for_AGOL_per_100k.gpkg`	GeoPackage	~26 MB	One feature per census tract — block predictions aggregated to tract level with distributional summaries. The zoomed-out dashboard layer.
`hurricane_<TIER>_summary_per_100k.csv`	CSV	~460–485 MB	Tier-wide distributional summary of block predictions (min, mean, median, max, p75/p90/p95/p99) in three unit scales.
`hurricane_<TIER>_summary_by_risk_class_per_100k.csv`	CSV	~1.4 KB	The same distributional summary broken out by the five risk classes. Five rows.

tropical_storm/ contains five files rather than six; no _blocks_FULL_ GeoPackage was produced for that tier. The _tilework/ folder in the source directory is an empty build-scratch directory for the dashboard's PMTiles pipeline and is not part of the deposit.

Spatial reference and extent

Coordinate reference system: NAD83 geographic (EPSG:4269), inherited from the Census TIGER/Line source geometry. Geometry type is MULTIPOLYGON in all GeoPackages.
Geometry source: Census TIGER/Line 2020 tabulation blocks, 2024 vintage files (tl_2024_<ss>_tabblock20), recorded per record in MERGE_SRC.
Extent: census blocks whose centroid falls within 25 miles of the U.S. Atlantic or Gulf coastline, in the coastal states from Texas through Maine, plus Puerto Rico and the U.S. Virgin Islands.
Row count: approximately 1,186,384 blocks per tier. The block universe is identical across the six tiers; only the predicted values change, so the files can be stacked or joined tier-to-tier on GEOID20.

Relationship between the files

All six files in a tier describe the same underlying prediction. The block attributes CSV and the two block GeoPackages are the same table (the GeoPackages add geometry and a fid primary key). The tract overview GeoPackage is that table aggregated to census tracts. The two summary CSVs are descriptive statistics computed over the block table — they contain no new information and are provided so that legend breaks and map scaling can be reproduced without reading the full 1 GB table.

Missing data and value conventions

Missing values are the literal string NA in the CSVs and SQL NULL in the GeoPackages.
FIPS-derived identifiers (STATEFP20, COUNTYFP20, TRACTCE20, BLOCKCE20, GEOID20, tract_id, county_fips, state_fips, JOIN_DIGITS, geoid_for_fips) are text, not numeric — leading zeros are significant and will be destroyed if the CSVs are opened in a spreadsheet application with default type inference. Read them with explicit character types (readr::read_csv(col_types = ...) in R, dtype=str in pandas).
Mortality fields are carried in three parallel unit scales throughout: unsuffixed (population-standardized deaths per person), _per_10k (× 10,000), and _per_100k (× 100,000). mortality_display_units records which scale the dashboard renders; in this release it is "per 100,000" for every record.
risk_class / risk_quantile are quintiles of the within-tier distribution of avg_mortality_clean: 1 = Very Low, 2 = Low, 3 = Moderate, 4 = High, 5 = Very High. Because they are quintiles, each class holds roughly one-fifth of blocks within a tier; classes are not comparable across tiers as absolute risk levels.

Data dictionary

1. `hurricane_<TIER>_block_attributes_no_geometry_per_100k.csv` — 89 fields

The two block GeoPackages carry these same 89 fields, preceded by fid (integer primary key) and geom (MULTIPOLYGON geometry).

Identifiers and place labels

#	Field	Type	Definition
1	`dashboard_row_id`	text	Unique row key for the dashboard: `<GEOID20>_<Storm_Category>` with spaces replaced by underscores.
2	`block_dashboard_id`	text	15-digit census block GEOID used as the dashboard's block key. Equal to `GEOID20`.
3	`tract_id`	text	11-digit census tract GEOID (state + county + tract).
4	`county_fips`	text	5-digit county/county-equivalent FIPS code.
5	`state_fips`	text	2-digit state/territory FIPS code.
6	`state_name`	text	Full state or territory name.
7	`state_abbr`	text	Two-letter USPS state or territory abbreviation.
8	`county_equiv_name`	text	County, parish, municipio, or island-area name.
9	`county_equiv_type`	text	Type of the county-equivalent unit (e.g., "County or county-equivalent").
10	`place_label`	text	Human-readable label used in dashboard tooltips, `"<county_equiv_name>, <state_abbr>"`.

Storm tier

#	Field	Type	Definition
11	`Storm_Category`	text	Saffir–Simpson tier of this surface: `Tropical Storm`, `Category 1` … `Category 5`.
12	`category_order`	numeric	Ordinal rank of the tier for sorting and symbology (tropical storm = 0, Category 1–5 = 1–5).
13	`short_label`	text	Compact tier label used in the dashboard UI (`TS`, `Cat1` … `Cat5`).
14	`mortality_field`	text	Name of the source field from which this surface was built (`ts_avg_mort`, `cat_1_avg_mort` … `cat_5_avg_mort`). Provenance pointer to fields 83–88.

Predicted mortality and risk classification

#	Field	Type	Definition
15	`avg_mortality`	numeric	Primary model output. Mean of the storm-specific boosted-tree predictions across all storms in this tier, expressed as population-standardized deaths per person per block. May be negative — see caveats.
16	`avg_mortality_clean`	numeric	`avg_mortality` with negative predictions floored at zero. This is the field used for mapping and for the risk-class quintiles.
17	`avg_mortality_per_100k`	numeric	`avg_mortality_clean` × 100,000 — predicted deaths per 100,000 population.
18	`avg_mortality_per_10k`	numeric	`avg_mortality_clean` × 10,000 — predicted deaths per 10,000 population.
19	`mortality_display_value`	numeric	The value shown in dashboard tooltips, in the units named by `mortality_display_units`.
20	`mortality_display_units`	text	Unit label for `mortality_display_value`; `"per 100,000"` throughout this release.
21	`map_value`	numeric	Value driving the choropleth fill, raw scale. Equal to `avg_mortality_clean` in this release.
22	`map_value_per_100k`	numeric	`map_value` × 100,000.
23	`map_value_per_10k`	numeric	`map_value` × 10,000.
24	`risk_class`	text	Quintile label of `avg_mortality_clean` within this tier: `Very Low`, `Low`, `Moderate`, `High`, `Very High`.
25	`risk_quantile`	integer	Numeric form of `risk_class`, 1 (Very Low) through 5 (Very High).

Census TIGER/Line 2020 block attributes

Fields 26–43 are carried through unmodified from the Census TIGER/Line 2020 tabulation-block files; consult the Census TIGER/Line technical documentation for authoritative definitions.

#	Field	Type	Definition
26	`JOIN_DIGITS`	text	15-digit block GEOID used as the join key during table assembly. Equal to `GEOID20`.
27	`STATEFP20`	text	2020 state FIPS code.
28	`COUNTYFP20`	text	2020 county FIPS code (3 digits, within state).
29	`TRACTCE20`	text	2020 census tract code (6 digits, within county).
30	`BLOCKCE20`	text	2020 census block code (4 digits, within tract).
31	`GEOID20`	text	15-digit 2020 census block identifier (state + county + tract + block). Primary join key to external Census products.
32	`GEOIDFQ20`	text	Fully qualified Census GEOID, e.g. `1000000US230299557001037`.
33	`NAME20`	text	Census block name, e.g. `Block1037`.
34	`MTFCC20`	text	MAF/TIGER Feature Class Code; `G5040` for tabulation blocks.
35	`UR20`	text	2020 urban/rural indicator: `U` = urban, `R` = rural.
36	`UACE20`	text	2020 urban area code, `NA` for rural blocks.
37	`FUNCSTAT20`	text	Census functional status code (`S` = statistical entity).
38	`ALAND20`	numeric	Land area of the block, square meters.
39	`AWATER20`	numeric	Water area of the block, square meters.
40	`INTPTLAT20`	text	Latitude of the block's internal point, signed decimal degrees, zero-padded string.
41	`INTPTLON20`	text	Longitude of the block's internal point, signed decimal degrees, zero-padded string.
42	`HOUSING20`	integer	2020 Census housing unit count for the block.
43	`POP20`	integer	2020 Census total population count for the block.

Assembly and quality-control fields

These fields record how the block table was built. They are retained for transparency and reproducibility and are generally not needed for reuse.

#	Field	Type	Definition
44	`MERGE_SRC`	text	Source TIGER/Line shapefile the geometry came from, e.g. `tl_2024_23_tabblock20`.
45	`MERGE_SRC_`	text	Secondary merge-provenance field from the ArcGIS Pro merge; largely `NA`.
46	`MERGE_SRC1`	text	Secondary merge-provenance field; largely `NA`.
47	`MERGE_SR_1`	text	Secondary merge-provenance field; largely `NA`.
48	`Shape_Leng`	numeric	Perimeter of the block polygon in CRS units (decimal degrees), from ArcGIS Pro.
49	`Shape_Area`	numeric	Area of the block polygon in CRS units (square decimal degrees), from ArcGIS Pro. Not a valid area measure — use `area_miles` or `ALAND20`.
50	`area_miles`	numeric	Block area in square miles, computed on a projected surface.
51	`UATYPE20`	text	2020 urban area type code; `NA` for rural blocks.
52	`POP20pd`	numeric	Population density, `POP20` per square mile.
53	`popfinal`	numeric	Population value used in model fitting after reconciliation of Census and interpolated counts.
54	`popfinalup`	numeric	Upper-bound / updated variant of `popfinal` used in sensitivity checks.
55	`STATE_CHECK`	text	State FIPS re-derived from `GEOID20` as a QC check against `STATEFP20`.

Aggregate model outputs

#	Field	Type	Definition
56	`avg_pred_boosted_weighted`	numeric	Population-weighted mean of the boosted-tree predictions across all 23 storms, all tiers pooled.
57	`cdr_pred_weighted`	numeric	Predicted crude death rate, population-weighted.
58	`cdr_pred_notweighted`	numeric	Predicted crude death rate, unweighted.

Per-storm predictions (fields 59–82)

One field per fitted storm model, giving that model's predicted population-standardized mortality for the block, on the raw (per-person) scale. These are the inputs averaged into the tier surfaces.

#	Field	Storm
59	`adjusted_deathsallison_pred`	Tropical Storm Allison (2001)
60	`adjusted_deathlee_pred`	Tropical Storm Lee (2011)
61	`adjusted_deathberyl_pred`	Hurricane Beryl (2024)
62	`adjusted_deathflorence_pred`	Hurricane Florence (2018)
63	`adjusted_deathirenepr_pred`	Hurricane Irene (2011), Puerto Rico
64	`adjusted_deathmatthew_pred`	Hurricane Matthew (2016)
65	`adjusted_deathsandy_pred`	Hurricane Sandy (2012)
66	`adjusted_deathfloyd_pred`	Hurricane Floyd (1999)
67	`adjusted_deathfrances_pred`	Hurricane Frances (2004)
68	`adjusted_deathfrances_pred_2`	Hurricane Frances (2004), second model run — see caveats
69	`adjusted_deathgeorges_pred`	Hurricane Georges (1998)
70	`adjusted_deathike_pred`	Hurricane Ike (2008)
71	`adjusted_deathsally_pred`	Hurricane Sally (2020)
72	`adjusted_deathivan_pred`	Hurricane Ivan (2004)
73	`adjusted_deathkatrina_pred`	Hurricane Katrina (2005)
74	`adjusted_deathrita_pred`	Hurricane Rita (2005)
75	`adjusted_deathwilma_pred`	Hurricane Wilma (2005)
76	`adjusted_deathharvey_use_pred`	Hurricane Harvey (2017)
77	`adjusted_death_ian_pred`	Hurricane Ian (2022)
78	`adjusted_deathida_pred`	Hurricane Ida (2021)
79	`adjusted_deathirmaprusvi_pred`	Hurricane Irma (2017), Puerto Rico and U.S. Virgin Islands
80	`adjusted_deathmariapr_pred`	Hurricane Maria (2017), Puerto Rico
81	`adjusted_deathsandrew_pred`	Hurricane Andrew (1992)
82	`adjusted_deathsmichael_pred`	Hurricane Michael (2018)

Landfall date, minimum central pressure, maximum sustained wind, landfall coordinates, and the Saffir–Simpson category assigned to each storm are given in Table 2A of the associated manuscript.

Tier-averaged surfaces and final key

#	Field	Type	Definition
83	`cat_1_avg_mort`	numeric	Mean of the per-storm predictions for storms assigned to Category 1.
84	`cat_2_avg_mort`	numeric	Mean of the per-storm predictions for storms assigned to Category 2.
85	`cat_3_avg_mort`	numeric	Mean of the per-storm predictions for storms assigned to Category 3.
86	`cat_4_avg_mort`	numeric	Mean of the per-storm predictions for storms assigned to Category 4.
87	`cat_5_avg_mort`	numeric	Mean of the per-storm predictions for storms assigned to Category 5.
88	`ts_avg_mort`	numeric	Mean of the per-storm predictions for the tropical storm tier.
89	`geoid_for_fips`	text	15-digit block GEOID retained as a trailing join key for FIPS-based lookups. Equal to `GEOID20`.

All six tier fields (83–88) are present in every file; the one matching the folder's tier is the field copied into avg_mortality, as recorded in mortality_field. This makes any single file sufficient to compare a block's risk across all six tiers.

2. `hurricane_<TIER>_tract_OVERVIEW_for_AGOL_per_100k.gpkg`

One feature per census tract. Layer name: hurricane_<TIER>_tract_OVERVIEW_per_100k.

Field	Type	Definition
`fid`	integer	GeoPackage primary key.
`geom`	MULTIPOLYGON	Census tract geometry, NAD83 (EPSG:4269).
`tract_id`, `county_fips`, `state_fips`, `state_name`, `state_abbr`, `county_equiv_name`, `county_equiv_type`, `place_label`	text	Tract identifiers and labels, as defined above.
`Storm_Category`, `category_order`, `short_label`, `mortality_field`	text / numeric	Tier identification, as defined above.
`n_blocks`	integer	Number of census blocks in the tract contributing to the aggregate.
`n_nonmissing`	integer	Number of those blocks with a non-missing prediction.
`map_value`, `map_value_per_100k`, `map_value_per_10k`	numeric	Tract-level value driving the choropleth, in three unit scales.
`mortality_display_value`, `mortality_display_units`	numeric / text	Tooltip value and its unit label.
`risk_class`, `risk_quantile`	text / integer	Tract-level risk quintile, defined as for blocks.
`min_mortality`, `mean_mortality`, `median_mortality`, `p75_mortality`, `p90_mortality`, `p95_mortality`, `p99_mortality`, `max_mortality`	numeric	Distribution of block-level predictions within the tract, raw scale.
`min_mortality_per_100k` … `max_mortality_per_100k`	numeric	The same eight statistics × 100,000.
`min_mortality_per_10k` … `max_mortality_per_10k`	numeric	The same eight statistics × 10,000.

3. `hurricane_<TIER>_summary_per_100k.csv` — 31 fields

Tier-wide distributional summary of the block-level predictions.

Field	Type	Definition
`Storm_Category`, `category_order`, `short_label`, `mortality_field`	text / numeric	Tier identification, as defined above.
`mortality_display_units`	text	Unit label for the dashboard display scale.
`n_blocks`	integer	Number of census blocks in the tier (~1,186,384).
`n_nonmissing`	integer	Number of those blocks with a non-missing prediction.
`min_mortality`, `mean_mortality`, `median_mortality`, `max_mortality`, `p75_mortality`, `p90_mortality`, `p95_mortality`, `p99_mortality`	numeric	Distribution of block-level predictions, raw scale.
`min_mortality_per_100k` … `p99_mortality_per_100k`	numeric	The same eight statistics × 100,000.
`min_mortality_per_10k` … `p99_mortality_per_10k`	numeric	The same eight statistics × 10,000.

4. `hurricane_<TIER>_summary_by_risk_class_per_100k.csv` — 20 fields, 5 rows

The same summary, computed separately within each of the five risk classes. Fields are as in file 3, with risk_class and risk_quantile added and mortality_field and all percentile columns omitted:

Field	Type	Definition
`risk_class`	text	`Very Low`, `Low`, `Moderate`, `High`, or `Very High`.
`risk_quantile`	integer	1–5, matching `risk_class`.
`n_blocks`, `n_nonmissing`	integer	Block counts within the class (approximately one-fifth of the tier each, by construction).
`min_mortality`, `mean_mortality`, `median_mortality`, `max_mortality` and their `_per_100k` / `_per_10k` variants	numeric	Distribution within the class.

These five rows are the authoritative source for the dashboard's legend breaks: the min_mortality_per_100k and max_mortality_per_100k of each class define the class boundaries.

Known issues and caveats

Please read before reuse.

Negative predicted values. Boosted-tree predictions are unconstrained and a substantial share of low-risk blocks receive small negative values in avg_mortality and in the per-storm prediction fields. avg_mortality_clean floors these at zero and is the field that should be used for mapping, classification, and any downstream summary. avg_mortality is retained unmodified for transparency.
Modeled, not observed. Every mortality field in this dataset is a model prediction for a hypothetical storm of the given category, not a count of deaths. Observed mortality came from NOAA and NWS sources and is known to underestimate true excess deaths, particularly indirect and delayed mortality; that bias propagates into these predictions.
Duplicated Frances field. adjusted_deathfrances_pred and adjusted_deathfrances_pred_2 both correspond to Hurricane Frances (2004) and are valid.
summary_per_100k.csv is highly redundant. These files are ~460–485 MB but contain a single tier-level summary record repeated once per block. Any single row is sufficient; the summary_by_risk_class file carries the same information in 1.4 KB. The full-length version is deposited unmodified so that it matches what the dashboard pipeline consumed.
No _blocks_FULL_ GeoPackage for the tropical storm tier. Use hurricane_tropical_storm_blocks_SIMPLIFIED_for_AGOL_per_100k.gpkg, or rejoin hurricane_tropical_storm_block_attributes_no_geometry_per_100k.csv to TIGER/Line 2020 blocks on GEOID20 for full-resolution geometry.
Simplified geometry is for display. The _SIMPLIFIED_ layers are generalized for web rendering. Do not use them for area calculations, overlay analysis, or any measurement; use the _FULL_ layers or ALAND20 / area_miles.
Risk classes are within-tier quintiles. A "Very High" block in the tropical storm surface and a "Very High" block in the Category 5 surface are not at comparable absolute risk. Compare the continuous *_per_100k fields across tiers, not the class labels.
Ecological inference. All results are population-level at block resolution. They do not support inference about individual risk, and block-level associations should not be read as individual-level effects.
Small-population blocks. Rates per 100,000 computed on blocks with very small populations are unstable and produce the extreme upper tail visible in the summary files (e.g., a maximum of ~13,996 per 100,000 in the Category 5 tier against a median of ~34). Filter on POP20 or popfinal before mapping rates.
Leading zeros. See "Missing data and value conventions" above — read all FIPS-derived fields as text.

Sharing/Access information

Links to other publicly accessible locations of the data:

Interactive dashboard: https://www.hurricanedashboard.org
Preprint: Mullins S, Uelmen J. Modeling Population Vulnerabilities to Climate Change-Driven Hurricanes and Tropical Storms in the North Atlantic Basin. medRxiv. https://www.medrxiv.org/content/10.64898/2026.08.16.26360517v1
Analysis code and the dashboard build pipeline are archived alongside these files in this same Dryad deposit (see Code/Software below)

Data were derived from the following sources:

NOAA National Hurricane Center, Tropical Cyclone Reports and Data Archive — storm tracks, landfall parameters, wind swaths, rainfall, storm-surge inundation, and reported mortality: https://www.nhc.noaa.gov/data/
NOAA National Centers for Environmental Information — costliest and deadliest U.S. tropical cyclone tabulations: https://www.ncei.noaa.gov/
NOAA National Weather Service Storm Events Database: https://www.ncdc.noaa.gov/stormevents/
U.S. Census Bureau, 2020 Decennial Census, Demographic and Housing Characteristics file: https://data.census.gov/
U.S. Census Bureau, American Community Survey (2020 and 2023): https://www.census.gov/programs-surveys/acs
U.S. Census Bureau, TIGER/Line Shapefiles, 2020 tabulation blocks (2024 vintage): https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html
CDC/ATSDR Social Vulnerability Index, 2022: https://www.atsdr.cdc.gov/place-health/php/svi/
U.S. Bureau of Economic Analysis / Bureau of Labor Statistics, county-level GDP, 2001–2023: https://www.bea.gov/data/gdp/gdp-county-metro-and-other-areas

Methods

Detailed methods are given in the associated manuscript. In brief:

Storm selection. Twenty-four of the deadliest and costliest North Atlantic tropical cyclones making U.S. landfall between 1992 and 2024 were selected from NOAA's tabulation of costliest U.S. tropical cyclones, sampling up to the top five storms within each Saffir–Simpson tier. Only two Category 5 landfalls and three qualifying tropical storms fell within the study period. Tropical Storm Imelda (2019) was dropped because it recorded no mortality within the study buffer, leaving 23 fitted models.

Study area. Census blocks within 25 miles of the U.S. Atlantic or Gulf coastline, including Puerto Rico and the U.S. Virgin Islands. Block resolution is roughly two orders of magnitude finer than census tract resolution.

Hazard layers. NHC wind swath and rainfall products were digitized and rasterized in ArcGIS Pro using maximum-likelihood and machine-learning image classifiers. Storm-surge inundation rasters were extracted and averaged per block by mean-pixel calculation. Hazard covariates entering the models were wind swath category (tropical-storm-force or hurricane-force), mean rainfall (inches), and mean predicted inundation (feet).

Outcome. Reported deaths were population-standardized to the block by multiplying the unadjusted death total by the block's population divided by the total population of the reporting state or county.

Covariates. 2020 Decennial Census age, sex, race, ethnicity, household size, and tenure; ACS household income quartiles and homeownership; CDC/ATSDR SVI percentile rank; county-level GDP (2001–2023 storms only); and the three hazard layers above. Multicollinearity was not treated as a modeling constraint because boosted trees do not estimate coefficients.

Models. One boosted-tree model per storm, with population-standardized mortality as the outcome, fitted with training and validation partitions. Performance was assessed by R² and root average squared error (RASE); see Table 4A of the manuscript.

Surfaces. Block-level predictions from all storms within a Saffir–Simpson tier were averaged to produce the six category surfaces in this deposit. Negative predictions were floored at zero (avg_mortality_clean), values were converted to rates per 10,000 and per 100,000, and blocks were classified into within-tier quintiles.

Code/Software

Software	Version	Role
ArcGIS Pro (Esri)	2026 release	Hazard raster digitization and classification, zonal statistics, block/tract joins, geometry merge and simplification, GeoPackage export
JMP Student Edition (SAS Institute)	18.2.1	Data management, boosted-tree model fitting, training/validation partitioning, prediction export
R / RStudio	RStudio 2025.08.0-357	Raster processing, spatial extraction, table assembly, summary and risk-class computation
ArcGIS Online	—	Hosting of the `_for_AGOL_` layers
PMTiles	—	Vector tile packaging for the public dashboard

The GeoPackages open without additional software in ArcGIS Pro, QGIS, R (sf::st_read()), and Python (geopandas.read_file()). The block-level CSVs are approximately 1 GB each; read them with a chunked or lazy reader (arrow, duckdb, readr::read_csv_chunked(), or pandas.read_csv(chunksize=...)) rather than loading them whole.

Analysis code and the dashboard build pipeline are archived with this Dryad deposit. Dryad routes software submissions to Zenodo, which mints a separate software DOI; that DOI is listed on the Dryad landing page for this dataset.

Citation

Please cite both the dataset and the associated manuscript:

Mullins, S., & Uelmen, J. (2026). Block-level tropical cyclone mortality risk surfaces for the U.S. Atlantic and Gulf coasts, Puerto Rico, and the U.S. Virgin Islands [Data set]. Dryad. https://doi.org/10.5061/dryad.qv9s4mwwq

Mullins, S., & Uelmen, J. (2026). Modeling population vulnerabilities to climate change-driven hurricanes and tropical storms in the North Atlantic Basin. medRxiv. https://doi.org/10.64898/2026.08.16.26360517

Author and contact information

Shaelyn Mullins — Department of Population Health Sciences, School of Medicine and Public Health, University of Wisconsin–Madison

Johnny Uelmen (corresponding) — Department of Population Health Sciences, School of Medicine and Public Health, University of Wisconsin–Madison — Uelmen Disease Ecology Lab, https://www.diseaseecologylab.org

Funding. This research received no specific external funding.

Conflicts of interest. The authors declare no conflicts of interest relevant to this work.

License. Data files in this deposit are released under a CC0 1.0 Universal public domain dedication, per Dryad policy. Accompanying code is released under CC BY 4.0 via the linked Zenodo software record.

Date of data generation. Derived surfaces generated August 2026 from source data spanning 1992–2024.