The Drought Relief Gap

Drought does more damage to American crops than any other weather peril. Across twenty-five years of weekly measurement in 3,211 counties, it produced zero federal disaster declarations. The relief runs through a different system entirely — one that never has to be asked.

Abstract

We compare the United States Drought Monitor's county-week record against the federal disaster declaration record over 2000–2024: 4,190,355 measured county-weeks and 80,275 county-years against 5,252 declarations. Not one of the 80,275 county-years carries a FEMA drought declaration, while 29,721 (37.0%) carry a declaration of some other kind. The same join resolves 38 earlier drought declarations to 1,336 county rows, the last in 1993, so the zero is a property of the record and not of the method. The only drought declaration anywhere inside the measurement window went to Chuuk, in the Federated States of Micronesia. Over the same period the measured record crossed the USDA Livestock Forage Program's D2-for-eight-consecutive-weeks threshold in 23,560 county-years (29.4%), touching 3,085 of the 3,211 counties — an instrument whose trigger is the measurement itself. Drought also reaches the declaration record indirectly. Scoring each declaration's opening week against its own county's same-month drought history, the 1,309 fire declarations sit at the 61st percentile (95% CI 0.595–0.624); collapsing the fire sieges that generate many grants at once, the conservative bound is the 57th percentile (state-year clusters, n = 328, CI 0.542–0.593, p = 7.2 × 10⁻⁶). Of eight meteorological perils tested, fire is the only one materially above the median, and six of the remaining seven sit at or below it.

1. Why this question

Drought is the most expensive thing that happens to American crops. In the recorded storm-damage ledger it accounts for $29.12 billion of $46.95 billion, or 62.0% of all crop damage — more than freeze, hail, wind, flood and tornado combined.

It is also the peril least visible in the instrument most people reach for first. The federal disaster declaration is the public record of "something bad happened here," consulted by anyone reconstructing what a place went through. If that record and the measured record disagree about drought, then a great many downstream conclusions drawn from declarations alone are wrong in a specific and predictable direction.

So: does relief follow the measurement?

2. Data and scope

Measured drought. The US Drought Monitor county record, 2000-01-04 to 2024-12-31 — 4,190,355 county-weeks across 3,211 counties in 51 states. Each cell carries the percentage of county area in each severity class D0 through D4.

Declarations. The FEMA disaster declaration record, 5,252 declarations from FY1953 to FY2026, each carrying its incident type, dates, and the list of areas designated under it.

Both are administrative records of realised conditions rather than model output. The USDM is an expert weekly assessment synthesising measured precipitation, streamflow, soil moisture and reservoir indicators; a declaration is an administrative act recording a determination that was made. Neither is a forecast, a projection, or a reanalysis field, and no modelled quantity enters this analysis.

Three instruments, not one. The 5,252 declarations are not interchangeable. They comprise 2,926 major disaster declarations (DR), 1,674 fire management assistance grants (FM), and 652 emergency declarations (EM). FM is a faster, narrower, lower-threshold program that reimburses wildfire suppression cost. Pooling them would make the fire count look like a disaster count, so instrument is carried through every result below.

3. Method

Resolving declarations to counties. FEMA designates areas by display name, not FIPS — Cabell (County), WV — so declarations were joined to counties by normalised name and state. Three traps were found by inspecting failures rather than assuming success: the county record spells Saint Louis where FEMA writes St. Louis; nested parentheticals appear (New London (County)(in (P)MSA 3280,5520), CT); and independent cities share names with their surrounding counties, so Baltimore (County), MD and Baltimore, MD are different places with different FIPS and must not be merged.

112,860 of 115,531 area designations resolved (97.7%). Statewide designations were expanded to every county in that state; 17 declarations in the window are of this kind. The unresolved remainder is not counties: Maine unorganised territories (Misery Gore, Batchelders Grant), tribal statistical areas, Alaska native villages, and Pacific territories outside the USDM's coverage. None of these has a county FIPS, so none can be joined, and each is counted rather than dropped quietly.

Severity at the moment of declaration. Each declaration-county pair was matched to the most recent USDM map published on or before the incident began — 46,297 pairs across 3,051 declarations. The county's D2-or-worse area that week was then scored as a percentile of that same county's own history, using mid-ranks because the distribution has heavy ties at zero.

The baseline is season-matched. Comparing an August declaration against a full-year baseline containing wet winters would manufacture a drought signal out of the calendar alone. Each declaration is therefore scored only against its own county's same calendar month across all 25 years — a median of 111 comparison weeks. Under the null, the expected percentile is exactly 0.5. The all-weeks figures are reported alongside so the size of the seasonal correction is visible.

Which perils are tested, by rule. Every meteorological incident type with at least 20 matched declarations in the window: fire, severe storm, hurricane, flood, snowstorm, severe ice storm, tornado and tropical storm. The only type clearing that bar and excluded is Biological (105 declarations — pandemic, not weather). Corrections are Bonferroni across all eight.

The unit of analysis, twice over. One declaration designates many counties, so county rows are not independent; every test is run on the per-declaration mean. But declarations are not fully independent either — a single fire siege in one state produces many separate FM grants in the same week. Results are therefore reported at three levels: per declaration, collapsed to state-week, and collapsed to state-year. The state-year figure is the conservative bound.

A note on the D2+ area column. The USDM's category percentages are cumulative — d2 means "D2 or worse". Verified across all 4,190,355 rows: none + d0 = 100 exactly, and d0 ≥ d1 ≥ d2 ≥ d3 ≥ d4. D2+ area is therefore d2 alone. A stored convenience column that sums d2+d3+d4 triple-counts and reaches 300%; it is not used here. The Livestock Forage eligibility flags are unaffected, as they test only whether D2+ area exceeds zero, and our independent recomputation reproduces the stored flag exactly (23,560 = 23,560).

4. The gap

Zero of 80,275 county-years carry a FEMA drought declaration.

Against 29,721 county-years (37.0%) carrying a declaration of some other kind. Within the window the record contains 1,316 Fire, 856 Severe Storm, 320 Hurricane, 241 Flood and 167 Biological declarations.

The zero is not a broken join. A pipeline that silently failed to resolve drought declarations would produce exactly this result, so the check that matters is whether the same machinery finds drought declarations when they exist. It does: it resolves 38 drought declarations to 1,336 county rows, every one of them before 1994. Whatever it can do in 1977 it can do in 2012.

Drought declarations exist — 46 of them, all-time — but they stop. Thirty-eight touched one of these 51 states, and the last was EM-3113, Texas, 10 September 1993, which designated 149 counties under the title "EXTREME FIRE HAZARD." The single drought declaration falling inside the measurement window, EM-3276 in 2007, designated one county-equivalent: Chuuk, in the Federated States of Micronesia.

The gap is not a quiet decline in a noisy series. It is a flat zero across a quarter century that includes 2012, when 1,859 counties — 58% of every county measured — met the livestock-forage drought threshold (Figure 1A).

5. What fired instead

The measured record did not go unanswered; it was answered by a different mechanism.

Over the same window, the USDM crossed the Livestock Forage Program's D2-for-eight- consecutive-weeks threshold in 23,560 county-years (29.4%), and the higher D3 tier in 17,820. 3,085 of the 3,211 counties — 96.1% — qualified at least once. The measured record contains 628,348 county-weeks at D2 or worse, 290,105 at D3 or worse, and 87,587 at D4. The longest unbroken D2+ run reaches 296 consecutive weeks: five and a half years.

What this instrument does and does not cover. The Livestock Forage Program compensates grazing loss — it pays livestock producers whose pasture failed, not the row-crop or orchard growers whose damage dominates §1. It is not the whole federal answer to agricultural drought, and it is not the crop-side answer. Federal crop insurance indemnities, the Noninsured Crop Disaster Assistance Program, and Secretarial Disaster Designations all sit on the crop side, and none of them is measured here. LFP is singled out for one reason: its trigger is the measurement itself. Eight consecutive weeks at D2 on the USDM map is the eligibility rule, which makes it the one drought instrument that can be read directly off the same record this paper is testing. The crop-side instruments are loss-based rather than measurement-triggered, and quantifying them is a separate study.

That distinction is the substantive point. The declaration system waits for a governor to request and a president to approve. The livestock-forage mechanism fires on the map: nobody has to ask. A peril routed to mechanisms of the second kind will look, from inside the first, exactly like a peril that never happened.

The threshold is not doing the work. The eligibility flag counts any D2+ area in the county, which is permissive. Requiring 25% of county area still yields 20,333 eligible county-years; requiring 50% yields 18,508; requiring 75% of the entire county in D2 or worse for eight straight weeks still yields 16,685 county-years, 20.8% of the panel. The finding does not depend on where the line is drawn.

6. Drought in the record, under another name

Fire and drought are physically coupled — dry fuel burns — and fire is the largest declaration category in the window. If drought reaches the declaration record at all, it should reach it through fire.

It does. Scored against their own counties' same-month history, the 1,309 fire declarations open at the 61st percentile of drought severity (mean 0.610, 95% CI 0.595–0.624; Bonferroni-corrected p = 3.3 × 10⁻⁴⁸, rank-biserial +0.47). Of the eight meteorological perils tested, fire is the only one materially above the median:

Declaration type	n	Mean percentile	95% CI	Opens in a D2+ week
Fire	1,309	0.610	0.595–0.624	54.4%
Tropical storm	25	0.542	0.498–0.596	15.3%
Snowstorm	92	0.499	0.479–0.524	9.1%
Tornado	38	0.491	0.463–0.529	4.0%
Severe storm	814	0.490	0.482–0.498	7.2%
Flood	229	0.484	0.470–0.497	8.1%
Hurricane	288	0.456	0.448–0.464	5.7%
Severe ice storm	62	0.449	0.435–0.464	9.9%

Table 1. Drought severity at the opening week of a declaration, as a percentile of the same county's same-month record, 2000–2024. Bonferroni-corrected across eight perils.

54.4% of fire declarations begin in a week with D2 drought or worse somewhere in the county, against 4.0–15.3% for every other peril.

One control does not sit below the median. Tropical storm comes in at 0.542, above 0.5. With n = 25 its interval spans the null (0.496–0.596) and its corrected p is 1.00, so it is not distinguishable from chance — but it is reported rather than dropped, and it is the reason the claim here is "fire is the only peril materially above the median" rather than "every other peril is below it."

The effect is not an artefact of the fire-grant instrument. Fire management assistance grants are 95.4% of the fire count (1,249 of 1,309), and they reimburse suppression cost rather than compensating an agricultural loss. If the result were an FM artefact it would vanish in the major-disaster subset. It does not: FM sits at 0.607 (p = 4.2 × 10⁻⁴⁶) and the 46 DR fire declarations sit higher, at 0.683 (CI 0.619–0.745, p = 4.9 × 10⁻⁵).

Fire sieges inflate the headline, and the effect survives their removal. One siege in one state in one week can produce a dozen separate grants, which are not a dozen independent observations. Collapsing to state-week clusters moves fire from 0.610 to 0.586 (n = 837, CI 0.568–0.604, p = 2.1 × 10⁻²¹); collapsing to state-year moves it to 0.567 (n = 328, CI 0.542–0.593, p = 7.2 × 10⁻⁶). Every interval excludes 0.5. Notably the controls barely move under the same treatment — severe storm goes 0.490 → 0.490 → 0.492 — because only fire arrives in sieges. The conservative reading of this paper is the 57th percentile, not the 61st.

The controls are what make this readable. They do not merely fail to show the effect — most run the other way. Hurricane and severe ice storm declarations open in weeks significantly drier-than-typical, which is what one would expect of perils that arrive with water. A method that manufactured drought signal from seasonality or from declaration-prone geography would have pushed every category up together. Instead one category separates and seven do not.

Seasonality accounts for a small share and not the effect: the all-weeks baseline puts fire at 0.631, the season-matched baseline at 0.610.

7. Limitations

Two different kinds of thing, compared on purpose. A declaration is a political and statutory instrument with eligibility thresholds, a request step and an approval step. The USDM is an assessment of conditions. They are not two measurements of one quantity, and the gap between them is not evidence that either is wrong. It is evidence about routing.

The zero is explained, not merely observed. The Stafford Act is built around discrete events producing sudden infrastructure damage. Drought produces neither, and agricultural drought relief was routed to USDA programs. The finding here is the size and completeness of that routing, not a claim that anyone failed.

Only one of the federal drought instruments is measured. LFP is quantified because its trigger is the USDM itself. Crop insurance indemnities, NAP and Secretarial Disaster Designations — the instruments that answer the crop damage in §1 — are not in this analysis, and the paper makes no claim about their adequacy.

Our eligibility flag is not an FSA determination. It encodes the D2-for-eight-weeks rule against USDM county area. Actual determinations involve grazing periods, pasture type and normal carrying capacity. The flag is an upper bound on eligibility, which is why the area-floor sensitivity above matters.

Relabelled describes the accounting, not an intent, and not a transfer. Fire declarations open in dry weeks because drought produces fire. FM money reimburses suppression cost to state agencies; it does not reach the drought-damaged farm. Nothing here shows anyone choosing a label, and nothing here should be read as drought relief arriving in a fire envelope.

Alternative explanations for §6 that this design does not exclude. The FM program's issuance rate has changed since its 2001 creation; fire-season length rather than drought severity could drive both sides; and states differ in how aggressively they request assistance. Within-county, same-month scoring removes geography and calendar but not program-era effects.

Coverage. The USDM county record begins in 2000, so this window cannot extend earlier. 2.3% of area designations do not resolve to a county and are enumerated in §3. One declaration in the record carries a truncated area list.

8. What it means

For anyone reconstructing what a place has been through, the practical consequence is narrow and sharp: the federal declaration record is not a drought record, and has not been one since 1993. A county can spend five consecutive years at D2 or worse and generate no declaration at all. Reading declarations as a proxy for weather burden will systematically erase the single most expensive peril in the ledger, and will do so most severely in exactly the places where drought is most persistent.

Two records exist. The one that responds to drought is the measurement, and it responds without being asked.

Provenance

Measured drought — US Drought Monitor county record, drought_category YearLocationdex, 4,190,355 county-weeks, 2000-01-04 to 2024-12-31.
Declarations — FEMA OpenFEMA Disaster Declarations Summaries, fema Eventdex, 5,252 declarations, FY1953–FY2026.
Prior crop-damage figures — NCEI Storm Events, as reported in Crop Damage: Nature's Number-One Offender (crop-loss-by-peril).

Related prior work. Drought-Wildfire-AQI Cascade (drought-wildfire-aqi-cascade- multi-month) tested a drought→fire→air-quality chain at state-month resolution and was limited by an AQI window that excluded the western fire season; the drought→fire coupling relied on here is consistent with its framing but is established independently, at county resolution, on the declaration record rather than on air quality. ENSO Phase vs US Drought Severity (enso-drought-correlation) found no significant ONI–drought correlation at n = 17 state-months; nothing in this paper depends on an ENSO term.

Scripts: scripts/inventory.py, scripts/extract.py, scripts/analyze.py, scripts/visualize.py, scripts/verify_draft.py. All numbers in data/results.json. Bootstraps are seeded (20260831). Reproducibility is a tested property, not an assumption: analyze.py was run five times and produced byte-identical results.json. Getting there required sorting every group-by that feeds a resampled array — a group mean is a floating-point sum whose value depends on row order within the group, and polars does not guarantee that order, so an unsorted frame moves the published intervals between runs.