Repeated disaster exposure barely moves county-level net migration through 2022 — but a real, recent outflow from high-risk counties has emerged since 2024

Summary

Very little, in the broad national data through 2022 — and a real but only recently visible amount in 2024-2025. When we count every FEMA disaster declaration a U.S. county received and test whether counties hit repeatedly then lose population faster than counties hit rarely, using IRS county-to-county migration records (2011-2022) with county and year fixed effects (i.e., comparing each county to its own history, not to other counties), the estimated effect of an extra trailing-3-year disaster declaration on net out-migration is -0.012 people per 1,000, statistically indistinguishable from zero (p=0.92, n=31,885 county-years across 3,082 counties). A naive comparison across counties (not controlling for each county's own baseline) actually points the opposite way — repeatedly-hit counties look like net in-migration magnets — but that is an artifact of Florida and Texas being simultaneously disaster-prone and fast-growing; it is not robust to excluding either state. Independently, Census Bureau domestic-migration data analyzed by Redfin show a real and recent reversal: high-flood-risk counties posted a net outflow for the first time in five years in 2024 (-34,099) and that outflow nearly doubled in 2025 (-63,357), while low-flood-risk counties gained 69,857 people. Academic literature finds real but modest and hazard-specific effects (largest for high-fatality disasters and hurricanes) that a blunt all-hazard county count does not isolate. Bottom line: disaster-driven out-migration is a real, growing, and now measurable phenomenon at the national level as of 2024-2025, but it has not yet shown up as a broad, robust county-fixed-effect pattern in the government migration data available through 2022, and most of what does show up in raw cross-sectional comparisons is confounded by Sun Belt growth.

What the research literature says

Several independent academic studies establish that disasters do increase out-migration, but the effect is concentrated in severe events and specific hazard types, not a blanket effect of "any FEMA declaration":

• Boustan et al. (2020) (cited in the Natural Hazards literature review below) find that disasters causing 25+ fatalities are associated with about 1.5 percentage points higher net out-migration in the affected county — i.e., only the most severe events register.
• A gravity-model study using IRS migration data (Natural Hazards, 2024) finds natural hazards in the origin county are associated with larger outflows, with hurricanes showing the largest effect, followed by floods and severe storms — and that people who do leave tend to move to nearby counties, which can leave them still exposed to the same hazard.
• Research summarized via Congress.gov/CRS and related work finds disasters raise a household's propensity to move out of its county, and further out of its metro area, by up to about one percentage point, with low-income households — who receive less FEMA aid — less likely to move despite repeat exposure.
• Out-migration probabilities in disaster-affected counties have been measured rising from roughly 6.4% to 7.7% pre- to post-disaster in some studies, but most disaster-hit places recover their population over a longer horizon through "recovery migration" — people moving back in — which is exactly the dynamic that washes out a simple net-migration measure.
• Effects are unequal: racially minoritized and lower-income households show more residential instability after a disaster but are, on net, less likely to permanently leave high-risk areas than white, higher-income households — for financial and family-tie reasons (Population and Environment, 2025; PMC, unequal impact of disasters).

None of these studies find that pooling all FEMA declaration types and severities together (as a raw repeat-count) produces a large, uniform out-migration signal — which is exactly what our own county-panel analysis below confirms.

The recent, real signal: 2024-2025 Census/Redfin data

A Redfin analysis of Census Bureau domestic migration data (June 2026) finds the clearest evidence yet of disaster-driven flight, and it is very recent: high-flood-risk counties (top 10% of homes facing high flood risk, per First Street climate-risk scores) had a net outflow of 34,099 people in 2024 — the first net outflow from these counties in five years — which nearly doubled to a net outflow of 63,357 in 2025. Over the same period, low-flood-risk counties gained 69,857 more people than they lost, the largest gain since 2018. Miami-Dade County alone lost a net 72,254 residents in 2025, its largest net outflow on record, with Pinellas, Collier, and Monroe counties (all Florida) also among the top ten flood-prone counties losing residents. Redfin attributes this to a combination of repeated flooding/storm damage, soaring insurance costs, and other non-climate factors (general housing costs, politics) that are hard to fully separate from climate risk in this data alone.

This is an important vintage distinction: the Redfin/Census reversal is measured for 2024-2025, which is after the last year (2022) covered by the IRS SOI county migration data used in our own regression below. It is plausible the effect has strengthened and become newly visible in exactly the years our government migration data does not yet reach.

Our own analysis: FEMA declarations x IRS county migration, 2011-2022

Method. We built county-year measures from two askamerica tables: disasters.disaster_declarations (FEMA, all declaration types, one row per county per disaster) and fiscal.county_migration_flows (IRS SOI, county-to-county migrant counts derived from year-over-year address changes on tax returns, one row per origin-destination-year). For each county-year we computed: (1) a trailing-3-year distinct-disaster-declaration count (declarations in years Y-3 through Y-1, to precede rather than coincide with the migration measured in year Y), and (2) a net out-migration rate = (out-migrants minus in-migrants) x 1,000 / county population (from census.acs_population), so a positive value means net population loss via migration and a negative value means net gain. We restricted to counties with population > 1,000. This produced 31,885 county-year observations across 3,082 counties, 2011-2022 (the full window IRS SOI migration data covers in this corpus).

Naive (pooled) result. A single-predictor regression pooling all counties and years finds a coefficient of -0.165 (p=0.01) on the trailing disaster count — meaning, read literally, that MORE disasters associate with MORE net in-migration, the opposite of the expected direction. A leave-one-state-out sensitivity check (refitting with each state's counties removed one at a time) shows this result is not robust: excluding Florida drops the coefficient to -0.114 (p=0.08, no longer significant at 0.05), excluding Texas has a similar effect, and the coefficient's magnitude swings across states by more than 50% depending on which single state is dropped. This is the Sun Belt growth confound: Florida, Texas, and similar fast-growing states are simultaneously among the most disaster-prone (hurricanes, severe storms) and the strongest population magnets for unrelated reasons (jobs, no state income tax, retirees, lower cost of living historically) — so a raw cross-county comparison mixes up disaster exposure with regional growth and gets the sign backwards.

Fixed-effects result (the credible design). Using two-way (county + year) fixed effects — which compares each county's own migration rate to its own historical baseline, absorbing everything structurally different about Florida or any other state, plus any nationwide year shock — the coefficient on trailing 3-year disaster count falls to -0.012 (SE 0.116, t=-0.11, p=0.92), statistically indistinguishable from zero. In other words: when a given county goes through more FEMA-declared disasters than its own recent average, we cannot detect a corresponding shift in its net migration rate in this national, all-hazard, 2011-2022 dataset. A quantile dose-response check (binning counties into 6 groups by trailing disaster count, without fixed effects) shows a numerically negative trend (bin means run from -0.66 to -1.73 net-out per 1,000, trend p=0.011) but the direction is again "more disasters, more net inflow," consistent with the same unadjusted Sun Belt confound, and the pattern is not monotonic.

A direct check on long-run population change. Comparing total population change 2012-2022 (Census ACS) for counties bucketed by cumulative major/all disaster count 2011-2019: counties with 10+ disaster declarations grew by an average of 1.30% over the decade, moderately-hit counties (4-9) grew 0.30%, and low-exposure counties (0-3) grew 0.93% — again no sign of higher disaster exposure translating into net population loss over this window, reinforcing the fixed-effects null result.

Reconciling the two findings

The literature-documented effect (up to ~1.5 percentage points higher out-migration for the most severe, high-fatality disasters; hurricanes and floods showing the largest hazard-specific effects) is real but concentrated and small relative to the noise in an all-hazard, all-severity county panel — our blunt trailing-declaration-count measure pools minor emergency declarations together with catastrophic hurricanes, which likely dilutes any true effect below what a two-way fixed-effects model with 3,082 counties can detect as statistically distinguishable from zero. The 2024-2025 Redfin/Census reversal, by contrast, is a national, aggregate, very recent signal that postdates our migration data's 2022 cutoff — it is plausible that disaster-driven flight has genuinely strengthened in the last two to three years (rising insurance costs and repeated storm damage being frequently cited proximate causes) in a way not yet visible in the IRS SOI series available here. We would need 2023-2025 IRS county migration data (not yet published; IRS SOI publishes on roughly a 3-4 year lag) to test directly whether the recent reversal shows up as a genuine within-county fixed-effects signal rather than another cross-sectional artifact.