Exodus of the affluent? Examining climate hazards, migration, and household income in the U.S.

Abstract

With rising global temperatures come greater temperature and precipitation variability, contributing to more frequent and severe climate hazards that can upend lives and displace families. Lower-income households are often disproportionately impacted, so it is important to understand how climate hazards influence human migration patterns across income levels. There has been limited research on climate migration within the United States (US), particularly with respect to its economic impacts, like the associated transfer of household resources and incomes, or “income migration.” Here, we investigate spatial and temporal patterns of US domestic migration across income brackets between 2011 and 2021. We then investigate the role of climate hazards in shaping migration and income migration across US counties using panel data for the years 1995–2021. We found that lower-income households moved at higher rates overall but had less net migration across state lines, while higher-income households moved in a more directed fashion towards the most popular migration destinations. We also found an uptick in migration and income migration after the onset of the COVID-19 pandemic, particularly among higher income brackets. Property damage from climate hazards had small but significant relationships with migration. More destructive hurricanes were associated with reduced net migration and income migration nationally and in the South and Northeast. Flood damage was associated with reduced net income migration (greater outflow and/or reduced inflow of aggregate household income from migration) but had minimal effects on net migration overall, suggesting higher-income households (whose moves have a larger impact on net income migration) may be more likely to leave or avoid counties impacted by flooding. This work provides valuable new insights on the roles of both climate hazards and income levels in shaping domestic migration.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11111-025-00496-5.

Introduction

The United States (US) has already begun to experience the effects of climate change in the form of more intense heat waves, storms, and wildfires. Such climate-related natural hazards are only expected to become more frequent and severe (Marvel et al., 2023; Radeloff et al., 2018). Domestic patterns of human migration have also driven population growth and development in areas where climate change is increasing natural hazards, like the South’s hurricane-prone coasts and the West’s fire-prone Wildland-Urban Interface (WUI) (Radeloff et al., 2023; Winkler & Rouleau, 2020). Climate change and migration combine to put more people and infrastructure in harm’s way, increasing the potential for devastating outcomes.

The impacts of increasing climate-related hazards are not equally distributed but tend to disproportionately affect the most vulnerable, particularly low-income households. Hazard resilience requires the resources to adequately prepare before severe events, carry sufficient insurance, repair or rebuild after damage, or potentially move out of high-risk areas altogether (Cattaneo et al., 2019; Hunter, 2005). Such resources could include economic, political, and social resources, including things like social networks or language proficiency (Davies et al., 2018; DeWaard et al., 2023). We focus here on economic resources, as measured by household income.

Migration within the US is influenced by a wide range of factors, including social networks, economic opportunities, natural amenities, and the impacts of climate hazards (Clark et al., 2022; Winkler & Rouleau, 2020). While opportunities and amenities can act as “pull factors,” attracting potential migrants, climate hazards can sometimes act as “push factors,” deterring migrants or spurring current residents to leave, either due to direct displacement (e.g., destroying homes), or indirect effects such as impacts on insurance or quality of life (e.g., frequent wildfire smoke or nuisance flooding) (Clark et al., 2022; McConnel et al., 2024).

When households move, they take economic resources with them in the form of wealth, non-earnings income, and human capital, as well as their spending and labor, which contribute to the local tax base and economy (DeWaard et al., 2023; Plane, 1999; Vias & Collins, 2003). These changes have implications for the hazard resilience of both origin and destination communities. If relatively high-income households join or leave hazard-prone communities at disproportionate rates, it could increase or decrease the economic resources (e.g., tax base) available for things like hazard preparedness or response (DeWaard et al., 2023; Shumway et al., 2014). This means if households are displaced or choose to leave following a destructive hazard event, it could leave communities even more vulnerable to future events. Such post-hazard migration can be viewed as a potential “vector of economic losses,” on top of direct losses from property damage or business interruptions (DeWaard et al., 2023).

Here, we quantify this transfer of economic resources using rates of net “income migration,” or the net flow of aggregate household income associated with the migration of people between states or counties (Plane, 1999). Net income migration is calculated as the total household income of all households moving into a given county minus the total household income of all households leaving that county. We then scale it to give the percent change in that county’s total household income due to migration. This reflects both the net change in aggregate household income due to a net population change and any change due to the substitution of higher-income households for lower-income households, or vice versa (Plane, 1999).

There has been limited study of the relationships between climate hazards, migration, and that migration’s economic impacts in a US context. Shumway et al. (2014) looked at income migration and natural hazards, finding that the most hazard-impacted counties might be losing aggregate household income due to migration, and that high earners might be becoming concentrated in the least impacted counties. DeWaard et al. (2023) found increased economic losses from migration in the year of and following “disaster”-level hazard events. In contrast, Raker (2020) found that after severe tornadoes, affected neighborhoods became more socioeconomically advantaged. Fussell et al. (2010) also found that residents of lower socioeconomic status were slower to return to New Orleans after Hurricane Katrina, primarily due to their neighborhoods sustaining greater housing damage. These diverging results reflect two prominent hypotheses in the literature: the concentration hypothesis, where advantaged groups leave while disadvantaged groups are “stuck in place”, and the displacement hypothesis, where disadvantaged groups are displaced while advantaged groups rebuild (Raker, 2020). These outcomes may vary with the hazard type, geographic scale, and characteristics of the affected community (Ibid).

Outside of economic impacts, there is more literature on climate hazards and migration (e.g., Clark et al., 2022; Winkler & Rouleau, 2020), but it has tended to focus on single hazards like hurricanes (Fussell et al., 2017, 2022) or tornadoes (Raker, 2020), or to use a combined metric of all natural hazards (Elliott, 2015; Shumway et al., 2014). Findings also conflict as to whether these hazards generally result in net out-migration, in-migration, or no change. It may depend on the hazard type, degree of warning, spatial scale, degree of destruction, and the study timescale (Raker, 2020). Residents who evacuate may or may not eventually return. Meanwhile, newcomers may be attracted to employment opportunities associated with recovery efforts, and large-scale investment in rebuilding and recovery may even spur long-term growth (Elliott, 2015; Fussell et al., 2017). This highlights the need for further research on migration, income migration, and other economic impacts of migration in response to those hazards exacerbated by climate change, as well as the need to differentiate among hazards since their impacts vary.

We address these gaps by leveraging new county-level data analyses to examine whether exposure to various climate hazards affects domestic migration patterns and associated income migration. We investigated two questions. First, what are recent spatial and temporal patterns of migration and income migration across the US, overall and by income bracket? Given the timing of our study, we also examined how these patterns changed during the COVID-19 pandemic. Second, what are the socioeconomic and environmental drivers of migration and income migration, particularly the role of natural hazards affected by climate change?

We first investigated patterns of net migration and net income migration across US counties from 2011 through 2021, highlighting counties that gained or lost higher-income households due to migration, and examining how migration varied across income brackets. Second, we compiled a novel panel dataset of socioeconomic factors, natural amenities, and climate hazard damages, and used spatially and temporally explicit Dynamic Spatial Autoregressive Models (dynamic SARs) to investigate how these covariates influenced annual rates of net migration and net income migration across US counties during the years 1995–2021. This is the first study we are aware of to examine the relationships between income migration and multiple climate hazards, and the first to include annual data through the onset of the COVID-19 pandemic.

Methods

Our analysis had two goals. The first was to investigate spatiotemporal patterns of domestic migration and relationships to household income. This was accomplished through a thorough analysis of migration rates across income brackets, and a spatial analysis of county-level net migration rates for the contiguous US (CONUS) 2011–2021 (a shorter timeframe that avoids long-term averages and data corrections).

Our second goal was to investigate major drivers of county-level net migration and net income migration, particularly the role of climate-related natural hazard events. Migration is affected by dozens of different factors, not all of which are easy to measure. While data limitations meant we could not account for every possible driver, we used publicly available data to compile a novel panel dataset of historical climate hazard damages, natural amenities, and important socioeconomic characteristics for CONUS counties, annually for 1995–2020. These variables were used as covariates (Table 1) in a suite of spatially and temporally explicit dynamic SAR models for two dependent variables: rates of net people migration and of net income migration for migration intervals 1995–1996 through 2020–2021 (the migration data’s full period of record).

Table 1.

Model coefficients for Net People Migration models with current-year hazards

Variable	1. National	2. Northeast	3. Midwest	4. South	5. West
Ln(Flood Damage)‡	− 0.002	− 0.005	–0.003	− 0.003	0.031*
Ln(Hurricane Damage)‡	− 0.019**	− 0.094†	—	− 0.018*	—
Ln(Storm Damage)‡	0.003	− 0.039***	−0.003	0.013†	0.009
Ln(Tornado Damage)‡	0.009†	− 0.006	0.005	0.008	—
Ln(Wildfire Damage)‡	0.022	—	—	—	− 0.045
Warm Winters+	0.38***	(omitted)	(omitted)	0.255***	0.39*
Cold Winters+	− 0.416***	− 0.08	− 0.148***	(omitted)	− 0.729***
Humid Summers+	0.297***	0.04	− 0.144	0.227*	− 0.196
Non-Humid Summers+	0.025	(omitted)	− 0.219*	− 0.594***	0.189
Far from Water Bodies+	− 0.271***	− 0.079	− 0.264***	− 0.216***	− 0.262
Near to Water Bodies+	0.264***	0.098	0.21**	0.282***	0.353*
Rugged Terrain+	− 0.041	− 0.004	− 0.337***	− 0.182*	− 0.097
Flat Terrain+	− 0.179***	0.049	0.063	− 0.241***	− 0.463*
Forest Cover	0.955***	− 0.109	1.141***	0.59***	0.313
Forest Cover Squared	− 0.665***	0.035	− 0.612***	− 0.463***	− 0.024
Ln(Income per Capita)‡	0.342***	0.506***	0.349***	0.375***	0.137
Ln(Unemployment Rate)‡	− 0.271***	− 0.103†	− 0.081**	− 0.323***	− 0.58***
Ln(Population Density)‡	0.198**	− 0.102	0.351**	0.8***	− 0.065
Ln(Population Density) Squared	− 0.585***	− 0.476***	− 0.614***	− 1.05***	− 0.556***
Metro+	− 0.033†	—	—	—	—
Population 65 and Over	0.135***	0.066†	0.098***	0.071***	0.346***
Ln(Black, NH)	− 0.013	− 0.09†	− 0.035†	− 0.223***	0.058
Ln(Hispanic)	− 0.052***	0.089*	− 0.044*	− 0.079***	0.003
Ln(Other)	− 0.039***	− 0.193***	0.002	− 0.017	− 0.084†
Agriculture+	− 0.148***	(omitted)	− 0.124***	− 0.106***	− 0.278***
Mining+	− 0.024	0.115	− 0.073	− 0.008	− 0.118
Manufacturing+	− 0.071***	− 0.05	− 0.093***	− 0.055*	− 0.06
Government+	− 0.094***	− 0.095†	− 0.063	− 0.111***	− 0.149*
COVID-19 Onset+	0.92***	− 0.754	0.377**	0.674***	1.247***
IRS Methodology Change+	− 0.821**	2.195**	− 0.497***	− 0.87***	− 0.909†
Spatially lagged error ($\mathit{\eta }$)	− 0.388***	0.017	− 0.169***	− 0.427***	− 0.135*
σc (cross-sectional heterogeneity)	0.799***	0.408***	0.548***	0.797***	0.863***
σe (idiosyncratic variation)	0.874***	0.459***	0.751***	0.909***	1.099***
n years	24	24	24	24	24
n counties	2901	217	995	1324	365
Pseudo R2	0.11	0.22	0.14	0.20	0.11

^‡Each model included interaction terms for the indicated variables. Socioeconomic interaction coefficients for Model 1 are plotted in Figures S3, S5, and S7. Other interaction results are not shown

⁺Categorical variables

^†*p ≤ 0.1, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001

Data

Our analyses are based on 2020 county boundaries for the contiguous US (US Census Bureau, 2021a). To use a consistent set of boundaries throughout, counties that changed names had their name and FIPS code updated, and counties that underwent boundary changes affecting populated areas (14 counties) were excluded in the years before the change, following the example of Winkler and Rouleau (2020).

Since our dynamic SAR models required strongly balanced panel data, any counties with incomplete sets of migration observations (mainly sparsely populated counties with small migration flows) had to be excluded from our models (102 counties, including those with boundary changes). Our models also excluded 50 counties and independent cities in Virginia, where independent cities are treated as county equivalents, but form small islands within other counties. This atypical separation of urban centers from their surroundings could result in unrepresentative migration rates (Winkler & Rouleau, 2020).

Migration and household income

Domestic migration data and associated household incomes were from the Internal Revenue Service (IRS; IRS, 2023). Data are based on address changes between filers’ income tax returns in two consecutive years, from the period 1995 to 1996 through 2020 to 2021. Filers whose address changed county are considered movers, while those staying in the same county are non-movers. The total number of non-migrant, inflow, and outflow tax returns (the number of returns is used as a proxy for the number of households) are given annually for each state and county, along with the associated number of non-migrant, inflow, and outflow exemptions (a proxy for the number of people) and the aggregate household income of non-migrants, in-migrants, and out-migrants in the form of aggregate Adjusted Gross Income (AGI). In other words, the data include estimates for the total amount of people, households, and aggregate household income present in each county and year, and also estimates for the in- and outflows of people, households, and aggregate household income for each county between consecutive years. All AGI values were inflation adjusted to 2020 dollars. State and national migration data are also given by income bracket.

The inclusion of AGI data makes this dataset uniquely valuable and allows us to calculate not only the net migration rates of people and households, but also “Net Income Migration”, defined here as the percent change in a state or county population’s aggregate household income (as AGI) due to the net gain or loss of households and their associated incomes due to migration. Of course, this metric is not the literal movement of income since many moves are associated with a change in job and associated earnings; however, it can be interpreted as the movement of human capital and unearned income (Plane, 1999). Large changes to a county’s aggregate AGI due to migration may have serious implications for the long-term health of its economy and tax base (Vias & Collins, 2003).

One limitation of the IRS data is that it only includes households that filed taxes, consistently estimated at about 87% (Winkler & Rouleau, 2020). It therefore underrepresents low-income households, the elderly, university students, and workers paid under the table. It only captures moves across county lines, thus ignoring more local moves. It also only captures moves as address changes between subsequent tax filings, therefore missing temporary moves, multiple moves within a single year, and movers who continue to use a tax preparer in their old county. The data are therefore more likely to undercount than overcount moves.

The IRS changed methodology beginning with data for the migration interval 2011–2012 in order to capture all returns filed in a calendar year rather than those filed by September, as it had before. This captured approximately 4.7% more returns, particularly higher-income households, which are more likely to file for extensions (Pierce, 2015). While this change increased the captured counts of in- and out-migrants, it also increased the total captured population, so it should have minimal effects on annual migration rates, particularly for net migration, where increases in in- and out-migration cancel out. We therefore focus our analysis on net migration rates and include a dummy variable in our dynamic SAR models to flag this change beginning in 2011–2012.

Data for 2014–2015 and 2016–2017 were anomalous (i.e., DeWaard et al., 2022), and this appears to be due to a known issue in the IRS’ internal data pipeline, which affected those two years, but left other years unaffected (K.K. Pierce, IRS, personal communication, November 17, 2023). Those two years were therefore excluded from our analyses (using imputed data for those years had little impact on the results; Tables S4-S5). Despite these limitations, the IRS migration dataset remains one of the most comprehensive sources of annual, county-level migration data for the US, and one of the only datasets to include information on movers’ household incomes.

We used the IRS data to conduct two sets of analyses. In the first set, we analyzed spatiotemporal patterns of migration across income brackets for the 2011–2021 period after the IRS methodology change. Here, we investigated relationships between income and migration at the national level, examining the percent of households in each income bracket that moved (Fig. 1). We also investigated domestic net migration rates of people, households, and AGI at the state and county level. Net migration rates were calculated as the percent change in each county’s (or state’s) population (or its total number of households, or total AGI) between two consecutive years, as follows:

$$\text{Net Migration Rate}=100\times \frac{\left(\text{inflow},-,\text{outflow}\right)}{\text{nonmigrant}+\text{outflow}}$$

Fig. 1.

Percentage of households in each income bracket that moved across county lines, showing a means for 2011–2021 with the breakdown of in-state vs cross-state moves, and b annual values for 2011–2021, showing a general uptick after the onset of the COVID-19 pandemic in 2020, particularly among the highest income brackets. Data are plotted by the second year in each two-year window (e.g., 2011–2012 moves are shown as 2012)

Domestic net migration rates of households and AGI were mapped at the county level (Figs. 2, 3). At the state level, net migration rates were also calculated and mapped separately for each income bracket (Fig. 4).

Fig. 2.

Mean annual net migration of households before (2011–2019, top) and after (2020–2021, bottom) the onset of the COVID-19 pandemic

Fig. 3.

Mean annual net migration of income (AGI) before (2011–2019, top) and after (2020–2021, bottom) the onset of the COVID-19 pandemic

Fig. 4.

Net migration rates of individuals in each income bracket for US states before (2011–2019, left) and after (2020–2021, right) the onset of the COVID-19 pandemic

In the second set of analyses, net migration rates of people and AGI (“Net People Migration” and “Net Income Migration”) were used as the dependent variables in two annual county-level dynamic SAR models for 1995–2021, the longest possible timeframe. To combat high kurtosis, we identified extreme outliers in each dependent variable, which we defined as values lying more than 6 interquartile distances from that variable’s interquartile range. Because the dynamic SAR requires a strongly balanced panel with no missing observations, we then dropped counties that had any extreme outliers, determined separately for each variable. This dropped 1.7% of observations for Net People Migration and 5.4% for Net Income Migration, including some counties that had large pulses of net out-migration associated with extreme events like Hurricane Katrina. This means our models may tend to underestimate relationships between climate hazards and migration.

Natural hazards

We represented the impacts of climate-related natural hazards using historical data on five hazards: flooding, hurricanes, storms, tornadoes, and wildfires. These were chosen for their widespread and destructive impacts, which might directly displace people or deter them through impacts to quality of life. Hazard data were annual, county-level property damage per capita values in 2020 dollars from SHELDUS, the Spatial Hazard Events and Losses Database for the United States (ASU CEMHS, 2022). SHELDUS splits damages from multi-hazard events equally across involved hazards, so our five hazard variables are mutually exclusive. Our storms variable is the sum of damages from four SHELDUS hazards which we found to be closely correlated: Severe Storms/Thunderstorms, Wind, Hail, and Lightning. We did not include heat because, while it is one of the deadliest weather-related hazards in US, it is more likely to cause hospitalizations and deaths than property damage.

SHELDUS is a loss and damage database that details specific hazard events that caused direct losses such as fatalities, injuries, or property damage. It covers 18 hazards, 1960–present. It only includes events that caused recorded losses, so it is not representative of underlying hazard frequencies or probabilities, only their historical impact on existing communities and infrastructure. Recorded property damage will be higher where there is more property to damage, or higher-value property. This means property damage will not always have a one-to-one relationship with impact; if a low-income family loses their mobile home to flooding, it could have a similar dollar value to a large house having its windows blown out, but a very different impact on the affected family. SHELDUS’ loss estimates are highly conservative, using the low end when damages are estimated as a range, and averaging state damages across counties when county-level estimates are not available. Despite these limitations, it remains one of the most comprehensive datasets on annual, county-level hazard events for the US.

Cumulative (1991–2020) county-level property damage per capita is mapped for each of our five hazards in Fig. 5. While we pursued a county-level analysis due to data limitations, it is important to note that county-level data can mask substantial sub-county variation in hazard impacts, such as between low-lying and elevated areas, across the rural–urban continuum, and between low- and high-income communities. Hazard events are rare, and damages can vary by several orders of magnitude, resulting in highly skewed distributions. Damages were therefore natural-log transformed for use in our dynamic SAR models, using the formula ln(x + 10e-4).

Fig. 5.

Maps of cumulative property damage per capita (1991–2020) in 2020 dollars from five different hazards

Natural amenities

Desirable environmental factors like a pleasant climate or scenic landscape — known as natural amenities — can play an important role in migration, attracting new residents and encouraging existing ones to stay (McGranahan 1999, 2008; Clark et al., 2022). We represented natural amenities using a suite of five county-level variables, including two climate variables (thirty-year means for winter temperature and summer humidity), and three landscape variables (mean distance to water, mean terrain ruggedness, and percent forest cover and its square) (Clark et al., 2022). Since landscape features are relatively stable and our climate variables change only gradually, we used fixed values for these variables over time, except for forest cover and its square, which we calculated annually.

We prepared the climate variables and distance to water according to Clark et al. (2022) using data from Abatzoglou (2012) and Pekel et al. (2016) but used updated terrain and forest data. We represented terrain using the mean Terrain Ruggedness Index (TRI), which summarizes the difference in elevation between a given grid cell and its neighbors. We calculated the mean TRI for each county (Figure S1) using a 1 km resolution TRI raster (Amatulli et al., 2018). The resulting county means were natural log transformed to reduce skewness.

Since forest cover changes over time (e.g., deforestation, harvest, regrowth), we calculated counties’ percent forest cover annually for 1995–2020 using land cover data from the US Geological Survey’s (USGS) LCMAP collection (2022; 30 m grid cells). We calculated forest cover as the percent of each county’s area in LCMAP’s Primary Landcover “Tree Cover” class. We also included forest cover squared, as studies have found a quadratic relationship with migration, indicating people are drawn to intermediate levels of forest cover (Clark et al., 2022; McGranahan, 2008).

Each natural amenity variable was standardized into a z-score using its mean and standard deviation to make the model coefficients in the subsequent SAR models more directly comparable. Natural amenity variables that were fixed over time (climate variables, distance to water, and TRI) caused multicollinearity issues in our spatiotemporal models, so they were converted to categorical variables based on their Z-scores, each with three categories: low (< − 1), medium (− 1 to 1), and high (> 1). High and low categories were included in our dynamic SAR models with reference to the medium category.

Socioeconomic variables

Socioeconomic factors can be some of the most important drivers of migration. Jobs, wages, affordability, population density, and access to social services have all been linked to migration, as have harder-to-measure factors like social networks, culture, and family ties (Clark et al., 2022; Czaika & Reinprecht, 2020; McGranahan, 2008; Roback, 1982). Migration patterns can also vary across demographics like age, race, and ethnicity (Johnson et al., 2013; Nelson et al., 2009; Winkler & Johnson, 2017). While annual, county-level data were not available for all potentially relevant socioeconomic variables, we included as many as possible while minimizing collinearity.

Our final socioeconomic variables included income per capita, unemployment rates, population density and its square, a flag for metropolitan (metro) vs nonmetropolitan counties, the population 65 and over (%), populations of four race and ethnicity groups (%), and an economic typology. We also included a flag to control for the onset of the COVID-19 pandemic in 2020, and a flag from 2011 onwards to control for the IRS methodology change. Data sources are summarized in Table S1. While our analysis was conducted at the county level, it is worth noting that socioeconomic characteristics can vary widely within a given county, particularly across the rural–urban continuum, and that county values may be weighted towards their major population centers.

We represented employment opportunities using income per capita (adjusted to 2020 dollars), and unemployment rates, expecting movers to be drawn to areas with higher incomes and lower unemployment (Clark et al., 2022; McGranahan, 2008). These were natural log transformed to reduce skewness, as was population density. The latter has been found to have a quadratic relationship with migration, suggesting people prefer intermediate population densities, so we included both the logged values and their squares (Ibid). We also include a flag for metro counties since proximity to metro areas can be important for access to jobs and other amenities, and population density varies widely across (and within) US metro areas.

Given the importance of retirement in shaping domestic migration patterns (Johnson et al., 2013; Nelson et al., 2009), we included the population 65 and over as a percent of each county’s population. Since migration patterns can also vary with race and ethnicity (Beale & Fuguitt, 2011; Nelson et al., 2009; Winkler & Johnson, 2017), with co-ethnic communities sometimes acting as a migration “pull” factor (Elliott, 2015), we also considered the percent of the population in each of four non-overlapping race and ethnicity groups: non-Hispanic (NH) white, NH Black, Hispanic (All Races), and other (calculated as the total population minus the other three groups). We used these four broad categories to improve comparability across decades, since Census data include six racial groups in the 2000 s but only four in the 1990s. Since these categories sum to 100%, we only included the three minority groups in our SAR models. Since minority populations tend to be rare, their distributions were highly skewed, and so were natural log transformed using ln(x + 0.01).

The economic typology was included to control for major employment sectors. It includes five mutually exclusive classes based on counties’ primary sectors — farming, mining, manufacturing, government, or non-specialized. Other classes that were not comparable over time were combined with the non-specialized class.

All continuous socioeconomic variables were standardized as z-scores using their mean and standard deviation in order to make the magnitudes of their model coefficients more directly comparable: coefficients can be interpreted as the increase in y associated with a 1 standard deviation (SD) increase in x.

Dynamic SAR models

We used county-level Spatial Autoregressive Models for panel data (dynamic SARs) with random effects to examine how socioeconomic and environmental factors— including annual property damages from climate-related natural hazards — influence Net People Migration and Net Income Migration. Dynamic SAR models are panel regression models that explicitly account for spatial dependence between observations (Kelejian & Prucha, 2001). This is important because many of our variables exhibit spatial autocorrelation: two counties that are closer together are more likely to have similar values when it comes to factors like migration rates, winter temperatures, or hurricane damage.

Our dynamic SAR model is given by:

$$y_{\mathit{nt}}=x_{\mathit{nt}}\beta +{\alpha (x}_t\ast t)+u_{\mathit{nt}}$$

$$u_{\mathit{nt}}=\eta M{u}_{\mathit{nt}}+{c_n+v}_{\mathit{nt}}$$

where: y_nt = vector of dependent variables (n × 1) for time t and n number of panels.

$X_{\mathit{nt}}$ = covariates with n panels in time t

$\beta$ = vector of coefficients of the covariates

$\alpha$ =interaction term coefficient for the interaction between covariate x_t and time t

$u_{\mathit{nt}}$ = spatially lagged error term.

$n$ = spatial error coefficient

$M$ = contiguity spatial weight matrices for the lagged dependent variable and error terms, respectively

$c_n$ = vector of panel effects for each panel n. Panel effects are normal and are independent and identically distributed (i.i.d.) with a variance of ${\sigma }_c^2$.

$v_{\mathit{nt}}$ = vector of error terms for panel n. The idiosyncratic error terms are i.i.d. across panels and time with variance ${{\sigma }_v}^2$.

Our spatial weight matrix, M, was a contiguity matrix weighting first-order neighbors. It assigned positive weight to neighboring counties and a 0 weight to all other counties. Terms $c_n$ and $u_{\mathit{nt}}$ account for panel-specific effects and time-specific effects and control for unobserved heterogeneity. This approach ensures that the model accounts for panel-specific traits that do not vary over time, and for time-specific shocks affecting all panels, such as the Great Recession and the COVID-19 pandemic. (For a more detailed discussion of dynamic SARs and our modeling choices, see the Supplementary Information.)

One of the strengths of the dynamic SAR is its ability to determine the trend of the non-stationary impact of a covariate on the dependent variable over time. Covariates may not have constant impacts over time due to structural changes, evolving spatial dependencies, or external shocks. Each trend is estimated through an interaction term of the time and a covariate — ${\alpha (x}_t\ast t)$. The interaction coefficient, α, can vary systematically with time, reflecting the change of the effect of x_t over time. We included interaction terms in each model for the five hazard variables, income per capita, unemployment rates, and population density. The inclusion of these interaction terms means that the “main” coefficient for each of the interacted variables indicates the initial relationship at the beginning of the study period, while the annual interaction coefficients indicate how the relationship deviates from that “main” coefficient in each year. Time (year) was also included as a linear covariate since it captures general time trends in the data that affect all units similarly.

Our study was concerned with the drivers of migration and roles of climate hazards nationally; however, not all climate hazards affect the entire country. We therefore included both national and regional models to investigate whether hazards that had minimal effects at the national level might have stronger effects in the most affected regions (such as wildfires in the West or hurricanes in the South and Northeast). For each dependent variable, we ran one “National” model using all available CONUS counties and a model for each of four Census regions: the Northeast, Midwest, South, and West. This resulted in ten annual, county-level dynamic SAR models: five for Net People Migration (Models 1–5 in Table 1) and five for Net Income Migration (Models 6–10 in Table 2). The final number of counties and years for each model are indicated in Tables 1 and 2.

Table 2.

Model coefficients Net Income Migration models with current-year hazards

Variable	6. National	7. Northeast	8. Midwest	9. South	10. West
Ln(Flood Damage)‡	− 0.014*	− 0.022	− 0.015†	− 0.018†	0.039†
Ln(Hurricane Damage)‡	− 0.021*	− 0.216*	—	− 0.018†	—
Ln(Storm Damage)‡	0	− 0.041*	− 0.008	0.017†	−0.003
Ln(Tornado Damage)‡	0.008	− 0.05	0.01	0.003	—
Ln(Wildfire Damage)‡	0.015	—	—	—	− 0.027
Warm Winters+	0.467***	(omitted)	(omitted)	0.337***	0.246
Cold Winters+	− 0.269***	0.185	− 0.114†	(omitted)	− 0.098
Humid Summers+	0.604***	0.286	− 0.266	0.47***	0.373
Non-Humid Summers+	0.152†	(omitted)	− 0.198†	− 0.993***	0.759**
Far from Water Bodies+	− 0.364***	−0.094	− 0.328***	− 0.313***	− 0.3
Near to Water Bodies+	0.759***	0.354***	0.79***	0.79***	0.576*
Rugged Terrain+	0.048	−0.049	− 0.528***	− 0.205†	0.206
Flat Terrain+	− 0.153*	0.315†	0.008	− 0.245*	− 0.645*
Forest Cover	1.457***	−0.266	1.611***	0.75***	0.915**
Forest Cover Squared	− 0.968***	0.294	− 0.805***	− 0.505***	− 0.247
Ln(Income per Capita)‡	0.43***	0.847***	0.451***	0.536***	0.072
Ln(Unemployment Rate)‡	− 0.236***	−0.167	− 0.001	− 0.288***	− 0.519***
Ln(Population Density)‡	0.358***	0.054	0.56***	1.046***	0.085
Ln(Population Density) Squared	− 0.921***	−0.874***	− 1.039***	− 1.494***	− 0.7**
Metro+	− 0.018	—	—	—	—
Population 65 and Over	0.215***	0.35***	0.099***	0.197***	0.506***
Ln(Black, NH)	0.014	− 0.071	− 0.033	− 0.23***	0.232**
Ln(Hispanic)	0.093***	0.238***	0.049†	0.075**	0.205*
Ln(Other)	− 0.05**	− 0.22**	0.011	− 0.049†	− 0.164*
Agriculture+	− 0.171***	(omitted)	− 0.126**	− 0.188***	− 0.262*
Mining+	− 0.176***	0.083	− 0.39***	− 0.078	− 0.366**
Manufacturing+	− 0.06**	− 0.021	− 0.063*	− 0.072*	− 0.239
Government+	− 0.144***	− 0.254**	− 0.107†	− 0.147***	− 0.255**
COVID-19 Onset+	1.115***	− 11.71***	0.298†	0.436*	1.462**
IRS Methodology Change+	− 0.691†	14.644***	− 0.499***	− 0.871***	− 0.783
Spatially lagged error ($\mathit{\eta }$)	− 0.235***	− 0.002	− 0.223***	− 0.323***	− 0.185***
σc (cross-sectional heterogeneity)	1.057***	0.591***	0.737***	1.061***	1.145***
σe (idiosyncratic variation)	1.141***	0.829***	0.971***	1.176***	1.46***
n years	24	24	24	24	24
n counties	2787	214	972	1270	331
Pseudo R2	0.14	0.18	0.17	0.19	0.19

^‡Each model included interaction terms for the indicated variables. Socioeconomic interaction coefficients for Model 6 are plotted in Figures S4, S6, and S8. Other interaction results are not shown

⁺Categorical variables

^†p ≤ 0.1, *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001

In Models 1–10 (Tables 1, 2), observations of Net People and Net Income Migration — which were for consecutive year pairs (e.g., 2020–2021) — were matched with covariate observations from the earlier year (i.e., 2020). Our goal was to capture any potential pulse of migration occurring in the months directly following a hazard event, even if this meant capturing a certain amount of baseline migration throughout the year. (For a discussion of the data’s temporal structure, see the Supplementary Information. See Figure S2 for a conceptual diagram.)

As an additional sensitivity analysis, we ran an additional ten models using hazards that had been lagged by an additional year (i.e., migration for 2020–2021 matched with flood damage for 2019). In these models (Models 11–15 for Net People Migration in Table S2, and Models 16 − 20 for Net Income Migration in Table S3), the goal was to capture relationships with moves that either did not occur right away (e.g., caused by indirect and/or cumulative hazard impacts), or that started out as temporary moves and were only documented in the IRS data after an additional year. We will refer to the first set (Models 1–10 in Tables 1 and 2) as models with “current-year” hazards, and the second set (Models 11–20 in Tables S2 and S3) as models with lagged hazards.

Results

Migration patterns and household income

Our analysis of recent IRS migration data by income bracket for 2011 through 2021 showed that, nationally, households in lower income brackets were more likely to move, with an average of 9.2% of households in the lowest income bracket moving each year, compared to just 4.8% in the highest bracket (Fig. 1a). Overall, households that moved were about as likely to move in-state as cross-state, but across income brackets, the highest income bracket was slightly more likely to move cross-state, while lower income brackets were slightly more likely to stay in-state (55% of moves in the lowest bracket were in-state vs 46% in the highest income bracket) (Fig. 1a).

Moving rates were fairly steady over time across income brackets but showed a sudden uptick after the onset of the COVID-19 pandemic in 2020, particularly among higher income brackets (Fig. 1b). In fact, moving rates among the highest income bracket were 29% higher in 2021 than before the pandemic.

We mapped county-level net migration rates of households (Fig. 2) and their aggregate household income (AGI, Fig. 3) to investigate spatial migration patterns across the country and see how patterns differed before and after the COVID-19 pandemic. In general, before the onset of the pandemic (Figs. 2, 3, top), counties in the Northeast, Midwest, areas along the Mississippi River, the Great Plains, and California tended to see net losses of households and aggregate household income (in purple) while parts of the Southeast, West, Florida, and Texas tended to see net gains (in green). It is important to note that we consider domestic migration only. International migration can be a significant contributor to population growth in some locations, meaning net domestic migration might be negative even where total net migration is positive (Golding & Winkler, 2020).

While net migration rates of households and aggregate household income exhibited similar spatial patterns, these patterns tended to be stronger and more pronounced for the movement of income (Fig. 3) than for that of households (Fig. 2).

Comparing spatial migration patterns before (Figs. 2, 3, top) and after (Figs. 2, 3, bottom) the onset of the pandemic, many existing migration patterns appear to grow stronger (deeper shades of green and purple), but new patterns also emerge. The spatial footprint of attractive areas appears to expand, growing to include more counties across the West, South, Texas, and Florida. For example, before the pandemic, three “donut cities” can be seen in Texas around Austin, Dallas, and Houston, with net out-migration from the urban centers and in-migration to the surrounding counties, whereas after the onset of the pandemic, these “donuts” merged into a larger region experiencing net in-migration around the three urban “donut holes” experiencing out-migration. We also see new regions attracting broader, stronger net in-migration, such as the Ozarks, northern Great Lakes, northern New England, and counties just north of New York City. While more counties saw net in-migration (green), this was balanced out by stronger net out-migration (purple) from urban centers with small geographic footprints such as New York City, San Francisco (Bay Area), Boston, and Washington DC.

We also mapped state-level net migration rates of individuals by income bracket (Fig. 4) and found that net migration rates tended to be stronger (more positive or more negative) among higher income brackets, in contrast to the higher total migration rates seen for lower income brackets (Fig. 1).

Drivers of migration and income migration

Results of our annual, county-level dynamic SAR models with current-year climate hazards are reported for Net People Migration (Models 1–5) in Table 1 and for Net Income Migration (Models 6–10) in Table 2. Results for models with lagged climate hazards are reported for Net People Migration (Models 11–15) in Table S2 and for Net Income Migration (Models 16 − 20) in Table S3. Outside of the natural hazard results, we focus our discussion on the current-year models (Models 1 − 10), since the current-year and lagged models had very similar results for non-hazard variables (apart from some differences for the COVID-19 Onset and IRS Methodology Change dummy variables). All twenty models included interaction terms for three socioeconomic variables and the five climate hazards, but we focus here on the interaction results for our primary models: national models with current-year hazards (Models 1 and 6; Figures S3–S8). Climate hazard interaction coefficients were not significant in most years, suggesting that the “main” hazard coefficients are largely representative of the entire study period.

Our five climate hazard variables all shared common units (annual, county-level property damage per capita in 2020 dollars) and underwent the same natural log transformation, meaning we can make direct comparisons among the relative magnitudes of their coefficients, but not between hazard and non-hazard coefficients. For these log-transformed hazard variables, results are best interpreted by considering the effect of a 10% increase in hazard damages, found by multiplying the reported coefficient by ln(1.10) or approximately 0.1. For example, at the national level, the effect of a 10% increase in annual hurricane damage per capita on Net People Migration is given by − 0.019*ln(1.10), which gives − 0.002, interpreted as a − 0.002% reduction in annual Net People Migration, equivalent to 2 fewer net migrants per 100,000 residents (Table 1).

All other continuous covariates were converted into z-scores to make them more directly comparable. Their coefficients can be interpreted as the change in the dependent variable associated with a 1 standard deviation (SD) change in the covariate. For example, a 1 SD increase in a county’s forest cover was associated with a 0.955% increase in county-level Net People Migration (Table 1), equivalent to about 955 more net migrants per 100,000 residents.

Some categorical variables were omitted from certain regional models where there was insufficient representation to include them, such as a lack of counties with “warm winters” in the Northeast and Midwest, a lack of “cold winters” in the South, and a lack of “non-humid summers” or primarily agriculture-dependent counties in the Northeast (Tables 1, 2).

Natural hazards

In national models with current-year climate hazards (Models 1–10), hurricane damage was the only hazard to have a significant negative relationship with Net People Migration (Table 1, p < 0.05), although there was a marginally significant positive relationship with tornado damage (p < 0.1). Hurricane and flood damage both had significant negative relationships with Net Income Migration (Table 2). In regional models, flood, hurricane, and storm damage had significant relationships with both Net People and Net Income Migration in certain regions, but wildfire damage had no significant effects.

While our analysis was conducted at the county level, it is important to remember that there can be wide variation in hazard damages and exposure within a given county, potentially resulting in some degree of local migration to the least impacted areas within counties. Our study can only capture cross-county migration. Spatial distributions of county-level hazard damages are shown in Fig. 5. While flooding and storms affect most of the country, tornadoes mostly affect the Midwest and South. Hurricane damage is concentrated along the Atlantic and Gulf coasts, and wildfire damage is concentrated in the West and Florida.

Hurricane damage had small but significant negative relationships with both Net People Migration (Table 1) and Net Income Migration (Table 2) in the current-year models. In both cases, there was a coefficient of about − 0.02 for the National and Southern models, meaning a 10% increase in hurricane damage was associated with a net loss of about 2 migrants per 100,000 county residents, and a net loss of about − 0.002% of counties’ aggregate household income. These relationships were significant despite some counties affected by Hurricane Katrina being excluded as migration outliers, meaning our results may underestimate the influence of hurricane damage nationally and in the South.

In the current-year Northeast models, hurricane damage had a coefficient of − 0.094 for Net People Migration and − 0.216 for Net Income Migration, meaning a 10% increase in hurricane damage was associated with a net loss of − 0.009% or 9 migrants per county 100,000 residents, and a net loss of − 0.02% of counties’ aggregate household income. This relationship was more than twice as strong for Net Income Migration as for Net People Migration, meaning equivalent levels of hurricane damage were associated with a disproportionately large percent loss of aggregate household income due to migration relative to the associated percent loss in population. For a county to lose aggregate household income at a faster rate than population, it must be losing households with relatively high incomes: the average household income of out-migrants must be higher than that of in-migrants. These results suggest that higher-income households may be disproportionately likely to leave or avoid counties affected by hurricane damage, at least in the Northeast.

Nationally, current-year flood damage had no significant effect on Net People Migration (Model 1, Table 1) but did have a small negative relationship − 0.014) with Net Income Migration (Model 6, Table 2), equivalent to a net loss of about − 0.001% of counties’ aggregate household income. As with hurricanes in the Northeast, this stronger relationship with Net Income Migration compared to Net People Migration suggests that higher-income households might be more likely to leave or avoid counties affected by flooding. Regionally, current-year relationships with Net Income Migration were marginally significant in the Midwest (− 0.015) and South (− 0.018), where they were very similar to the national effect. Intriguingly, current-year flooding’s only significant relationship with Net People Migration was in the Western model, where there was a small positive relationship (0.031, Table 1), equivalent to about 3 additional net migrants per 100,000 county residents.

Current-year storm damage had no significant relationship with either Net People or Net Income Migration in the national models, perhaps because it is so ubiquitous and tends to do less property damage than other hazards (Fig. 5). Interestingly, current-year storm damage had significant negative relationships with Net People (− 0.039) and Net Income Migration (− 0.041) in the Northeast, but marginally significant positive relationships with those outcomes in the South (0.013, 0.017).

Tornado damage had only one small, marginally significant (p < 0.1) positive relationship with Net People Migration in the current-year national model (0.009, Table 1), equivalent to just under one additional net migrant per 100,000 county residents.

Wildfire damage had no significant relationships in any of our dynamic SAR models.

Coefficients for the five hazard interaction terms with the year were not significant in most years (not shown), suggesting that relationships with these hazards did not change much over time. In years when these coefficients were significant, they pointed to weaker relationships with hazards, whether those relationships were positive or negative to begin with.

Compared to our primary models with current-year hazards (Models 1–10, Tables 1, 2), models with lagged hazards (Models 11–20, Tables S2, S3) had weaker results. Relationships with lagged hurricane damage were weaker and not significant, suggesting that most of the migration response to destructive hurricanes occurs in the short term. Relationships with lagged flood damage were insignificant nationally, but significantly negative in the Northeast and Midwest for both Net People and Net Income Migration, which was not the case in the corresponding current-year models. This suggests a more delayed migration response in these cases, or a delay in documenting moves as permanent. Meanwhile, lagged storm damage had significant positive relationships with Net People and Net Income Migration, both nationally and in the South. Lagged tornado and wildfire damage were not significant.

Natural amenities

Natural amenities generally had the expected relationships with Net People and Net Income Migration, with some exceptions. We found that people and aggregate household income were more likely to move to counties that had warmer winters, were closer to water bodies, and had intermediate levels of forest cover (Tables 1, 2), consistent with previous studies (Clark et al., 2022; McGranahan, 2008). In contrast to those studies, we found people (and income) were more likely to move to counties with higher summer humidity, a relationship driven by high migration to the South. Interestingly, we also found that people (and income) seemed to be drawn to intermediate levels of terrain ruggedness. Previous studies had assumed a linear relationship with terrain/topography (Clark et al., 2022; McGranahan, 2008), but our results suggest there may be a quadratic one.

Forest cover also had a quadratic relationship with migration across models, indicated by positive coefficients for forest cover and negative coefficients for its square. Interestingly, forest cover had the strongest effects of any covariate at the national level: a 1 SD increase was associated with a nearly 1% increase in Net People Migration and a 1.5% increase in Net Income Migration. Relationships with distance to water and forest cover also seemed to be stronger (more positive or negative) for Net Income Migration than for Net People Migration, suggesting higher-income households can afford to put more weight on these factors when choosing where to move, and can perhaps afford proximity to these amenities even where lower-income households have been priced out.

Socioeconomic variables

Our socioeconomic variables generally had the expected relationships with Net People and Net Income Migration. As in previous studies (Clark et al., 2022; McGranahan, 2008), people (and aggregate household income) were more likely to move to counties with economic opportunities (higher-income per capita, lower unemployment rates), and intermediate population density (indicated by a positive relationship with population density and a negative relationship with its square), as well as moving to counties with a larger population 65 and over (Tables 1, 2). Our metro variable was only marginally significant (p < 0.1), and only for Net People Migration, perhaps due to shifting preferences for metro vs. nonmetro counties over our long study period.

Annual coefficients for the socioeconomic interaction terms were consistently significant (p < 0.05) across years in both national models (Figures S3-8). They pointed to the positive relationships with income per capita (Tables 1, 2) that weakened over time (negative trend in a positive relationship), particularly during the first year of the pandemic (Figures S3-4). This might suggest that moving to urban centers with high wages became less important with the rise of remote work during the pandemic. Our results also pointed to the negative relationships with unemployment rates (Tables 1, 2) strengthening over time (negative trend in a negative relationship), particularly during the Great Recession (Figures S5-6, 2008–2011), suggesting that finding work became a more important factor during this period, even though migration rates overall were suppressed during the Recession (Johnson et al., 2017). Finally, positive relationships with population density strengthened over time (positive trend in a positive relationship), although less so during the pandemic (Figures S7-8).

Nationally, relative to non-Hispanic (NH) white populations, higher NH Black populations had no effects and higher “other” populations had small negative coefficients, perhaps reflecting out-migration from some diverse urban counties and/or fewer economic opportunities in American Indian reservations and other historically marginalized communities. Interestingly, larger Hispanic populations were associated with reduced Net People Migration but increased Net Income Migration — the only covariate to have such diverging relationships.

Relative to Non-Specialized counties, other economic typology classes had negative relationships, when significant. These categories were designed for rural counties, so it is not surprising that they are no longer preferred given long-standing out-migration from rural areas and the general transition towards a service economy (Johnson & Lichter, 2019; Moeller, 2020).

Our COVID-19 onset flag highlighted the much higher rates of Net People and Income Migration in most counties after the onset of the pandemic nationally, in the Midwest, South, and especially in the West, but not in the Northeast, where there was exceptionally strong net out-migration of aggregate household income. This indicates disproportionate out-migration of higher-income households from counties in the Northeast early in the pandemic, presumably from cities along the Northeast seaboard.

Discussion

Migration patterns and household income

In our analysis of migration by income bracket, we found that households in lower income brackets were more likely to move than those in higher income brackets (Fig. 1a). It is possible that this trend could be partially explained by age dynamics, since young adults early in their careers tend to have lower incomes and move more frequently than older adults (Plane, 1999; Shumway et al., 2014). Adults aged ~ 20–30 move at much higher rates than the rest of the population as they leave the family home, pursue educational or job opportunities, and set up their own households (Plane, 1999; Rogers, 1990). However, the IRS migration data do not include a breakdown by age at the county level, so we did not analyze the role of age or life stage in driving this pattern. It is possible that relationships between income and housing tenure could also contribute to this trend: if lower-income households are more likely to rent, they might be more likely to move, while homeowners might be more settled in place. However, we did not examine relationships with housing tenure.

We also found that, relative to lower income brackets, households in the highest income bracket were slightly more likely to move cross-state (Fig. 1a). One possible explanation is if long-distance moves are more expensive than local moves, it is also possible that higher-income professionals may be more likely to move long distances for career opportunities than blue-collar or service workers.

While moving rates were fairly stable before the COVID-19 pandemic, they showed a sudden uptick after its onset, especially among households in the highest income brackets (Fig. 1b). One possible explanation is if pandemic-related job losses and the rise of remote work spurred people to re-evaluate where they wanted to live and work, contributing to increased migration. Many workers lost their jobs early in the pandemic, and while many benefited from expanded unemployment benefits, it is possible this contributed to more moves, either for new job opportunities or to save money by moving in with family. More remote work — particularly among higher-income white-collar workers — may have also given many workers more flexibility to move down the urban hierarchy or even across the country in search of more space and access to nature.

We found similar spatial patterns between our maps of county-level net migration rates of households (Fig. 2) and aggregate household income (Fig. 3) but found that migration rates were stronger (more positive or negative) for the movement of income than for households. If counties are gaining and losing disproportionate amounts of aggregate household income relative to their net gains or losses of people, it could have the effect of concentrating higher earners in certain areas over time, while exacerbating the economic effects of population decline in others (Shumway et al., 2014; Vias & Collins, 2003). Spatial migration patterns also grew stronger (more positive or negative) after the onset of the pandemic (Figs. 2, 3), and the spatial footprints of attractive regions expanded.

Finally, in mapping state-level migration by income bracket (Fig. 4), we found stronger (more positive or negative) net migration rates among higher income brackets. Combined with our finding that households in lower income brackets were more likely to move overall (Fig. 1), this suggests that lower income brackets may have more “churn” with in- and out-migration cancelling each other out and yielding less net migration. Meanwhile, households in higher income brackets might be moving in a more directed fashion (e.g., towards desirable areas), yielding more net migration.

Natural hazards as drivers of migration and income migration

Our county-level dynamic SAR models with current-year hazards found that, nationally, hurricane damage had significant negative relationships with both Net People and Net Income Migration. In contrast, flood damage had a significant negative relationship with Net Income Migration nationally, but no significant relationship with Net People Migration. In national models with lagged hazards, relationships with hurricane and flood damage were not significant, but there were significant positive relationships with storm damage.

Hurricane damage had the strongest effect of any of the hazards tested in current-year models. This was not unexpected given that hurricanes are the most devastating of our hazards in terms of spatial scale and total property damage (NOAA, 2024). While other climate hazards may affect areas smaller than a single county, hurricanes often inflict widespread destruction across multiple states. This means that in the short term, residents may need to evacuate relatively far from home to fully escape the threatened or affected area. Once the storm has passed, many will return home, but others may not (Fussell et al., 2022). For example, a year after Hurricane Katrina, 47% of pre-storm residents of the New Orleans metro area had failed to return home, of which most were living in other states (Sastry & Gregory, 2014).

In contrast, we found only a minimal relationship with current-year tornadoes, which occur at a much smaller spatial scale. This was consistent with the findings of Raker (2020), who found tornadoes had no effect on population size, even at the block-group level. Tornadoes generally have a very small spatial footprint relative to US counties, so those who are displaced may find it relatively easy to find new housing within their home county. Any positive relationships with migration may be due to recovery investment stimulating the local economy (Elliott, 2015; Fussell et al., 2017).

For current-year flood damage, the combination of a negative relationship with Net Income Migration and no relationship with Net People Migration suggests that higher-income households may be more likely to leave or avoid counties affected by flooding while lower-income households stay put — or are even attracted to affected counties by opportunities like construction jobs or cheaper housing (Elliott, 2015; Fussell et al., 2017). This is consistent with expectations that well-off households may have more resources to respond to hazards than poor households (Masozera et al., 2007; Watts, 2017). It is hard to say whether those who stay do so by choice or due to a lack of resources to cover moving costs or potentially higher housing costs in less flood-prone areas — or whether they are staying put at all, since they may make more localized moves to safer areas within affected counties, which our data would not capture. Shu et al. (2023) found a negative association between flood risk and population growth across census blocks over a similar study period (2000–2020), so population redistribution may be occurring at finer spatial scales. Hazards may also vary in how their impacts play out over time. In some regions, we found negative relationships with lagged flood damage even where there were no significant relationships with current-year flooding, suggesting there are some contexts where it may take longer to react to hazards or implement a permanent move.

Nationally, we also found positive relationships with lagged storm damage despite finding no significant current-year relationships. This raises the possibility that over time, investment in repairs to heavy or cumulative storm damage could create opportunities that attract newcomers (Elliott, 2015; Fussell et al., 2017).

We found no relationships between wildfire damage and county-level migration rates. This contrasted with previous studies that found county-level net migration was suppressed by the most destructive wildfires (McConnell et al., 2021; Winkler & Rouleau, 2020). One explanation is that Winkler and Rouleau (2020) used a binary variable indicating counties affected by a “disaster-level” wildfire (as designated by FEMA), while we looked at a continuous measure of wildfire damage per capita, and most damages were relatively low.

Although wildfires are increasing with climate change, they remain rare events, and disastrous ones are even rarer. In most cases, wildfire footprints remain much smaller than a Western county, so after less destructive fires, displaced people may be able to find new housing within their home county. However, if an entire neighborhood is displaced by a disaster-level wildfire, some displaced residents might need to move further away, perhaps across county lines, to find housing that has not already filled up or burned down.

Outside of disasters, wildfire-prone areas can also be highly attractive to migrants since they tend to be high in natural amenities (Clark et al., 2022; McConnell et al., 2021). The wildfire-prone Wildland Urban Interface (WUI) has seen dramatic growth in both population and housing in recent decades (Radeloff et al., 2018; McConnell et al., 2021), a long-term trend which might cancel out any smaller pulses of out-migration after severe wildfires in our analysis.

There is value in considering the full range of hazard events (including less destructive ones), but in future it would also be useful to identify relevant thresholds of property damage for wildfire and other hazards so that research can focus on events that are severe enough to displace or deter people. Finer spatial scales can also help to detect more localized migration responses. McConnell et al. (2021) found increased out-migration after wildfires at the census tract level, but found most moves remained in-county, meaning our county-level analysis might miss most of this effect.

Limitations and future research

Our study faced several limitations. First, migration patterns are driven by a wide variety of social, cultural, economic, and environmental factors, many of which are hard to measure. While we tried to account for as many as possible, our final suite of variables was limited by data availability. Our geographic unit of analysis, US counties, may also be too large to fully capture migration dynamics following certain climate hazards. These dynamics may be better studied at finer geographic scales such as zip codes or census block groups, although these often have even more limited data availability. At the county level, it may only be possible to detect a response to the most destructive hazard events (e.g., those exceeding some threshold of damage per capita). Questions also remain about the effects of repeated hazard exposure and cumulative damage, and more research is needed to understand how hazard migration responses play out over multiple years.

There is a need for continued research on the relationships between climate hazards, migration, and socioeconomic outcomes, particularly across demographics and income levels. Future studies should consider the demographics of movers and non-movers and the impacts of climate migration on origin and destination communities. It would be particularly useful to account for age since both migration rates and incomes can be age dependent. Although it was beyond the scope of our study, in-, out-, and effective migration could also be considered separately to give an even more nuanced understanding of the migration response to natural hazards. Finally, it will be important to investigate the effects of rising insurance premiums and increased difficulty finding coverage as insurance companies begin to pull out of states seeing rising claims due to increases in severe weather with climate change.

Conclusions

We found that, while lower-income households were more likely to move, state-level net-migration rates were stronger (more positive or negative) for higher-income households. This suggests more “churn” among lower-income households, with in- and out-migration cancelling each other out, while higher-income households show more directed migration towards the most popular migration destinations, attracted by economic opportunities and natural amenities like warm winters, water bodies, and intermediate levels of forest and terrain ruggedness. We also found county-level rates of income migration tended to be stronger than those of household migration, indicating that the most attractive parts of the country may be gaining a disproportionate amount of aggregate household income relative to the flow of households, while the least attractive areas lose a disproportionate amount, potentially impacting local economies and tax bases. There was also a sharp uptick in moves among households in higher income brackets after the onset of the COVID-19 pandemic, contributing to more migration and income migration. This hugely disruptive event acted as a catalyst, spurring households that could afford it to move at higher rates.

Our study also demonstrated the diverse effects that different climate hazards can have on county-level migration. Nationally, hurricane damage was associated with reductions in both county-level net migration and net income migration. Hurricanes also had the strongest effects of any hazard variable, presumably because they tend to do more damage and have larger spatial footprints, driving people further from home in search of safety. Flood damage had no relationship with net people migration but did have a small negative relationship with net income migration, suggesting higher-income households may be more likely to leave for higher ground or avoid affected areas, at least at the county level. This finding is consistent with the “concentration hypothesis,” where advantaged households tend to leave impacted areas while disadvantaged households remain, potentially leaving affected communities with fewer resources to prepare for or recover from future hazard events. Future research should consider hazards separately and investigate them at appropriate spatial and temporal scales, since, depending on the hazard and location, a county may be too large to fully capture all hazard-related migration dynamics of interest for local planning and hazard management. Hazard migration responses can also occur over different timelines in different contexts.

As climate change brings more frequent and destructive hazards, it will be important to understand their implications for migration and community resilience across income levels. Our study highlights how events that disturb our social systems — such as climate hazards and pandemics — can increase migration, particularly among higher-income households, suggesting a need for emphasis on adaptation options for lower-income households that stay in place.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

Conceptualization and design: Mahalia Clark (MC), Ephraim Nkonya (EN), and Gillian Galford (GG); data preparation and visualization: MC; methodology and formal analysis: MC and EN; writing — original draft preparation: MC; writing — analysis, editing, and review: MC, EN, and GG; all authors read and approved the final manuscript.

Funding

This research was supported by the National Science Foundation Research Traineeship Program (QuEST, Award No. 1735316), the USDA Economic Research Service (FAIN 58–6000-2–0066), the NASA Socioeconomic Assessment Program (Award No. 80NSSC22K1802), the U.S. Department of Commerce National Institute of Standards and Technology (Financial assistance award 70NANB24H080), a USDA Economic Research Service Pathways Internship, the Gund Institute for Environment (University of Vermont), the Rubenstein School of Environment and Natural Resources (University of Vermont), the University of Vermont Food Systems Graduate Program, and the Phil Lasalle/Norman Foundation. Additional funding was provided by the University of Vermont Food Systems Research Center via a cooperative agreement with the USDA Agricultural Research Service and NASA Land-Cover/Land-Use Change grant #80NSSC23K0537.

Data availability

Original data generated by this study are presented in the article and Supplementary Information. Public data sets used in this study are cited in the article and Supplementary Information. Further inquiries may be directed to the corresponding author.

Declarations

Conflict of interests

The authors declare no competing interests.

Disclaimer

This research was supported by the U.S. Department of Agriculture, Economic Research Service and Agricultural Research Service. The findings and conclusions in this publication are those of the authors and should not be construed to represent any official USDA or U.S. Government determination or policy.