Tornadoes, poverty and race in the USA: A five-decade analysis

Abstract

Purpose

The study aims to analyze the roles of poverty and African American status in terms of vulnerability to tornado damages and barriers to recovery afterward.

Design/methodology/approach

Using five decades of county-level data on tornadoes, the authors test whether economic damages from tornadoes are correlated with vulnerability (proxied by poverty and African American status) and wealth (proxied by median income and educational attainment), controlling for tornado risk. A multinomial logistic difference-in-difference (DID) estimator is used to analyze long-run effects of tornadoes in terms of displacement (reduced proportions of the poor and African Americans), abandonment (increased proportions of those groups) and neither or both.

Findings

Controlling for tornado risk, poverty and African American status are linked to greater tornado damages, as is wealth. Absent tornadoes, displacement and abandonment are both more likely to occur in urban settings and communities with high levels of vulnerability, while abandonment is more likely to occur in wealthy communities, consistent with on-going forces of segregation. Tornado damages significantly increase abandonment in vulnerable communities, thereby increasing the prevalence of poor African Americans in those communities. Therefore, the authors conclude that tornadoes contribute to on-going processes generating inequality by poverty/race.

Originality/value

The current paper is the first study connecting tornado damages to race and poverty. It is also the first study finding that tornadoes contribute to long-term processes of segregation and inequality.

Citation

Kashian, R.D., Buchman, T. and Drago, R. (2022), "Tornadoes, poverty and race in the USA: A five-decade analysis", Journal of Economic Studies, Vol. 49 No. 7, pp. 1304-1319. https://doi.org/10.1108/JES-06-2021-0287

1. Introduction

While most tornadoes affect a limited area, some are large and the net impact can be substantial. For example, 2011 was a particularly bad year for tornadoes, with 59 resulting in 552 deaths, and over $US23b in economic losses (Simmons and Sutter, 2012). Therefore, the effects on economic development could be substantial.

To conceptualize the role of poverty and race in analyzing the economic effects of tornado, consider two sequential causal paths. First, poverty and minority race/ethnicity may be related to vulnerability, such that the immediate economic effects of tornadoes tend to be more severe. Second, those same characteristics may be associated with less robust recovery efforts. Both coming into and going out of a tornado may, therefore, each disadvantage relevant communities. This study tests for such effects using five decades of tornado and census data, which were measured at the county level.

2. Background

Research on the economic effects of natural disasters tends to focus on patterns of economic growth and factors associated with growth. Skidmore and Toya (2002) find that climactic disasters, such as tsunamis, floods, hurricanes or extreme heat or snow, are associated with long-run growth, while geologic disasters, including earthquakes and volcanic eruptions, are correlated with long-run decline. Loayza et al. (2012) find that mild or moderate disasters can promote economic growth, while severe disasters are related to decline. Other research finds that African Americans tend to live in areas subject to Atlantic hurricanes or in densely populated areas which become urban heat islands during heat waves (Wilson et al., 2010), such that these climate-related effects disproportionately impact the disadvantaged. More generally, poverty increases the short- and long-term ill effects of natural disasters (Benson and Clay, 2003).

For tornadoes, these overall effects are split into two sequential causal paths, immediate vulnerability and barriers to recovery. For each path, the lens of poverty and race is applied.

2.1 Vulnerabilities

A key distinction for this analysis regards risk and vulnerability (Ashley and Strader, 2016). Risk is the probability of a tornado striking a particular geography with a given force, duration, width and length of track. Historically, “Tornado Alley” was presumed to represent a place with frequent tornadoes, mainly in the Great Plains region between the Rocky and Appalachian Mountains, although it has been supplemented with the “Dixie Alley” in the southeast, noting that the largest number of tornadoes occur in the state of Mississippi (Dixon et al., 2011), with generally increasing prevalence in the Mississippi river valley and Midwest (Ashley, 2007; Gensini and Brooks, 2018).

Tornado risk is not randomly distributed. For example, as of 2015, the state of Mississippi had a poverty rate of 22.0%, compared to a national average of 14.7% (Bishaw and Benson, 2017). That same year, African Americans represented 37.7% of Mississippi residents but only 12.1% of USA residents (Kaiser Family Foundation, n.d.). Given high tornado prevalence in that state, a positive correlation between tornadoes, poverty and race may exist, with the finding of a positive correlation between poverty and tornado prevalence also reported by Lim et al. (2017). We do not treat poverty or race as causal in terms of risk in what follows and control for risk in the vulnerability analysis to avoid misinterpreting correlation as causation.

Vulnerability is the economic and human harm caused by a given tornado. This analysis focuses on economic damages because fatalities are relatively rare [1]. Studies suggest there are racial differences in risk perceptions, awareness and responses. One well-established finding is that white males perceive lower levels of environmental risk (Flynn et al., 1994; Johnson, 2002; Saterfield et al., 2004). Those perceptions suggest that non-whites and non-males either are or, at least, perceive themselves to be more vulnerable to environmental hazards, such as tornados.

Attitudes may not necessarily be reflected in responses. For example, in a survey of 102 people in the immediate aftermath of a severe tornado which struck Indianapolis, Indiana, on September 20, 2002, Mitchem (2003) found that African Americans were more likely to confuse the terms “tornado watch” and “tornado warning,” relative to white respondents, but were not less likely to take shelter. Donner (2007) found the proportion of African American residents (census tract level data) had no significant correlation to tornado fatalities or injuries, but economic deprivation, comprised poverty, low levels of education and high levels of disabilities, was significantly related to tornado injuries. Using county-level data, Simmons and Sutter (2013) found that tornado fatalities were positively related to the percentage of non-whites in the population. Using similar data, Lim et al. (2017) found that socio-economic status was negatively related to tornado fatalities, with individual components such as income, poverty, education, female-headed households and mobile homes each providing some explanatory power.

Economic damages are likely to be positively correlated with vulnerability, whether in terms of race or poverty. However, damages are also likely to rise with levels of wealth. This relationship is in contrast to fatalities, which are solely related to vulnerability. For relevant examples, a mobile home is more likely to be totaled in a tornado relative to a brick home (Lim et al., 2017), but it is more expensive to replace a late-model sports car than a decade-old Ford sedan. The analysis of vulnerability should control for these contradictory effects.

2.2 Barriers to recovery

Race and poverty may further represent barriers to recovery. In one well-known example, after Hurricane Katrina devastated New Orleans, Louisiana, in 2005, a major federal response in terms of the Community Development Block Grant program was administered in ways which discriminated against both African Americans and the poor (Gotham, 2014, 2015). Moreover, public housing, which had covered 10% of New Orleans residents, was curtailed, such that the African American population shrank (Scoopetta, 2016). The percentage of the population living in poverty in New Orleans parish shrank from 28% in 1999 to 23% as of 2019, and the African American population shrank by 92,974, compared to 8,289 for whites, with the African American share declining from 67 to 59% over the same time period (Data Center, 2020). Katrina (along with Hurricane Rita) provides an example of a disaster leading to displacement of vulnerable populations (Curtis et al., 2015).

There are reasons to believe these barriers may be relevant to tornado recovery efforts. First, an analysis of Federal Emergency Management Agency (FEMA) responses to severe storms, flooding and tornadoes in Missouri during 2008 revealed that individual assistance (IA) grants were more likely to be given as the percentage of owner-occupied housing increased (Kousky, 2013), and as of the first quarter 2021, rates of householder home ownership were 73.8% among non-Hispanic whites but only 45.1% among African Americans (US Census, 2021). Additionally, it has historically been true that low income families are less likely to either become homeowners or maintain that status (Boehm and Schlottmann, 2004). Together, these facts suggest that low rates of home ownership may represent a barrier to recovery for both African Americans and the poor.

Second, Squires (2003) summarizes evidence suggesting that property insurance companies often discriminate against African Americans by not selling policies to them, not offering full replacement insurance or either charging more for the same policy purchased by an African American. Discrimination in terms of sales or less than replacement insurance might each serve as barriers to African American homeowner recovery from tornadoes.

As mentioned above with regard to Katrina, some evidence suggests displacement of the vulnerable following a natural disaster (Curtis et al., 2015). Relatedly, Elliot and Howell (2017) analyzed property damage from all natural disasters from before 1999 to 2011 and found later that residential instability is a positive correlate, with larger effects for African American and Latina women. Of immediate relevance here, Raker (2020) analyzed census block-group level data on severe tornadoes (F3 and above on the Fujita scale) using three decades of data. He found that the composition of residents became both more white and with lower rates of poverty, suggesting that African Americans and the poor were unable to recover in place, consistent with displacement.

Recovery may instead involve abandonment or what Raker (2020) labels concentration, wherein the vulnerable are left behind. Some evidence supports net abandonment: following 32 hurricanes, Logan et al. (2016) found whites and young adults leaving communities and leaving African Americans and the elderly behind, with population growth rates significantly reduced, particularly where poverty rates were low (e.g. suggesting those more able to leave were more likely to do so). A more comprehensive study by Boustan et al. (2017) covered ten decades of disaster data, applied to the county level, and found net declines in both economic growth and population following disasters involving, at least, 25 fatalities, but especially for more than 500 fatalities.

On the contrary, in both Raker's (2020) study and Fussell et al.'s (2017) study of population change following hurricanes and tropical storms from 1980 to 2012, minimal changes in population were identified (although most measured effects were negative in the latter). In the prior study, this result might be due to restricting the effects of severe tornadoes to affected census block groups, which would miss broader effects. The latter finding might be due to different methodology relative to Logan et al. (2016).

To summarize, if resources pour into an area post-tornado but are concentrated among whites, homeowners and the non-poor, then displacement will likely occur. If, however, resources are insufficient or not forthcoming at all, then abandonment may follow. In this regard, it is important to recognize that governmental resources are not the only relevant resource. As Kousky (2013) notes for 2008 disasters in Missouri, including tornadoes, relevant IA amounts provided by FEMA are typically small as in a few thousand dollars. The Small Business Administration (SBA) disaster loan program, for which businesses of any size are eligible, was reformed after the poor response to Katrina and other severe hurricanes, which provides some additional funding (Lindsay, 2015). Nonetheless, private insurance is a major source of recovery funding as are resources from local and national non-profit organizations (Smiley et al., 2018).

Both the finding that homeowners tend to be privileged in FEMA recovery efforts (Kousky, 2013) and that African American are both less likely to be homeowners and tend to be underinsured when they are (Squires, 2003) suggest that displacement is more likely. However, there is one contrary finding: Kousky (2013) found that the probability of approvals by FEMA rose slightly in zip codes with more African Americans in tandem with slightly lower levels of aid among approvals.

Given conflicting reports on population change following natural disasters, we focus the recovery analysis on three outcomes: (1) displacement, where the proportion of African Americans and the poor declines; (2) abandonment, where those proportions rise and (3) where neither occurs. Given those outcomes, the tornado recovery analysis asks where tornado property damage in one decade is followed by displacement, abandonment or neither using measures of change from the decade prior to the decade after tornadoes. Further, we can ask how pre-existing vulnerabilities function in terms of displacement and abandonment, both directly and in tandem with tornado damage.

3. Data

Location data on tornadoes occurring in the USA states and the District of Columbia, for 1970–2019, were compiled by the Centre for Research on the Epidemiology of Disasters (CRED) (https://www.emdat.be/). Tornadoes are classified along the enhanced Fujita scale from F0 to F5, with the latter being the most severe. Tornadoes below F2 rarely cause significant property damage or fatalities, so they are ignored, as in Lim et al. (2017) or Raker (2020). A total of 10,979 tornadoes occurred during these years and were assigned to counties where they first touched down by CRED. Of those, CRED includes data on 789 cases (7.19%) where the tornado tracked into a second county. For these cases, property damages are allocated half-and-half to each county [2]. The tornado variable captures the number of F2 or above (or F3 and above) in a county during five decades, beginning with 1970–1979, and ending with 2010–2019.

Counties are divided into urban, micropolitan and rural categories using the 2013 National Center for Health Statistics (2017) classification, with county land size measured in miles as of 2011 (US Bureau of the Census, 2011) [3].

The distribution of tornadoes and damages by decade is shown in Table 1, noting that tornado damage is deflated to $US1983 using the CPI-U. CPI-U provides a measure of the average change (over time) in the prices paid by urban consumers. This is based on a market basket of consumer goods and services. Additionally, we exclude 93 outliers in terms of tornado damage being above $US15.5m in $US1983 (the top 5% of positive damages observation) because we suspect that these events are both more well-publicized and receive greater attention in terms of public and private recovery efforts. We later check for the effects of this exclusion and those results support the approach that was used. There is a gradual decline from the 1970s to the 2000s in the number of, at least, F2 or, at least, F3 tornadoes, with a spike in the 2010s. Damages followed a similar pattern before a substantial increase in the 2000s and particularly the 2010s, implying that the average damage per tornado rose in recent decades, which were consistent with the findings of Ashley and Strader (2016).

Demographic data for the first four decades are from the US Census as found in Lim et al. (2017), with interpolation to estimate mid-decade figures. For the latest decade, poverty and income data from 2015 are from the US Bureau of the Census (2016), while race, education and age data for 2015 are from the US Bureau of the Census (2020). Median income is also deflated to $US1983 using the CPI-U. The proportions African American and white in all counties include Hispanics, who are not measured separately in these data.

Averaging characteristics across all decades yield the descriptive statistics reported in Table 2, either for counties experiencing, at least, an F2 or above tornado during a decade, as opposed to those which do not. As it can be seen from the sample size figures at the bottom, approximately one-third of counties experience, at least, one tornado in each decade. Beyond the tornado figures, poverty and African American race are discussed extensively in the literature described above and here the proxy vulnerability is discussed. The proportion above the age of 65 years is included because the elderly have been found to be more vulnerable to tornadoes (Ashley, 2007). Median income is often included in addition to poverty in tornado vulnerability studies (Fricker and Elsner, 2019; Lim et al., 2017) and is here used in conjunction with the proportion of adults with, at least, a bachelor's degree to proxy wealth. Although median income does not reflect income at the top of the distribution, it is generally above the poverty line, so it captures higher income [4]. Additional controls are included for population density (i.e. population/land area) and urban/suburban/rural status, as in Donner (2007).

As expected, the proportions living in poverty and African Americans are slightly higher in counties experiencing tornadoes. Relatedly, median income and the prevalence of bachelor's degrees are slightly lower where tornadoes strike. Land area being smaller where tornadoes occur is likely related to tornadoes being most prevalent in suburban, followed by urban and with the lowest figure for rural areas.

4. Methods

The analyses begin with vulnerability, then turns to barriers to recovery, where the two analyses are combined. Considering vulnerability, as argued earlier and as the figures provided in Table 2 highlight, it is important to control for risk so that where vulnerable people live is not conflated with vulnerability. To do so, the number of, at least, F2 and, at least, F3 tornadoes are used; these measures capture realized risk. Given that context, the random effects estimator is used as follows:

(1)F2damaget=α+B1(F2+t)+B2(F3+t)+B3(Vulnerabilityt-1)+B4(Wealtht-1)+B5(Densityt-1)+B6(Locationt)+µc+εct,

where, α is a constant, the Β terms are coefficients or vectors of coefficients, vulnerability is proxied by Pov, AA and 65+, wealth is proxied by Inc and Bach, with further controls for population density and with two location dummy variables for the three categories, μ_c is the random error term associated with each county and ε_ct is the observation-specific error term. Initially, all of the vulnerability and wealth variables are entered separately. Subsequently, the pov and AA variables are multiplied to check whether the term better captures underlying vulnerability effects (noting that older populations are not highly correlated with either variable). Similarly, the two wealth variables are combined to ascertain whether they capture an underlying wealth effect. A fixed effects alternative to the random effects specification was considered but rejected because many of the independent variables tend to change slowly if at all. As an alternative specification of the dependent variable, damages are divided by population to estimate per capita damages. For the two key vulnerability variables, simulations are performed with the variables either one standard deviation below or above the mean (AA at 0 and 0.2365, Pov at 0.07148 and 0.20796). Parallel simulations are performed for the two wealth proxies (Inc at $US16,058.6 and $US27,657, Bach 0.06259 and 0.20513).

For the recovery analysis, two-decade change measures for poverty and race are used Pov_t–Pov_t-2 and AA_t–AA_t-2. The measure of displacement or abandonment, disaban, is then 0 if both terms are below a certain negative cut-off (displacement), 2 if both are above a positive cut-off (abandonment) and 1 otherwise, noting that the variable is not ordered given small shifts, no shift and mixed shifts are all classified as “1.” The cut-offs are necessarily somewhat arbitrary, so −0.01 and 0.01, −0.025 and 0.025 and −0.05 and 0.05, respectively, are each applied.

Given the outcomes are not ordered, a multinomial logit estimator is applied to ascertain whether tornado damages are associated with displacement, abandonment and both or neither. With the base “1” as the omitted category, the multinomial logit estimates two equations as follows:

(2)U0=B0(F2damaget-1)+ε0,

(3)U2=B2(F2damaget-1)+ε2,

where, U is a latent variable which rises as the probability of outcome “0” or “2” rises; there are separate coefficients and error terms for each equation with the same tornado damage variable used in each. These latent variables are transformed into probabilities such that, for each county/decade, the probabilities for each of the three outcomes sum to unity. Note that (2) and (3) in fact comprise a DID estimator, given the outcome is measured by change from the decade prior to the decade after tornado damages occur. As Card and Krueger (1994) note, these estimators have the desirable characteristic of controlling for various omitted variables, which also influence the dependent variable.

Regardless of the initial recovery results, it is worth knowing with greater specificity where tornado damage is most likely to lead to displacement, abandonment and both or neither. Toward that end, the damage variable is interacted with the combined vulnerability variable and, separately, with the combined wealth variable and two of the three urban/suburban/rural dummy variables. For consistency with the approach used to estimate vulnerability (1), given the damages variable is lagged once in (2) and (3), the combined term is lagged twice, with the interaction picking up the influence of the combined term on damages. The reason for separate estimates is that no assumption of independent effects is required. All estimates are performed using Stata 15.1.

5. Results

Table 3 provides various regression results for tornado damages, beginning with the specification with separate vulnerability and wealth terms (1) that the same specification using per capita damages as the dependent variable (2) and then replicating each after combining Pov and AA and, separately, Inc and Bach in (3) and (4). As expected, the number of, at least, F2 or F3 tornadoes is positively correlated with damages in all specifications and significant in 7 of 8 cases, which were consistent with serving as controls for risk. In all four specifications, the coefficients on all three (in [1] and [2]) or two (in [3] and [4]) vulnerability variables are positive and significant as expected. The same pattern holds for the wealth proxies as also predicted. Population density is surprisingly negatively associated with damages, while damages in suburban and rural areas show a mixed pattern, though with significantly higher per capita damages in rural areas. However, the overall R² statistic is around twice as large for specifications (1) and (3), suggesting these are preferred to the per capita estimates.

Simulation results are provided at the bottom of Table 3 for the key coefficients (noting all results are significant). For specification (1), increasing Pov and AA from one standard deviation below to one standard deviation above their respective means is projected to increase tornado damage more than two-fold, with a more than four-fold increase for high levels of Inc and Bach. For specification (2), the same exercise yields an approximate tripling of (per capita) damages for Pov and AA and for Inc and Bach. Turning to specification (3), the combined vulnerability term suggests an increase of approximately one-quarter for higher vulnerability and an almost doubling for the combined wealth variables. Results in the final specification suggest slightly larger vulnerability effects and smaller wealth effects. Specification (3) is treated, as preferred given model fit is superior to specification (4); it provides the most conservative estimates for size effects related to vulnerability.

As a check on whether the F2+ and F3+ variables are picking up risk related to poverty and African American status (highlighted in Table 2), the preferred regression (3) was replicated after excluding those terms (results not shown). As expected, the coefficient on Pov × AA almost doubled, rising from 1,377,000 to 2,932,679, which supports the analytical approach used here.

Turning to the recovery analysis, full results for the multinomial regressions are reported in an Appendix. For simplicity, most of the discussion focuses on results using the 0.025 (and −0.025) cut-offs for counties/decades where displacement or abandonment occur. In part, the 0.025 cut-off is ideal because it yields a nearly balanced number of cases of displacement (368) and abandonment (371), while the 0.05 cut-off favors displacement (152) over abandonment (74) and the 0.01 cut-off favors abandonment (1,073) over displacement (687).

Regressing the displacement/abandonment measure on the damages variable (the simplest DID estimate) yields a significant negative coefficient for displacement and a significant positive coefficient for abandonment (see Table A1, first numeric column). Simulations from the regressions are presented in Table 4, where the tornado damage term is set at either $US0 (the median value) or $US670,799 (the mean for positive values). The first row, labeled “overall,” is for the initial regression and suggests the probability of displacement declines from 0.0427 to 0.0331 (0.96% point difference) when tornado damage is introduced. The probability of abandonment rises from 0.0386 to 0.0453 in the same circumstance (0.67 increase). These figures imply that average tornado damage results in abandonment – with rising Pov and AA around a third more often than displacement – with falling Pov and AA (i.e. 0.0453/0.0331).

The next specification adds the combined vulnerability term (Pov × AA) term and interacts it with the damages variable, with “*”s denoting significant interaction terms. The tornado damages coefficients are now insignificant for displacement but remain positive and significant for abandonment. The vulnerability term is significant and positive for both displacement and abandonment, with a positive and significant interaction for abandonment (Table A2). Applying the same simulated values as before and considering displacement, note first that the high values of the vulnerability terms (compared to low values) are associated with an approximate doubling of the probability of either displacement or abandonment. Considering just low values of Pov × AA, introducing tornado damages reduces the probability of displacement from 0.0275 to 0.0222 (0.53 reduction). For high values of the combined term, the same exercise results in the probability of displacement falling from 0.0579 to 0.0404 (1.75 decrease). Considering abandonment, for low values of the combined term, tornado damages raise this probability from 0.0324 to 0.0365 (0.41 decrease), and for high values, the parallel figures are 0.0485 and 0.0599 (1.14 increase), with the larger effect size for the latter reflecting the positive interaction for abandonment and vulnerability. These results suggest that the presence of poor and African American people almost doubles the probability of displacement and of abandonment, with tornado damages muting the displacement effect and particularly enhancing the abandonment effect in poor communities of color.

Switching to the regressions using the wealth term (Inc × Bach), as in the prior case, the tornado damage effect is positive and significant for abandonment but insignificant for displacement. The direct wealth effect is negative for displacement and positive for abandonment, both with significance, and no interaction terms attract significance (Table A3). For low values of the wealth term, introducing tornadoes drops the probability of displacement from 0.1047 to 0.0916 (1.31 reduction) and for high values that probability declines from 0.0005 to 0.0003 (0.02 reduction). For abandonment, low values of the wealth terms are associated with a slight increase for tornado damage from 0.0233 to 0.0274 (0.4 increase) and for high values with a larger increase from 0.0612 to 0.0699 (0.87). In both cases, large effects are found in terms of wealth reducing displacement and increasing abandonment, particularly subsequent to tornadoes.

The next exercise involves the urban/suburban/rural split, which is again interacted with the tornado damages term. Here the damages coefficient remains significant and positive for abandonment, with suburban and rural locations significantly less likely to experience either displacement or abandonment with no significant interaction (Table A4). Overall, as shown in Table 4, displacement and abandonment are both most common in urban areas, with displacement least prevalent in suburban areas and abandonment least common in rural areas. The largest reduction in displacement related to tornado damage is in urban areas, which yields a decline from 0.0682 to 0.0509 (1.73). Similarly, the largest increase in abandonment associated with tornado damage is in urban areas, which shows an expansion from 0.0673 to 0.0787 (1.14).

Switching the cut-off from 0.025 to either 0.01 or 0.05 yields very similar results (see Tables A1 through A4), albeit significance tends to decline for the 0.05 cut-off, which is reasonable given the smaller number of displacement and abandonment cases for that cut-off. Moreover, the pattern (though not absolute size) of simulation results is also stable when the cut-off is switched (see Tables A5 and A6).

Finally, a couple of patterns behind the numbers are worth noting. First, while both displacement and abandonment are associated with urban areas, average population size diverges. For counties experiencing neither displacement nor abandonment, average population is 87,823. Where displacement occurred, that figure drops to 38,523 but rises to 286,448 for abandonment, a figure which is larger by an order of magnitude (7.5). This difference implies that abandonment affects far more people.

Second, consider the effects of outliers in terms of tornado damages. To understand these effects, Table 4 figures were replicated after adding back in the observations with damages of, at least, $US15.5m. The results (see Table A7) for the simple DID equation (“Overall” row) show the absolute negative effects of tornado damages on displacement rising slightly; instead of a 0.96% point decline, it is estimated to yield a 0.97% point decline. For abandonment, however, the difference declines from a 0.67% point increase to 0.14. Other substantial differences appear for abandonment as well. Whereas before the introduction of tornado damages in areas with high proportions of poverty and African Americans increased the probability of abandonment by 1.14% points, that difference here is only 0.38% points (0.0572–0.0534). Similarly, introducing tornado damages into urban areas yielded an estimated expansion of the probability of abandonment by 1.14% points, and those figures are now simulated to be 0.42% points. These substantial declines in the estimated effects of tornado damages on the probability of abandonment are consistent with disproportionately large expenditures of recovery resources in response to the largest tornadoes.

6. Discussion

This analysis sought to understand the role of race and poverty in terms of both vulnerability to tornadoes and barriers to recovery. In terms of economic damages associated with tornadoes, the evidence supports a positive correlation between vulnerability and damages and between wealth and damages while controlling for tornado risk. In the preferred specification ([3] in Table 3), the simulated wealth effect is associated with an almost doubling damages, while vulnerability in terms of high proportions of African Americans and the poor is linked to a simulated increase of approximately one-third in measured damages. These findings are clearer than the prior mixed findings regarding race, with Donner (2007) finding no correlation and Simmons and Sutter (2013) finding a correlation, which might be because those studies considered tornado fatalities and injuries instead of economic damages. The findings are consistent with poverty being linked to fatalities and injuries (Donner, 2007; Lim et al., 2017).

In relation to recovery, some earlier studies are consistent with displacement of the type associated with Hurricane Katrina (Curtis et al., 2015), suggesting that a decline in the African American and poor populations may be relevant to tornadoes (Elliot and Howell, 2017; Raker, 2020). Other studies predict that abandonment may instead occur (Boustan et al., 2017; Logan et al., 2016), with the proportions of African Americans and the poor rising due to wealthier whites leaving the area. A DID model tracking displacement or abandonment from one decade to two decades later finds that tornado damage in the intermediate decade is associated with large and significant reductions in displacement and small though significant increases in abandonment. An implication of this finding is that the substantial resources flowing to wealthier, whiter individuals and businesses after Katrina is not typical following tornadoes, with the communities tending to instead experiencing slight increases in poverty and the proportion of African Americans as a result.

Displacement and abandonment are both strongly and directly related to vulnerability in terms of the presence of high proportions of the poor and African Americans in the decade prior to any measured tornado damage. Wealth on the other hand, indicated by median income and the proportion of adults holding, at least, a bachelor's degree, is negatively related to displacement but positively correlated with abandonment. Urban areas are positively associated with both displacement and abandonment. The fact that it is urban areas with large populations, with high proportions of African Americans and the poor as well as above-average wealth, which are linked to abandonment, suggests the dynamics of hypersegregation, identified long ago by Massey and Denton (1989), are still operational. The findings here add to that the possibility that tornado damages play a small but significant role in maintaining those dynamics.

This study makes two clear advances over earlier research. First, by moving beyond fatalities and injuries to focus on economic damage, we found that both vulnerability and wealth are associated with larger tornado damages. Second, by focusing on tornadoes, we identify tornado-specific effects that broader studies of natural disasters have not identified (e.g. Boustan et al., 2017; Elliot and Howell, 2017; Kousky, 2013). By the same token, however, the current study cannot inform us as to the likely effects of other types of natural disasters, including hurricanes, but also wildfires, droughts and floods. That task is left to future research.

Table 1

Distribution of F2+ tornadoes and damages per county and decades

	1970s	1980s	1990s	2000s	2010s
F2+ number	1.01	0.700	0.594	0.577	0.721
F2+ damage	$131,976	$91,839	$49,492	$349,884	$666,668
F3+ number	0.321	0.210	0.200	0.149	0.195
F3+ damage	$75,713	$48,097	$28,760	$122,476	$247,280

Table 2

Average characteristics of counties from 1970s–2010s

	F2+ tornadoes	No F2+ tornadoes
F2+ F2+ number	1.86	0
F2damage F2+ damage	$655,579	$0
Vulnerabilities
Pov Poverty	0.141	0.138
AA African American	0.108	0.078
65+ Age 65 years or above	0.140	0.143
Wealth
Inc Median income	$29,975	$34,074
Bach At least a bachelor's degree	0.127	0.138
Population	83,849	88,046
Land area (sq. mi.s)	769	1,058
Dense Population density	144.6	363.2
Location
Urban	39.2%	60.8%
Suburban	41.2%	58.8%
Rural	37.2%	62.8%
N	5,789	9,138

Table 3

Random effects regressions for tornado damages, decades/counties

	(1)	(2)	(3)	(4)
Variables	F2damage	F2damage/pop	F2damage	F2damage/pop
F2+	440,727** (27,048)	17.80** (1.470)	443,728** (27,251)	17.75** (1.460)
F3+	87,921 (51,415)	6.572* (3.112)	84,548 (51,682)	6.466* (3.111)
Pov	1.178e+06** (153,371)	68.17** (15.43)
AA	319,766** (82,506)	19.69** (7.011)
65+	895,990** (225,697)	177.0** (22.00)	513,832* (224,860)	163.1** (22.57)
Inc	18.83** (2.702)	0.000656** (0.000138)
Bach	944,573** (188,172)	38.55** (10.52)
Dense	−6.112** (0.842)	−0.000214** (4.72e-05)	−2.878** (0.424)	−9.19e-05** (2.30e-05)
Suburban	−18,228 (27,358)	2.483 (1.276)	−28,487 (27,277)	2.253 (1.235)
Rural	−27,526 (23,800)	11.94** (1.515)	−37,515 (22,945)	11.95** (1.467)
Pov × AA			1.377e+06** (307,538)	137.2** (31.51)
Inc × Bach			35.50** (4.806)	0.00114** (0.000163)
Constant	−848,247** (77,359)	−58.69** (5.322)	−201,880** (35,595)	−32.20** (3.433)
Observations	11,867	11,867	11,867	11,867
Number of counties	2,999	2,999	2,999	2,999
R2 within	0.224	0.133	0.216	0.130
R2 between	0.295	0.158	0.295	0.162
R2 overall	0.243	0.135	0.238	0.134
chi2	490.2**	258.9**	452.1**	246.9**
Pov and AA high	$419,263	$21.56	$329,758	$18.08
Pov and AA low	$182,829	$7.59	$262,024	$11.34
Inc and Bach high	$456,877	$20.00	$371,891	$16.40
Inc and Bach low	$103,855	$6.90	$206,150	$11.06

Note(s): Robust standard errors in parentheses: **p < 0.01 and *p < 0.05

Table 4

Simulations for displacement and abandonment with 0.025 cut-off

	Odds of displacement		Odds of abandonment
	F2damage $0	F2damage $670,799	F2damage $0	F2damage $670,799
Overall	0.0431	0.0331	0.0386	0.0453
Pov × AA low	0.0275	0.0222	0.0324*	0.0365*
Pov × AA high	0.0579	0.0403	0.0485*	0.0599*
Inc × Bach 500	0.1047	0.0916	0.0233	0.0274
Inc × Bach 4980	0.0005	0.0003	0.0612	0.0699
Urban	0.0682	0.0509	0.0673	0.0787
Suburban	0.0238	0.0155	0.0302*	0.0355*
Rural	0.0308	0.0245	0.0172*	0.0193*

Note(s): **p < 0.01 and *p < 0.05 for interaction term with damages significant

Table A1

Multinomial logit results for displacement/abandonment and F2damage only

	(1)	(2)	(3)
Variables	Cut-off 0.025	Cut-off 0.01	Cut-off 0.05
Displacement results
F2damage	−4.01e-07* (1.92e-07)	−3.02e-07* (1.26e-07)	−5.10e-07 (3.41e-07)
Constant	−3.058** (0.0549)	−2.311** (0.0411)	−3.995** (0.0845)
Abandonment results
F2damage	2.33e-07** (3.33e-08)	2.20e-07** (2.99e-08)	2.61e-07** (4.91e-08)
Constant	−3.170** (0.0554)	−1.953** (0.0339)	−4.894** (0.124)
Observations	8,871	8,871	8,871
Pseudo-R2	0.00748	0.00587	0.00864
χ2	45.42**	65.40**	20.41**

Note(s): Robust standard errors in parentheses: **p < 0.01 and *p < 0.05

Table A2

Multinomial logit results for displacement/abandonment, F2damage and vulnerability variables

	(1)	(2)	(3)
Variables	Cut-off 0.025	Cut-off 0.01	Cut-off 0.05
Displacement results
F2damage	−3.23e-07 (2.42e-07)	−2.80e-07 (1.64e-07)	−3.99e-07 (4.28e-07)
(Pov × AA) × F2damage	−4.61e-06 (4.32e-06)	−3.05e-06 (2.93e-06)	−5.24e-06 (6.41e-06)
Pov × AA	16.19** (0.880)	17.37** (0.836)	15.54** (1.052)
Constant	−3.532** (0.0686)	−2.746** (0.0508)	−4.533** (0.108)
Abandonment results
F2damage	1.75e-07** (4.18e-08)	1.79e-07** (3.41e-08)	2.41e-07** (6.15e-08)
(Pov × AA) × F2damage	2.61e-06* (1.11e-06)	1.73e-06 (1.07e-06)	4.29e-07 (8.74e-07)
Pov × AA	9.244** (1.098)	7.917** (0.912)	11.02** (1.735)
Constant	−3.368** (0.0635)	−2.083** (0.0381)	−5.187** (0.146)
Observations	8,871	8,871	8,871
Pseudo-R2	0.0700	0.0517	0.0945
χ2	424.9**	575.7**	223.3**

Note(s): Robust standard errors in parentheses: **p < 0.01 and *p < 0.05

Table A3

Multinomial logit results for displacement/abandonment, F2damage and wealth variables

	(1)	(2)	(3)
Variables	Cut-off 0.025	Cut-off 0.01	Cut-off 0.05
Displacement results
F2damage	−1.17e-07 (6.16e-07)	−4.28e-07* (2.10e-07)	−5.67e-07 (6.03e-07)
(Inc × Bach) × F2damage	−9.79e-11 (3.54e-10)	1.16e-10 (6.85e-11)	1.33e-10 (2.52e-10)
Inc × Bach	−0.00116** (8.42e-05)	−0.00112** (6.05e-05)	−0.00162** (0.000148)
Constant	−0.955** (0.134)	−0.250* (0.100)	−1.300** (0.203)
Abandonment results
F2damage	2.32e-07** (6.02e-08)	1.47e-07* (6.00e-08)	1.80e-07 (1.02e-07)
(Inc × Bach) × F2damage	−0 (0)	0 (0)	0 (0)
Inc × Bach	0.000191** (1.49e-05)	0.000224** (1.29e-05)	0.000149** (2.54e-05)
Constant	−3.814** (0.0823)	−2.672** (0.0567)	−5.392** (0.167)
Observations	8,871	8,871	8,871
Pseudo-R2	0.0892	0.0927	0.109
χ2	541.6**	1032**	258.4**

Note(s): Robust standard errors in parentheses: **p < 0.01 and *p < 0.05

Table A4

Multinomial logit results for displacement/abandonment, F2damage and location variables

	(1)	(2)	(3)
Variables	Cut-off 0.025	Cut-off 0.01	Cut-off 0.05
Displacement results
F2damage	−4.47e-07 (2.40e-07)	−3.09e-07 (1.62e-07)	−4.83e-07 (3.56e-07)
Suburban × F2damage	−2.02e-07 (7.69e-07)	−6.58e-08 (4.09e-07)	−2.28e-07 (1.37e-06)
Rural × F2damage	9.91e-08 (4.60e-07)	−3.48e-09 (2.99e-07)	−9.95e-07 (1.60e-06)
Suburban	−1.142** (0.179)	−1.047** (0.132)	−1.421** (0.297)
Rural	−0.892** (0.123)	−0.652** (0.0891)	−1.211** (0.205)
Constant	−2.539** (0.0731)	−1.864** (0.0593)	−3.372** (0.103)
Abandonment results
F2damage	2.22e-07** (4.29e-08)	2.12e-07** (4.19e-08)	2.55e-07** (5.83e-08)
Suburban × F2damage	1.47e-08 (9.13e-08)	−4.14e-08 (7.68e-08)	7.22e-08 (1.29e-07)
Rural × F2damage	−5.78e-08 (1.04e-07)	−2.66e-08 (7.53e-08)	−8.70e-06 (8.24e-06)
Suburban	−0.893** (0.155)	−0.654** (0.0880)	−0.878* (0.361)
Rural	−1.460** (0.145)	−1.467** (0.0870)	−0.963** (0.309)
Constant	−2.552** (0.0701)	−1.342** (0.0462)	−4.366** (0.159)
Observations	8,871	8,871	8,871
Pseudo-R2	0.0427	0.0436	0.0452
χ2	258.9**	485.3**	106.8**

Note(s): Robust standard errors in parentheses: **p < 0.01 and *p < 0.05

Table A5

Simulations for displacement and abandonment and 0.01 cut-off

	Odds of displacement		Odds of abandonment
	F2damage $0	F2damage $670,799	F2damage $0	F2damage $670,799
Overall	0.0799	0.0650	0.1143	0.1320
Pov × AA low	0.0540	0.0446	0.1048*	0.1176*
Pov × AA high	0.1129	0.0848	0.1377*	0.1647*
Inc × Bach 500	0.1887	0.1575	0.0647	0.0742
Inc × Bach 4980	0.0011	0.0012	0.1978	0.2240
Urban	0.1095	0.0883	0.1845	0.2111
Suburban	0.0457	0.0354	0.1142	0.1275
Rural	0.0708	0.0578	0.0528	0.0602

Note(s): **p < 0.01 and *p < 0.05 for interaction term with damages significant

Table A6

Simulations for displacement and abandonment with 0.05 cut-off

	Odds of displacement		Odds of abandonment
	F2damage $0	F2damage $670,799	F2damage $0	F2damage $670,799
Overall	0.0179	0.0128	0.0073	0.0087
Pov × AA low	0.0106	0.0081	0.0055	0.0065
Pov × AA high	0.0223	0.0145	0.0093	0.0112
Inc × Bach 500	0.0506	0.0383	0.0050	0.0058
Inc × Bach 4980	0.0000	0.0000	0.0105	0.0124
Urban	0.0328	0.0239	0.0121	0.0145
Suburban	0.0082	0.0051	0.0052	0.0065
Rural	0.0101	0.0038	0.0048	0.0000

Note(s): **p < 0.01 and *p < 0.05 for interaction term with damages significant

Table A7

Simulations for displacement and abandonment with 0.025 cut-off and outliers included

	Odds of displacement		Odds of abandonment
	F2damage $0	F2damage $670,799	F2damage $0	F2damage $670,799
Overall	0.0427	0.0330	0.0423	0.0437
Pov × AA low	0.0271	0.0222	0.0350**	0.0354**
Pov × AA high	0.0575	0.0402	0.0534**	0.0572**
Inc × Bach 500	0.1043	0.0908	0.0253	0.0260
Inc × Bach 4980	0.0005	0.0003	0.0666	0.0684
Urban	0.0673	0.0506	0.0731	0.0773
Suburban	0.0236	0.0154	0.0336*	0.0344*
Rural	0.0306	0.0246	0.0180*	0.0183*

Note(s): **p < 0.01 and *p < 0.05 for interaction term with damages significant

Notes

1. Specifically, of the F2 and above tornadoes analyzed below, 93.7% do not involve fatalities but 34.2% involve property damage.

2. Cases where more than two counties were affected exist but are rare. The National Weather Service has a list of the 42 longest track tornadoes in the USA, each of which covered at least 3 counties. Of the 42 tornadoes, 19 occurred during our study time frame. (See https://www.weather.gov/jan/30_longesttortracks)

3. There are slight changes in land area over time which is ignored here, but these are not relevant to any of the key variables in the analysis, which are updated each decade (e.g. population, race).

4. In these data, out of 14,996 observations, there are only 5 cases where the poverty rate is above median income at 50% or above (0.03% of observations).