The Effects of Disasters on State Population

The Effects of Disasters on State Population

Francis Sanzari

College of the Holy Cross

Advisor: Professor Baumann

Abstract

Disasters have long been an object of study and discussion in academic, popular, and political spheres throughout the United States. A recent rise in the number and intensity of natural disasters has magnified such discussion. This paper analyzes the relationship between disasters and state population change. The results of a number of econometric specifications reveal that, in the time period studied, there is no statistically significant impact of disasters on a state’s population. Hurricane Katrina alone, the most destructive natural disaster in the time period studied, is an exception to these findings.

Introduction

Disasters have long been an object of study. Concerns about the human and economic costs of disasters havecaptured the attention of academics, policymakers, and the public. In particular, disasters with high human costs tend to draw immense national and even international attention.

According to the National Oceanic and Atmospheric Administration (NOAA), the deadliest hurricane to make landfall in the United States hit Galveston, Texas in 1900. The storm caused an estimated 8,000 to 12,000 deaths, or approximately four times the human costs of the next deadliest hurricane in U.S. history (NOAA). Officials of the time attributed much of the high death toll to insufficient structural protection, such as sea walls, for the City of Galveston (Cline). Clearly other factors, such as a lack of communication, transportation, and meteorological technology in the time period played a significant role in the immense human losses of the storm.

Since the Galveston disaster in 1900, enormous strides have been made in disaster preparedness. The early 20th century brought advances in weather observation and forecasting technology, methods of mass communication, and professional organization of meteorologists (NOAA). Such advances have enhanced the ability of government agencies to forecast destructive disasters, and to communicate such forecasts to the population well in advance of the disaster’s impact.

The second half of the twentieth century was marked by an increase in coordination among government agencies responsible for weather prediction and disaster relief. NOAA was formed in 1970, consolidating and synchronizing a number of federal weather bureaus (NOAA). In 1979, President Jimmy Carter authorized the formation of the Federal Emergency Management Agency, merging a vast array of federal disaster-response organizations. The formations of both of these organizations signified increased efforts by the federal government to predict, prepare for, and respond to disasters on U.S. soil.

To a large extent, technological and organizational developments have been successful in mitigating the deadly effects of hurricanes and other natural disasters. For example, since 1928, only one hurricane has caused more than 1,000 deaths (Katrina). The vast majority of modern disasters on U.S. soil have relatively low human costs, in comparison to the death counts of disasters from earlier in history.

This is not to say, however, that disasters no longer have an impact. Although deaths have been reduced, severe disasters continue to cause economic, sociological, and political problems. I investigate whether disasters have an impact on interstate migration. States that are prone to disasters, for example gulf states to hurricanes, face higher risks of catastrophe, costs associated with disaster readiness, and costs of redevelopment after a storm. Negative population effects could, in turn, trigger adverse economic and political effects, as decreased populations could hinder economic development, and storm refugees could induce political instability.

I estimate the impact of disasters on domestic migration patterns and state civilian population. My results suggest that migration patterns are not affected by the presence or magnitude of disasters. In the time periods studied, there appears to be no statistically significant impact of disasters upon migration or population patterns. The only notable exception to these findings is Hurricane Katrina, which is discussed at greater length below.

Literature Review

Although the specific literature regarding the impact of disasters on migration is thin, there is one article that is similar to my own investigation. Rossi, Wright, Wright and Weber-Burdin (1978) explore the demographic impacts of natural disasters, and develop an econometric model to assess the long-run effects of natural disasters on population change. They use a variety of independent variables, including total population, median age, lagged median income,percent of housing over 20 years old, percent non-white, area of location, SMSA population, SMSA population change, percent unemployed, geographic controls, and dummy variables for several disaster types. Ultimately, the authors find within their analysis that there were “no discernable effects of the natural disaster events occurring in [1960-1970] which materially altered population… trends.” (Rossi et. al., 127)

While Rossi, Wright, Wright and Weber-Burdin provide a useful econometric model, they also have a number of shortcomings. First, the authors focus on long-term impacts of natural disasters by using ten-year population changes (1960-1970) in each county. Such a broad dependent variable may obscure significant effects of disasters on populations that occur in shorter time periods. Additionally, the dated nature of the research poses significant issues for a modern application of the authors’ econometric equation. While still broadly relevant, some of the ideal independent variables may have changed over the 32 years since publication.

Beyond specific analyses of disaster and population, othereconometric papers examine population change. Carlino and Mills (1987) investigate the determinants of county growth using an empirical approach. The authors include lagged variables of employment density, population density, percent black, local government taxes per capita, median family income, median school years attained, and crime rate, as well as leading interstate highway density and regional dummy variables to account for geographic effects. Ultimately, the authors find that climate and employment prospects have a significant impact on population, while variables which are dependent on public policy, including taxes and crime rates, are relatively insignificant.

Other authors focus on the effect of specific variables on population or migration. Treyz, Rickman, Hunt and Greenwood (1993) find that internal migration is affected by relative economic opportunities and amenity differentials. In a fairly extensive review of the literature concerning migration and welfare benefits, Moffitt (1992) writes that some studies current to the 1990’sfind positive and significant relationships between the two variables. However Moffitt qualifies this finding with concerns about methodology and endogeneity between welfare benefits and cross-sectional variation in residential location, indicating that econometric estimates may be biased. Beyond welfare benefits, other amenities exist, but many, such as “topological, climatological, and environmental amenities… may be at least partly reflected in labor and land markets.” (Greenwood 1985, 527)

Economic indicators are particularly important in determining migration patterns. Many studies incorporate some measure of employment or unemployment, with the expectation that employment opportunities (or lack thereof) affect migration to or from an area. For example, Davies, Greenwood and Li (2001) find that a destination-to-origin unemployment rate ratio is statistically significant in a conditional logit investigation of migration. In his survey of internal migration literature, Greenwood (1975) finds that income generally has a positive impact on migration decisions. In another model of interregional migration, Gabriel, Shack-Marquez, and Wascher (1992) find that differentials in housing prices between two regions are important determinants of migration. Schachter and Althaus (1989) find that “taxes…have an adverse effect on migrants.” (Schachter and Althaus, 156)

Demographic characteristics also appear frequently in migration studies. Gallaway (1969) finds that older individuals generally require a greater level of increased compensation to migrate than younger individuals. Schwartz (1973) finds a similar result, arguing that the probability that one will migrate decreases with age until retirement. Schwartz also finds that education has a strong positive effect on the probability of migration. Other articles focus on the influence of the family unit on migration decisions. Mincer (1978) finds that married people are far less mobile than those who are single. In a survey of internal migration literature, Greenwood (1975) finds that “determinants and consequences of nonwhite migration differ appreciably from those with white migration.” (Greenwood 1975, 407) According to Davies, Greenwood and Li (2001), individuals are more likely to move from relatively less-populated states to states with larger populations.

Method 1—State-Annual

In order to assess the potential relationship between disasters and net domestic migration, I estimate a model using net domestic migration as the dependent variable and disaster count and several demographic controls as independent variables. Many of my control variables are motivated by the independent variables found in the extant literature. I also use state-level and year-level dummy variables, or fixed-effects, to control for time-invariant factors of domestic migration.

A second model uses net domestic migration as a percentage of the state’s total population for the dependent variable, and again uses disaster count and demographic controls as independent variables. I include this specification because more populous states tend to have larger absolute changes in net domestic migration. State- and year-invariant factors of domestic migration are again controlled for using dummy variables.

Data

My state-level annual data come from a variety of sources. The dependent variable, net domestic migration, is from releases of the Population Estimates Program, U.S. Census Bureau. Disaster counts arederived from FEMA expenditure data, which were provided by FEMA finance administrators. FEMA disaster expenditures are dependent upon a federal disaster declaration. Therefore, disaster counts are the product of a disaster declaration request by a state’s governor, and a subsequent approval of such declaration by the President of the United States (FEMA). Unemployment figures by state and year are from the Bureau of Labor Statistics “Local Area Unemployment Statistics” releases. Per capita personal income is taken from the Bureau of Economic Analysis’ “State Annual Personal Income” statistics. Education statistics are from the Economics and Statistics Administration of the U.S. Department of Commerce. Crime statistics are from statistical releases of the Federal Bureau of Investigation’s Uniform Crime Reports. State and local taxes are also from the U.S. Census Bureau. Finally, race/ethnicity andage distributions statistics are from the Current Population Survey.

Descriptive statistics for the state-level annual data discussed above are provided in Table 1. With one exception (violent crime) there are a total of 714 observations of each variable, over the years 1993-2006. Not all states reported violent crime data for every year of my sample frame. The mean of net domestic migration variable is zero by construction since a net domestic migration gain in one state must come at the expense of another. The relatively highstandard deviations of median house price and violent crimes per 100,000 people indicate that these controls vary significantly by state, year, or both.

A few variable pairs have notably high correlations. Specifically, median house price and per capita income (0.8074), property and violent crime levels (0.7138), college education levels and per capita personal income (0.7509), and college education levels and median house price (0.6720) are all highly correlated pairs of variables.

Results

The results of the fixed effect regression with net domestic migration as the dependent variable are presented in Table 2. The standard errors use White’s formula clustered at the state level to adjust for state-level heteroskedasticity. First, the number of disasters variable is positive but insignificant. Thus, this estimation suggests that migration patterns are not sensitive to disasters.

In other important findings, the unemployment rate estimate is negative and statistically significant. The coefficient of total population is positive and statistically significant, indicating that larger states attract more domestic migrants than smaller states. Finally, the age distribution controls suggest that an increase in the percent younger than 18 has a positive effect on net domestic migration, though not all controls are significant.

Table 3 presents results using percent domestic migration as the dependent variable. Similar to the net domestic migration estimation, the number of disasters has an insignificant impact on the dependent variable. Theunemployment rate remains negative and significant.

There are a few differences in the estimates between Tables 2 and 3. First, in Table 3 the coefficients of all the age group variables are negative and significant. Since the age group 0-17 is omitted, these results indicate thatall other age groups have a weaker positive or more negative impact on percent domestic migration. The coefficient of total population in Table 3 is no longer statistically significant, likely due to the incorporation of total population into the dependent variable by making net domestic migration a percentage of total population. Finally, in the second regression, percent Hispanic is negative and significant. All other independent variables are statistically insignificant.

Disastercount yields an insignificant coefficient for either dependent variable. Leading the dependent variable by one year for both dependent variables in separate fixed effect regressions yields similar insignificance. Thus, lagging the independent variables by one year relative to the dependent variable still yields an insignificant disaster count coefficient. According to these estimations, disasters as declared by FEMA do not affect domestic movement choices.

There are a number of potential reasons for this result. First, the independent variable representing disasters could be too broadly defined. Types of disasters included in the count variable include severe snowstorms, forest fires, droughts, and a number of other types of disasters that many would consider less fatal than other, more severe natural disasters. Furthermore, federal expenditures on declared disasters in the studied time period range from $0 for some minor disastersto over $25 billion for Hurricane Katrina in Louisiana. This vast range exemplifies the stark differences in the magnitude of declared disasters. It could be that severe natural disasters have a significant impact on domestic migration decisions. If this were the case, further investigation which limited disasters to only the most severe might yield a significant impact on population.

It is also possible that the concept of moral hazard explains the insignificant effect of natural disasters on migration. Generally defined, moral hazard is “the tendency of people to expend less effort protecting those goods that are insured against theft or damage.” (Frank and Bernanke, 390) Well before the first year considered in these regressions (1993), the federal government had developed a pattern of intervention in relief efforts for the victims of natural disasters. Given this social safety net, incentive structures to move from an area might have been materially altered. Such considerations are substantiated, for example, by the arguments ofBaen and Dermisi (2007), who assert on theoretical grounds that there is a positive relationship between federal relief programs and redevelopment patterns in a disaster area.

Alternatively, individuals could be moving within their state as a result of disasters rather than moving to another state. If this were the case, more geographically specific data are necessary. Unfortunately, the availability of such data, for a number of the control variables, and more importantly, FEMA disaster data, precludes such an investigation.

Finally, it is possible that my concerns about the length of time in Rossi, Wright, Wright & Weber-Burdin’s dependent variable may apply to my annual analysis as well. That is, a natural disaster might have a negative impact on a state’s population, but the state’s populationmay recover within a year.

Method 2—State-Monthly

My second method uses monthly data to determine if yearly data are too broad to analyze disaster effects on population changes. This requires some changes to my variable specification. First, monthly net domestic migration data are not available, but civilian non-institutional population data by state and month are. Therefore, total civilian non-institutionalpopulation is the dependent variable within my state-monthly model. Second, of all of the independent control variables included in Method 1, only the unemployment rate is readily available in state-monthly form. Along with dummy variables to indicate specific disasters or disaster seasons, these data constitute the independent variables of the state-monthly model.

Autocorrelation is a serious concern in monthly data, particularly for dependent variables like population change. In my estimations, Durbin-Watson tests suggest the presence of autocorrelation. I use an autoregressive model, which includes a lagged dependent variable to account for unexplained changes in civilian population. Furthermore, considering the broad nature of the “disaster count” variable in my fixed effect model, I study population impacts of specific severe disasters. I consider the largest disaster of the last 30 years—Hurricane Katrina, as well as the 10 most expensive hurricanes(in nominal terms) to hit Florida in the time period studied: Katrina, Andrew, Wilma, Charley, Ivan, Frances, Jeanne, Opal, Dennis, and Georges.

Data

Monthly statewide employment data—including civilian non-institutional population and the unemployment rate—arefrom the Bureau of Labor Statistics, Local Area Unemployment Statistics. Disaster-specific dummy variables were generated based upon the month and year that the storm was active in a given state.

Descriptive statistics of the state-level monthly data are included below in Table 4. There are a total of 21,684 monthly observations, for all 50 states plus the District of Columbia and a subset for New York City. The sample frame is January 1976 until June 2010. Since the dataset spans several economic expansions and recessions, the wide range of values of the unemployment rate is to be expected.