SOCIO-ECONOMIC DETERMINANTS OF TORNADO FATALITIES IN THE UNITED STATES  By  Jungmin Lim                      A THESIS  Submitted to Michigan State University in partial fulfillment of the requirements for the degree of  Agricultural, Food and Resource Economics - Master of Science  2016       ABSTRACT  SOCIO-ECONOMIC DETERMINANTS OF TORNADO FATALITIES  IN THE UNITED STATES  By  Jungmin Lim  Tornadoes are the most frequent of the natural hazards in the United States, causing significant human and economic losses every year. Given the potential destructive power of tornado events and their largely unpredictable nature, it is critical to identify the major determinants of vulnerability. To date, only a limited number of studies have empirically investigated the determinants of tornado-induced deaths. Based on a conceptual framework where risk is considered to be a function of physically defined natural hazards and socially constructed vulnerability, this study extends previous empirical studies by examining a wider range of potential socio-economic, governmental, and housing factors that determine tornado-induced fatalities. Using detailed county level data for years 1980-2014, I find that counties with higher per capita income and per capita government spending on public safety and welfare have fewer deaths, whereas counties with greater income disparity are more vulnerable to tornadoes. This study explores which aspects of poverty seem most associated with fatalities. Housing quality (measured by mobile homes as a proportion of housing units) is a critical factor in explaining tornado-induced fatalities.    iii ACKNOWLEDGEMENTS   I would like to sincerely thank my major supervisor, Mark Skidmore, for his patient guidance, encouragement, and precious advice he has provided throughout the development of this thesis. I have been extremely lucky to have such an incredible mentor, to have the opportunity to work with him, and to learn so much from him. I also gratefully acknowledge the help and insight from my thesis committee members: Dr. Scott Loveridge, Dr. Robert Shupp, and Dr. Soren Anderson. I appreciate their thoughtful comments and suggestions on this study.   I would not have been able to pursue graduate studies at Michigan State University without the love and support of my family. I want to thank my husband and best friend, Jongwoo, who has gone through all the ups and downs of the academic life at MSU with me together, and my son, Kyle, my pure happiness. I would also like to thank my parents and my family-in-law for their endless support, constant encouragement, and their faith in me.    iv TABLE OF CONTENTS  LIST OF TABLES .........................................................................................................................v  LIST OF FIGURES ..................................................................................................................... vi  1. INTRODUCTION .....................................................................................................................1  2. EMPIRICAL STUDIES ON THE DETERMINANTS OF DISASTER IMPACTS ................4  3. TORNADO RISK IN THE UNITED STATES ........................................................................8 3.1  Tornado Frequency ........................................................................................................8 3.2  Tornado Intensity ...........................................................................................................9  4. DETERMINANTS OF TORNADO VULNERABILITY.......................................................11 4.1  Motivation ...................................................................................................................11 4.2  Conceptual Framework................................................................................................12 4.3  Hypotheses on the Determinants of Tornado Vulnerability ........................................13 4.3.1  Income/Wealth and Income Distribution ..........................................................13 4.3.2  Gender and Female-Headed Households ..........................................................15 4.3.3  Human Capital...................................................................................................16 4.3.4  Housing Choice .................................................................................................17 4.3.5  Local Government Investment ..........................................................................19  5. EMPIRICAL ANALYSIS .......................................................................................................20 5.1  Data Description ..........................................................................................................20 5.2  Empirical Model ..........................................................................................................21  6. RESULTS ................................................................................................................................27 6.1  Richer counties experience fewer tornado-induced deaths .........................................29 6.2  Human capital plays an important role in reducing tornado vulnerability ..................30 6.3  Mobile homes residents experience more tornado fatalities .......................................33 6.4  Female-headed households are more vulnerable to tornadoes ....................................33 6.5  Government spending in public safety and welfare mitigates losses from tornadoes .34 6.6  Additional Analyses ....................................................................................................35   7. CONCLUSION ........................................................................................................................39  APPENDIX ..................................................................................................................................42  REFERENCES ............................................................................................................................45   v LIST OF TABLES  Table 1: Mobile Homes in the United States ...............................................................................17  Table 2: Tornadoes and Resulting Impacts by Fujita-scale (1980-2014) ....................................21  Table 3: List of dependent and explanatory variables in the model ............................................23  Table 4: Fatalities induced by Strong Tornadoes (F2-F5), 1980  2014 .....................................24  Table 5: County Summary Statistics............................................................................................26  Table 6: Socio-economic Characteristics and Disaster Impacts         Poisson Random Effect Regressions Results .................................................................31  Table 7: Socio-economic Characteristics and Disaster Impacts         Negative Binomial Random Effect Regressions Results ...............................................32  Table 8 : Socio-economic Characteristics and Disaster Impacts         Additional Regressions Results .....................................................................................38  Table A1: Socio-economic Characteristics and Disaster Impacts          Poisson Fixed Effect Regressions Results ...................................................................43     vi LIST OF FIGURES  Figure 1: Global Tornado Activity ................................................................................................8  Figure 2: Average Annual Number of Tornadoes during 1980-2014 ............................................9  Figure 3: Total Number of Strong/Violent Tornadoes (F2-F5), 1980-2014 ................................11  Figure 4: Total Number of Fatalities from Strong/Violent Tornadoes (F2-F5), 1980-2014 .......11  Figure 5: Proportion of Households Living in Mobile Homes, 2010 ..........................................18    1 1. INTRODUCTION Natural disasters such as tornadoes result in the significant loss of human life, as well as substantial economic damages. For example, in 2011 there were a record breaking 1,701 tornadoes in the United States resulting in 551 deaths (the most in the 62-year period for which we have records) and estimated total economic damages of over 28 billion U.S. dollars1. Given the recent demonstrations of the destructive power of tornado events and their largely unpredictable nature, improving our understanding of the factors that determine tornado-induced fatalities will help identify ways to potentially reduce losses. Surprisingly, to date there are relatively few studies that have empirically investigated the determinants of tornado impacts. This paper adds to this literature in several ways. First, this study considers a broader array of socio-economic factors that influence vulnerability. In particular, a range of alternative measures of poverty, including housing quality are considered. I also consider factors such as family structure as well as local government spending on emergency services.  As a prelude to full analysis, I find that counties with higher per capita income and per capita government spending on public safety and welfare have fewer deaths, whereas counties with greater income disparity and more female-headed households are more vulnerable to tornadoes. Perhaps of most importance, housing quality as measured by mobile homes as a proportion of housing units is a critical factor in explaining tornado-induced fatalities. It might seem that tornado fatalities are simply a function of location  living in an area with a high risk of tornadoes increases the chances that one would die from a tornado. While this is certainly true, other factors are also at play. Blaikie et al. (1994) argue that Disaster = Risk + Vulnerability,                                        1 NOAA National Climatic Data Center, State of the Climate: Tornadoes for Annual 2011, published online December 2011, retrieved on January 6, 2015 from http://www.ncdc.noaa.gov/sotc/tornadoes/2011/13.  2 where vulnerability depends on community and socio-economic variables in addition to location. Similarly, Cutter et al. (2003) discuss the interaction between social and biophysical vulnerabilities that determine overall place vulnerability. Overall, numerous scholars assert that underlying socio-economic factors such as poverty, access to social protection and security, as well inequalities with regard to gender, economic position, age, or race play an important role in determining disaster vulnerability (Aptekar and Boore 1990; Albala-Bertrand 1993, Cannon 1994, Blaikie et al. 1994; Cutter 1996; Enarson and Morrow 1998; Peacock et al. 1997; Morrow 1999).  A number of empirical studies of disasters sought to identify the major determinants of direct disaster impacts, where several focus on the role economic development plays in reducing disaster impacts using multi-national disaster data obtained from EM-DAT (Kahn 2005, Toya and Skidmore 2007, Stromberg 2007, Raschky 2008, Gaiha et al. 2013). Some of the above mentioned studies evaluate the role of governmental conditions and structure, inequality, and education in determining disaster impacts. I build upon a study by Simmons and Sutter (2013), which uses U.S. county level tornado data from 1984-2007 to evaluate factors that determine vulnerability. They find that tornado characteristics such as timing, magnitude, and length are the major drivers of tornado-induced fatalities, but also find that economic and demographic factors such as education, race, community, and housing type are important. As discussed in detail below, this study expands on Simmons and Sutter (2013) by using data from a longer period of time as well as considering a broader array of potential factors and, importantly, accounting for potential interactions between tornado severity and the socio-economic factors that determine vulnerability.  3 Based on a conceptual framework where risk is considered to be a function of physical natural hazard characteristics as well as socially constructed factors, the present study uncovers a number of the socio-economic variables that make people and places more vulnerable to tornadoes. For the empirical examination, panel structured tornado data are used with observations at the sub-national level - 3,107 U.S. counties2 over the 1980-2014 period. The detailed data on tornado events in U.S. counties are collected from NOAA, while socio-economic, housing, and local government fiscal data are obtained from U.S. Bureau of the Census. Taking into consideration that tornadoes are localized events as opposed to other more geographically dispersed disasters such as hurricanes, or earthquakes, the county level data (as opposed to aggregated national level data) allow us to more accurately identify and thus better understand the determinants of disaster vulnerability. By identifying the factors influencing tornado-induced fatalities, with particular focus on which dimensions of poverty seem to contribute most, this study provides insight that will help policy makers to better prepare for future devastating events and reduce societal vulnerability to disasters. The following section offers a review of the empirical literature regarding the determinants of the impacts of natural disasters. Section III discusses tornado risks in the United States, and section IV describes the underlying theoretical foundation for my analysis and introduces the primary hypotheses. Sections V and VI present the empirical framework of the analysis and empirical results, respectively.                                           2 Alaska and Puerto Rico are excluded.  4 2. EMPIRICAL STUDIES ON THE DETERMINANTS OF DISASTER IMPACTS  While many sociologists, geographers and other social scientists have studied how social, economic, and political factors potentially affect a society's vulnerability to natural disasters (Aptekar and Boore 1990; Albala-Bertrand 1993, Cannon 1994, Blaikie et al. 1994; Cutter 1996; Enarson and Morrow 1998; Peacock et al. 1997; Morrow 1999), most of these studies are qualitative in nature in that they use subjective identification rather than quantitative methods to suggest statistical evidence.  In addition, economists have studied the economic impacts of natural disasters, estimating the economic consequences of significant disaster events. However, there are relatively few quantitative empirical studies that investigate the underlying determinants of disaster impacts. This literature review focuses on research that empirically examines the major factors associated with the disaster-induced losses.  Many of these studies focus on the relationship between income/wealth and disaster impacts. The overall argument is that economic development plays an important role in mitigating the disaster vulnerability of a society. One of the first studies to identify this relationship (Burton et al., 1993) compares the post-disaster responses of high-income and low-income countries and finds that the consequences of natural disasters such as drought, floods and tropical cyclones differ across countries not only by hazard, but also by income. Horwich (2000) draws a similar conclusion, to disaster is its level of wealth. He explains that a rise in income will provide not only general safety but also improved protection from natural disasters.   5 Many of the more recent empirical studies that examine the determinants of disaster vulnerability have been cross-national and use disaster data obtained from EM-DAT3. For instance, Kahn (2005) uses this data source to examine the relationship between disaster-induced death and explanatory factors such as income, geography, and national institutions in the context of multiple types of natural disasters in 73 nations from 1980 to 2002. He finds that while a n disaster events it experiences, higher levels of development reduce disaster-induced deaths. Kahn estimates that an increase in per capita GDP from $2,000 to $14,000 results in a reduction in natural disaster deaths from 9.44 to 1.80 per million people per year. He also finds that democracies and nations with less income inequality suffer fewer deaths from disasters. -safety-development relationship by including other socio-economic measures. Specifically, they use disaster impact data from EM-DAT and several other sources for 151 countries over 44 years (1960-2003). Their study confirms that economic development as measured by per capita GDP is inversely correlated with both disaster deaths and damages. However, they also find that higher levels of educational attainment, greater openness, and a stronger financial sector are also associated with fewer deaths and less damage. Other studies corroborate and expand on the cross-country link between economic development and disaster outcomes. For instance, Anbarci et al. (2005) in their study of earthquakes show that greater income inequality increases earthquake fatalities. Raschky (2008)                                        3 Emergency Events Database EM-DAT that has been maintained by the Centre for Research on the Epidemiology of Disasters (CRED) contains essential core data on the occurrence and effects of mass disasters in the world from 1900 to present.  6 also shows that economic development reduces disaster fatalities and losses, but this relationship is nonlinear. Economic development decreases disaster losses but with a diminishing rate. Kellenberg and Mobarak (2008) find a similar relationship between economic development and disaster vulnerability with losses increasing at first and then declining as GDP rises. Raschky also incorporates a national government stability measure and finds that more stability is associated with fewer losses. Similarly, Stromberg (2007) finds that greater wealth and government effectiveness (World Bank, 2006) are associated with fewer disaster fatalities. Finally, Gahia et al (2013) find that poorer and larger countries suffered more disaster related fatalities, but that experience from past disasters and more resources targeted to disaster prevention and mitigation can dramatically reduce deaths.  One cross-country study that does not find a significant link between GDP/income inequality and disaster vulnerability is Brooks et al. (2005). In an effort to develop national-level indicators of vulnerability and present a set of socio-economic, political and environmental variables that correlate with mortality from disasters, they include many additional socio-economic factors beyond GDP into their analysis. They find that including factors such as sanitation, life expectancy, government effectiveness, and literacy are significant predictors of disaster fatalities, whereas GDP and income inequality are not. However, their significant factors may serve as proxies for GDP.  As noted earlier, most of the research discussed above incorporates multiple types of natural disasters across multiple countries and relies primarily on the multi-national EM-DAT data set as their source of information on disasters and their impacts. In contrast, this study focuses on a specific disaster type within a single country. As previously noted, the study most closely related to my study is that by Simmons and Sutter (2013); they employ detailed U.S.  7 county level tornado data from National Oceanic and Atmospheric Administration (NOAA) over the period 1984-2007 to examine the societal impacts of tornadoes. In this book, the authors examine the patterns in tornado casualties over time, by state and Fujita Scale rating, and provide a regression analysis on the potential determinants of tornado casualties. Using a Poisson estimation method, they show that not only do the elements of tornado hazards (timing, magnitudes, and length of incidence) determine tornado impacts, but that economic and demographic factors such as level of education, percentage of non-white and rural population, and percentage of mobile homes contribute to tornado vulnerability. However, the authors offered little evidence that income, poverty and income distribution were important determinants of disaster impacts. The present study extends this line of research by examining a wider range of potential socio-economic factors using U.S. county level data over the 1980-2014 period.    8 3. TORNADO RISK IN THE UNITED STATES 3.1  Tornado Frequency  As shown in Figure 1, the United States is the most tornado-prone country worldwide, with an average of 1,200 recorded tornado events each year. Canada is a distant second with around 100 tornadoes per year.4 Focusing on the United States, the average annual number of tornadoes (all intensities) by state for years 1980-2014 is presented in Figure 2. The darker green area shown in Figure 2 spanning from Texas to South Dakota is called "Tornado Alley"5 because of the disproportionately high frequency of tornadoes. Figure 1.  Global Tornado Activity                                           4 NOAA National Climatic Data Center, U.S. Tornado Climatology, retrieved on November 6, 2014 from http://www.ncdc.noaa.gov/climate-information/extreme-events/us-tornado-climatology 5 Although the boundaries of Tornado Alley are not clearly defined, for this analysis I define the states of Texas, Oklahoma, Kansas, Colorado, Nebraska, South Dakota, Iowa, Illinois, Missouri, and Arkansas as the Tornado Alley.  9 3.2  Tornado Intensity  In addition to tornado frequency, the magnitude and intensity of tornadoes are also important in determining impacts. According to National Climatic Data Center (NOAA), over the 1950 to 2010 time period the vast majority of tornadoes (about 77%) in the United States were categorized as weak (i.e., Fujita Scale6 F0 or F1). Thus, nearly a quarter of tornadoes are classified as significant or strong/violent (F2 and above), with only 0.1% achieving F5 status (winds over 200 mph, resulting in near complete destruction of everything in its path). Given that, on average, about 1,200 tornadoes occur in the United States each year, about 276 will be classified as strong/violent, with perhaps one being F5. These strong/violent tornadoes account for the vast majority of tornado-induced fatalities and damage. For example, in May of 2013, a severe tornado produced catastrophic damage in Moore, Oklahoma and adjacent areas.  Figure 2:  Average Annual Number of Tornadoes during 1980-2014                                         6 Note that in 2007-2008 NOAA introduced and began using the Enhanced Fujita scale for measuring tornado intensity. We use the term Fujita scale throughout the paper since the majority of the data falls under this category.  10 This F5 rated tornado was the most deadly and devastating tornado of the year, claiming 24 lives and injuring 377 people. The tornado destroyed approximately 1,150 homes, and caused more than $2 billion in damage (Insurance Journal, 2013). Another recent example is the tornado outbreak that occurred during April 2528, 2011. This 4-day period included hundreds of tornadoes that struck communities across the southern plains and southeastern United States and was the largest and the deadliest tornado outbreak since formal record keeping began in 1950. In total, the National Weather Service (NWS) confirmed 351 tornadoes of which four were rated F5. In the four-day period 316 people died, more than 2,400 were injured, and economic damages totaled over $4.2 billion7.                                           7 National Oceanic and Atmospheric Administration. Service assessment: the historic tornadoes of April 2011. Silver Spring, MD: U.S. Department of Commerce, National Oceanic and Atmospheric Administration; 2011. Available at http://www.nws.noaa.gov/om/assessments/pdfs/historic_tornadoes.pdf.  11 4. DETERMINANTS OF TORNADO VULNERABILITY 4.1  Motivation  While it is clear that some places are simply more prone to tornadoes due to climactic reasons, this does not fully explain the differences in fatalities across the regions. For example,  Figure 3:  Total Number of Strong/Violent Tornadoes (F2-F5), 1980-2014  Figure 4:  Total Number of Fatalities from Strong/Violent Tornadoes (F2-F5), 1980-2014   12 Figure 3 and 4 shows the differences between tornado frequencies and fatalities. The map in Figure 3 presents the total number of F2 or higher rated tornadoes (strong/violent) over the period 1980  2014 by state, whereas the map in Figure 4 shows total fatalities from these tornadoes over the same period. As is clear, the areas with relatively high tornado fatalities do not necessarily match up with the areas with the highest tornado intensities. For example, though tornado activity is relatively modest in Missouri, this state experienced a relatively high number of fatalities per year. The present research is in part motivated by this observation. Note that these differences could be driven by many things including that there may have been a higher ratio of violent (F4 and F5) events in Missouri relative to say Texas. My analysis below takes this into account and yet I still find significant evidence that specific socio-economic factors appear to be, at least in part, driving these differences. 4.2  Conceptual Framework As highlighted earlier, Cutter et al. (2003) discuss the possible interactions between social and biophysical vulnerabilities that determine overall place vulnerability. They explain that the hazard potential is either moderated or enhanced via a combination of geographic factors hazards, and its ability to respond to, cope with, recover from, and adapt to hazards, which in turn are influenced by socio-economic status, demographics, and housing characteristics. In their model, disaster fatalities are largely determined by socio-economic factors that shape a  Similarly, Blaikie et al. (1994) note that vulnerability, in the disaster context, is a person's or group's "capacity to anticipate, cope with, resist, and recover from the impact of a natural  13 group's vulnerability against natural hazard is shaped by human components (O'Keefe et al. 1976; Hewitt 1983). In the same vein, Cannon (1994) asserts that economic systems and class structures allocate income and access to resources, and this affects people's ability to cope with and recover from hazards. In general, it has been argued by many scholars that structural factors such as poverty, access to social protection and security, and inequalities with regard to gender, economic position, age, or race, cause or exacerbate vulnerability (Cannon 1994, Aptekar and Boore, 1990; Albala-Bertrand 1993, Enarson and Morrow 1998; Peacock et al., 1997; Morrow 1999). Fothergill et al. (2004) point out that -analyses of minority, gender, and inequality issues in the context of disasters. 4.3  Hypotheses on the Determinants of Tornado Vulnerability Based on a conceptual framework where risk is considered to be a function of physically defined natural hazards and socially constructed vulnerability, this study seeks to identify key elements of tornado fatalities through empirical analysis using detailed data on tornado events and socio-economic data for 3,107 U.S. counties from 1980 through 2014. In addition to controlling for primary factors such as county population, lagged tornado frequency, and tornado magnitude (Fujita scale), I hypothesize that there are a number of demographic, socio-economic, housing, and governmental factors that may also play significant roles in determining tornado-induced deaths. 4.3.1  Income/Wealth and Income Distribution First, as one of the well-known determinants of disaster impacts. The robustness of the hypothesis is tested that the level of community's income/wealth plays significant role in  14 vulnerability of disasters. Researchers such as Wildavsky (1988) contends that greater income and wealth translates to a safer society. Safety can be viewed as a natural product of a growing market economy since higher income places have a higher demand for safety and more resources to invest in risk reduction measures, which in turn leads to reduced vulnerability to disasters. The income/wealth hypothesis has been supported by many empirical studies (Kahn 2005, Toya and Skidmore 2007, Stromberg 2007, Raschky 2008, Gaiha et al. 2013). Note that these studies use cross-country data where GDP is used as a measure of income/wealth, whereas in this study, U.S. county per capita income is used. In addition to per capita income, I also include the county top ten percentile income level and county poverty rates in my analysis as measures of income distribution. If income distributions are similar across all counties and over time, the top ten percentile income level measure should be closely correlated with per capita income. However, since income disparity in the United States has increased over the sample period and more so in some counties than others, I speculate that controlling for per capita income, the top ten percentile income variable will capture the role income disparity plays in determining disaster vulnerability. Similarly, I hypothesize that societies with a higher concentration of poverty might encounter higher tornado-induced human losses. According to Fothergill et al. (2004), the poor in the United States are more vulnerable to natural disasters due to such factors as place and type of residence, building construction, access to information, low quality infrastructure, and social exclusion. Furthermore, Moore (1958) highlighted the relationship between socio-economic status and warning response, reporting that lower income groups were less likely to take the warnings of impending natural disasters seriously. Gladwin and Peacock (1997) reported in their study of warnings and evacuation for Hurricane Andrew that lower income people were less able and thus  15 less likely to evacuate, mostly due to constraints placed by a lack of transportation and affordable refuge options. Similarly, an empirical study of natural disasters in Fiji, (Lal et al., 2009) finds evidence that the level of poverty (measured by the HDI) negatively affects disaster outcomes. The authors argue that those living in poverty are more sensitive to disasters because they have lower economic and social conditions; that is, they are unable to invest in adequate preparedness and risk reduction measures. 4.3.2  Gender and Female-Headed Households I also hypothesize that female-headed households are likely to be among the most vulnerable. According to the 2012 Census, families headed by a single adult are more likely to be headed by women, and these female-headed families are at greater risk of poverty and deep poverty; 30.2% of families with a female householder where no husband is present were poor and 16.9% were living in deep poverty. In addition, a study by Neumayer and Plumper (2007) suggests that for both social and physiological reasons, females are more vulnerable in disaster situations than men and therefore suffer higher mortality rates. While the this study attempts to shed light on the direct impacts of disasters on female-headed households, the vulnerability of female-headed households in a longer-run framework is highlighted in the literature. Researchers focusing on post-disaster outcomes indicate the degree of disaster impacts vary by gender not only in terms of direct physical loss, but also during the periods of emergency response, recovery, and reconstruction. For example, Blaikie et al. (1994) argue that women have a more difficult time during the recovery period than men, often due to sector-specific employment, lower wages, and family care responsibilities. Similarly, two years after Hurricane Andrew, thousands of poor families headed by minority women were still living in substandard temporary housing (Morrow and Enarson, 1996).  16 4.3.3  Human Capital The third hypothesis is that human capital as measured by percentage of population aged 25 and over holding  is one of the major characteristics defining social vulnerability. Several cross-country studies found significant correlations between level of educational attainment and reduced fatalities (see Skidmore et al., 2007). Education attainment is linked to the emergency decision-making process; education influences one's ability to understand warning information and perform evacuation or other necessary actions. Cutter et al. (2003) explain that while education is clearly linked to socio-economic status (higher educational attainment resulting in greater lifetime earnings), lower education may also constrain the ability to understand warning information and access to recovery information. Additionally, they argue that those with higher levels of education are more likely to choose safer locations and homes constructed with more durable materials, thus resulting in fewer fatalities. In a recent study, Muttarak and Lutz. (2014) argue that education can directly influence risk perceptions, skills and knowledge and indirectly reduce poverty, as well as promote access to information and resources. These factors contribute to higher adaptive capacity and vulnerability reduction. The authors collect empirical evidence from a series of studies contained in a special issue aimed at investigating the role of education in vulnerability reduction; the authors provide consistent and robust findings on the positive impact of formal education in reducing vulnerability.     17 4.3.4  Housing Choice The fourth hypothesis is that communities with a higher proportion of households living in mobile homes or trailers will suffer increased levels of tornado casualties. Aptekar (1991) argues that it is more likely that disasters adversely affect those with lower socio-economic status largely because of the types of housing they occupy. Logically, people living in mobile homes are more vulnerable to natural events such as tornadoes because mobile homes typically have no foundation or basement and can more be easily destroyed. From 1996 to 2000, about half of tornado-induced deaths in the United States were in mobile homes8, even though mobile homes accounted for less than 8% of the nation's housing during the same period, according to the National Oceanic and Atmospheric Administration and the U.S. Census Bureau. Historical data on tornado fatalities (1975-2000) tell us that the rate of death from tornadoes in mobile homes is about 20 times higher than that in site-built homes7. Table 1:  Mobile Homes in the United States Year Mobile Homes (%) in U.S. housing units Total Mobile Homes in U.S. housing units Total U.S. housing units 1950 0.7% 315,218 45,983,398 1960 1.3% 766,565 58,326,357 1970 3.1% 2,072,887 68,679,030 1980 5.1% 4,401,056 88,411,263 1990 7.2% 7,399,855 102,263,678 2000 7.6% 8,779,228 115,904,641  2010* 6.7% 8,684,414 130,038,080 Source: U.S. Census Bureau, Housing and Household Economic Statistics Division  *2010 data are estimates produced by American Community Survey while data for years 1950-2000 are from Decennial Census.                                        8 Brooks, H., & Doswell III, C. A. (2001). A brief history of deaths from tornadoes in the United States. Weather and Forecasting, 1-9. http://www.nssl.noaa.gov/users/brooks/public_html/deathtrivia/  18  As shown in Table 1, the proportion of households living in mobile homes increased significantly since 1950. While the quality of these homes is probably higher than in the past, they still lack structural characteristics (e.g. foundations and basements) that make other types of construction more resistant to tornadoes. Importantly, mobile home living is very high in many rural counties across the Unites States. As shown in Figure 5, in 2010 many rural counties had more than a third of households living in mobile homes. The increase in the U.S. population living in mobile homes is likely to have important policy implications for disaster management in the context of tornadoes and other high wind events (Brooks 2001, Merrell et al. 2005, Kusenbach et al. 2010, Fothergill and Peek 2004, Schmidlin et al. 2009).  Figure 5:  Proportion of Households Living in Mobile Homes, 2010      19 4.3.5  Local Government Investment  My last hypothesis is that communities where local governments invest more resources in safety, protection and welfare will experience fewer fatalities. This type of expenditure number is not readily available so I construct a measure of government spending on public safety and welfare by aggregating local government expenditures on fire/police protection and protective inspections/regulations and housing/community development, and public welfare. Local government resources devoted to public safety services such as fire/police protection and protective inspection and regulation should lead to better preparedness and faster responses to disaster events, which, in turn, may play critical roles in reducing fatalities. It is also possible that allocating more resources to public welfare may reduce disaster vulnerability. In the context of local government, welfare services are not direct cash assistant (this comes from state ance abuse treatment and the like.      20 5. EMPIRICAL ANALYSIS 5.1  Data Description  The county level panel data in the analysis consists of: (1) data on tornadoes from NOAA (1980-2014) used to develop detailed tornado information on locations, magnitudes and deaths, (2) data from U.S. Decennial census of population for the major socio-economic and housing factors in 3,107 counties from 1980 to 2010, and (3) local government fiscal data from the U.S. Census of Governments (1982 to 2012). Note that the Census of Population data are only available every ten years, whereas local government fiscal data are reported every five years (years ending in 2 or 7). Also, since, at the county level, the tornado data has many zero observations, the panel data is organized such that it contains county level tornado observations across seven time blocks between 1980-2014 (in five year intervals) : '80-84, '85-'89, '90-'94, -'99, '00-'04, '05-'09, '10-'14. The detailed tornado data are aggregated and rearranged to form county level observations and the tornado variables are averaged over each time block and are demographic and housing variables are interpolated to obtain data in 1982, 1987,.., 2012. Lastly, averaged tornado data and the interpolated census data are merged with the local government fiscal data. Overall, seven time-blocks are constructed for each of the 3,107 counties9. Thus, the unit of observation of this study is counties, not tornado event.  When I average tornado data across time blocks, I include only strong/violent tornadoes rated F2 or greater for the main analysis or, for the additional analysis F3 or greater. Accordingly,                                        9 Given that county level socio-economic variables are only available every ten years, I use averaged tornado data in time intervals to avoid using interpolated data for all the socio-economic variables for all years except for years ending in 0, and interpolated government fiscal data for most time periods as well. By having a county as a unit of observation in this study, I am able to retain and explore a long-term variation in county socio-economic and government fiscal factors more accurately whose role in disaster events is the main interest of this study.  21 the dependent variable is the average number of deaths10 caused by tornadoes rated F2-F5 (or F3-F5 in additional analysis). As noted earlier and shown in Table 2, most tornadoes are classified as F0 or F1 and those tornadoes commonly lead to very few deaths or do not claim lives at all. Since these types of tornadoes are effectively non-disasters, they are excluded for the analysis. As a result, county level panel data for my empirical estimation contains 2,120 counties that have experienced tornadoes of F2+ at least once over the study period. Table 2 presents the total number of tornadoes and resulting fatalities and injuries by F-scale over the years 1980-2014. Table 2:  Tornadoes and Resulting Impacts by Fujita-scale (1980-2014)*  Tornado Fatalities Injuries F-scale Obs. % Total Avg. Total Avg. F0 22,028 51.31  12 0.001 536 0.024 F1 11,977 27.90  128 0.011 3,945 0.329 F2 3,907 9.10  330 0.084 8,427 2.157 F3 1,193 2.78  880 0.738 13,586 11.388 F4 301 0.70  869 2.887 13,055 43.372 F5 27 0.06  639 23.667 4,567 169.148 Total 42934 100 2447 0.057 39877 0.929 * Only F2-F5 tornadoes are examined in this study.   5.2  Empirical Model  The dependent variable in this analysis is the average number of fatalities per tornado and thus, non-negative value. I employ Poisson model which properly treats the non-negative                                        10 For example, a county A experienced two tornadoes each rated F2 and F0, having fatalities of 3 and 0 respectively, in a time block B, then county A in year B is assigned 3 for its average fatalities per tornadoes F2 or higher. I exclude and do not count F0 and F1 tornadoes when I generate Avg. Fatalities_F2-F5 or Avg.Fscale_F2-F5 variables.  22 variables within the county level panel data framework (Wooldridge, 1991)11. Also, considering the large portion of zeros in the dependent variable, I repeat the analysis using a Negative Binomial model as a robustness check. In this study, many of the county socio-economic characteristics do not change much over time. Thus, there is little within-county variation for many of the explanatory variables. Given this, the fixed effects model is not necessarily preferred to random effects model.12 In his multi-national disaster study, Kahn (2005) points out the presence of sluggish adjustment and long latency in economic development, which makes the inclusion of country fixed effects problematic. Taking the same stance as Kahn, I estimate the model using both random and fixed effects Poisson, but mainly discuss the random effects estimates.13  The regression analysis is characterized by the following equation:  where  is the average deaths per tornado in county during time block ,  is a vector of socio-economic and housing variables affecting deaths in county at time , is local government spending on public/safety,  is the dummy variable for Tornado Alley,  is the average F-scale or the share of tornadoes of each F-scale levels (F2-F5) occurred in a county at time ,  is the number of tornadoes in county at time ,  represents a series of time indicator variables,  is a time-invariant effect for county j, and  is the unobservable error term. The detailed explanation for the variables in the model is provided in Table 3.                                        11 The dependent variable is an average value and can be non-integer. However, the Poisson (quasi-MLE) model is robust to distributional assumptions; it can be applied to any nonnegative outcome, either continuous or integer valued (Wooldridge, 1991). 12 Wooldridge (2010) also discusses that when the key explanatory variables do not vary much over time, fixed effects methods can lead to imprecise estimates. 13 The result of Fixed Effects Poisson is presented in the Appendix.  23 Table 3:  List of dependent and explanatory variables in the model Dependent Variable   Avg. Deaths from tornadoes  Explanatory Variables  Demographic Log (Population size)  Log ( Land Area) Percent of population over 65 Percent of population under 18 Percent of people aged 25 and over holding Bachelor's degree Percent of female-headed households Economic Log (Per capita Income) Log (Top 10 percentile income level) Poverty rate Housing Percent of mobile homes in total housing units Government Log (Local government expenditures on public safety/welfare)  Tornado Magnitude of tornadoes (Avg. magnitude OR Percent of tornadoes of F2, F3, F4, and F514)  Lagged tornado frequency of F2+ Tornado alley  Time Dummy 1987, 1992, 1997, 2002, 2007, 2012     Table 4 shows that over the 35 years from 1980 to 2014, a total of 5,428 tornadoes of F2 or greater occurred and caused 2,718 deaths and 39,635 injuries; 4,733 of these tornado events resulted in zero fatalities (Table 4). I aggregate tornado data into the aforementioned five-year intervals and form a panel structure. The county level panel data for this study contains 4,757 county-year observations15 with at least one strong/violent tornado rated F2 or higher and 1,016 observations had fatalities from those events. Using these data, I estimate equation (1) using a Poisson and Negative Binomial estimation procedures.                                        14 For a robustness check, I repeat my analysis using the percent of tornadoes of each F-scale among F2-F5  tornadoes that occurred ( or among F3-F5 tornadoes for severe tornado analysis), instead of using the average F-scale as in my main analysis. The result is presented in Table 8. 15 County-year observations without any experience of tornadoes of F2+ are excluded.  24 Table 4:  Fatalities induced by Strong Tornadoes (F2-F5), 1980-2014* Fatalities Freq. Percent 0 4,733 87.20  1-5 577 10.63  6-15 86 1.58  16-30 26 0.48  31-158 6 0.11  Total 5,428 100.00  * For this information, yearly tornado data from NOAA is used. However, this study exploits a panel data with county-year observations.   Eight specifications are estimated to test my hypotheses. The dependent variable is the average number of deaths per tornado (of Fujita Scale 2-5) in each county in a particular time block. Some of the socio-economic determinants are highly correlated with each other, which may result in multicollinearity. To address this possibility, I conduct preliminary analyses using more parsimonious model specifications as shown in columns (1) to (7) of Table 6 and 7. Each hypothesized potential determinant of tornado impacts  for example, poverty rate, education level, female-headed household, and mobile homes  are examined separately but with a consistent set of control variables. Given that many prior studies found income level to be one of the most important factors, per capita income is included in every specification. Government spending on public safety and welfare also appears in every specification because this is the only variable that represents the role of government, although government spending might be weakly related to the economic variables discussed above. The last specification includes all the poverty-related potential determinants, testing them in a single specification. In all specifications the following variables are included as controls: average tornado magnitude, population size, land area, percent of population over age 65 and under 18, lagged tornado frequency, and a categorical variable for counties located in the Tornado Alley region.  25  The EM-DAT data used in most of the prior studies discussed do not contain information on disaster magnitude on many of the recorded disaster events, so most studies using those data are unable to control for disaster magnitude. The tornado data from NOAA, however, does provide a magnitude measure for each tornado (F-scale), and thus I can more readily distinguish impacts on fatality due to disaster magnitude versus other explanatory variables I wish to explore. Specifically, I use the average magnitude of all tornadoes of F2-F5 that occurred in a particular county in a given period because the unit of observation of this study is counties, not tornado event.   Also, considering that Tornado Alley regions are more highly prone to tornadoes than other regions, I introduce a dummy variable in the model. (if the county j is in this geographic region and , otherwise) along with lagged tornado frequency of F2-F5 (or F3-F5 in additional analysis on severe tornadoes). These variables allow us to test whether greater familiarity with this type of emergency makes the area more able to cope (e.g., building codes, population behavior during the event).       26 Table 5:  County Summary Statistics  Mean Standard Deviation Min Max Number of Obs. Dependent Variables        Avg. Tornado Deaths (F2-F5) 0.29  1.34  0  52.67  4757   Avg. Tornado Deaths (F3-F5) 0.77  2.45  0 52.67  1884 Independent Variables        Avg. Fscale (F2-F5) 2.40  0.58  2  5  4757   Avg. Fscale (F3-F5) 3.25  0.44  3 5 1884   Pct Tornado of F2 68.03  42.78  0 100 4757   Pct Tornado of F3 24.50  39.31  0 100 4757   Pct Tornado of F4 6.95  22.91  0 100 4757   Pct Tornado of F5 0.53  5.76  0 100 4757   Lagged Freq. of F2-F5 0.58  0.96  0  9  4757   Lagged Freq. of F3-F5 0.20  0.53  0 5 1884   Tornado Alley Dummy 0.44  0.50  0  1  4757   Log (Land Area) 6.46  0.52  3.13  9.91  4757   Log (Population) 10.38  1.30  4.37  15.91  4757   Pct Over 65 14.01  3.93  3.06  35.99  4757   Pct Under 18 26.01  3.28  11.20  45.16  4757   Log (Per Capita Gov Expenditure        on Public Safety & Welfare) -1.55  0.70  -5.90  1.11  4757   Log (Per Capita Income) 9.79  0.25  8.80  10.93  4757   Log (Top 10% Income ) 11.52  0.28  10.73  12.07  4757   Poverty Rate 15.97  6.90  0  58.18  4757   Pct BA Degree 15.04  6.90  4.12  55.35  4757   Pct Mobile Home 12.46  8.05  0.05  57.21  4757   Pct Female-Headed Household 10.54  4.30  2.88  35.46  4757 * Statistics are from observations with F2-F5 tornado experience that are used for the main regressions. For the additional regressions using severe tornadoes of F3-F5, only tornado statistics (Avg. Tornado Deaths, Avg. Fscale) are presented.     27 6. RESULTS  Table 6 and Table 7 presents the results of the regressions using F2 or higher tornado observations recorded in counties over 1980-2014 and a set of demographic, socio-economic, housing, and government fiscal factors as presented in Table 5. I mainly discuss the results of Random Effects Poisson and Negative Binomial16 specifications here; however, the Fixed Effects specification estimates outcomes are provided in the Appendix for the interested reader.  Before discussing the primary findings as they relate to the hypotheses, consider the estimated effects of the control variables. The F-scale variable which is an indicator of the average magnitude of tornadoes within a given time period, has a strong association with the number of deathsin all specifications. As expected, the analysis confirms the magnitude of the tornado is a critical physical determinant of the tornado fatalities. The estimated coefficient of the average F-scale in column (8) in Table 7 implies that an increase in F-scale to the next level increases expected tornado fatalities by a factor of 4.21 (. Both lagged tornado frequency and tornado alley variables are estimated to be negatively correlated with fatalities in all specifications. Counties in tornado alley region who experience tornadoes relatively often are estimated to experience 13% ( lower fatalities than counties outside of the tornado-prone area, all other conditions being equal. This result supports the idea that there might be some kind of learning effects from risk history, where counties that suffered more tornado outbreaks tend to put more efforts to reduce their vulnerability and be better prepared for disasters and in turn, better able to mitigate the societal impacts. McEntire (2001) asserts that                                        16 I discuss both Poisson and Neg. Binomial regressions results here, however, t (dispersion parameter) = 0 strongly rejects the null hypothesis that the errors do not exhibit overdispersion. Thus, the Poisson regression model is rejected in favor of its generalized version, the Neg. Binomial regression model. When explaining the estimated effects of explanatory variables, I refer to the results of Neg. Binomial model in Table 7.  28 beliefs and activities play a major role in the creation of vulnerabilities and past disaster lessons reduce future consequences.  As a measure of density, both county population and land area are included in logarithmic terms17. The results show that counties with greater populations and smaller land area experience more deaths when tornadoes strike - together implying the higher the density, the larger the tornado impacts. The estimates suggest that for two counties of equal land area, if one has 10 percent more population, the expected fatalities increase by 4 percent. Also, as a control, proportions of the population over the age of 65 and under 18 are included. In all estimates it is shown that counties with greater proportions of elder and young experience fewer fatalities. In my initial assessment I expected that these population groups would be more vulnerable rather than less. One possible explanation is the older people and families with children may be more risk averse and thus heed tornado warnings, thus reducing exposure. It could also be caused by higher proportions of these individuals being in environments (schools, retirement communities) where warnings are more easily distributed.  Let's now turn to the primary interest in the role that the various dimensions of poverty, and social vulnerability play in determining tornado impacts. I begin this portion of the discussion by considering the factors that align with my first hypothesis regarding the role of income/wealth in determining vulnerability.                                          17 Note that Population Density=Population/Land Area. Also, Log(Density)=Log(Population)-Log(Land Area). Thus, the estimated coefficients of Log(Population) and Log(LandArea) variables are similar in magnitude but opposite in sign.  29 6.1.  Richer counties experience fewer tornado-induced deaths    Consistent with most other empirical studies, I find that per capita income is a key determinant of tornado-related deaths. The negative relationship between income and tornado fatalities is significant and robust in both Poisson and Negative Binomial models, indicating that higher county per capita income results in fewer tornado-induced fatalities. The estimated coefficient on the log of per capita income suggests that a one percent increase in county per capita income is expected to reduce tornado fatalities by one percent18. As Anbarci et al. (2005) and Kahn (2005) argued in their studies, it is also found in this study that income distribution (as measured by the top ten percentile income level) a significant factor. Holding other factors constant, per capita income and the poverty rate, higher top ten percentile income level means larger share of lower-middle income group, which indicates wider income disparity in the community. The estimates suggest that greater income inequality tends to exacerbate the impacts of disasters. In addition, controlling for income, the poverty rate is not a statistically significant factor. However, this result is largely due to multicollinearity as per capita income and the poverty rate are highly correlated. Consider the estimates in column 4 in both Table 6 and 7, where the poverty rate is included but not income per capita in the specification. In this regression we see that the poverty rate is positive and statistically significant as expected. The estimated coefficient in column 4 in Table 7 suggest that one percentage point increase in poverty rate is estimated to increase tornado fatalities by 3percent.                                           18 The estimated coefficients of log transformed variables can be interpreted as elasticities.  30 6.2.  Human capital plays an important role in reducing tornado vulnerability  The regression results indicate that human capital as measured by the proportion of the population aged 25 and over with a Bachelor (or higher) degree is also a significant determinant of tornado fatalities. As presented in specifications (5) and (8), the percent of bachelor degree holders is found to be negatively associated with the likelihood of deaths in disaster situations, though only statistically significant in specification (5). A one percentage point increase in the proportion of the Bachelor degree holder in a county is associated with 1.6 percent reduction in expected tornado fatalities. Educational attainment may be linked to emergency decision-making processes such as the ability to quickly comprehend warning information and perform evacuation or other necessary actions or to have work functions located inside, with more solid construction (e.g., office building versus pole barn). Thus, those with lower education attainment may be more vulnerable to disaster shocks. The estimated results are consistent with previous studies (e.g., Skidmore et al., 2007, Muttarak and Lutz, 2014). However, again, education and other economic variables such as income levels and poverty measures are highly correlated; thus, the insignificance of education in column (8) is likely the result of multicollinearity.    31 Table 6:  Socio-economic Characteristics and Disaster Impacts Poisson Random Effect Regressions Results  Dependent variable: Deaths from F2-F5 tornadoes Independent variables (1) (2) (3) (4) (5) (6) (7) (8) Fscale_F2+ 1.551*** 1.550*** 1.553*** 1.555*** 1.550*** 1.575*** 1.554*** 1.576***  (0.061) (0.061) (0.061) (0.060) (0.061) (0.060) (0.061) (0.061) Lag_Tornado_F2+ -0.003 -0.005 -0.004 -0.003 -0.002 -0.003 -0.006 -0.006  (0.045) (0.045) (0.045) (0.045) (0.045) (0.046) (0.045) (0.046) Tornado Alley -0.461*** -0.457*** -0.454*** -0.467*** -0.432*** -0.213** -0.421*** -0.206*  (0.105) (0.105) (0.106) (0.106) (0.107) (0.107) (0.107) (0.108) Log(Land Area) -0.039 -0.045 -0.037 -0.017 -0.018 -0.131 -0.016 -0.142  (0.094) (0.095) (0.094) (0.092) (0.095) (0.095) (0.095) (0.098) Log(Population) 0.323*** 0.322*** 0.312*** 0.260*** 0.337*** 0.443*** 0.294*** 0.432***  (0.059) (0.058) (0.060) (0.054) (0.059) (0.059) (0.060) (0.062) Pct Over65 -0.044** -0.034 -0.043** -0.042** -0.053** -0.014 -0.039* -0.001  (0.020) (0.022) (0.020) (0.020) (0.022) (0.019) (0.020) (0.024) Pct Under18 -0.042** -0.044** -0.044** -0.042** -0.048** -0.012 -0.047** -0.011  (0.021) (0.021) (0.021) (0.021) (0.022) (0.021) (0.022) (0.023) Log(PerCapita GovtExp onPublicSafetyWelfare) -0.317*** -0.320*** -0.324*** -0.355*** -0.309*** -0.221** -0.344*** -0.232** (0.085) (0.086) (0.085) (0.084) (0.086) (0.086) (0.090) (0.092) Log (PerCapita Income)  -1.367*** -1.916*** -1.020**  -1.058*** -0.539 -1.013*** -1.041 (0.304) (0.527) (0.489)  (0.357) (0.336) (0.391) (0.854) Log (Top 10% Income)   0.777      0.668  (0.596)      (0.592) Poverty Rate   0.010 0.030***    0.001    (0.012) (0.007)    (0.017) Pct BA degree     -0.017*   0.008      (0.010)   (0.013) Pct Mobile home      0.053***  0.054***       (0.007)  (0.008) Pct Female-Headed       0.025 0.004        (0.015) (0.019) Dummy 1987 0.377** 0.318 0.336* 0.225 0.352* 0.160 0.319 0.101  (0.192) (0.197) (0.190) (0.185) (0.192) (0.192) (0.201) (0.196) Dummy 1992 0.278 0.204 0.198 -0.029 0.239 -0.133 0.154 -0.218  (0.176) (0.180) (0.202) (0.169) (0.177) (0.187) (0.196) (0.227) Dummy 1997 0.718*** 0.611*** 0.605** 0.281 0.657*** 0.179 0.544** 0.063  (0.192) (0.200) (0.237) (0.177) (0.199) (0.209) (0.230) (0.281) Dummy 2002 0.973*** 0.809*** 0.839*** 0.454*** 0.909*** 0.343 0.762*** 0.169  (0.204) (0.230) (0.252) (0.170) (0.207) (0.222) (0.251) (0.308) Dummy 2007 1.106*** 0.868*** 0.956*** 0.548*** 1.059*** 0.469* 0.874*** 0.222  (0.215) (0.266) (0.279) (0.196) (0.219) (0.242) (0.277) (0.354) Dummy 2012 1.246*** 0.921*** 1.079*** 0.645*** 1.212*** 0.619*** 0.998*** 0.285  (0.220) (0.307) (0.288) (0.204) (0.224) (0.237) (0.276) (0.380) Constant 4.685 1.197 1.356 -8.350*** 1.945 -5.603 1.275 -8.441  (3.053) (4.117) (4.829) (1.264) (3.426) (3.437) (3.823) (6.838) No. of Observations 4,759 4,759 4,757 4,757 4,759 4,759 4,759 4,757 No. of Counties 2,121 2,121 2,120 2,120 2,121 2,121 2,121 2,120 Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1    32 Table 7:  Socio-economic Characteristics and Disaster Impacts Negative Binomial Random Effect Regressions Results  Dependent variable: Deaths from F2-F5 tornadoes Independent variables (1) (2) (3) (4) (5) (6) (7) (8) Fscale_F2+ 1.408*** 1.409*** 1.411*** 1.413*** 1.407*** 1.436*** 1.409*** 1.437***  (0.051) (0.051) (0.051) (0.051) (0.051) (0.051) (0.051) (0.051) Lag_Tornado_F2+ -0.005 -0.006 -0.006 -0.006 -0.003 -0.007 -0.007 -0.010  (0.038) (0.038) (0.038) (0.038) (0.038) (0.038) (0.038) (0.038) Tornado Alley -0.369*** -0.364*** -0.361*** -0.372*** -0.342*** -0.146 -0.327*** -0.133  (0.094) (0.094) (0.094) (0.094) (0.095) (0.097) (0.097) (0.099) Log(Land Area) -0.082 -0.086 -0.080 -0.059 -0.061 -0.159* -0.056 -0.163*  (0.087) (0.087) (0.087) (0.086) (0.088) (0.087) (0.088) (0.091) Log(Population) 0.315*** 0.315*** 0.304*** 0.253*** 0.326*** 0.422*** 0.283*** 0.408***  (0.049) (0.049) (0.050) (0.044) (0.049) (0.050) (0.052) (0.054) Pct Over65 -0.029* -0.021 -0.028* -0.027* -0.038** -0.001 -0.024 0.010  (0.015) (0.017) (0.015) (0.015) (0.016) (0.015) (0.016) (0.018) Pct Under18 -0.038** -0.041** -0.039** -0.036** -0.044** -0.009 -0.044** -0.009  (0.017) (0.017) (0.017) (0.017) (0.018) (0.018) (0.018) (0.019) Log(PerCapita GovtExp onPublicSafetyWelfare) -0.301*** -0.305*** -0.307*** -0.336*** -0.292*** -0.196*** -0.334*** -0.213*** (0.072) (0.072) (0.072) (0.071) (0.072) (0.073) (0.074) (0.076) Log (PerCapita Income)  -1.325*** -1.817*** -0.994**  -1.035*** -0.526* -0.929*** -0.921 (0.273) (0.480) (0.443)  (0.321) (0.291) (0.344) (0.723) Log (Top 10% Income)   0.685      0.490  (0.551)      (0.549) Poverty Rate   0.010 0.029***    -0.002    (0.011) (0.006)    (0.014) Pct BA degree     -0.016*   0.007      (0.009)   (0.011) Pct Mobile home      0.050***  0.051***       (0.006)  (0.007) Pct Female-Headed       0.027* 0.010        (0.014) (0.017) Dummy 1987 0.246* 0.194 0.211 0.109 0.220 0.047 0.176 -0.000  (0.148) (0.153) (0.152) (0.145) (0.149) (0.149) (0.152) (0.161) Dummy 1992 0.185 0.121 0.108 -0.114 0.148 -0.199 0.044 -0.270  (0.156) (0.165) (0.176) (0.145) (0.158) (0.163) (0.173) (0.196) Dummy 1997 0.627*** 0.538*** 0.519** 0.205 0.566*** 0.107 0.429** 0.014  (0.168) (0.182) (0.203) (0.146) (0.171) (0.179) (0.197) (0.238) Dummy 2002 0.869*** 0.728*** 0.743*** 0.373** 0.808*** 0.273 0.635*** 0.132  (0.182) (0.214) (0.225) (0.153) (0.185) (0.195) (0.220) (0.275) Dummy 2007 1.033*** 0.827*** 0.890*** 0.494*** 0.985*** 0.406** 0.770*** 0.209  (0.177) (0.243) (0.234) (0.153) (0.180) (0.193) (0.225) (0.307) Dummy 2012 1.110*** 0.826*** 0.952*** 0.532*** 1.073*** 0.506*** 0.835*** 0.245  (0.176) (0.289) (0.243) (0.155) (0.178) (0.191) (0.229) (0.349) Constant 7.076*** 4.094 3.873 -5.615*** 4.539 -2.871 3.282 -4.697  (2.730) (3.641) (4.349) (1.012) (3.098) (2.993) (3.383) (5.977) No. of Observations 4,759 4,759 4,757 4,757 4,759 4,759 4,759 4,757 No. of Counties 2,121 2,121 2,120 2,120 2,121 2,121 2,121 2,120 Standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1   33 6.3.  Mobile homes residents experience more tornado fatalities  The fourth hypothesis is that mobile home living results in more tornado fatalities. The regression estimates in specifications (6) and (8) show that the percent of mobile homes in a county is positively related to tornado fatalities, and the estimates are robust. The results confirm that more mobile homes in a county results in greater vulnerability to tornadoes. The estimated coefficient implies that one percentage point increase in the proportion of mobile homes in total housing units is expected to increase tornado-related deaths by 5.2 percent (). Further, as noted earlier more households are choosing this type of housing arrangement over time, and thus vulnerability may be increasing. This finding may have important policy implications in the context of developing approaches to reduce tornado vulnerability. For example, mobile home parks could potentially provide common tornado shelter areas to be used in the event of a tornado watch or warning. 6.4.  Female-headed households are more vulnerable to tornadoes  The second hypothesis is that female-headed households are more vulnerable to tornadoes. This hypothesis is examined in specifications (7) and (8) in the Poisson and Negative Binomial models. These regressions show that female headed households and tornado-induced fatalities weakly have a positive correlation. The estimate in specification (7) shows that a one percentage point increase in the proportion of the female-headed households in a county is expected to increase tornado fatalities by 2.7 percent. It is implied that all else equal, places with more female-headed households are more vulnerable, perhaps because female-headed households have limited access to resources during high risk events. The result is consistent with the previous arguments by sociologists (Enarson and Morrow 1998; Enarson, Fothergill, and Peek 2006). However, the estimated effect only achieves significance in specification (8).  34 6.5.  Government spending in public safety and welfare mitigates losses from tornadoes   Finally, I test the degree to which local government plays a role in reducing the potential tornado fatalities. The regression results show a significant and negative relationship between tornado fatalities and per capita government spending on public safety, protection, and welfare. Such local government expenditures appear to improve overall safety/welfare of a community,  For example, 10 percent increase in government per capita spending, which is $27.10 on average in our sample (in 2009 dollars), is estimated to have about 3 percent decrease in tornado-induced fatalities. Given the parameter estimate, if governments in each county had allocated 50% more funds to safety, protection, and welfare over the study period 1980-2014, 268 lives would have been saved from tornados19. However, considering the limited government resources available for public services, I offer an evaluation of whether it would be worthwhile for local government to allocate more funds to public safety, protection, and welfare, with the goal of reducing tornado fatalities. Specifically, I perform a straightforward cost-benefit analysis by comparing the amount of extra funds required to save a life in local governments from severe tornadoes with the benefit in terms of the value of life. On this benefit side, I follow the practice of giving an economic value to mortality - a value of a statistical life (VSL). The VSL that is currently being used in the U.S. government agencies when they appraise the benefits of regulations ranges from $8.2 to $9.5 million (in 2009 dollars) (Viscusi 2014). The cost-benefit comparison reveals that in order to save a life from severe tornado, each county would need to spend additional $508 per capita, on average, which is                                        19 The expected number of lives that could have been saved by increasing per capita government expenditures by 50% has been calculated across all counties who had experienced FS2+ tornadoes over the study period and added up.  35 approximately $30 million in extra burden to local governments20. Altogether, it does not appear that increasing government expenditures is a cost-effective way of achieving tornado fatality reduction, even after taking into consideration that the life-saving benefit is just one component of the multiple benefits that may arise from such government spending. My empirical analysis suggests that general increase in government funds on public safety, protection, and welfare is linked to the goal of mitigating tornado impacts to some extent but the cost-benefit analysis reveals that it is not an effective policy scheme for mitigating tornado fatalities in most counties. In this regard, further research is needed to investigate to better target which set of public services provided by local governments most effectively mitigates the degree to which their citizens are exposed to tornadoes.  6.6.  Additional Analyses  In Table 8, I present the results of additional analyses. The second set of regressions consist of four specifications: (1) and (2) use very strong tornadoes of Fujita-scale 3 or greater (F3-F5) in the same framework as in Table 6 and 7, and (3) and (4) exploit the magnitude of tornadoes in a different way compared to the main analysis presented in Table 6 and 7.  First two columns show Poisson and Negative Binomial regressions result for severe tornadoes rated F3 or higher. Focusing on the larger events reduces the number of tornado events by 3,907, leaving just 1,507 severe tornadoes. Thus, in my sample, only 1,245 counties are used                                        20 Let Per Capita Gov Expenditure on Public Safety & Welfare =, Avg. Fatalities per FS2+ tornado=, and Yearly fatalities from FS2+tornado=. The extra funds needed to save a life is calculated using the estimated relationship between  and  where  holds ( is the estimated coefficient of ), and the relationship (Yearly No.FS2+ tornado). The expression  is derived such that  (i.e. yearly fatalities from FS3+ tornado decrease by one unit) using the sample mean ( and  from the observations with . I obtain = $508, which implies if local governments that suffered at least one death every year from tornado increase per capita spending by $508, on average, one death would be avoided in each county every year. The average extra burden to local governments, $30M is obtained by multiplying county population by the per capita extra expenditure, $508.  36 for the analysis excluding those counties without any experience of tornadoes of F3+ during 1980-2014. The estimates reported in columns (1) and (2) in Table 8 are very consistent with those in Table 6 or 7 with the exception of a few differences. When we consider only the very strong and more destructive tornadoes, the significance of the estimated coefficients on some of the socio-economic variables disappears or weakens in magnitude. However, tornado vulnerability related variables such as Tornado Alley, lagged tornado frequencies, and population density measure take on greater importance, while the coefficient on the mobile home variable remains statistically significant and similar in magnitude compared to the estimation results using F2+ tornadoes. Taken together, these findings suggest that the stronger tornadoes extend vulnerability to a broader array of people in a community such that social-economic status becomes less important whereas the intensity of natural force and physical factors become more significant in determining vulnerability. In addition, the larger coefficients on tornado alley and lagged tornado frequencies compared to the estimate from F2+ tornadoes suggests that the previous experiences of severe disasters bring a stronger learning effect in case where the community is hit by a severe tornado.    Let's now turn to the result presented in columns (3) and (4) in Table 8. I perform an additional analysis as a robustness check to my main analysis, by including the fractions of tornadoes of each F-scale among F2 to F5 tornadoes (or among F3 to F5 in specification (4))21 , instead of using the average F-scale variable. As shown in Table 2, tornadoes of different magnitudes can have widely differing degrees of impact. For instance, the average death from F5 tornadoes is 280 times larger than that of F2 tornadoes in my sample. In this additional analysis, I try to account for such differentiated impacts of each level of F-scale events. The results from                                        21 The fraction of F2 tornadoes (or F3 tornadoes in specification (4)) is a reference point.  37 the Poisson model22 using tornadoes of F2+ and F3+ are presented in column (3) and (4), respectively. As expected, tornadoes of different levels of F-scale have largely different impacts on expected fatalities. The estimates in column (3) suggest that one percentage point increase in the share of F3 tornadoes (while having a one percentage point decrease in the share of F2 tornadoes instead) increases expected fatalities by 2 percent, F4 by 3 percent, and F5 by 6 percent. For instance, if we consider a worst case scenario where the percent of F5 tornadoes changes from 0% to 100% with an F5 tornado occurrence in a county without any other tornado event, the county is expected to suffer 365 times () as many fatalities as that from F2 tornadoes. After accounting for the magnitudes of tornadoes in a detailed way, I obtain results that are mostly similar to my main analysis. The estimation results in column (3) again demonstrate that housing quality, population density, income levels, and government spending on public safety and welfare are critical factors in determining tornado vulnerability.                                            22 Due to a convergence difficulty in Negative Binomial estimation process, Poisson model is only employed for the estimation.  38 Table 8 : Socio-economic Characteristics and Disaster Impacts Additional Regressions Results  (1) Poisson (2) Neg. Binomial (3) Poisson (4) Poisson Dependent variable Deaths from F3+ Deaths from F3+ Deaths from F2+ Deaths from F3+ Fscale_F3+ 1.508*** 1.339***    (0.096) (0.078)   Pct F3 tornado   0.018***     (0.001)  Pct F4 tornado   0.029*** 0.011***    (0.001) (0.001) Pct F5 tornado   0.059*** 0.041***    (0.004) (0.004) Lag_Tornado_F2+   -0.010     (0.045)  Lag_Tornado_F3+ 0.070 0.057  0.074  (0.081) (0.075)  (0.091) Tornado Alley -0.296** -0.191* -0.191* -0.265**  (0.125) (0.109) (0.107) (0.125) Log(Land Area) -0.219* -0.272*** -0.158 -0.146  (0.124) (0.102) (0.099) (0.124) Log(Population) 0.459*** 0.423*** 0.411*** 0.422***  (0.076) (0.061) (0.061) (0.075) Pct Over65 0.017 0.033* -0.006 0.004  (0.027) (0.020) (0.023) (0.028) Pct Under18 -0.012 -0.015 -0.014 -0.033  (0.027) (0.022) (0.023) (0.029) Log(Per Capita Govt Exp on Public Safety & Welfare) -0.232** -0.195** -0.199** -0.227** (0.113) (0.089) (0.089) (0.111) Log (PerCapita Income)  -1.282 -0.654 -0.783 -1.091 (1.035) (0.828) (0.840) (1.039) Log (Top 10% Income) 1.073 0.526 0.404 1.069  (0.713) (0.608) (0.586) (0.773) Poverty Rate 0.006 -0.000 0.003 0.008  (0.020) (0.016) (0.017) (0.020) Pct BA degree 0.009 0.007 0.007 0.005  (0.016) (0.012) (0.012) (0.015) Pct Mobile home 0.053*** 0.053*** 0.055*** 0.048***  (0.010) (0.008) (0.008) (0.010) Pct Female-Headed -0.012 -0.001 0.003 -0.004  (0.022) (0.019) (0.019) (0.022) Dummy 1987 -0.047 -0.351* 0.117 -0.069  (0.255) (0.194) (0.197) (0.258) Dummy 1992 -0.419 -0.537** -0.270 -0.456  (0.274) (0.221) (0.227) (0.278) Dummy 1997 -0.126 -0.266 -0.006 -0.085  (0.359) (0.272) (0.275) (0.364) Dummy 2002 0.054 -0.116 0.160 -0.067  (0.372) (0.316) (0.307) (0.383) Dummy 2007 0.190 0.032 0.233 0.034  (0.433) (0.350) (0.346) (0.451) Dummy 2012 -0.022 -0.023 0.216 -0.161  (0.460) (0.390) (0.380) (0.506) Constant -10.482 -7.880 -4.305 -6.849  (8.154) (6.672) (6.713) (8.433) No. of Observations 1,884 1,884 4,757 1,884 No. of Counties 1,245 1,245 2,120 1,245 Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1   39 7. CONCLUSION   While tornado activity is exogenously determined by natural forces, it is also true that socio-economic factors are critical in determining vulnerability. This study seeks to uncover these underlying factors. To this end, I investigate the relationship between tornado fatalities and the potential determinants of tornado impacts within U.S. counties over the period from 1980 to 2014. Findings of the study enable us to identify which societal characteristics exacerbate or mitigate vulnerability to hazards, which in turn allow us to suggest policies that may help mitigate human losses from such events.  The empirical analysis of this study consistently demonstrates that income level is a crucial determinant of tornado fatalities; this finding is consistent with an array of previous studies, but this study offers more detail on how the various expressions of poverty may contribute to deaths. The analysis also suggests that income inequality is a significant factor that may exacerbate the impacts of disasters. Also, counties with higher poverty rate and more female-headed households tend to be more vulnerable, while the higher the education level, the lower the vulnerability. In general, households most affected by disasters are those with weaker economic and social bases. The information presented here may help to target the most vulnerable households and provide improved access to safety resources.  In addition, my analysis offers evidence that per capita government spending on public safety and welfare is negatively related to death tolls. This suggests that increased government spending in critical areas such as safety, protection, and welfare, reduce overall vulnerability within a community. For some counties with frequent tornado occurrences and higher fatality rates, extra funds on safety, protection, and welfare might mitigate the impacts and save lives effectively. However, a cost-benefit analysis that compares the estimated extra government  40 expenditures required to save a life from severe tornado on average and a value of a statistical life (VSL) as a benefit reveals that generally increasing government funds on safety, protection, and welfare is not a cost-effective policy scheme for most local governments for reducing tornado fatalities. Nevertheless, it may be useful for policy makers to consider allocating resources on specific public services that improve safety and reduce tornado vulnerability. In this regard, further research is needed to investigate which particular public service provided by local government mitigates the degree to which their citizens are affected by tornadoes. Another key finding is that the number of mobile homes in a county is critical factor in explaining tornado fatalities. This finding implies that housing quality is perhaps the most important factor in determining tornado vulnerability. Importantly, the proportion of households living in mobile homes has increased nearly three-fold since the 1970s, with much of this increase occurring between 1970 and 1980 (prior to the period of analysis). Though mobile homes offer a relatively inexpensive but comfortable housing alternative, it appears that this trend has made the United States more vulnerable to tornadoes over time. Given this trend and my findings, it is critical that federal, state and local policy makers consider alternatives to reduce vulnerability for those living in this type of housing arrangement. Policies aimed at strengthening the ability of mobile homes to withstand high winds and flying objects and more systematically require communal tornado shelters may be effective at reducing tornado fatalities. In particular, mobile homes are commonly classified and taxed as personal property placing lower tax burden to home owners. This tax advantage makes mobile home living economically more attractive, but at the same time the tax policy is in fact encouraging more people to live in housing that is more vulnerable to tornados. The external cost of being exposed to greater tornado risks may be ignored when households choose to live in mobile homes due to  41 affordability. One potential policy scheme to internalize this social cost would have governments i) require communal shelters in mobile home parks and communities23, ii) impose a higher tax rate to mobile homes where tornado shelter/safe room are unavailable, and iii) redirect the tax revenue raised from step ii) towards additional government funds for the local communities' safety/protection. In this way, local governments could broaden their tax base and target the revenue from that source to further mitigate human losses from future tornado events. Overall, this study reveals which types of households tend to have more difficult time when disaster occurs, thus informing policies targeted at reducing tornado fatalities. More generally, addressing the root of the issue by improving the conditions of those with lower socio-economic status would reduce vulnerability over time. I expect that these findings will increase our understanding of the socio-economic nature of tornado impacts and enable decision-makers to improve mitigation efforts.                                         23 There are communities that already require all mobile home parks to provide storm shelters for their residents, including the State of Minnesota, and some individual counties (e.g. Sedgwick County and Butler County in KS, St. Joseph County, MO, etc.)  42                        APPENDIX    43 APPENDIX   Table A1:  Socio-economic Characteristics and Disaster Impacts Poisson Fixed Effect Regressions Results  Dependent variable Deaths from F2+ Deaths from F3+ Fscale_F2+ 1.777***   (0.094)  Lag_tornado_F2+ -0.030   (0.059)  Fscale_F3+  1.774***   (0.153) Lag_tornado_F3+  0.042   (0.099) Log(Population) 0.253 -0.503  (0.448) (0.805) Pct Over65 -0.046 -0.171  (0.076) (0.111) Pct Under18 0.025 0.041  (0.071) (0.089) Log(Gov Exp on Public Safety & Welfare) -0.197 -0.189 (0.206) (0.283) Log (PerCapitaIncome)  -3.375* -3.846 (1.963) (2.694) Log (Top 10% Income) 2.199** 1.835  (1.011) (1.397) Poverty Rate -0.084** -0.070  (0.038) (0.053) Pct BA degree 0.064 0.153  (0.059) (0.099) Pct Mobile home 0.012 0.020  (0.028) (0.036) Pct Female-Headed -0.159* -0.186  (0.088) (0.118) Dummy 1987 0.471 0.498  (0.312) (0.405) Dummy 1992 0.553 0.520  (0.473) (0.520) Dummy 1997 0.616 0.677  (0.686) (0.780) Dummy 2002 0.867 1.077  (0.799) (0.855) Dummy 2007 0.786 1.091  (0.881) (0.959) Dummy 2012 1.100 1.157  (0.988) (1.128) No. of Observations 2,026 722 No. of Counties 629 288 Robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.1   44  Though I do not offer a detailed discussion of the fixed effects estimates presented here, in general the statistical significance of the socio-economic variables is greatly reduced. Few of the variables are significant, but this is not too surprising given that within county changes over the 1980-2014 period are typically small for most of these variables. Note that in Table A1 we observe a reversal of sign on most of socioeconomic variables except for income levels and government spending. However, those counterintuitive results are not robust as they are mostly insignificant in both columns (1) and (2). As noted by Kahn (2005) the fixed effects approach be problematic, given the presence of sluggish adjustment and long latency in economic development. Nevertheless, I present these estimates in Appendix for the interested reader.      45                      REFERENCES    46 REFERENCES   Albala-Bertrand, J.M. 1993. Political Economy of Large Natural Disasters. New York: Oxford University Press.  Anbarci, N., Escaleras, M., and Register, C. A. 2005. Journal of Public Economics 89(9): 1907-1933.  Aptekar, L. 1991. The psychosocial process of adjusting to natural disasters. Natural Hazards Research and Applications Information Center, University of Colorado.  Aptekar, L., and Boore, J.A. 1990. "The emotional effects of disaster on children: A review of the literature."International Journal of Mental Health: 77-90.  Blaikie, P., Cannon T., Davis, I., and Wisner, B. 1994. Vulnerability, and Disasters. New York: Routledge.  brief history of deaths from tornadoes in the United Weather and Forecasting, 1-9.  adaptive capacity at the national level and the implications for adaptationGlobal Environmental Change 15(2): 151-163.  Burton, I., Kates, R.W. and White, G.F. 1993. The Environment as Hazard. New York City, NY: The Guilford Press.   Disasters, Development and Environment, 13-30.  Social Science Quarterly 84(2): 242-261.  Cutter, S. L. 1996. "Societal responses to environmental hazards." International Social Science Journal 48(150), 525-536.  Handbook of Disaster Research, edited by H. Rodriguez, E. L. Quarantelli, and R. R. Dynes, pp. 130-146. New York: Springer.  Women's International Hurricane Center. Laboratory for Social and Behavioral Research, 81-94.   47 Fothergill, A., Peek, L. A. 2004 Natural Hazards 32: 89110.  Gaiha, R., Hill, K., Thapa, G., and KulkarniBrooks World Poverty Institute Working Paper, 181.  Hurricane Andrew: Ethnicity, Gender, and the Sociology of Disasters, ed. W. G. Peacock et al., 52-74. New York: Routledge.  Hewitt, K., ed. 1983. Interpretations of calamity from the viewpoint of human ecology. No. 1. Taylor & Francis.  d Censored Econometrica 60: 533-567.  Economic Development and Cultural Change 48: 521-542.  Retrieved June 12, 2015 from www.insurancejournal.com/news/southcentral/2013/05/23/293129.htm   Kahn, E. M. 2005. The death toll from natural disasters: The role of income, geography, and institutions The Review of Economics and Statistics 87(2): 271-284.   Journal of Urban Economics 63(3): 788-802.  Natural Hazards 52: 79-95.  Lal, P. N., Singh, R., and Holland, P. 2009. "Relationship between natural disasters and poverty: a Fiji case study." SOPAC Miscellaneous Report 678.  Disaster Prevention and Management: An International Journal 10 (3): 18996.   Land Economics 81, 87-99  Moore, H. E. 1958. Tornadoes over Texas, Austin: University of Texas Press.  Int J of Mass Emerg and Disasters 14(1): 5-22.  48 Disasters 23(1): 1-18.  Ecology and Society 19 (1): 42.  NOAA. 2011a. National Climatic Data Center, State of the Climate: Tornadoes for Annual 2011, published online December 2011, retrieved on January 6, 2015 from http://www.ncdc.noaa.gov/sotc/tornadoes/2011/13.  NOAA 2011b. National Climatic Data Center, U.S. Tornado Climatology, retrieved on November 6, 2014 from http://www.ncdc.noaa.gov/climate-information/extreme-events/us-tornado-climatology  NOAA. 2011c. Service assessment: the historic tornadoes of April 2011. Silver Spring, MD: U.S. Department of Commerce, National Oceanic and Atmospheric Administration; 2011. Available at http://www.nws.noaa.gov/om/assessments/pdfs/historic_tornadoes.pdf.  Neumayer, E., and Plümper, T. 2007. "The gendered nature of natural disasters: The impact of catastrophic events on the gender gap in life expectancy, 19812002." Annals of the Association of American Geographers 97(3) : 551-566.  Nature 260, 566-567.  Peacock, W. G., Morrow, B. H., and Gladwin, H. 1997. Hurricane Andrew: Ethnicity, gender, and the sociology of disasters. Psychology Press.  Natural Hazards and Earth System Science 8(4): 627-634.  -seeking Natural Hazards 48(2): 191-201.  Simmons, K., and Sutter, D. 2013. Economic and societal impacts of tornadoes. Springer Science & Business Media.  Skidmore, M., and Toya, H. 2013. "Natural Disaster Impacts and Fiscal Decentralization." Land Economics 89(1): 101-117.  The Journal of Economic Perspectives 21(3): 199-222.  Economics Letters 94: 20-25.   49 Viscusi, W. Kip. 2014, "The value of individual and societal risks to life and health." Handbook of the Economics of Risk and Uncertainty, 385-452.  Wildavsky, A. B. 1988. Searching for Safety. Vol. 10. New Brunswick, NJ: Transaction Books.  Wooldridge, J.M. 1991." Specification testing and quasi-maximum-likelihood estimation", Journal of Econometrics 48(1-2), pp.29-55.  Wooldridge, J. M. 2010. Econometric analysis of cross section and panel data. MIT press.