SOCIO - POLITICAL NATURE OF DISASTER IMPACT : TORNADOES, FLOODS, AND EXTREME HEAT By Jungmin Lim A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Agricultural, Food and Resource Economics Doctor of Philosophy 201 8 ABSTRACT SOCIO - POLITICAL NATURE OF DISASTER IMPACT: TORNADOES, FLOODS, AND EXTREME HEAT By Jungmin Lim Severe weather and climate events such as floods, heat waves, and tornadoes, are the most frequent and devastating extreme events among all types of natural disasters in the United States. Climate sc ientists predict that extreme weather phenomena are expected to increas e in both frequency and intensity under ongoing global climate change . Given the anticipated growing risks and detrimental impacts on people of weather extremes, it is imperative to inv estigate past disaster incidents and uncover community characteristics that reflect vulnerability and resilience, in order Socio - Political Nature of Disaster examines three types of extreme weather events in each of three chapters to investigate the determinants of community vulnerability to disasters and evaluate the life - saving benefits of disast er mitigation measures and practices . Each of three chapters empirically examine tornadoes, floods, and extreme heat events at the subnational level I consider the disaster experiences in about 3,100 counties in the contiguous United States. The integ rated view of the physical, social, economic, and political elements of multi - faceted disaster vulnerability guides the empirical analyses. Each chapter employs different types of panel methods to address the county heterogeneity and potential simultaneity between government al actions and disaster vulnerability such as Poisson Fixed Effects , the Control - function (CF) approach with in the Correlated Random Effects (CRE) framework , and Throughout the three chapters, I present evidence that people most vulnerable to disasters are those who have weaker economic and social bases; lower income, poverty, lower education , and poor housing quality increase disaster vulnerability. Also, I find that u rbanization intensifies disa ster vulnerability while learning from past experiences enhance s coping capacity . In the case of heatwaves, vulnerability is greater in counties with higher proportions of elderly, the very young, and non - white populations. Findings suggest th at the socially isolated elderly and the elderly living in poverty are the most heat - vulnerable population sub - groups. M y dissertation pa ys special attention to the examination of the degree to which local government plays a role in reducing the potentia l disaster fatalities. The first chapter on tornadoes and the second chapter on floods shed light on the role of local government resources devoted to public safety , protection, and welfare in mitigating disaster fatalities . The second chapter on floods also provides a new evaluation of the role of the National Flood Insurance Program (NFIP) in preventing and reducing the loss of human life from flooding as an important ex - ante disaster management scheme. The third chapter provides significant evidence on the benefits of the government - initiated H eat I sland M itigation (HIM) measures in lowering heat intensity as well as reducing the loss of life from extreme heat. Taken together , increase s our understanding of the Moreover, t underscore s the need for more proactive and precautionary public measures and policies to counter the potential harmful effects of the growing risk of weather extremes. guidance to future disaster mitigation policies at the local, state and national levels . Copyright by JUNGMIN LIM 2018 v ACKNOWLEDGEMENTS First, and most of all, I would like to express my deepest appreciation to my major advisor and dissertation supervisor, Mark Skidmore, for his patient guidance, support, encouragement, and inspiration he has provided throughout my entir programs at MSU. 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. His dedication and support made the completion of this dissertation possible. I al so owe my gratitude to my external committee member, Dr. Jeffrey Wooldridge, who has always provided excellent advice on the empirical strategies for my research. His feedback and suggestions have contributed much in shaping my dissertation. I also gratefu lly acknowledge the guidance and insight from my other committee members: Dr. Scott Loveridge, Dr. Robert Shupp, and Dr. Soren Anderson. I appreciate their detailed comments and thoughtful suggestions on this study. I would not have been able to complete this journey of the graduate studies without the love , support , and encouragement of my family . My special gratitude goes to my husband, Jongwoo, who has gone through all the highs and lows of the academic life in AFRE at MSU together with me and helped me stay strong and sound through the Ph.D. program. I am also indebted to my son, Kyle, pure happiness and joy, who gives endless love and energy that enabled me to overcome obstacles and challenges during the entire period of my graduate studies. I would al so like to thank my parents and my father - and mother - in - law for their loving and tireless support, constant encouragement, and their faith in me. vi TABLE OF CONTENTS LIST OF TABLES ................................ ................................ ................................ ..................... vi ii LIST OF FIGURES ................................ ................................ ................................ ....................... x KEY TO ABBREVIATIONS ................................ ................................ ................................ ...... xi INTRODUCTION ................................ ................................ ................................ ......................... 1 CHAPTER 1 ................................ ................................ ................................ ................................ .. 5 SOCIO - ECONOMIC DETERMINANTS OF TORNADO FATALITIES IN THE UNITED STATES: DIMENSIONS OF POVERTY, HOUSING QUALITY, AND GOVERNMENT ...... 5 1.1 INTRODUCTION ................................ ................................ ................................ ........... 5 1. 2 EMPIRICAL STUDIES ON THE DETERMINANTS OF DISASTER IMPACTS ...... 8 1. 3 TORNADO RISK IN THE UNITED STATES ................................ ............................ 11 1. 3.1 Tornado Frequency ................................ ................................ ............................... 11 1. 3.2 Tornado Intensity ................................ ................................ ................................ .. 1 2 1. 4 DETERMINANTS OF TORNADO VULNERABILITY ................................ ............. 1 4 1. 4.1 Motivation ................................ ................................ ................................ ............. 1 4 1. 4.2 Conceptual Framework ................................ ................................ ......................... 15 1. 4.3 Hypotheses on the Determinants of Tornado Vulnerability ................................ . 1 6 1. 5 EMPIRICAL ANALYSIS ................................ ................................ ............................. 23 1. 5.1 Data Description ................................ ................................ ................................ ... 2 3 1. 5.2 Empirical Model ................................ ................................ ................................ ... 2 5 1. 6 RESULTS ................................ ................................ ................................ ...................... 30 1.6 .1 Richer Counties Experience Fewer Tornado - induced Deaths .............................. 32 1.6 .2 Human Capital Plays an Important Role in Reducing Tornado Vulnerability ..... 3 3 1.6 .3 Mobile Homes Residents Experience More Tornado Fatalities ........................... 3 6 1.6.4 Female - Headed Households Are More Vulnerable to Tornadoes ........................ 3 6 1.6.5 Government Spending in Public Safety Mitigates Losses from Tornadoes ......... 3 7 1. 6.6 Additional Analyses ................................ ................................ .............................. 3 8 1. 7 CONCLUSION ................................ ................................ ................................ .............. 42 APPENDIX ................................ ................................ ................................ ................................ .. 4 5 REFERENCE S ................................ ................................ ................................ ............................ 4 8 CHAPTER 2 ................................ ................................ ................................ ................................ 53 FLOOD FATALITIES IN THE UNITED STATES: THE ROLES OF SOCIO - ECONOMIC FACTORS AND THE NATIONAL FLOOD INSURANCE PROGRAM ................................ 53 2.1 INTRODUCTION ................................ ................................ ................................ ......... 53 2.2 LITERATURE REVIEW ................................ ................................ ............................... 56 2.2 .1 Socio - political Nature of Disasters ................................ ................................ ....... 56 2.2 .2 Economic Development and Disaster Impacts ................................ ..................... 57 2.2 . 3 Severe Weather Events in the United States and Disaster Vulnerability .............. 59 vii 2.3 CONCEPTUALIZING HUMAN AND ENVIRONMENTAL COMPONENTS OF FLOOD VULNERABILITY ................................ ................................ ............................ 60 2.3 .1 Disaster - specific Determinants of Flood Vulnerabil ity ................................ ........ 61 2.3 .2 Area - specific Physical and Environmental Determinants of Flood Vulnerability 62 2.3 .3 Socio - political Determinants of Floods Vulnerability ................................ .......... 63 2.4 EMPIRICAL ANALYSIS ................................ ................................ ............................. 69 2 . 4. 1 Data Description ................................ ................................ ................................ ... 69 2.4 .2 Empirical Model ................................ ................................ ................................ ... 72 2.4 .2 .1 Base Mode ................................ ................................ ................................ ...... 72 2.4 .2 .2 Instrumental Variable Model ................................ ................................ ......... 76 2.5 RESULTS ................................ ................................ ................................ ...................... 7 9 2.5 . 1 Base Model Results from ZINB Estimation ................................ ......................... 79 2. 5.2 Instrumental Variable Model Results from IV Poisson CRE approach ............... 82 2.6 CONCLUSION ................................ ................................ ................................ .............. 89 APPENDIX ................................ ................................ ................................ ................................ .. 91 REFERENCE S ................................ ................................ ................................ ............................ 94 CHAPTER 3 ................................ ................................ ................................ ................................ 99 GROWING HEAT VULNERABILITY OF AGING SOCIETY: THE POTENTIAL ROLE OF HEAT ISLAND MITIGATION MEASURES ........................... 99 3.1 INTRODUCTION ................................ ................................ ................................ ......... 99 3.2 RISK OF EXTREME HEAT IN THE U.S. ................................ ................................ . 103 3.3 COMMUNITY HEAT ISLAND MITIGATION ACTIONS ................................ ...... 108 3.4 LITERATURE REVIEW ................................ ................................ ............................. 111 3.5 CONCEPTUALIZING MULTI - FACETED HEAT VULNERABILITY ................... 1 12 3.5 . 1 Major C omponents of Heat Vulnerability ................................ .......................... 1 15 3. 5. 2 Institutional Efforts for Mitigation and Adaptation ................................ ............ 119 3.6 EMPIRICAL ANALYSIS ................................ ................................ ........................... 120 3 . 6. 1 First - Phase: Heat Hazard Mitigation Model ................................ ....................... 120 3. 6 .2 Second - Phase: Heat Vulnerability Fatality Model ................................ .......... 125 3. 6 .3 Heat Island Mitigation Actions and Heat Fatality: A Direct Estimation ............ 129 3.7 RESULTS ................................ ................................ ................................ .................... 131 3.7 . 1 First - Phase: Heat Hazard Mitigation Model ................................ ....................... 131 3.7 . 2 Second - Phase: Heat Vulnerability Fatality Model ................................ .......... 134 3. 7 .3 Heat Island Mitigation Actions and Heat Fatality ................................ .............. 140 3.7 . 3.1 First and Second Phase Models Combined: A Mediated Effect .................. 140 3. 7 . 3 . 2 A Direct Estimation of the Effect ................................ ................................ . 142 3.8 CONCLUSION ................................ ................................ ................................ ............ 144 APPENDIX ................................ ................................ ................................ ................................ 145 REFERENCE S ................................ ................................ ................................ .......................... 149 viii LIST OF TABLES Table 1. 1: Mobile Homes in the United States ................................ ................................ ......... 20 Table 1. 2: Tornadoes and Resulting Impacts by Fujita - scale (1980 - 2014) * ............................ 2 4 Table 1. 3: List of dependent and explanatory variables in the model ................................ ...... 2 6 Table 1. 4: Fatalities induced by Strong Tornadoes (F2 - F5), 1980 - 2014* ................................ 2 7 Table 1. 5: County Summary Statistics ................................ ................................ ...................... 2 9 Table 1. 6 : Socio - economic Characteristics and Disaster Impacts Poisson Random Effect Regressions Results ................................ ................................ ............... 34 Table 1. 7: Socio - economic Characteristics and Disaster Impacts Negative Binomial Random Effect Regressions Results ................................ ............................. 3 5 Table 1. 8: Socio - economic Characteristics and Disaster Impacts Additional Regressions Results ................................ ................................ ................................ ... 41 Table 1. A1: Socio - economic Characteristics and Disaster Impacts Poisson Fixed Effect Regressions Results ................................ ................................ ................... 4 6 Table 2.1 : Fatalities and Injuries by Disaster Events, 1996 2015 ................................ .......... 70 Table 2.2 : List of dependent and explanatory variables in the ZINB model ............................ 74 Table 2.3 : Summary Statistics of Variables in ZINB Model ................................ .................... 75 Table 2.4 : Summary Statistics of variables in Poisson IV model ................................ ............. 78 Table 2.5 : Determinants of Flood Fatalities Zero - Inflated Negative Binomial Regressions Results ................................ ................................ 80 Table 2.6 : Determinants of Flood Fatalities Poisson IV CF / CRE Estimates of Key Explanatory Variables ................................ .................. 83 Table 2.7 : Predicted Death Counts by NFIP Participation Rates and t he Total Value of Lives Saved during 1996 - 2015 ................................ ................................ ................................ .............. 88 Table 2. A 1 : Determination of Flood - category Events ................................ .............................. 92 Table 2. A 2 : Variable Definitions and Sources ................................ ................................ ......... 93 ix Table 3. 1: Heat Island Mitigation Actions List by Initiation Year (1985 - 2017) .................... 109 Table 3.2 : List of Variables in the Heat Hazard Mitigation Model ................................ ........ 123 Table 3.3 : Heat Vulnerability and Heat Island Mitigation Actions by Metropolitan Status .. 1 24 Table 3.4 : List of Variables in the Heat Fatality Model ................................ ......................... 1 28 Table 3.5 : Summary Statistics ................................ ................................ ................................ 1 30 Table 3.6 : Heat Hazard Model : T he Role of Heat Island Mitigation Actions Panel Fixed Effects & Random Trend Model Results ................................ ............................... 132 Table 3.7 : Heat Vulnerability Fatalities Model Zero - Inflated Negative Binomial Regressions Results ................................ .............................. 135 Table 3.8 : Increase in Heat Fatalities Given the Projected Growth of the Elderly Population in 2030 and 2040 ................................ 1 39 Table 3.9 : The Mediated Effect of Heat Mitigation Actions on Heat Fatalities ..................... 1 41 Table 3.10 : A Direct Estimation of the Effect of Heat Island Mitigation Actions on Heat Fatalities Poisson FE and RE Model Key Results ................................ .................... 1 43 Table 3. A 1 : Determination of Heat and Excess Heat ................................ ............................. 1 46 Table 3. A 2: Heat Hazard Model : The Role of Heat Island Mitigation Actions Alternative Specifications: Fixed Effects OLS ................................ ................................ .......... 1 47 Table 3. A 3 : Heat Hazard Model : The Role o f Heat Island Mitigation Actions Random Trend Model Using Heat Wave Days as an Alternative DV ................................ ...... 1 48 x LIST OF FIGURES Figure 1 .1 : Global Tornado Activity ................................ ................................ ........................ 12 Figure 1.2 : Average Annual Number of Tornadoes during 1980 - 2014 ................................ .... 13 Figure 1. 3: Total Number of Strong/Violent Tornadoes (F2 - F5), 1980 - 2014 .......................... 1 4 Figure 1. 4: Total Number of Fatalities from Strong/Violent Tornadoes (F2 - F5), 1980 - 2014 . 1 4 Figure 1. 5: Proportion of Households Living in Mobile Homes, 2010 ................................ .... 21 Figure 2.1 : Key Elements of Disaster Vulnerability ................................ ................................ . 61 Figure 2.2 : Total Number of Floods by County, 1996 - 2015 ................................ .................... 71 Figure 2.3 : Total Deaths from Flood by County, 1996 - 2015 ................................ ................... 71 Figure 3.1 : Heat Index Chart ................................ ................................ ................................ .. 1 04 Figure 3.2 : Summertime Average Maximum Heat Index vs. Daily Maximum Temperature (1998 - 2011) ................................ ................................ ............... 1 04 Figure 3.3 : Total Number of Heat Waves by State (1996 - 2015) ................................ ............ 1 07 Figure 3.4 : Heat - Induced Fatalities per Year by State (1996 - 2015) ................................ ....... 1 07 Figure 3.5: Conceptualizing Multi - Faceted Heat Vulnerability ................................ ............. 1 13 xi KEY TO ABBREVIATIONS PFE CF IV CRE RTM ZINB NCEI NOAA NWS CRED FS VSL FEMA NFIP SFHA CRS NID IPCC EPA HIM NDHS Poisson Fixed Effects Control Function Instrumental Variable Correlated Random Effects Random Trend Model Zero - Inflated Negative Binomial National Center for Environmental Information National Oceanic and Atmospheric Administration National Weather Service Centre for Research on the Epidemiology of Disasters Fujita Scale V alue of a S tatistical L ife Federal Emergency Management Agency National Flood Insurance Program Special Flood Hazard Area Community Rating System National Inventory of Dams Intergovernmental Panel on Climate Change Environmental Protection Agency Heat Island Mitigation Net Daily Heat Stress 1 INTRODUCTION Severe weather and climate events such as floods, heat waves, and tornadoes, are the most frequent and devastating extreme events among all types of natural disasters in the United States. In 2017 alone , weather - related disasters caused more than $300 bill ion in total damage s and 508 fatalities . Climate scientists predict that extreme weather phenomena are increasing in both frequency and intensity under ongoing global climate change . Given this trend, it is likely that the economic and human losses from these climatic events will be even greater in the coming decades. Notably, w e have learned that Evaluation of d evastating natural disasters have revealed significant differentials in terms of impacts across differen t population segments, depending on socio - economic and political status. Given the anticipated growing risks and detrimental impacts of weather extremes, it is imperative to investigate past disaster incidents to uncover characteristics that make communiti es more or less vulnerable. Within this context, it is particularly important to examine the role of government in mitigating adverse impacts, in order . To this end, my Socio - Political Nature of Disaster Impact: examines three types of extreme weather events in each of three chapters to investigate the factors that make communities more or less vulnerable to disas ters and to evaluate the life - saving benefits of disaster mitigation measures and practices . The three chapters empirically examine tornadoes, floods, and extreme heat events , respectively using subnational data about 3,100 counties in the contiguous Unit ed States. The integrated view of the physical, social, economic, and political elements of multi - faceted disaster 2 vulnerability guides the empirical analyses. Detailed disaster data are collected from National Center for Environmental Information (NCEI) o f National Oceanic and Atmospheric Administration (NOAA). M ajor socio - economic , housing, and local government finance data at the county level available from U . S . Bureau of the Census are also used in the analysis . Each chapter employs different types of panel methods to address the county heterogeneity and potential simultaneity between government decision s and disaster risk Poisson/Neg. Binomial Random Effects and Poisson Fixed Effects approach in the first chap ter on tornados, the Control - function (CF) approach with in the Correlated Random Effects (CRE) framework for the flood analysis, and original event data set in a Cross - Sectional - Time - Series stru cture with the over - dispersed non - negative count outcomes. Throughout the three chapters, I present evidence that disaster - specific physical factors such as intensity, location, and timing of event s , the built - environment as well as socio - economic characteristics such as demograp hic characteristics, income level, poverty, education, and housing quality determine the overall disaster fatalities . I consistently find that disaster - induced fatalities are greater in communities with weaker economic and social bases; lower income, poverty , and lower education increase disaster vulnerability. H ousing quality is also a critical factor in explaining disaster - induced fatalities; l iving in mobile homes or rental homes increases vulnerability to climatic shock s. Urbanization intensifies disaster vulnerability. Also, results confirm the existence of learning effects from past experiences, where counties that suffered more disaster s in the recent past tend to enhance their coping capacity against disasters and in turn are better able to mitigate the societal impacts. I n the case of extreme heat, population 3 composition is an important factor; heat vulnerability is greater in counties with higher proportions of elderly, young, and non - white populations. Findings sug gest that the socially isolated elderly and the elderly living in poverty are the most heat - vulnerable population sub - groups. Notably, heightened heat vulnerability due to the growing elderly population is predicted to generate a two - fold increase in heat fatalities by 2030. M y analyses pa y a special attention to the degree to which local government plays a role in reducing the potential disaster fatalities. The first chapter on tornadoes and the second chapter on floods shed lights on the role of local government resources devoted to public safety , protection, and welfare in mitigating disaster fatalities . The empirical analyses indicate that such local government expenditures appear to lead to better preparedness and faster responses to disaster events and improve overall safety/welfare of a community, thus reducing fatalities. The second chapter on floods also provides a new analysis of the role of National Flood Insurance Program (NFIP) in reducing the loss of life from flooding as an important ex - ante disaster management scheme. Communi ty participation in the NFIP program requires the participating communities to implement floodplain management requirements for flood risk and damage reduction. My findings provide an empirical evidence that flood - prone community participation help communi ties become more flood - resistant. My evaluation also shows that the life - saving benefits of the NFIP over the 20 - year study period are estimated to be substantial fiscal soundness of the program necessitate a thoughtful reform of the NFIP, which must balance the affordability of flood insurance with financial solvency of the program. In this redesign 4 process, the life - saving benefits of the disaster management of the NFIP ought to be taken into account. In terms of extreme heat, the most relevant public efforts for heat mitigation and adaptation currently undertaken by st ate and local government are the government - initiated community Heat Islands Mitigation (HIM) activities ( i.e. trees / vegetation, green/cool roofs, cool pavements). HIM strategies act primarily as heat - hazard mitigation measures by helping communities that are at higher risk of heat exposure to manage the fundamental meteorological risk of high temperature s . However, there has been no prior heat study that seeks to determine the extent to which government - initiated HIM measures have reduced heat - related fata lities . The third chapter provides new evidence on the benefits of HIM measures in terms of reducing heat intensity as well as reducing the loss of life from extreme heat. My estimate indicates that a Taken together , increase s our understanding of the Moreover, t underscore s the need for more proactive and precautionary public measures and policies to counter the potential harmful effects of the growing risk of weather extremes. guidance to future disaster mitigation policies at the local, state and national levels . 5 CHAPTER 1 SOCIO - ECONOMIC DETERMINANTS OF TORNADO FATALITIES IN THE UNITED STATES: DIMENSIONS OF POVERTY, HOUSING QUALITY, AND GOVERNMENT 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 re sulting 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. dollars 1 . 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 in fluence 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 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 . 6 tornadoe s. 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 a n 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 , where vulnerability depends on com munity 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 underlyin g 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 - Ber trand 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 th e 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 vulne rability. 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 importa nt. As discussed in detail 7 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, account ing for potential interactions between tornado severity and the socio - economic factors that determine vulnerability. 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 observat ions at the sub - national level - 3,107 U.S. counties (excludes Alaska and Puerto Rico ) 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 un derstand 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 re duce societal vulnerability to disasters. The following section offers a review of the empirical literature regarding the determinants of the impacts of natural disasters. Section 1.3 discusses tornado risks in the United States , and section 1.4 describes the underlying theoretical foundation for my analysis and introduces the primary hypotheses. Sections 1.5 and 1.6 present the empirical framework of the analysis and empirical results, respectively. 8 1.2 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 focus es 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 identif y 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 ha zard, but also by income. Horwich (2000) to disaster is its level of wealth. He explains that a rise in income will provide not only general safety but also i mproved protection from natural disasters. 9 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 - DAT 2 . For instance, Kahn (2005) uses this data so urce 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 ation 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 reductio n 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. saster - 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 e conomic development as measured by per capita GDP is inversely correlated with both disaster deaths and damages. However, they also find that h igher 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) 2 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. 10 also shows that economic development reduces disaster fatalities and losses, but this relationship is nonlinear. Economic devel opment 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 in corporates 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 fatal ities. 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, politica l 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 incorporate s multiple types of natural disasters across multiple countries and rel ies 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. A s previously noted, the study most closely related to my study is that by Simmons and Sutter (2013); they employ detailed U.S. 11 county level tornado data from National Oceanic and Atmospheric Administration (NOAA) over the period 1984 - 2007 to examine the so cietal 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. T he 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. 1.3 TORNADO RISK IN THE UNITED STATES 1. 3.1 Tornado Frequency As shown in Figure 1. 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. 3 Focusing on the United States, the average annual number of tornadoes (all intensities) by state for years 1980 - 2014 is presented in Figure 1. 2. The darker green area shown in Figure 1. 2 spanning from Texas to South Dakota is called "Tornado Alley" 4 because of the disproportionately high frequency of tornadoes. 3 NOAA National Climatic Data Center , U.S. Tornado Climatology , retrieved on November 6, 201 4 from http://www.ncdc.noaa.gov/climate - information/extreme - events/us - tornado - climatology 4 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. 12 Fig ure 1 .1 : Global Tornado Activity 1. 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 Scale 5 F0 or F1). Thus, nearly a quarter of tornadoes are classified as significant or strong/violent (F2 and above), with only 0.1% achievin g F5 status (winds over 200 mph, resulting in near complete destruction of everything in i ts 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 F 5. 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. 5 Note that in 2007 - 2008 NOAA introduced and began using the Enhanced Fujita scale for measuring tornado intensity . I use the term Fujita scale throughout the paper since the majority of the data falls under this category . 13 Figure 1. 2: Average Annual Number of Tornadoes during 1980 - 2014 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 25 28, 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 peopl e died, more than 2,400 were injured, and economic damages totaled over $4.2 billion 6 . 6 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 . 14 1.4 DETERMINANTS OF TORNADO VULNERABILITY 1. 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 1. 3 and 1. 4 shows the differences between frequencies and fatalities of strong tornadoes . Figure 1. 3: Total Number of Strong/Violent Tornadoes (F2 - F5), 1980 - 2014 Figure 1. 4: Total Number of Fatalities from Strong/Violent Tornadoes (F2 - F5), 1980 - 201 4 15 The map in Figure 1. 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 1. 4 shows total fatalities from these tornadoes over the same period. As is clear, the a reas with relatively high t ornado 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 re search 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. 1. 4.2 Conceptual Framework As highlighted earlier, Cutter et al. (2003) discuss the possible interactions between s ocial and biophysical vulnerabilities that determine overall place vulnerability. They explain that t he hazard potential is either moderated or enhanced via a combination of geographic factors and the social fabric of the place. This social fabric 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 fat alities 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 o r group's "capacity to anticipate, cope with, resist, and recover from the impact of a however, a group's vulnerability against natural hazard is shaped by human components 16 (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 Boor e, 1990; Albala - Bertrand 1993, Enarson and Morrow 1998; Peacock et al., 1997; Morrow 1999 ). Fothergill et al. (2004) point out that disaster researchers increasingly - research, conducting analyses of minority, gender, and inequality issues in the context of disasters. 1. 4.3 Hypotheses on the Determinants of Tornado Vulnerability B ased 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 - economi c, housing, and governmental factors that may also play significant roles in determin ing tornado - induced death s. 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 vulnerability of disasters . Researchers such as Wildavsky (1988) contends that greater income and wealth translates to a safer society. Safety 17 can be viewed as a natural product of a growing mar ket 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 stu dies (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 add ition 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 play s in determining disaster vulnerability. Similarly, I hypothes ize t hat societies with a higher concentration of poverty might encounter higher tornado - induced human losses. According to Fothergill et al. (200 4 ), t he 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 Peac ock (1997) reported in their study of warnings and evacuation for Hurricane Andrew that lower income people were less able and thus less likely to evacuate, mostly due to constraints placed by a lack of transportation and 18 affordable refuge options. Similar ly, 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 beca use they have lower economic and social conditions; that is, they are unable to invest in adequate preparedness and risk reduction measures. Gend er 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 h usband 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 situation s than men and therefore suffer higher mortality rates. While 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 wome n 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). 19 Human Capital The third hypothesi s 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 i n 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 m ore 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 a s 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. 20 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 lowe r 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 easi ly destroyed. From 1996 to 2000, about half of tornado - induced deaths in the United States were in mobile homes 7 , even though mobile homes accounted for less than 8% of the nation's housing during the same period, according to the National Oceanic and Atmo spheric 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 homes 7 . Table 1 .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 f or years 1950 - 2000 are from Decennial Census. 7 Brooks, H., & Doswell III, C. A. (2001). A brief history of deaths from tornadoes in the United States. Weather and Forecasting , 1 - 9. h ttp://www.nssl.noaa.gov/users/brooks/public_html/deathtrivia/ 21 As shown in Table 1 .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 char acteristics (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 1. 5, in 2010 many rural countie s 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 200 4 , Schmidlin et al. 2009). Figure 1. 5: Proportion of Households Living in Mobile Homes, 2010 22 Local Government Investment My la s t 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 govern ment 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 fatalit ies. 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 government) but are for services like chi homes or payments to vendors for substance abuse treatment and the like. 23 1.5 EMPIRICAL ANALYSIS 1. 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 - eco nomic 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 fis cal 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 bet ween 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 an d 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 counties 8 . Thus, t he unit of observation of this study is counties, not tornado event. 8 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 interpo lated 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 i n disaster events is the main interest of this study. 24 When I average tornado data across time blocks, I include only s trong/violent tornadoes rated F2 or greater for the main analysis or, for the additional analysis F3 or greater. Accordingly, the dependent variable is the average number of deaths 9 caused by tornado es rated F2 - F5 (or F3 - F5 in additional analysis). As note d earlier and shown in Table 1. 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 pane l data for my empirical estimation contains 2, 1 2 0 counties that have experience d tornadoes of F2+ at least once over the study period. Table 1. 2 presents the total number of tornadoes and resulting fatalities and injuries b y F - scale over the years 1980 - 201 4. Table 1. 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 0.0 11 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. 9 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 p er 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. 25 1. 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 variables within the county level panel data framework (Wooldridge, 1991) 10 . Als o, 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. 11 In his multi - national disaster study, Kahn (2005) points out the presence of sluggish adjustmen t 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. 12 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 10 The dependent variable is an average value and can be non - integer. However, t he Poisson (quasi - MLE) model is robust to distributional assumptions; it can be applied to any nonnegative outcome, either continuous or integer valued (Wooldridge, 1991). 11 Wooldridge (2010) also discusses that when the key explanatory variables do not vary much over time, fixed effects methods can lead to imprecise estimates. 12 The result of Fixed Effects Poisson is presented in the Appendix. 26 time , is the number of tornado es 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 1. 3. Table 1. 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 p opulation under 18 Percent of people aged 25 and over holding Bachelor's degree Percent of f emale - headed household s Economic Log ( Per capita I ncome ) Log ( Top 10 percentile income level ) Poverty rate Housing Percent of m obile 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 F5 13 ) Lagged tornado frequency of F2+ Tornado alley Time Dummy 1987, 1992, 1997, 2002, 2007, 2012 Table 1. 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 1. 4). I aggregate tornado data into the afo rementioned 13 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 1. 8. 27 five - year intervals and form a p anel structure. The county level panel data for this study contains 4 , 7 57 county - year observations 14 with at least one strong/violent tornado rated F2 or higher and 1,0 16 observations had fatalities from those events. Using these data, I estimate equation (1) using a Poisson and Negative Binomial estimation procedures. Table 1. 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 1. 6 and 1. 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 facto rs, 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 14 Count y - year observations without any experience of tornadoes of F2+ are excluded. 28 variable that represents the role of government, although government spending might be w eakly related to the economic variables discussed above. The last specification includes all the poverty - related potential determinants, testing them in a single specification. I n all specification s the following variables are included as controls: average t ornado 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. The EM - DAT data used in most of the prior studies discus sed 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 torn ado (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 cou nty 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). 29 Table 1. 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 1.3 0 4 .3 7 1 5 . 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 Exp enditure 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 Fe male - Headed Household 10.54 4.30 2.88 35.46 4757 * Statistics are from observations with F2 - F 5 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. 30 1.6 RESULTS Table 1. 6 and Table 1. 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 1. 5. I mainly discuss the results of Random Effects Poisson and Negative Binomial 15 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, th e analysi s 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 1. 7 implies that an increase in F - scale to the next level increases expec ted 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 tornadoe s relatively often are estimated to experience 13 % ( lower fatalities than counties outside of the tornado - prone area, all other conditions being equal. This result support s 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 mitig ate the societal impacts. McEntire (2001) asserts that 15 I discuss bot h 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 N eg. B inomial regression model. When explaining the estimated effects of explanatory variables, I refer to the results of Neg. Binomial model in Table 1. 7. 31 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 c ounty population and land area are incl uded in logarithmic term s 16 . 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 fo r 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. I n all estimates it is shown that count ies 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 n ow 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 regardin g the role of income/wealth in determining vulnerability. 16 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. 32 1. 6.1 Richer C ounties E xperience F ewer T ornado - induced D eaths 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 percent 17 . As Anbarci et al. (2005) and Kahn (2005) argued in their studies, it is also found in this study tha t 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 . Ho wever, 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 1. 6 and 1. 7 , where the poverty rate is included but not income per capita in the specif ication. In this regression we see that the poverty rate is positive and statistically significant as expected. The estimated coefficient in column 4 in Table 1. 7 suggest that one percentage point increase in poverty rate is estimated to increase tornado f atalities by 3 percent . 17 The estimated coefficients of log transformed variables can be interpreted as elasticities. 33 1. 6.2 Human C apital P lays an I mportant R ole in R educing T ornado V ulnerability The regression results indicate th at 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 i s found to be negatively associated with the likelihood of deaths in disaster situation s , 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 previo us 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. 34 Table 1. 6 : Socio - economic Characteristics and Disaster Impacts Poisson Random Effect Regressions Results Dependent variable: D eath s from F 2 - 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_ T ornado _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 Gov t Exp onPu blic SafetyWelfare) - 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 In come) - 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 s tandard errors in parentheses , *** p<0.01, ** p<0.05, * p<0.1 35 Table 1. 7: Socio - economic Characteristics and Disaster Impacts Negative Binomial Random Effect Regressions Results Dependent variable: D eath s from F 2 - 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_ T ornado _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 Gov t Exp onPu blic SafetyWelfare) - 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 In come) - 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 36 1.6.3 Mobile Homes Residents Experience More Tornad o Fatalities The fourth hypothesis is that mobile home living results in more tornado fatalities. T he 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 vul nerability to tornadoes. Th e e stimated 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, mo bile home parks could potentially provide common tornado shelter areas to be used in the event of a tornado watch or warning. 1.6.4 Female - Headed Households Are More Vulnerable t o 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 . T hese regression s show that female headed households and tornado - induced fatalities weakly have a positive correlat ion . 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 f emale - hea de d 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 P eek 2006) . However, the estimated effect only achieve s significance in specification (8). 37 1.6.5 Government Spending i n Public Safety Mitigates Losses f rom 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 safe ty , 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 my 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 wel fare over the study period 1980 - 2014, 268 lives would have been saved from tornados 18 . However, considering the limited government resources available for public services, I offer an evaluation of whether it would be worthwhile for local government to alloc ate 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 agenci es 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 c apita, on average, which is 18 The expected number of lives that could have been saved by increasing per capita government expenditure s by 50% has been calculated across all counties who had experienced FS2+ tornadoes over the study period and added up. 38 approximately $ 30 million in extra burden to local governments 19 . Altogether , it does not appear that increasing government expenditures is a cost - effective way of achie ving tornado fatality reduction , even after taking into consideration that the life - saving benefit is just one component of the multiple benefit s that may arise from such government spending. My empirical analysis suggests that general increase in government funds on public safety, prot ection, 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 n eeded 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. 1. 6.6 Additional Analyses In Table 1. 8 , I present the results of additional analyses. The second set of regressions consist of four specifications: (1) and (2) us e very strong tornadoes of Fujita - scale 3 or greater (F3 - F5) in the same framework as in Table 1. 6 and 1. 7, and (3) and (4) exploit the magnitude of tornadoes in a different way compared to the main analysis presented in Table 1. 6 and 1. 7 . First two columns show Poisson and Negative Binomial regressions result for severe tornadoes rated F3 or higher. F ocusing on the larger events reduces the number of tornado events 19 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 t he 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 aver age extra burden to local governments, $ 30 M is obtained by multiplying county population by the per capita extra expenditure, $508. 39 by 3,907, leaving just 1,507 severe tornadoes. Thus, in my sample, only 1,2 45 counti es are used for the analysis excluding those counties without any experien ce of tornadoes of F3+ during 1980 - 2014. The estimates reported in columns (1) and (2) in T able 1. 8 are very consistent with those in Table 1. 6 or 1. 7 with the exception of a few differences. When I 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, torna do vulnerability related variables such as Tornado Alley , lagged tornado frequencies, and population density measure take on greater importance , while t he coefficient on the mobile home variable remains st atistically significant and similar in magnitude compared to the estimation results using F 2 + tornadoes. Taken together, these findings suggest that the stronger tornadoes extend vulnerability to a broader array of people in a community such that social - ec onomic status becomes less important whereas the intensity of natural force and physical factors become more significant in determining vulnerability. In addition, t he larger coefficient s on tornado alley and lagged tornado frequencies compared to the estimate from F2+ tornadoes suggest s 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 1. 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)) 20 , instead of using the average F - scale variable. A s shown in Table 1. 2, tornadoes of different magnitudes can have widely differing degrees of impact. For instance, the average death from F5 20 The fraction of F2 tornadoes (or F3 tornadoes in s pecification (4)) is a reference point. 40 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 o f each level of F - scale events. The results from the Poisson model 21 using tornadoes of F2+ and F3+ are presented in column (3) and (4), respectively. As expected, t ornadoes of different levels of F - scale have largely diffe rent 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. F o r 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 estimat ion results in column (3) again demonstrate that housing quality , population density, income levels, and government spending on public safety and welfare are critical factor s in determining tornado vulnerability . 21 Due to a convergence difficulty in Negative Binomial estimation process, Poisson model is only employed for the estimation. 41 Table 1. 8: Socio - economic Characteristics and Disaster Impacts Additional Regressi ons Results (1) Poisson (2) Neg. Binomial (3) Poisson (4) Poisson Dependent variable D eath s from F 3+ D eath s from F 3+ D eath s from F 2+ D eath s from F 3+ Fscale _F 3 + 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_ T ornado _F2+ - 0.010 (0.045) Lag_ T ornado _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 Gov t Exp on Pu blic Safety & Welfare) - 0.232** - 0.195** - 0.199** - 0.227** (0.113) (0.089) (0.089) (0.111) Log (PerCapita In come) - 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 42 1.7 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 bet ween tornado fatalities and the potential determinants of tornado impacts with in 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 stu dy 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 death s. 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 f emale - headed households tend to be more vulnerable , while the higher the education level, the lower the vulnerability. I n general , households most affected by disasters are those with weaker economic and social bases. The information present ed 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 s uggests that increased government spending in critical areas such as safety, protection, and welfare , reduce overall vulnerability within a community. F or some counties with frequent tornado occurrences and higher fatality rates, extra funds on safety, pro tection, and welfare might mitigate the impacts and save lives effectively. However, a cost - benefit analysis that compares the estimated extra government 43 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 c onsider 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 wh ich their citizens are affected by tornadoes. A nother 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 proportio n 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 U nited 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. I n 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 44 affordability. One potential policy scheme to internalize t his social cost would have governments i) require communal shelters in mobile home parks and communities 22 , 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 a t 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 s ocio - economic nature of tornado impacts and enable decision - makers to improve mitigation efforts. 22 There are communities that already require all mobile home parks to provide storm shelters for their resident s, including the State of Minnesota , and s ome individual counties (e.g. Sedgwick County and Butler County in KS, St. Joseph County , MO , etc.) 45 APPENDIX 46 APPENDIX Table 1. A1 : Socio - economic Characteristics and Disaster Impacts Poisson Fixed Effect Regressions Results Dep endent variable D eath s from F 2+ D eath s from F 3+ 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 Pu blic Safety & Welfare) - 0.197 - 0.189 (0.206) (0.283) Log (PerCapita In come) - 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 47 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. N ote that in Table 1. A1 we observe a revers al of sign on most of socioeconomic variabl es 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. 48 REFERENCES 49 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 revie w 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 adaptation Global 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. Disaster s, 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 - 14 6. New York: Springer. Women's International Hurricane Center. Laboratory for Social and Behavioral Research, 81 - 94. 50 Fothergill , A., P eek , L . A. 200 4 Disasters in the United States: A Review of Natural Hazards 32: 89 110. Brooks 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 c alamity from the viewpoint of human ecology. No. 1. Taylor & Francis. 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 institutio ns The Review of Economics and Statistics 87(2) : 271 - 284. Journal of Urban Economics 63 (3): 788 - 802. Kusenbach 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): 189 96. Merrell, Land Economics 81, 87 - 99 Moore, H. E. 1958. Tornadoes over Texas , Austin: University of Texas Press. Morrow, B. H. and Enarson, E Int J of Mass Emerg and Disasters 14 (1): 5 - 22. 51 Disasters 23 ( 1): 1 - 18. Muttarak 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/ext reme - events/us - tornado - climatology NOAA. 2011 c . 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, 1981 2002." 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 behavior and tornado shelter options among mobile home residents in the United Natural Hazards 48 (2): 191 - 201. Simmons, K., and Sutter, D. 2 013. 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. 52 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 q uasi - 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. 53 CHAPTER 2 FLOOD FATALITIES IN THE UNITED STATES: THE ROLES OF SOCIO - ECONOMIC FACTORS AND THE NATIONAL FLOOD INSURANCE PROGRAM 2.1 I NTRODUCTION Over the 20 - year period from 19 96 - 201 5 , a total of 107,743 floods resulted in 1,563 direct fatalities and over $ 167 billion in damages in the United States (US) . Over the last 30 years floods kill an average of 84 people annually. Floods rank second in terms of resulting fatalities among the different types of life - threatening weather - related events; floods claimed more lives than high intensity disasters such as hurricanes or tornadoes. Importantly, climate scientists predict increases in climate variability and frequency of weather extremes in the coming years. Floods are no exception; Milly et al. (2002) find that the trend of increasing risk of significant floo ds was substantial during the twentieth century and their climate model suggests that the trend will continue. Floods are one of the costliest natural hazard types in the United States, imposing a financial burden on a large number of households and communities. A large portion of property damage during massive hurricanes such as Katrina in 2005 and Harvey in 2017 were the result of f looding triggered by those storms. Given the lack of flood coverage in the private insurance market, the National Flood Insurance Act of 1968 established the National Flood Insurance Program (NFIP) to provide an insurance option priced below actuarial risk - based rates. However, the subsidized premiums of the insurance program have resulted in operating deficits. After a series of devastating hurricanes and superstorms since 2005 (e.g. Katrina, Rita, Sandy, Harvey, 54 bt level to the U.S. Treasury has increased substantially 23 . To improve the financial solvency of the program, government officials attempted to increase the policy premiums to match actuarial rates (Biggert - Waters Flood Insurance Reform Act of 2012), but d ue to the strong opposition by policy holders, the premiums were not increased to the level where claims can be paid without continuing to rely on federal subsidies 24 . There have been mounting concerns and criticisms over the fiscal sustainability of the pr ogram 25 . While the problems of the program have been debated, some of the beneficial components of the program have not been fully evaluated; to my knowledge no existing studies have empirically examined the NFIP as a disaster management scheme . It should be noted that - funded coverage for floods as well as helping to guide and manage community implementation of floodplain management and mitigation practices. In disaster management, ex - ante hazard prevention a nd damage mitigation is at least as important as ex - post recovery efforts, but only the former can help to avert irreversible societal damages and fatalities. By design, the prevention and mitigation efforts of the NFIP are interconnected with the provisio n of flood insurance; the NFIP enables property owners in participating communities to purchase insurance in exchange for the mandatory implementation of floodplain management ordinances for flood risk and damage reduction. In this regard, I 23 . As of September 2017, the NFIP owes $24.6 billion to the U.S. Treasury. In October 2017, the Additional Supplemental Appropriations for Disaster Relief bt is $20.5 billion. 24 The Homeowner Flood Insurance Affordability Act (HFIAA) was enacted in 2014 and reinstated certain premium subsidies and slowed down certain premium rate increases that had been included in the Biggert - Waters Act. 25 NFIP has been identified by U.S. Government Accountability Office - federal program since 2006 as a result of its substantial financial expo sure and operational challenge s. 55 hypothesize that the NFIP has played a substantial role in preventing and reducing the adverse impacts of floods through ex - ante flood risk management. To explore and test the extent to which proactive disaster management practices regulated by the NFIP h elps to reduce disaster impacts, I rely on the integrated view of the physical, social, economic, and political elements of disaster vulnerability. Lethal disasters in the United States such as the Loma Prieta Earthquake in 1989 in California, Hurricane An drew in 1992, and Hurricane Katrina and Rita in 2005 and many other disasters reveal significant differential impacts across different population segments, depending on socio - economic and political status. That is, research is showing that natural disaster Consequently, the socio - political nature of disasters is increasingly the focus of attention in studies of disaster vulnerability. Given that those with lower socio - economic status are more likely to experience the greatest impact s from natural hazards, I use a framework where th e underlyin g social and institutional factors determine vulnerability to floods . To this end , I investigate the relationship between flood - induced fatalities and a wide range of vulnerability indicators such as demographic, socio - economic, and housing characteristics, as well as institutional factors. Considering that there might be a bi - directional process between the disaster - related government activities and dis aster impacts, I test the endogeneity of institutional variables and implement the instrumental variable (IV) estimator by using the Control - function (CF) approach (Wooldridge, 2014). This study explore s yearly flood events that occurred over the 1996 to 2 015 period using US county level data . In my empirical analysis, I control for disaster - specific physical factors (e.g. timing , duration of the incidence ) and area - specific environmental characteristics (e.g. urbanization, the number of dams, and past flood experiences) to assess the socio - economic and 56 institutional factors that increase (or reduce) vulnerability to floods. Using county - year panel structur ed data, I employ the correlated random effects (CRE) framework combined with the Control - function approach, that allows unobserved county heterogeneity to be correlated with observed covariates (Wooldridge, 2010). This study 's contributions to the litera ture are as follows. First, my study provide s a robust assessment of a broad array of structural and social components of disaster vulnerability, including urbanization, past flood experience, education, and housing quality, while controlling for the uniqu e attributes of counties. Second, I examine the role of local government public safety and protection services in mitigating flood impacts, and I do so in a way that corrects for potential simultaneity bias by applying the IV method. Third, this study pres ent s new evidence showing the National Flood Insurance Program (NFIP) has significantly reduced flood - related fatalities. T h e present study reveals range of factors that influence flood vulnerability, which can shows the importance of the proactive mitigation measures and helps policymakers better prepare for future flood events. 2. 2 LITERATURE REVIEW 2. 2.1 S ocio - political Nature of Disasters In general, it has been argued by many scholars that structural, social, political 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 ( Aptekar and Boore 1990, Albala - Bertrand 1993, Cannon 1994, Blaikie et al. 1994, Cutter 1996, Peacock et al. 1997, 57 Enarson and Morrow 1998, Morrow 1999 ). Blaikie et al. (1994) note that vulnerability, in a disaster c ontext, is a person's or group's "capacity to anticipate, cope with, resist, and recover from the impact of a natural hazard". While disaster risk is imposed exogenously by natural forces, the vulnerability of people to naturally occurring extreme events i s influenced by human activity (O'Keefe et al. 1976, Hewitt 1983). Cannon (1994) asserts that people's ability to cope with and recover from hazards depends on economic systems and class structures that allocate resources and access to resources. Cutter et al. (2003) discuss the possible interactions between social and biophysical vulnerabilities that determine overall place vulnerability. In their model, disaster fatalities are largely determined by socio - disaster vulnerability ; hazard potential is either moderated or enhanced via a combination of geographic factors and the social fabric of the place that are influenced by socio - economic status, demographics, and housing characteristics. With special attention to th e institutional component, my study uses this conceptual framework where disaster risk is defined by the combination of bio - physical hazards of nature and societal vulnerability which is shaped by social conditions and structure. 2. 2.2 Economic Developme nt and Disaster Impacts Most disaster studies addressing social vulnerability are qualitative in nature, but there are several quantitative empirical studies that investigate the major factors associated with the disaster - induced fatalities. The relationship between the level of economic development and disaster consequences are primary focus of this research. Burton et al. (1993) argue that the impacts of natural hazard (drought, floods, and tropical cyclones) vary across countries by income level. Similarly, Horwich (2000) asser ts that higher income enables an increase in 58 general safety of society as well as an improvement in protection against natural disasters. An economy's resilience and response to disasters are largely determined by its level of wealth. More recent empirical studies on the determinants of disaster vulnerability use cross - country disaster data obtained from EM - DAT 26 . The work of Kahn (2005), Toya and Skidmore (2007), Strömberg (2007), Kellenberg and Mobarak (2008), Raschky (2008), and Gahia et al. (2013) examin e the role of economic and institutional factors in determining disaster - induced fatalities. Kahn (2005) investigates the relationship between disaster deaths and income, geography, and institutions. He finds that disaster fatalities are negatively correla ted with the level of development. Also, his research shows that democracies and nations with less income inequality tend to suffer fewer deaths from disasters. Another early study on the disaster - safety - development relationship is that of Toya and Skidmor e (2007). Using disaster data from EM - DAT for 151 countries over 44 years (1960 - 2003) and other measures of socio - economic fabric, the study confirms that economic development as measured by per capita GDP is inversely correlated with both disaster deaths and damages. Strömberg (2007) finds that greater wealth and government effectiveness are associated with fewer disaster fatalities. Raschky (2008) and Kellenberg and Mobarak (2008) find a nonlinear relationship between economic development and disaster impacts; economic development reduces disaster losses but with a diminishing rate. Also, Gahia et al. (2013) show that poorer and larger countries suffer more disaster related fataliti es. Brooks et al. (2005) assess vulnerability to climate - related events by developing national - level indicators of vulnerability 26 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. 59 and adaptive capacity. They find that socio - economic, political and environmental factors such as civil and political rights, l ife expectancy, government effectiveness and accountability, and literacy are significant predictors of disaster vulnerability. 2. 2.3 Severe Weather Events in the United States and Disaster Vulnerability Unlike the abovementioned studies on multiple typ es of natural disasters across multiple countries using the EM - DAT data set, there are a few quantitative studies that discuss U.S. natural disasters and the role of various demographic, economic, and political factors. Addressing this research gap, my stu dy focuses on U.S. county level flood events, within the context of socio - economic and political vulnerability. Most empirical studies on U.S. natural disasters examine flood, tornado, and hurricane events. Simmons and Sutter (2013) and Lim et al. (2017) u se detailed U.S. county level tornado data from National Oceanic and Atmospheric Administration (NOAA) to examine the societal determinants of tornado vulnerability. Both studies show that the physical elements of tornado hazard (e.g. tornado intensity) an d socioeconomic and demographic conditions of localities are key determinants of tornado fatalities. Simmons and Sutter (2013) find that education, percentage of non - white and rural population, and percentage of mobile homes are key factors. Lim et al. (20 17) expand the findings of Simmons and Sutter (2013), showing that local governments can and do play a significant role helping to reduce fatalities. The study also finds that income inequality and various dimensions of poverty intensify societal vulnerabi lity to tornadoes, and confirms the existence of learning effects from tornado risk history. In terms of flooding, Zahran et al. (2008) analyze flood events in Texas counties from 1997 - 2001 to examine whether areas with higher concentration of socially vulnerable populations suffer greater fatalities from flood events. They construct an index of s ocial 60 vulnerability using measures of poverty, median income, and race. Their empirical analyses indicate the built - environment and social vulnerability significantly contribute to the degree to which localities are affected by flood events. They consider FEMA rating scores of Texas counties based on the Community Rating System (CRS) and show that FEMA premium discount provides incentives for flood mitigation, reducing the flood casualties. This study further extends and adds to the work of Zahran et al. (2 008) by investigating the US nationwide flood vulnerability over the 20 - year period with particular focus on the roles of institutional factors NFIP and local government in mitigating flood vulnerability. 2. 3 CONCEPTUALIZING HUMAN AND ENVIRONMENTAL C OMPONENTS OF FLOOD VULNERABILITY Based on the conceptual framework where risk is considered to be a function of physically defined natural hazards and socially constructed vulnerability, I hypothesize that three key elements determine the degree of disaster impacts: i) disaster - specific climatic factors, ii) area - specific physical and environmental factors, and iii) socio - political conditions within communities. As shown in Figure 2.1, thes e three conditions together contribute to the overall place vulnerability to natural disasters. Although the third element of the disaster vulnerability is of my main interest , a ll three elements are integrated in order to conduct a robust examination of t he socio - political determinants of flood - induced deaths. 61 Figure 2.1 : Key Elements of Disaster Vulnerability 2. 3.1 Disaster - specific Determinants of Flood Vulnerability First, I control for disaster - specific factors, such as timing and the duration of the events. It is expected that the degree of disaster impacts would be different across the time of day when an event occurred. As in Simmons and Sutter (2011), each flood event is categorized by the time of day: overnight (12:00 - 5:59 AM), morning (6:00 - 11:59 AM), early afternoon (12:00 - 3:59 PM), late afternoon (4:00 - 7:59 PM), and evening (8:00 - 11:59 PM). The time of the day is related to the degree of vulnerability since people are better able to receive warnings, promptly respond, and take actions during the daylight hours. As a climatic element of flood vulnerability, I also control for the month of the flood event. One key factor is the duration of the event; the longer the expos ure to the flood hazard, the more intense are the flood impacts. 62 2. 3.2 Area - specific Physical and Environmental Determinants of Flood Vulnerability I also incorporate area - specific characteristics that capture pre - existing physical and hydrological vulnerability to flood hazards . Those factors are the total number of dams (of all purposes) as well as those for flood control and storm water management l ocated in a county, the percent of urban population, and the flood experience in the previous two years. Dams are constructed for various purposes, for example, water supply, irrigation, power generation, water flow control, and/or flood prevention, etc. I n flood - prone areas, dams are constructed for the specific purpose of flood control and storm water management. The existence of such dams can aid controlling the flow of water during flood events. I thus hypothesize that the number of dam structures play expected that the risk of flooding would be lowered in a county if more dam structures are available for water flow management in flood situations. Another important envi ronmental condition related to flood vulnerability is the pattern of land use and land cover as a result of urbanization. The idea is that urban areas are more likely to be covered by building structures and paved surfaces, and as a result infiltration cap acity of the land is greatly reduced, causing greater surface water runoff. Consequently, urbanization may magnify the risk of flooding (Hollis, 1975). The percent of population living in urban areas (including urbanized areas and urban clusters) in a coun ty is included as a measure of urbanization in the empirical analysis. In addition, the frequency of fatal floods (i.e. number of floods that resulted in one or more fatalities) in the previous two years are included as a measure of flood hazard history of the area. Flooding is one of the most frequent disaster types that occur in most U.S. counties, but most floods are not large - scale events nor deadly. Thus, frequently occurring small - scale flood 63 events are not likely to alarm a community as a whole. Ho wever, significant events with fatalities are more likely to capture the attention of residents through local news and media about the dangers of floods and thus influence their perceptions and behaviors. Significant events may also stimulate local governm ents to increase efforts toward disaster prevention and management. critical role in reducing vulnerability and the future consequences of floods (McEntire 2001 ). 2. 3.3 Socio - political Determinants of Floods Vulnerability I also hypothesize that demographic, socio - economic, housing, and institutional factors including the National Flood Insurance Program, are critical in shaping the overall vulnerability of people and places to disasters. Each key factor is discussed next. Income The level of income within a community is a key factor that determines societal vulnerability to disasters . Communities with higher income and/or wealth have a greater demand for safety and can allocate more resources to safety and protective measures (Wildavsky, 1988). On the other hand, limited financial, physical, and social assets of the poor increase their susceptibility to disasters. The role of income (or wealth) and poverty in disaster contexts has been illustrated in many empirical studies (Kahn 2005, Toya and Skidmore 2007, Strömberg 2007, Raschky 2008, Lal et al. 2009, Gaiha et al. 2013, Lim et al. 2017) . I hypothesize that communities with lower income level (or higher r ates of poverty 27 ) suffer greater flood - related fatalities. In this study, per capita income of U.S. counties is included as a measure of economic status . 27 The correlation coefficient between poverty rate and per capita income level of county is - 0.71. Two measures both represent economic status of counties from different angles but considering the strong correlation between two measures, I only include per capita income measure in the empirical analysis. 64 Education Level Prior disaster studies suggest that education level is closely linked with disaster vulnerability (Brooks et al. 2005; Cutter et al. 2003; Lim et al. 2017; Simmons and Sutter 2013; Skidmore et al. 2007; Muttarak and Lutz 2014) . I include the share of degree holders among population aged 25+ in the estimation as a measure of education level of county population. Education enhances risk perception and promotes disaster preparedness against disasters ( Hoffmann and Muttarak 2017) , which are crit ical preconditions of disaster impact reduction at the individual level. More educated people who have better understanding of the hazard risks are more likely to take preventive measures and be better prepared for the shocks. Thus, I expect that counties with more educated population are less vulnerable in the face of flooding. Housing Quality I also hypothesize that communities with a higher proportion of households living in mobile homes will suffer increased flood - induced fatalities. People living in mobile homes face greater vulnerability due to the structural features of mobile homes that typically have no foundation and are less able to withstand shocks. Moreover, a higher proportion of households living in mobile home implies greater vulnerability in a different context because lower cost mobile homes are often occupied by those who have relatively limited financial resources. Scholars argue that disasters adversely affect people in lower socio - economic status largely because of the types and quali ty of housing they occupy ( Fothergill and Peak 2004). Similarly, minorities may be more likely to live in unsafe, substandard housing , and are thus at greater risk (Aptekar 1991, Phillips 1993, Pastor et al. 2006) . For these reasons, a higher proportion of households living in mobile homes within a county indicates greater physical and socio - economic vulnerability of the community. 65 Local Government Investment I also examine the degree to which local government plays a role in protecting citizens from flood hazards. The idea is that more government resources allocated in safety, protection and welfare can increase overall safety of the localities and strengthen th eir ability to resist the impact of natural hazards, which can lead to the reduction in societal losses and damages. Lim et al. (2017) provide empirical evidence that local government expenditure on emergency services and community protection is a critical factor in reducing tornado impacts in the United States. Following Lim et al. (2017), I also test the role of local government in mitigating the flood impacts by constructing a measure of local government spending on public safety, protection, and welfare, which includes expenditures on fire/police protection and protective inspections/ regulations and housing/community development, and public welfare 28 . However, this type of government expenditure and the disaster occurrence may have a bi - directional relati onship where frequent disaster events in a county would increase its spending on public safety, protection, and welfare. Acknowledging that such government expenditures may not be strictly exogenous, I apply two methods to address potential simultaneity. F irst, I construct a predetermined level of government expenditure by lagging the local government fiscal data, following the prior studies (Garcia - Mila and McGuire 1992, Cullen and Levitt 1999). The government fiscal data are reported every five years (in years ending in 2 and 7 within a decade), so I interpolate the expenditures (inflation - adjusted) and match the yearly data with flood event data set to reduce the possibility of capturing the reverse relationship between consequences of disasters and government activity. Second, I transform the original 28 In the context of local government, welfare services are not direct cash assistant (this comes from state treatment and the like. 66 event data set into a county - year panel structured data and apply the instrumental variable (IV) estimator by using the Control - function approach (Wooldridge, 2014 ) with in the Correlated Random Effects (CRE) framework . National Flood Insurance Program Lastly, as the main concern of this study, I evaluate the role of National Flood Insurance Program (NFIP) in preventing and reducing the loss of human life from flooding through ex - ante floodplain management and mitigation efforts. The NFIP is a Federal program established by the U.S. Congress throug h the National Flood Insurance Act of 1968. The NFIP enables property owners in participating communities to purchase insurance (administered by the government) as financial protection against flood losses, in exchange for the implementation of floodplain management ordinances for flood risk and damage reduction. Participation is based on a cooperative agreement 29 between communities and the Federal Government. In order to participate in the NFIP, communities must meet (or exceed) the minimum floodplain mana gement requirements, through building codes, zoning ordinances, subdivision regulations, health and safety codes, and stand - alone floodplain ordinances. The Federal Emergency Management Agency (FEMA) manages the NFIP and oversees the identification and m apping 30 of flood - prone communities, reviews community adoption and implementation of land use regulation and construction standards, determines flood insurance 29 Once the flood hazard has been identified and an NFIP map has been provided to a community, the identified flood - prone community must assess its flood hazard and determine whether flood insurance and floodplain management would benefit the community's residents and economy. 30 In support of the NFIP, FEMA identifies flood hazards nationwide and publishes flood hazard data such as Flood Hazard Boundary Maps (FHBMs), Flood Insurance Rate Maps (FlRMs) , and Flood Boundary and Floodway Maps management regulations. 67 rates for different mapped zones of risk, provides flood insurance, and funds mitigation project s. The identification of flood hazards is an essential process as it creates an awareness of the hazard and provides communities with the information needed for land use planning, floodplain development, and for emergency management. The 1994 NFIP amendm ent implemented through the National Flood Insurance Reform Act of 1994 directs FEMA to develop a standard form for determining whether the building or mobile home is located in the Special Flood Hazard Area (SFHA) 31 ; in these areas for acquisition and/or c onstruction of buildings, purchasing flood insurance as well as complying with specific building restrictions are mandatory as a condition of Federal or Federally related financial assistance. The floodplain management requirements within the SFHA are desi gned to prevent new development from increasing the flood threat and to protect new and existing buildings from anticipated flood events. The National Flood Insurance Reform Act of 1994 also strengthened the program by enacting a Community Rating System (C RS) that recognized and encouraged community floodplain management activities exceeding the minimum standards of the NFIP. With the CRS, the NFIP further incentivizes communities with discounts on flood insurance premiums to conduct mitigation and outreach activities that further increase safety and resilience of the area. The NFIP also pays special attention to the vulnerability of mobile homes to flooding. FEMA P - 85 titled Protecting Manufactured Homes from Floods and Other Hazards (second edition, init ial edition of FEMA 85 published in 1985 and updated to FEMA P - 85 in 2009) 31 Special Flood Hazard Area (SFHA), which is defined as an area of land that would be inundated by a flood having a 1 pe rcent chance of occurring in any given year (also referred to as the base or 100 - year flood). Development within the SFHA must compl y with local floodplain management ordinances, which must meet the minimum Federal requirements. 68 provides guidance on foundation design and installation of mobile homes in floodplains; these guidelines are designed to make mobile homes less susceptible to floods (and other natu ral hazards). For example, the NFIP require manufactured homes located in Special Flood Hazard Areas be elevated and securely anchored to resist floatation, collapse, or lateral movement. - 85) that addresses the vulnera bility of mobile homes to natural hazards by establishing mandatory regulations and standards governing the mobile homes in hazard - prone areas contributes to the improvement in safety of mobile homes and the resilience of communities as a whole. Risk - transfer mechanisms such as mandatory catastrophe insurance alleviate the impacts of natural hazards, reducing the burden of recovery and welfare losses (Kunreuther 1996, Luechinger and Raschky 2009). However, my hypothesis here is that by identifying flood hazards across the states and promoting and enforcing proper floodplain management and safety standards to mitigate future consequences of floods, NFIP plays a vital role in enhancing resilience and thus reduces vulnerability in flood - prone communit ies. To empirically evaluate the potential life - saving role of the program, a measure of the NFIP participation rate at the county level is constructed . Within - county participation rate is determined by the percent of communities (city, town or township, v illage) 32 within a county that entered in the program at least two years before the year when a flood event occurred. I use lagged participation rates 32 The comprehensive list of the communities city, town or township , village within a county (or county equivalent) is from the list of su bcounty governments that are used for local government finance /employment data, where municipal and township governments are identified by gov ernment type code 2 and 3 , respectively . 69 because the implementation of the NFIP requirements would not take effect immediately in terms of enhancin g overall safety of the community. 33 . 2. 4 EMPIRICAL ANALYSIS 2. 4.1 Data Description T h e analysis uses individual flood event data within U.S. counties over the 199 6 - 201 5 period 34 . D ata on fatalities from floods in the United States are collected from NOAA National Centers for Environmental Information (NCEI) 35 . D etailed information on time, dates, and locations of the event s are also provided. M ajor socio - economic , housing, and government expenditure data at the county level are coll ected from U . S . Bureau of the Census 36 and merged with the flood data. Detailed data on locations and built years of dams are from National Inventory of Dams (NID) published by U.S. Army Corps of Engineers. National Flood Insurance Program (NFIP) participation status of communities is from Federal Emergency Management Agency (FEMA). Table 2.1 presents the total number of various type s of severe weather events in the United States and resulting fatalities and injuries by types of storm events during 1996 - 2015. At the county level, the total 175,863 storm events occurred over t he period. F lood s are the most frequent 33 One might be concern ed about reverse causality where greater flood impacts cause participation rates to increase . I performed the endogeneity tests for two institutional factors - the government expenditure on pub lic safety, protection, & welfare and the NFIP participation rates. The test results suggest that only the government expenditure variable is endogenous and the predetermined NFIP participation rates are non - endogenous once I correct the simultaneity bias resulted from government expenditure variable. I thus, use instrumental variables methods to address the endogeneity of the government expenditure, taking the predetermined NFIP participation rates a s exogenous. 34 Note that f lood events in 1996 and 1997 are only used for constructing the past 2 - county. 35 Data source: www.ncdc.noaa.gov/data - access/severe - weat her 36 Decennial census data for demographic and housing variables, and local government expenditure data are interpolated /extrapolated to obtain yearly data over the study period (1996 2015). 70 disaster type; flood s and flash floods account for 60% of the total climatic events in the United States. The next most frequent event type is the tornado with nearly 1,400 events per year. These storm events induced total of 7,342 deaths and 52,216 injuries over the last 20 years. Heat caused the highest number of fatalit ies , followed by tornado es and flash flood ing . In terms of the average fatality per event, rip current s recorded the highest deaths rate, with 0.7 f atalit ies per event. T ornado es caused the largest number of injuries in total, wh ereas heat and excessive heat together recorded the highest number of injur ies per event. This study focuses on flood - related fatalities 37 . County level flood frequencies and f atalities over the study period are presented in Figures 2.2 2.3 . Table 2.1 : Fatalities and Injuries by Disaster Events , 19 96 201 5 1. T his study explore s flood and flash flood events. The definition/determination of flood and flash flood are provided in Table 2. A 1 in the Appendix. 2. The Hurricane/Typhoon category data only include fatalities, injuries, and damage amounts associated with wind damage (the other hazards are reported in their respective categories.). 37 The historical storm data I have collected from NOAA contain injury data as well, however, the number of persons injured during flood events are not fully reported (whereas the number of persons killed by floods are extensively collected from various sources) and thus, county level in jury data do not represent exact injury count . For this reason, I do not conduct empirical analysis using injury data. Frequency Fatalities Injuries Event Type 1 Total % Total Per event Total Per event Flood 39,893 22.7% 419 0.011 2,320 0.058 Flash Flood 67,850 38.6% 1,144 0.017 6,282 0.093 Rip Current 807 0.5% 569 0.705 561 0.695 Hurricane 2 162 0.1% 6 0.037 17 0.105 HurricaneTyphoon 2 1,350 0.8% 86 0.064 921 0.682 Tornado 27,539 15.7% 1,680 0.061 23,089 0.838 Avalanche 427 0.2% 219 0.513 153 0.358 Heat 16,424 9.3% 1,966 0.120 8,956 0.545 Excessive Heat 5,897 3.4% 422 0.072 5,185 0.879 Debris Flow 429 0.2% 83 0.193 49 0.114 Lightning 15,085 8.6% 748 0.050 4,683 0.310 Total 175,863 100% 7,342 0.042 52,216 0.297 71 Figure 2.2 : Total Number of Floods by County , 199 6 - 2015 Figure 2.3 : Total Deaths from Flood by County , 199 6 - 2015 72 2. 4.2 Empirical Model 2. 4.2.1 Base Model I first analyze the flood vulnerability preserving its original event data structure (also called Cross - Sectional - Time - Series), which contains observations of multiple cross - sectional units over multiple time periods. Note, however, that this data structure is different than a panel (or longitudinal) data structure because it can contain multiple observations of a unit in a year (e.g. +1 observat ions for a county in a specific year is possible if +1 flood events occurred within the county in that year). By retaining this original individual flood event data structure in this base model, the detailed information on each flood can be included. The dependent variable is the number of fatalities caused from each flood event . Among total 107,708 flood events during the study period 1996 - 2015, only 1,067 events resulted in fatalities; for a large portion of observations, the dependent variable is zero. Thus, for the econometric analysis of the flood event data (to which conventional panel data methods cannot be applied), I employ Zero - Inflated Negative Binomial (ZINB) model which properly treats the non - negative count variables with the over - dispers ion ( excess zeros) problem (Long and Freese, 2006) . Because of the distributional features of disaster - induced fatalities, ZINB model is increasingly employed in disaster studies (e.g. Kahn 2005, Zahran et al. 2008). In the ZINB model, the excess zeros are cons idered to be generated by a separate process from the count values and the excess zeros are modeled independently. The ZINB model combines binary Logit model for zero outcomes and Negative Binomial model for event - counts. The ZINB regression analysis is ch aracterized by the following model: 73 (1) Log Likelihood: (2) (3) : the inverse of the logit link (4) : the set of flood observations for which the outcome ( : death) is zero. (5) : Inflation variables for the binary Logit model predicting whether an observation is in the always - zero group where (6) : Covariates for counts model (Negative Binomial) In my empirical analysis, the covariates for the count model of Negative Binomial include the following variables: , a vector of demographic and socio - economic characteristics, as well as institutional factors of the county that may determine fatalities of flood j , , the disaster specific characteristics , , a vector of physical and hydrologic characteristics of the county where the disaster occurred. To control for the unobserved statewide heterogeneity, I also include state fixed effects along with year fixed effects. The detailed list of the explanatory variables in the ZINB analysis is provided in Table 2. 2 . In addition, four key variables are selected from the explanatory variables to serve as inflation variables of ZINB model that determine the probability of being in the always - zero group: previous 2 - capita government expenditure on public safety/welfare. Each of these variables represent past flood history, socio - economic characteristics, and the institu tional components of the affected areas, respectively. 74 Dependent Variable Deaths from each Flood event Explanatory Variables Event - specific Begin time of the event: Overnight, Morning, Early Afternoon, Late Afternoon, Evening Duration (days) Season: Spring, Summer, Fall, Winter Event Type: Flood, Flash Flood Area - specific Environment No. Dams in Total within a county (all purposes) No. Dams for Flood Control and Storm Water Management Percent of Urban Population Past 2 - Demographic Population Size Percent of Population over 65 Percent of Population under 18 Socio - economic Per capita I ncome* Percent of Mobile Homes in Total Housing Units Government 2yr - lagged NFIP Participation Rates within a County* Year Fixed Effects Indicator variables for each year State Fixed Effects Indicator variables for each state S * These factors serve as inflation variables of the ZINB model. T he hypothesized socio - economic characteristics and institutional components are examined with a set of control variables: i) population characteristics: county population size, percent of population over age 65 and under 18 , ii) area - specific physical and env ironmental factors: the number of dams in total and for flood control, percent of urban population, and the previous 2 - iii) event - specific climatic factors: begin time of the day and month of the event, duration of the incidenc e , and the type of the flood events . Summary statistics for all variables included in the ZINB analysis are presented in Table 2.3 . The variable definitions and data sources are provided in Table 2. A 2 in the Appendix. 75 * Summary statistics of year dummies (1998 - 2015) and 50 state dummies that are included in the ZINB estimation are not presented here. Mean Standard Deviation Min Max Number of Obs. Dependent Variable Fatalities from individual event 0.014 0.188 0 20 97,416 Independent Variables Total No. Dams (all purposes) 32.879 39.091 0 331 97,416 No. Dams for Flood Control 6.228 14.675 0 230 97,416 Pct Urban population 50.773 31.526 0 100 97,416 Past 2 - 7.465 9.842 0 115 97,416 Ln(Population) 10.908 1.560 4.205 16.115 97,416 Pct Over65 14.068 3.631 2.148 43.641 97,416 Pct Under18 24.131 3.053 7.605 45.207 97,416 Pct Bachelor Degree 20.274 9.663 1.868 76.762 97,416 Pct Mobile home 11.878 8.662 0 61.29 97,416 Ln(Per Capita Income) 10.026 0.231 8.787 11.118 97,416 Lagged NFIP Participation Rates 62.937 32.366 0 100 97,416 Lagged Ln(Per Capita Gov Exp on Public Safety in thousands ) - 0.949 0.678 - 6.943 2.031 97,416 Duration (in days) 1.356 4.218 0 30.999 97,416 Flood 0.368 0.482 0 1 97,416 Flash Flood 0.632 0.482 0 1 97,416 Overnight 0.205 0.404 0 1 97,416 Morning 0.225 0.418 0 1 97,416 Early Afternoon 0.185 0.388 0 1 97,416 Late Afternoon 0.229 0.420 0 1 97,416 Evening 0.156 0.363 0 1 97,416 Spring 0.295 0.456 0 1 97,416 Summer 0.413 0.492 0 1 97,416 Fall 0.164 0.370 0 1 97,416 Winter 0.127 0.333 0 1 97,416 76 2. 4.2.2 Instrumental Variable Model As discussed above, disaster - related government activities might not be strictly exogenous, rather bi - directional. In consideration of the potential simultaneity bias, I implement the Poisson Instrumental Variable (IV) estimator by using the Control - function (CF) approach (Wooldridge, 2014) which accounts for both endogenous regressors and non - negative outcome variable. F irst , the endogeneity tests are performed for two i nstitutional factors - the government expenditure on public safety, protection, & welfare and the NFIP participation rates. Both the robust Hausman test and CF - based Hausman test results 38 suggest that only the government expenditure variable is endogenous. I could not find evidence of the endogeneity of the predetermined NFIP participation rates. Hence, the instrumental variables methods is used to correct the endogeneity of the government expenditure, taking the predetermined NFIP participation rates as exogenous. I consider two variables as an IV for the local government expenditure on public safety, protection & welfare. One is the number of government entities (e.g. city, town or townsh ip, village ) within a county and the other is the ratio of the highest and the lowest (census tract level) effective tax rates for real estate within a county 39 . As the county government expenditures are aggregated values that include expenditures of all su bdivisions located within the county, an increase in the number of governmental entities is likely to increase government spending. 38 For examining the endogeneity of the two institutional variables, I performed two tests. First, a test statistic is used, defined as the difference of two Sargan - Hansen statistics: one for the equation with the smaller set of instruments, where the suspe ct regressors are treated as endogenous, and one for the equation with the larger set of instruments, where the suspect regressors are treated as exogenous. This statistic is reported after ivreg2 in Stata, which are robust to various violations of conditi onal homoskedasticity (Baum et al. 2007). Second, based on the Control - function approach, I carried out the regression - based Hausman tests of whether the suspected endogenous variables are actually endogenous (Wooldridge, 2014). 39 Decennial Census and Amer ican Housing Survey data on the tract level aggregate real estate tax and the aggregate housing values are used to construct the effective tax rates of tracts. 77 However, changes in the number of local governmental entities is unlikely to have a direct impact on flood fatalities. A rat ionale for the use of the latter IV the highest/lowest ratio of the effective tract property tax rates within a county is that a greater difference in the effective tax rates within a county means higher inequality in property values and wealth as well as local government tax bases (and thus expenditures) within the county. Economic inequality and disproportionate police expenditures among nearby communities in a county might be a fostering ground for crime and generate negative spillovers across distri cts (Simon Hakim 1980; Furlong and Mehay 1981 ). The crime spillovers thus drive higher spending on public safety and protection within intra - county areas ( Stephen Mehay 1977; Hakim et al. 1979;). However, the tax rate differentials within a county is unlik ely correlated with flood fatalities. I formally test the validity of the two IVs by performing Weak identification test (Kleibergen - Paap Wald F statistic) and Overidentification test (Hansen J Statistic) as well as a regression - based correlation test betw een two IVs and the dependent variable (using Poisson CRE and Poisson FE). The tests of the validity of IVs I performed all suggest that the proposed instruments are reasonable. To deal with the county heterogeneity while at the same time handling the endogeneity of the government spending, I adopt the correlated random effects (CRE) framework, combined with the CF approach that allows unobserved heterogeneity to be correlated wit h observed covariates (Wooldridge, 2010). For the application of CF in a CRE setting, county - year panel structured data is constructed . Some of the flood event specific details are averaged or aggregated by year (e.g. average deaths per flood, total durati on of floods) while for event timing variables including time of the day and season, I generate shares of floods in each category by year (e.g. % of floods occurred in Spring, % of floods occurred in the morning, etc.). Summary statistics of the variables included in the Poisson IV (CF) model are presented in Table 2 .4 . 78 * Summary statistics of year dummies (1998 - 2015) and the mean values of the explanatory variables for the Correlated Random Effects (CRE) estimation are not presented here. Mean Standard Deviation Min Max Number of Obs. Dependent Variable Annual Avg. Fatalities per flood 0.013 0.119 0 6 29,680 Explanatory Variables Total No. Dams (all purposes) 29.811 37.128 0 331 29,680 No. Dams for Flood Control 5.407 14.136 0 230 29,680 Ln(Population) 10.578 1.410 6.104 16.115 29,680 Pct Over65 14.291 3.725 2.148 43.641 29,680 Pct Under18 24.124 3.125 7.721 44.185 29,680 Past 2 - 0.059 0.263 0 5 29,680 Pct Urban Population 45.565 30.880 0 100 29,680 Pct Bachelor Degree 19.042 8.987 1.868 76.762 29,680 Pct Mobile home 12.844 8.882 0 59.950 29,680 Ln(Per Capita Income) 10.001 0.221 8.997 11.118 29,680 Lagged NFIP Participation Rates 62.096 32.057 0 100 29,680 Ln(Per Capita Gov Exp on Public Safety &Welfare in thousands ) - 0.974 0.667 - 7.488 1.501 29,680 Duration (in days) 4.408 15.924 0 476.572 29,680 Pct Flash Flood 0.656 0.410 0 1 29,680 Pct Flood 0.344 0.410 0 1 29,680 Pct Overnight floods 0.194 0.309 0 1 29,680 Pct Morning floods 0.217 0.322 0 1 29,680 Pct Early Afternoon floods 0.185 0.306 0 1 29,680 Pct Late Afternoon floods 0.240 0.340 0 1 29,680 Pct Evening floods 0.164 0.292 0 1 29,680 Pct Spring floods 0.298 0.387 0 1 29,680 Pct Summer floods 0.414 0.421 0 1 29,680 Pct Fall floods 0.154 0.305 0 1 29,680 Pct Winter floods 0.134 0.291 0 1 29,680 Instrumental Variables (IVs) Number of Subdivisions 14.118 14.454 1 151 29,680 H/L Ratio of Real Estate Tax Rates 2.392 4.893 1 131.746 29,680 79 2. 5 RESULTS 2. 5.1 Base Model Results from ZINB Estimation I first present in Table 2.5 the estimates from the Zero - Inflated Negative Binomial (ZINB) model using individual flood event s recorded at the scale of count ies during 19 96 - 201 5 40 . The dependent variable is fatalities from each flood event. T he determinants of flood fatalities are estimated with three specifications, controlling for state and year fixe d effects in all specifications. As a part of ZINB model, the results of the logit model for predicting whether an observation is in the always - zero group are presented in columns (2), (4), and (6). The key policy variable of interest, the NFIP participation rate, is introduced into the second and third specifications 41 . The estimated effects of the vulnerability factors on flood fatalities from specifications A, B, and C are largely consistent in direction but differ in magnitude once NFIP partic ipation rate is included . In particular, comparing the specifications A and B, I find that t he coefficients on income level and government expenditure variable decrease in magnitude as the NFIP variable is incorporated into the model. However, except for t he per capita income, precision of the estimates of the other socio - economic factors and the government expenditure variable is low. 40 Note that the previous 2 - flood experience is incorporate d as an explanatory variable an d accordingly, flood observations in 1996 and 1997 are used as pre - sample data. Flood observations during 199 8 - 2015, total 97,416 flood events are used in the estimation procedure. 41 In specification C, I test whether the result is sensitive to a change in the choice of inflat ion variables for the first stage logit model and whether the inclusion of the lagged variable - previous flood experience (although it is not exactly a lagged dependent variable) causes any compl ications in the estimation process and leads to any notable changes in the estimates. 80 Table 2.5 : Determinants of Flood Fatalities Zero - Inflated Negative Binomial Regressions Results Model A Model B Model C Dependent Variable (1) (2) (3) (4) (5) (6) : Deaths from Floods ZINB Logit ZINB Logit ZINB Logit Overnight - 0.292** - 0.283** - 0.270** (0.125) (0.127) (0.127) Morning - 0.363*** - 0.344*** - 0.336*** (0.117) (0.120) (0.119) Early Afternoon - 0.244** - 0.228* - 0.232* (0.118) (0.120) (0.120) Late Afternoon - 0.409*** - 0.394*** - 0.387*** (0.114) (0.116) (0.116) Flash Flood 0.682*** 0.703*** 0.702*** (0.106) (0.108) (0.110) Spring 0.182** 0.193** 0.223** (0.089) (0.090) (0.091) Fall 0.322*** 0.333*** 0.332*** (0.106) (0.108) (0.108) Winter - 0.182 - 0.189 - 0.136 (0.128) (0.129) (0.131) Duration_days 0.029*** 0.030*** 0.026** (0.010) (0.010) (0.011) Total No. Dams (all purposes) - 0.004*** - 0.005*** - 0.005*** (0.002) (0.002) (0.002) No. Dams for Flood Control 0.008** 0.007** 0.007** (0.003) (0.003) (0.003) PastFloodExperiences_2yr - 0.009 0.025** - 0.006 0.028** (0.008) (0.011) (0.009) (0.011) Pct BA Degree - 0.002 - 0.002 - 0.002 (0.010) (0.010) (0.010) Pct Mobile home 0.008 0.009 0.009 (0.008) (0.009) (0.009) Ln (Per Capita Income) - 1.734*** - 2.975*** - 1.679*** - 2.817*** - 1.591** - 2.904*** (0.643) (0.879) (0.641) (0.849) (0.651) (0.892) Lagged NFIP Participation Rate - 0.018*** - 0.011** - 0.026*** - 0.011** - 0.026*** (0.003) (0.005) (0.006) (0.005) (0.006) Lagged ln(Govt Exp on Public - 0.141 - 0.435 - 0.053 - 0.321 - 0.037 - 0.258 Safety, Protection & Welfare) (0.290) (0.332) (0.281) (0.316) (0.295) (0.340) Constant 10.947* 31.743*** 10.963* 30.930*** 10.580 31.936*** (6.486) (9.048) (6.409) (8.669) (6.513) (9.152) Observations 97,416 97,416 97,416 97,416 97,416 97,416 1. Cluster - adjusted robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.10 2. The omitted categories are " Evening ", "Summer" , and "Flood" . 3. Logit estimation indicates that counties with less past experiences, higher income, higher NFIP participation rates, and more government expendit ures on public safety are less likely to be in the always - zero group . 4. The estimates of several control variables (pct_elderly, pct_young, lnpopulation, pct_urban) and state and year fixed effects are not reported here. 81 Disaster - specific Determinants of Flood Fatalities As the estimations in the base model examine individual flood events within the original event data structure, the relationship between flood event - specific details and fatalities can be more precisely estimated. Estimation results demo nstrate that event - specific factors such as duration, timing of the event, and flood types are key factors that affect the degree of flood impacts. As expected, a longer duration of flood ing significantly relates to the number of deaths from flood events . The estimates also show that the degree of flood impact is different across the time of day when an event begins to occur. My analysis suggests that the impact of a flood tends to be greater when it occurs in the evening. Estimates also show that fatalit ies from flood events are higher in the fall season (Sep. to Nov.) when the frequency of flooding is typically low and thus unexpected. Comparing to floods, flash floods are estimated to be deadlier. Demand for Safety The base model result is consisten t with the well - known argument that c ommunities with higher income have a greater demand for safety and allocate more resources to safety and protective measures, mitigating societal vulnerability to disasters. My estimation results support the previous fi ndings that the higher income is associated with the increase in safety against disasters. A negative association between income and flood fatalities is statistically and economically significant in all specifications. Moreover, I could also see a strong n egative relationship between the NFIP participation rates and flood impact. T he estimates from specification B and C indicate that flood - fatalities decrease by about 10.4%, on average, for a ten - percentage point increase in the within - county NFIP participa tion rate. The underlying mechanism is that a higher demand for safety against flooding within a community can be translated into the adoption of NFIP, promoting proper floodplain management and mitigation efforts, ultimately reducing flood fatalities. 82 The base model estimation using ZINB provides evidence for the role of income and the National Flood Insurance Program in mitigating flood vulnerability. However, I acknowledge that there are limitations in this estimation method. By retaining the original event data structure, panel methods that would allow us to deal with the county heterogeneity and the possible endogeneity of institutional factors could not be ap plied. Thus, the estimates here may not reflect the true relationships due to these issues, which I am not able to address in the current data configuration. In particular, the ZINB model exploits variation across counties, rather than the within - county, a nd hence, the estimates here may not be the basis of the causal inference. We proceed to the next subsection to discuss the bias - corrected results using the Instrumental Variable method in a CRE setting. 2. 5.2 Instrumental Variable Model Results from IV P oisson CRE approach I present in Table 2.6 the estimation results from the Poisson Instrumental Variable (IV) model using Control - function (CF) method, correcting the endogeneity of the government spending on public safety, protection, and welfare. I use two IVs - the number of local government entities and the highest/lowest ratio of the effective tract property tax rates within a county. I exploit the county - year panel data spanning from19 96 to 201 5 42 , controlling for the county heterogeneity within t he Correlated Random Effects (CRE) framework. I discuss in detail the results of the estimated relationship between flood fatalities and key explanatory variables such as area - specific characteristics, and socio - political vulnerability factors. 42 I incorporate the previous 2 - flood experience as an explanatory variable and accordingly, flood observations in 1996 and 1997 are used as pre - sample data. 83 Table 2.6 : Determinants of Flood Fatalities Poisson IV CF / CRE Estimates of Key Explanatory Variables Poisson IV (CF) Correlated Random Effects Estimator Dependent Variable Model A Model B Model C Model D : Avg Deaths per Floods (1) (2) (3) (4) Duration_days 0.047*** 0.057*** 0.057*** 0.055*** (0.014) (0.015) (0.014) (0.015) Total No. Dams (all purposes) 0.080 0.134 0.151 0.153 (0.138) (0.142) (0.142) (0.141) No. Dams for Flood Control - 0.173 - 0.243 - 0.260 - 0.253 (0.215) (0.221) (0.226) (0.226) Pct Urban Population 0.203*** 0.240*** 0.198*** 0.210*** (0.055) (0.063) (0.061) (0.065) Past 2 - - 6.138*** - 6.754*** - 6.756*** - 6.680*** (0.804) (0.900) (0.934) (0.930) Pct Mobile home 0.194*** 0.207*** 0.241*** 0.241*** (0.075) (0.074) (0.080) (0.084) Pct BA Degree - 0.316*** - 0.383*** - 0.380*** (0.102) (0.108) (0.114) Ln (Per Capita Income) 11.720*** 11.821*** (3.388) (3.468) Lagged NFIP Participation Rate - 0.029* - 0.028* - 0.024 (0.016) (0.016) (0.016) Ln (Govt Exp on Public Safety, - 3.433 - 4.487* - 4.698* - 6.804** Protection & Welfare) (2.492) (2.573) (2.684) (2.871) Constant - 75.436* - 90.287** - 85.866** - 113.841** (38.667) (39.916) (41.212) (44.3742) Exogeneity Test (p - value) 0.069 0.031 0.032 0.005 Observations 29,680 29,680 29,680 29,680 1. Cluster - adjusted robust standard errors in parentheses, *** p<0.01, ** p<0.05, * p<0.10 2. The unit of observation is a county - year. U.S. counties with more than one flood experience in a given year are included. 3. The estimation results not reported here are the first stage Control - function, the mean values of regressors for CRE, and the control variables estimates. The full results are available upon request. Area - specific Physical and Environmental Determinants of Flood Fatalities A s et of area - specific environmental and hydrologic characteristics of the affected area is considered such as past 2 - negative correlation between previous flood fatalities and current flood deaths implies that ning from their past experiences. The 84 results of all specifications consistently demonstrate that learning effects from risk history play an important role in increasing the coping capacity of communities, thus reducing disaster fatalities in areas that re cently experienced lethal flooding. Contrary to the strong positive estimates from the ZINB model, the number of dams for flood control variable now has a negative coefficient (although insignificant with p - value of 0.24 in specification 4). In the ZINB model, the positive coefficient of the number of dams for flood control appears to capture the positive correlation, rather than a causal relationship, between the level of flood risk and the number of structures for flood control. However, once I control for county heterogeneity and handle th e endogeneity issue in the IV/CRE model, the estimates presumably reveal a causal relationship. An increase in the availability of dams for flood management helps localities prevent massive water flow into human settlements, thus reducing the risks of life - threatening floods. Also, the positive coefficients of the percent urban population suggests the urbanization worsens the flood impacts. One explanation for the greater flood vulnerability of more urbanized counties is that paved surfaces and concentrated building structures tend to reduce infiltration capacity of the land and consequently intensify flood risk. Socio - political Determinants of Floods Fatalities In addition to the environmental and hydrologic factors of flood vulnerability, my findings rev eal a significant role of socio - economic and political factors in determining flood impacts. First, I find that the education closely linked with the flood vulnera bility. The results indicate that counties with a higher proportion of educated people experience fewer flood - related fatalities. This result is consistent with the findings of previous empirical disaster studies (Brooks et al. 2005; Lim et al. 2017; Simmo ns and Sutter 2013; Skidmore et al. 2007; Muttarak and Lutz 2014) . Higher education 85 level is associated with enhanced risk perception and proper disaster preparedness and responses. More educated people may have better understanding of the hazard risks and thus are likely to be better prepared for shocks, thus reducing disaster vulnerability. I also examine the degree to which housing quality, as measured by the percent of mobile homes in the county, is closely linked to flood vulnerability. The estimation s in Table 2.6 show a robust and significant result; the estimates of mobile homes are positive with statistical significance. The result confirms that housing quality is one of the more important determinants of flood impacts. Mobile homes are increasingl y filling a demand for affordable housing across the states; from 2006 to 2015 nearly half of new manufactured homes were shipped to the seven coastal states in the South region (Texas, Louisiana, Mississippi, Alabama, Florida, South Carolina, and North Ca rolina). Greater vulnerability of those living in mobile homes suggests important policy implications for disaster management and community vulnerability assessment (Fothergill and Peek 2004, Merrell et al. 2005, Schmidlin et al. 2009, Kusenbach et al. 201 0, Lim et al. 2017). Notably, once the IV method is applied to deal with the simultaneity between government resource allocation decision and disaster risk, I obtain a statistically significant evidence that local government spending in public safety, protection, and welfare plays a role in helping to mitigate human losses from floods. For example, a five percent increase in government spending is expected to r educe flood fatalities by about 20 percent. Consistent with the previous findings from an analysis of tornado impacts by Lim et al. (2017), it appears that overall safety of a county can be enhanced through local government public safety and protection ser vices. 86 Contrary to the results of the base model, once I correct the simultaneity bias and control for the unobserved heterogeneity, I find a positive relationship between income level and flood fatalities. The difference in the estimated results is first attributed to the fact that ZINB estimator mostly explains the variation across the cross - sectional units, while the CRE method exploit within variation. we should also note that income level/wealth can in fact influence flood outcomes through various pat hways, including those that are already take n into account education level, location and housing choices, as well as local government resource availability for emergency management and disaster mitigation. For instance, in the first stage Control - functio n estimation, the per capita income is estimated to be a dominant and significant factor of government expenditure decision. Thus, these pathways by which income translates to increased safety are included in the estimation, effectively capturing the effec t of individual and community level attitudes and efforts for safety and preparedness against flood within a county. Controlling for these other factors, the positive coefficient on income may reflect the idea that growing income translates to increased ho using in higher amenity areas such as near water where flooding risk is higher. The Role of NFIP on Floods Fatalities We now discuss the results of the main policy variable of this study the role of National Flood Insurance Program (NFIP) in helping communities become more aware of and better prepare for the risks of floods, and avoid the adverse impacts. The NFIP participation rate is consistently estimated to have a statistically sig nificant negative effect in both the ZINB and the IV Poisson estimations, while the magnitude of the effect is relatively larger in the IV estimation. In particular, the estimates from specification C indicate that flood - fatalities are reduced by about 24% on average if the within - county NFIP participation rate increases by ten percentage point, whereas the estimated effect in 87 the ZINB model was about a 10% reduction in fatalities. Biases from the simultaneity and county heterogeneity may account for the di fferences in magnitude of effects and statistical precision between the two model specifications. Comparing the column 3 and 4, we observe that the coefficient on the government spending on safety, protection, and welfare greatly rise in absolute magnitud e once the NFIP variable is excluded from the model. The enhanced resilience and increased safety of the place for public safety and protection services. One possi ble explanation for this result is that local government resources allocated to public safety and protection, and compliance with the flood mitigation measures required by the NFIP work towards the same goal of improved flood safety. Thus, when the NFIP pa rticipation rates are incorporated to explain the variability of flood fatalities, the explanatory power of the local government variable is reduced. The change in the estimated effect of this key institutional factor further highlights the importance of t aking into consideration the role of NFIP when analyzing the flood vulnerability. Post - estimation: Hypothetical NFIP Participation Rates and Predicted Fatalities T he average within - county participation rate was 27% in 1980 and doubled to 54% by 1990. Dur ing the study period from 1996 - 2015, the average participation rate has risen by 12% from 57% to 69%. To calculate the effect of the flood program in terms of saving lives, I predict the change in fatality rates using the estimation results of model 3 in T able 2.6 . I also compute the predicted outcome of several hypothetical cases: i) the participation rate had not grown at all from 1996 to 2015 in any county, remaining at the same rate as in 1996 for the whole study period, ii) the participation rates were 20% lower for all counties, iii) 30% lower, and iv) the participation rates were 50% lower (to reflect a NFIP participation rates of about zero). A 88 comparison of the predicted death rates in various scenarios and the actual flood deaths is presented in Ta ble 2.7 . Prediction Valuation NFIP Participation Rates Death Rate 1 Expected Total Death 2 Difference 3 : no. lives saved Value of Lives Saved 4 In - Sample .0137 1,563 Hypothetical Cases Remain at 1996 level .0155 1,814 251 $ 2.26 billion Lower by 20% 5 .0213 2,442 879 $ 7.92 billion Lower by 30% 5 .0254 2,972 1,409 $ 12.68 billion Lower by 50% 5 .0359 4,327 2,764 $ 24.87 billion 1. Death rate indicates the county - year average of predicted deaths per flood event. 2. The actual number of total flood occurrence in the U.S. counties during 1996 - 2015 is used for calculation. 3. The difference is between each hypothetical case and in - sample prediction. This indicates the potential loss of lives prevented by NFIP. 4. The value of a statistical life used in calculation is $9 million (Viscusi, 2014) 5. The standard deviation of the NFIP participation rate is 32.37% and the yearly avera ges range from 53% to 69%. Each row in Table 2.7 indicates the predicted death rates (i.e. fatality per flood event) and expected total death counts from floods by in - sample or hypothetical NFIP participation rates. Following the practice of giving an economic value to mortality a value of a statistic al life (VSL), I also perform a straightforward calculation of the benefit of the NFIP in saving lives from floods. 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). Table 2.7 shows that the predicted deaths increase as the NFIP participation rates fall. If the NFIP participation rate had not risen and remained at the 1996 level for 20 years for all counties, we would have suffered 251 more fatalities from flood events during t he 1996 - 2015 period. The estimated value of lives saved due to the expanded adoption and implementation of 89 the program since 1996 across the nation is $2.26 billion. Moreover, the impact would be greater by 20% relative to the actual rate; the calculations suggest that the same number of flood events would have resulted in additional 879 deaths during the period. For the last scenario - the NFIP participation rates were close to zero my calculations indi cate that we would have experienced 2,764 more deaths from floods during the 1996 to 2015 period, implying that the program has helped prevent flood - induced fatalities, which is valued at $25 billion ndings provide evidence that flood - prone communities become more flood - resistant due to the enforcement of floodplain management requirements of the NFIP, and in turn, the loss of human life induced by flooding is reduced in high flood risk areas across th e United States. 2. 6 CONCLUSION While floods are exogenously determined by climatic and environmental factors, this study shows that socio - economic and institutional factors are critical in determining vulnerability. T his paper seeks to uncover th e underlying factors that make people and places more vulnerable to floods in the United States . To this end , I investigate the relationship between flood fatalities and the potential human and institutional component s of disaster impacts with in US counties over the 1996 to 2015 period . The study f indings enable us to identify the societal characteristics and government factors that exacerbate or mitigate vulnerability to hazards and the extent to which different population groups are disproportionally affected by floods . This in turn allow s us to suggest policies that may help mitigate human losses from such events. The empirical analysis in this paper demonstrate s that people most affected by disasters like floods are primarily those who have weaker economic and social bases, those who are less 90 educated and have limited risk perception and preparedness, and those who are living in homes less resistant to shocks. People living in mobile homes are at greater risk due to the structural vulnerability of mobile homes as well as due to the lower socio - economic status. Another key finding is that urbanization and past disaster experience are critical components of flood vulnerability. The analysis using IV method shows that increased government spending in critical areas such as safety, protection, and welfare, is associated with reduced overall community vulnerability to floods. Above all, this paper provides new evaluation of the life - saving role of National Flood Insurance Program. To my knowledge, this is the first empirical study that presents evidence that the National Flood Insurance Program has played a vital role in reducing flood fatalities. My findings suggest that t he benefits of the NFIP in terms of saving lives over the 20 - year study period are and the public concerns regarding the fiscal soundness of the program nece ssitate a thoughtful reform of the NFIP, which must ensure a balance between the affordability of flood insurance and the financial solvency of the program. In this redesign process, the benefits of the proactive disaster management of the NFIP ought to be taken into account. Overall, this study reveals which population subgroups are most vulnerable to flooding in the United States, as well as local and federal government public actions that serve to reduce vulnerability. Generally, these findings increase our understanding of the socio - political nature of disaster impacts, enable decision - makers to better prepare for and respond to pending catastrophic events, and guide mitigation efforts at the local, state and national levels. 91 APPENDIX 92 APPENDIX Table 2. A 1 : Determination of Flood - category Events Determination by Event Type Flood Any high flow, overflow, or inundation by water which causes damage. In general, this would mean the inundation of a normally dry area caused by an increased water level in an established watercourse, or ponding of water, that poses a threat to life or pro perty. If the event is considered significant, it should be entered into Storm Data , even if it only affected a small area. Urban and small stream flooding commonly occurs in poorly drained or low - lying areas. These are types of areal flooding and are to b e recorded as Flood events, not Heavy Rain. Flash Flood A life - threatening, rapid rise of water into a normally dry area beginning within minutes to multiple hours of the causative event (e.g., intense rainfall, dam failure, ice jam). Ongoing flooding ca n intensify to the shorter - term flash flooding in cases where intense rainfall results in a rapid surge of rising flood waters. Every Flash Flood event that occurred and meets the criteria will be logged in Storm Data , regardless of whether or not a flash flood warning was issued. Source: N ational Weather Service Instruction 10 - 1605 (MARCH 23, 2016) Operations and Services Performance, Storm Data Preparation. ( http://www.nws.noaa.gov/directives/ ) 93 Table 2. A 2 : Variable Definitions and Sources VARIABLE DEFINITION SOURCE F lood Fatalities Direct deaths from a flood - category event NOAA NCEI* Overnight Begin time of a flood event is 12 :00 - 5:59 AM NOAA NCEI* Morning Begin time of a flood event is 6:00 - 11:59 AM NOAA NCEI* Early Afternoon Begin time of a flood event is 12:00 - 3:59 PM NOAA NCEI* Late Afternoon Begin time of a flood event is 4 :00 - 7:59 PM NOAA NCEI* Evening Begin time of a flood event is 8 :00 - 11:59 PM NOAA NCEI* Jan Dec Begin month of a flood event NOAA NCEI* Flood T ype NOAA NCEI* Flash Flood T ype NOAA NCEI* Duration _days Duration of a flood event in days NOAA NCEI* Total No. Dams (all purposes) Total number of dams within a county regardless of the main purposes of dams US Army Corps of Engineers No. Dams for Flood Control Total number of dams within a county for flood control and storm water management US Army Corps of Engineers Previous 2 - Flood Experience County level flood fatalities in the previous two years NOAA NCEI* Pct NFIP Participating Communities_Lag Percent of communities (city, town, village) within a county that entered in the National Flood Insurance Program at least two years prior to a flood event FEMA National Flood Insurance Program (NFIP) Pct Urban Population Percent of the county population living in urban areas (= urbanized areas and urban clusters) US Census: Geography Ln (Population) County population in natural logarithm US Census: Population Pct Over 65 Percent of population 65 years old and over US Census: Population Pct Under18 Percent of population 18 years old and under US Census: Population Pct BA degree Percent of people aged 25 and over holding Bachelor's degree US Census: Population Pct of Mobile Homes Percent of mobile/manufactured homes in housing units US Census: Housing Ln (Per Capita Income) County Per Capita Income which is derived by dividing the total income of a county by its total population in natural logarithm. US Census: Income Ln (Per Capita on Public Safety &Welfare) Local government spending (in thousands, 2009 $) on public safety, protection, and welfare in natural logarithm, which includes expenditures on fire/police protection, protective inspections/ regulations, housing/community development, public welfare US Census: Local Government Finances * NOAA NCEI: National Oceanic and Atmosphere Administration National Center for Environmental Information 94 REFERENCES 95 REFERENCES Albala - Bertrand, J.M. 1993. Political Economy of Large Natural Disasters. New York: Oxford University Press. Aptekar, L. 1991. The psychosocial process of adjusting to natural disasters . Natural Hazards Research and Applications Information Center, University of Colorado. International Journal of Mental Healt h : 77 - 90. Baum, C.F., M.E. Schaffer, and S. Stillman. 2007. ivreg2: Stata module for extended instrumental variables/2SLS, GMM and AC/HAC, LIML, and k - class regression. Boston College Department of Economics, Statistical Software Components S425401. Down loadable from http://ideas.repec.org/c/boc/bocode/s425401.html . Blaikie, P., Cannon T., Davis, I., and Wisner, B. 1994. Vulnerability, and Disasters. New York: Routledge. Global 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 , edited by A. Varley, Chichester, N ew York: John Wiley and Sons, 13 - 29. Cullen, J., and Levitt, S. 1999. "Crime, urban flight, and the consequences for cities." The Review of Economics and Statistics 81 (2): 159 - 169. Cutter, S.L. 1996. "Societal responses to environmental hazards." Inter national Social Science Journal 48 (150): 525 - 536. Social Science Quarterly 84 (2): 242 - 261. International Hurricane Center. Laboratory for Social and Behavioral Research, 81 - 94. FEMA P - ds A multi - 96 Natural Hazards 32: 89 110. Furlong, W.J., and Mehay, S. 14 (1): 44 57. https://doi.org/10.2307/134839 Brooks World Poverty Institute Working Paper, 181. Garcia - Mila, T., and McGuire, T.J. 1992. "The contribution of publicly provided inputs to states' econo mies." Regional Science and Urban Economics 22 (2): 229 - 241. Urban Studies 17 (3): 265 76. https://doi.org/10.1080/00420988020080581 . Land Economics 55 (2): 200 212. https://d oi.org/10.2307/3146062 . Hewitt, K., ed. 1983. Interpretations of calamity from the viewpoint of human ecology. No.1. Taylor & Francis. Education and Experienc World Development 96 (August): 32 51. https://doi.org/10.1016/j.worlddev.2017.02.016 . Water Resources Research 11 (3): 431 - 435. Economic Development and Cultural Change 48: 521 - 542. Ka The Review of Economics and Statistics 87 (2): 271 - 284. ris Journal of Urban Economics 63 (3): 788 - 802. Journal of Risk and Uncertainty 12 (2): 171 187. ility and evacuation readiness: Natural Hazards 52: 79 - 95. 97 Lal, P.N., Singh, R., and Holland, P. 2009. Relationship between natural disasters and poverty: a Fiji case study . SOPAC Miscellaneous Report 678. Journal of Public Economics 93 (3): 620 633. Lim, J., Loveridge, S., Shupp , R., and Skidmore, M. 2017. "Double danger in the double wide: Dimensions of poverty, housing quality and tornado impacts." Regional Science and Urban Economics 65 :1 - 15 Long, J.S. and Freese, J. 2006. Regression models for categorical dependent variable s using Stata. Stata press. McEntire, D. A. 2001. "Triggering agents, vulnerabilities and disaster reduction: towards a holistic paradigm." Disaster Prevention and Management: An International Journal 10 (3): 189 - 196. Mehay Southern Economic Journal 43 (3): 1352 59. https://doi.org/10.2307/1057793 . e determinants of tornado casualties and Land Economics 81: 87 - 99 Nature 415 (6871): 5 14 - 17 Disasters 23 (1): 1 - 18. Ecology and Society 19 (1). https://doi.org/10.5751/ES - 06476 - 190142 . NWS. 2016. National Weather Service Instruction 10 - 1605. Operations and Services Performance, NWSPD 10 - 16, Storm Data Preparation http://www.nws.noaa.gov/directives/ Nature 260: 566 - 567. Pastor, M., Bullard, R.D., Boyce, J.K., Fothergill, A., Mo rello - Frosch, R. and Wright, B. 2006. In the wake of the storm: Environment, disaster and race after Katrina . Russell Sage Foundation. Peacock, W.G., Morrow, B.H., and Gladwin, H. 1997. Hurricane Andrew: Ethnicity, gender, and the sociology of disasters. Psychology Press. 98 - term International Journal of Mass Emergencies and Disasters 11 (1): 99 - 110. Raschky, P.A. 2008. "Institutions and the losses from natural disasters." Natural Hazards and Earth System Sciences 8 (4): 627 - 634. Schmidlin - seeking behavior and tornado shelter options among mobile home residents in the United Natural Hazards 48 (2): 191 - 201. Simmons, K., and Sutter, D. 2013. Economic and societal i mpacts of tornadoes . Springer Science & Business Media. The Journal of Economic Perspectives , 199 - 222. Economics Letters 94: 20 - 25. Handbook of the Economics of Risk and Uncertai nty 1: 385 451. Wildavsky, A.B. 1988. Searching for Safety. Vol. 10. Transaction Publishers. Wooldridge, J.M. 2010. Econometric Analysis of Cross Section and Panel Data . MIT Press. - Maximum Likelihood Estimation and Testi ng for Nonlinear Journal of Econometrics , Causality, Prediction, and Specification Analysis: Recent Advances and Future Directions, 182 (1): 226 34. https://doi.org/10.1016/j.jeconom.2014.04.020 . Zahran, S., Brody, S.D., Peacock, W.G., Vedlitz, A., and Grover, H. 2008. "Social vulnerability and the natural and built environment: a model of flood casualties in Texas." Disasters 32 (4): 537 - 560. 99 CHAPTER 3 GROWING HEAT VULNERABILITY OF AGING SOCIETY: THE POTENTIAL ROLE OF HEAT ISLAND MITIGATION MEASURES 3.1 INTRODUCTION Greenough, 2001; Beniston and Stephenson, 2004) 100 43 Aubrecht 43 Tota l 172 statewide or community level actions database are publicly available from the U.S. Environmental Protection Agency (EPA) website https://www.epa.gov/heat - islands/heat - island - community - actions - database 101 102 103 3. 2 RISK OF EXTREME HEAT IN THE U.S. 104 Figure 3.1 : Heat Index Chart Figure 3.2 : Summertime Average Max imum Heat Index vs. Daily Maximum Temperature (1998 - 2011) 105 44 45 The defi nition/determination of heat and excessive heat are provided in Table 3. A 1 in the Appendix. 44 Service (NOSS NWS). Retrieved from https://www.weather.gov/phi/heat . 45 Storm Events Database available at h ttps://www.ncdc.noaa.gov/stormevents/ftp.jsp . 106 Greenough, 2001; Beniston and Stephenson, 2004) 107 Figur e 3.3 : Total Number of Heat Waves by State (1996 - 2015) Figure 3.4 : Heat - Induced Fatalities per Year by State (1996 - 2015) 108 3.3 COMMUNITY HEAT ISLAND MITIGATION ACTIONS 46 46 Total 172 statewide or community level actions are publicly available at the U.S. Environmental Protection Agency (EPA) website https://www.epa.gov/heat - islands/heat - island - community - actions - database 109 Table 3. 1: Heat Island Mitigation Actions List by Initiation Year (1985 - 2017) Trees & Vegetation Cool Roofs Green Roofs Cool Pavements Others (HVI , etc. ) Years Total 1985 - 1989 2 0 0 0 0 1990 - 1994 2 1 0 0 0 1995 - 1999 4 2 3 1 1 2000 - 2004 19 9 10 8 0 2005 - 2009 35 19 18 23 1 2010 - 2014 22 19 15 15 0 Total 84 50 46 47 2 Years Total 1995 - 1999 0 1 0 0 0 2000 - 2004 2 2 1 0 0 2005 - 2009 3 3 1 0 1 2010 - 2014 6 2 2 3 0 Total 11 8 4 3 1 110 111 3.4 LITERATURE REVIEW Aubrecht 47 47 Previous studies on heat vulnerability index for areas within the United States are only introduced. 112 3.5 CONCEPTUALIZING MULTI - FACETED HEAT VULNERABILITY 113 Figure 3.5 : Conceptualizing Multi - Faceted Heat Vulnerability 114 115 3. 5 .1 Major C omponents of Heat Vulnerability Heat Hazard Profile 116 Climatic and Environmental Conditions 48 Demographic & Socio - Economic Characteristics 48 As an alternative measure of the urbanization, the percent of urban population was also considered . However, due to the strong correlation between the percent of urban population and the population size (correlation coefficient = 0.84), I incorporate the urban density, controlling for the population size in the empirical analysis. 117 Åström 118 Previous studies discuss that p ( Extension: Poverty + Aging Society Klin enberg, 1999) 119 3. 5 . 2 Institutional Efforts for Mitigation and Adaptation 120 3. 6 EMPIRICAL ANALYSIS 3. 6 .1 First - Phase: Heat Hazard Mitigation Model 121 49 49 The number of heat wave days is computed at the county level, the totals show the number of heat wave days per county per year. When the geographic area spans more than one county, an extreme heat event is counted for each cou nty where measurable observations that met the heat event definition occurred. ( National Climate Assessment, 2015. Extreme Heat Events: Heat Wave Days in May - September for years 1981 - 2010 ) 122 50 51 50 In earlier years, there are very limited number of counties tha t adopted any HIM measures. For example, prior to 2006, counties with 2 HIM measures make up less than 1% of total observation. 51 For example, a well - known factor of Urban Heat Island urban growth is expected to be captured by the county trend effects. A continuing urbanization trends are found nation - wide, but the rate of the urban growth may vary across the counties. However, the county level urban population data are only available decennially. The interpolation method is commonly used in practice to treat the decennial data to obtain a monotonic interpolation of data, i.e. a linear trend. I include the urbanization measure in the FE specification (Table 3. A2) but do not in the RTM specification, as the trend effect in the RTM would capture county - specific urbanization trends that influence the heat hazard intensities. 123 Table 3.2 : List of Variables in the Heat Hazard Mitigation Model 124 Table 3.3 : Heat Vulnerability and Heat Island Mitigation Actions by Metropolitan Status Direct Fatalities Resulted from Heat Events (1996 2010) Heat Island Mitigation Actions Adoption Status Metro + Micro + Rural Metro + Micro Metropolitan only Avg. Obs. % Avg. Obs. % Avg. Obs. % No Actions taken 0.031 39,128 84% 0.050 22,190 82% 0.074 13,879 80% 1 or more Actions 0.094 7,447 16% 0.146 4,705 18% 0.196 3,461 20% Total 0.041 46,575 100% 0.067 26,895 100% 0.098 17,340 100% 125 3. 6 .2 Second - Phase: Heat Vulnerability Fatality Model 52 53 52 Data source: www.ncdc.noaa.gov/data - access/severe - weather 53 Decennial census data for years 1990 and 2000, and American Community Survey data for year 2015 are used for demographic and housing variables. They are interpolated to obtain ye arly data over the study period 1996 2011. 126 (a) (b) (c) (d) (e) (f) 127 128 Table 3.4 : List of Variables in the Heat Fatality Model 129 3. 6 .3 Heat Island Mitigation Actions and Heat Fatality: A Direct Estimation 130 Table 3.5 : Summary Statistics 131 3. 7 RESULTS 3. 7 .1 First - Phase: Heat Hazard Mitigation Model 54 54 I also estimate alternative specifications using the FE approach that allows only a level effect, but not a county - specific time trend. The result is presented in Table 3. A 2 in the Appendix. However, Wooldridge Test (2002) indicates that these FE models suffer from the presence of serial correlation, supporting the choice of the Random Trend Model with cluster - robust standard errors. 132 Table 3.6 : Heat Hazard Model: The Role of Heat Island Mitigation Actions Panel Fixed Effects & Random Trend Model Results 133 134 3.7 .2 Second - Phase: Heat Vulnerability Fatality Model Heat Hazard Profile 135 Table 3.7 : Heat Vulnerability Fatalities Model Zero - Inflated Negative Binomial Regressions Results 136 Table 3.7 Climatic and Environmental Conditions 137 Demographic & Socio - Economic Characteristics to the individual vulnerability of elderly and population characteristics of aging society as a whole . 138 139 55 56 Table 3.8 : Increase in Heat Fatalities Given the Projected Growth of the Elderly Population in 2030 and 2040 55 56 I am primarily interested in the projected increase in heat fatalities in relation with the growth in elderly in this calculation while considering the poverty rates among them. However. due to the assumption of Ceteris paribus of multiple regression, the coefficient of Pct Elderly in specification 3 means a partial effect of increase in Pct Elderly, holding other variables including the elderly poverty rates constant. Assuming the poverty rates among elderly will remain the same, I estimate the predicted increase in heat fatalities as a result of the aging population using the result in s pecification 3. 140 3. 7 .3 Heat Island Mitigation Actions and Heat Fatality 3.7.3.1 First and Second Phase Models Combined: A Mediated Effect 57 57 141 Table 3.9 : The Mediated Effect of Heat Mitigation Actions on Heat Fatalities 142 3.7.3.2 A Direct Estimation of the Effect 58 58 The heat days measure is closely related with the Heat Index but it only accounts for the temperature component of extreme heat, but not the humidit y component. 143 Table 3.10 : A Direct Estimation of the Effect of Heat Island Mitigation Actions on Heat Fatalities Poisson FE and RE Model Key Results 144 3.8 CONCLUSION 145 APPENDIX 146 Table 3. A 1 : Determination of Heat and Excess Heat Source: N ational Weather Service Instruction 10 - 1605 (MARCH 23, 2016) Operations and Services Performance, Storm Data Preparation. ( http://www.nws.noaa.gov/directives/ ) 147 Table 3. A 2 : Heat Hazard Model : The Role of Heat Island Mitigation Actions Alternative Specifications: Fixed Effects OLS 148 Table 3. A 3 : Heat Hazard Model : The Role of Heat Island Mitigation Actions Random Trend Model Using Heat Wave Days as an Alternative DV 149 REFERENCES 150 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 revie w 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. Albala - Bertrand, J.M. 1993. Political Economy of Large Natural Disasters. New York: Oxford University Press. Aptekar, L. 1991. The psychosocial process of adjusting to natural disasters . Natural Hazards Research and Applications Information Center, University o f Colorado. International Journal of Mental Health : 77 - 90. on Maturitas 69 (2): 99 105. https://doi.org/10.1016/j.maturitas.2011.03.008 . Aubrecht, Christoph, and Dilek Environment International 56 (June): 65 77. https: //doi.org/10.1016/j.envint.2013.03.005 . - Aburto, Luis A. Cifuentes, and - Related Mortality in Latin America: A Case - Crossover Study in Sao Pa International Journal of Epidemiology 37 (4): 796 804. 151 Global and Pla netary Change 44 (1 4): 1 9. Blaikie, P., Cannon T., Davis, I., and Wisner, B. 1994. Vulnerability, and Disasters. New York: Routledge. ility and Global Environmental Change , Adaptation to Climate Change: Perspectives Across Scales, 15 (2): 151 63. h ttps://doi.org/10.1016/j.gloenvcha.2004.12.006 . Browning, Christopher R., Danielle Wallace, Seth L. Feinberg, and Kathleen A. Cagney. 2006. - Related Mortality: The Case of the 1995 Chicago H American Sociological Review 71 (4): 661 678. Cutter, S. L. 1996. "Societal responses to environmental hazards." International Social Science Journal 48 (150): 525 - 536. rability to Social Science Quarterly 84 (2): 242 61. https://doi.org/10.1111/1540 - 6237.8402002 . Influencing U.S. Tornado Fatalities and Injuries, 1998 Demography 44 (3): 669 85. https://doi.org/10.1353/dem.2 007.0024 . International Hurricane Center. Laboratory for Social and Behavioral Research, 81 - 94. , Ethnicity and Disasters in the Disasters 23 (2): 156 73. Natural Hazards 32 (1 ): 89 110. Gago, E.J., J. Roldan, R. Pacheco - Renewable and Sustainable Energy Reviews 25 (September): 749 58. https://doi.org/10.1016/j.rser.2013.05.057 . Greenough, G, M McGeehin, S M Bernard, J Trtanj, J Riad, and D Engelberg Potential Impacts of Climate Variability and Change on Health Impacts of Extreme Environmental Health Perspectives 109 (Suppl 2): 191 98. 152 - Related Mortality: A R eview and Exploration of Journal of Epidemiology & Community Health 64 (9): 753 60. https://doi.org/10.1136/jech.2009.087999 . Hansen, Alana, Linda Bi, Arthur Saniotis, and Monika Nit Global Health Action; Jarfalla 6. http://dx.doi.org.proxy1.cl.msu.edu/10.3402/gha.v6i0.21364 . Harlan, Sharon L., Anthony J. Brazel, Lela Prashad, William L. Stefanov, and Larissa Larsen. Social Science & Medicine 63 (11): 2847 63. https://doi.org/10.1016/j.socscimed.2006.07.030 . Harlan, Sharon L., Juan H. Declet - Barreto, William L. Stefanov, and Diana B. Petitti. 2013. Vulnerability in Mari Environmental Health Perspectives (Online); Research Triangle Park 121 (2): 197. http://dx.doi.org.proxy2.cl.msu.edu/10.1289/ehp.1104625 . Hondula, David M., Rob ert E. Davis, Matthew J. Leisten, Michael V. Saha, Lindsay M. Veazey, - Scale Spatial Variability of Heat - Related Mortality in Philadelphia County, USA, from 1983 - 2008: A Case - Environmental Health 11 (Mar ch): 16. https://doi.org/10.1186/1476 - 069X - 11 - 16 . Pattern of Land Surface Temperatures, Land Cover and Neig hborhood Socioeconomic Journal of Environmental Management 92 (7): 1753 59. https://doi.org/10.1016/j.jenvman.2011.02.006 . Intergovernmental Panel on Climate Change (IPCC). 2013. Summary for Policymakers. In: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change [Stocker, T.F., D. Qin, G. - K. Plattner, M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y. Xia, V. Bex and P.M. Midgley (eds.)]. Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA. Johnson, Daniel P., Austin Stanforth, Vijay Lulla, and Geor Applied Extreme Heat Vulnerability Index Utilizing Socioeconomic and Environmental Applied Geography 35 (1 2): 23 31. https://doi.org/10.1016/j.apgeog.2012.0 4.006 . Review of Economics and Statistics 87 (2): 271 84. https://doi.org/10.1162/003 4653053970339 . Journal of Urban Economics 63 (3): 153 788 802. opsy of the 1995 Chicago Heat Theory and Society; Dordrecht 28 (2): 239 95. http://dx.doi.org.proxy2.cl.msu.edu/10.1023/A:1006995507723 . Kochanek, Kenneth D., Arialdi M. Mi niño, Sherry L. Murphy, Jiaquan Xu, and Hsiang - Ching Annual Review of Public Health 29 (1): 41 55. https://doi.org/10.1146/annurev.publhealth.29.020907.090843 . Lal, Padma Narsey, Padma Narsey Lal, Reshika Singh, and Paula Holland. 2009. Relationship between Natural Disasters and Poverty: A Fiji Case Study . SOPAC. Lim, Jungmi n Regional Science and Urban Economics 65 (Supplement C) : 1 15. https://doi.org/10.1016/j.regsciurbeco.2017.04.003 . Long, J.S. and Freese, J. 2006. Regression models for categorical dependent variables using Stata. Stata press. Loughnan, Marga Subtropical Brisbane, Australia: A Guide for Heatwave Preparedness and Health International Scholarly Research Notices 2014 (February): e821759. https://doi.org/10.1155/2014/821759 . Science 305 (5686): 994 997. Meschede GERONTOLOGIST , 51:322 322. OXFORD UNIV PRESS INC JOURNALS DEPT, 2001 EVANS RD, CARY, NC 27513 USA. asing Risk of Nature 415 (6871): 514 17. https://doi.org/10.1038/415514a . Disasters 23 (1): 1 - 18. M Societal Outcomes of Extreme Weather in a Changing Climate: An Integrated Annual Review of Environment and Resources 36 (1): 1 25. https://doi.org/10.1146/annurev - environ - 060809 - 100145 . 154 National Weather Service Instruction 10 - 1605 (MARCH 23, 2016) Operations and Services Performance, Storm Data Preparation. ( http://www.nws.noaa.gov/directives/ ) National Climate Assessment - Extreme Heat Events: Heat Wave Days in May - September for years 1981 - 2010 on CDC WONDER Online Database, released 2015. Accessed at http://wonder.cdc.gov/NCA - heatw avedays - historic.html North America Land Data Assimilation System (NLDAS) Daily Air Temperatures and Heat Index, years 1979 - 2011 on CDC WONDER Online Database, released 2013. Accessed at http://wonder. cdc.gov/NASA - NLDAS.html on Jan 11, 2018 1:37:10 PM Pastor, M., Bullard, R.D., Boyce, J.K., Fothergill, A., Morello - Frosch, R. and Wright, B. 2006. In the wake of the storm: Environment, disaster and race after Katrina . Russell Sage Foundation. Peacock, W. G., Morrow, B. H., and Gladwin, H. 1997. Hurricane Andrew: Ethnicity, gender, and the sociology of disasters. Psychology Press. Height, and Atmospheric Urban Forestry & Urban Greening 13 (3): 495 506. https://doi.org/10.1016/j.ufug.2014.03.003 . asters: Sheltering, housing, and long - term International Journal of Mass Emergencies and Disasters 11 (1): 99 - 110. Natural Hazards and Earth System Science 8 (4): 627 6 34. Ana V. Diez - Environmental Health Perspectives 117 (11): 1730. Rosenzwei g, Cynthia, William D. Solecki, Jennifer Cox, Sara Hodges, Lily Parshall, Barry Lynn, Richard Goldberg, Stuart Gaffin, Ronald B. Slosberg, and Peter Savio. 2009. Scientific E Bulletin of the American Meteorological Society 90 (9): 1297 1312. Climate Research 24 (3): 255 65. https://doi.org/10.3354/cr024255 . Amer. Meteor. Soc . 155 Stafoggia, Massimo, Francesco Forastiere, Daniele Agostini, Annibale Biggeri, Luigi Bisanti, - Related Mortality: A Multicity, Population - Based, Case - Epidemiology (Cambridge, Mass.) 17 (3): 315 23. https://doi.org/10.1097/01.ede.0000208477.36665.34 . Strader, Stephen M., Walker S. Ashley, Thomas J. Pingel, and Andrew J. Krmenec. 2017. Climatic Change 141 (2): 301 13. https://doi.org/10.1007/s10584 - 017 - 1905 - 4 . The Journal of Economic Perspectives 21 (3): 199 222. Journal of Applied Meteorology and Climatology 47 (11) : 2846 2856. Synnefa, Afroditi, Theoni Karlessi, Niki Gaitani, Mat Santamouris, D. N. Assimakopoulos, and B uilding and Environment 46 (1): 38 44. https://doi.org/10.1016/j.buildenv.2010.06.014 . Tan, Zheng, Kevin Ka - Lun Mitigating Daytime Urban Heat Island Effects in a High - Energy and Buildings , SI: Countermeasures to Urban Heat Island, 114 (February): 265 74. https://doi.org/10.1016/j.enbuild.2015.06.031 . Economics Letters 94 (1): 20 25. https://doi.org/10.1016/j.econlet.2006.06.020 . Uejio, Christopher K., Olga V. Wilhelmi, Jay S. Golden, David M. Mills, Sam P. Gulino, and - Urban Societal Vulnerability to Extreme Heat: The Role of Hea t Exposure and the Built Environment, Socioeconomics, and Neighborhood Health & Place , Geographies of Care, 17 (2): 498 507. https://doi.org/10.1016/j.healthplace.2010.12.005 . Report Number P23 - 212. Washington, DC. U.S. Environmental Protection Agency (EPA). 2008. Reducing urban heat islands: Compendium of strategies. Draft. https://www.epa.gov/heat - islands/heat - island - compendium . Zahran, Sammy, Samuel D. Brody, Walter Gillis Peacock, Arnold Vedlitz, and Himanshu nd Built Environment: A Model of Disasters 32 (4): 537 60. https://doi.org/10.1111/j.1467 - 7717.2008.01054.x . 156 nd quasi - maximum - Journal of Econometrics. 48 (1 2), 29 55. Wooldridge, Jeffrey M. 2010. Econometric analysis of cross section and panel data . Cambridge, MA: MIT press.