Government Effectiveness and Institutions As Determinants of Tropical Cyclone Mortality

Government effectiveness and institutions as determinants of tropical cyclone mortality

Elizabeth Tennanta,1 and Elisabeth A. Gilmoreb

aDepartment of Economics, Cornell University, Ithaca, NY 14853; and bDepartment of International Development, Community and Environment, Clark University, Worcester, MA 01610

Edited by Arild Underdal, University of Oslo, Oslo, Norway, and approved October 7, 2020 (received for review April 6, 2020)

Strong institutions as well as economic development are gener- ardous conditions. This can be represented as follows (e.g., refs. ally understood to play critical roles in protecting societies from 6–8): the adverse impacts of natural hazards, such as tropical cyclones. risk = f (hazard, exposure, vulnerability), [1] The independent effect of institutions on reducing these risks, where the risk, in this case, the probability of mortality from however, has not been confirmed empirically in previous global tropical cyclones, is a function of the hazard (the frequency and studies. As a storm’s path and intensity influence the severity of intensity of storms), exposure (the assets or population in the the damages and may be spatially correlated with human vul- hazard zone), and the vulnerability (susceptibility to harm) of nerabilities, failing to accurately capture physical exposure in an the exposed population. econometric analysis may result in imprecise and biased estimates Empirical efforts to relate vulnerability and risk will therefore of the influence of the independent variables. Here, we develop be confounded by hazard and exposure if these variables are not an approach to control for physical exposure by spatially interact- also accounted for. Studies of vulnerability that include multiple ing meteorological and socioeconomic data for over 1,000 tropical classes of hazard are unable to control for intensity and exposure, cyclone disasters from 1979 to 2016. We find evidence that higher as events of different types (i.e., earthquakes, storms, floods, and levels of national government effectiveness are associated with heat waves) are not directly comparable. As a result, estimates lower tropical cyclone mortality, even when controlling for aver- of socioeconomic risk factors for vulnerability will be imprecise. age income and other socioeconomic conditions. Within countries, Indeed, previous large-N empirical efforts that have pooled dif- deaths are higher when strong winds are concentrated over areas ferent types of hazards have been unable to provide statistical of the country with elevated infant mortality rates, an indica- evidence of the relative importance of different socioeconomic tor of institutional effectiveness through public service delivery. risk factors for natural disaster mortality (9, 10). Measures of These results suggest that policies and programs to enhance insti- democracy and the quality of institutions, including government tutional capacity and governance can support risk reduction from effectiveness, are found to be correlated with natural disaster extreme weather events. deaths, but these effects are not precisely estimated when con- sidered in combination with other possible explanatory variables tropical cyclones | disasters | institutions | vulnerability such as GDP per capita (10, 11). Furthermore, if hazard is correlated with socioeconomic conditions, the failure to control for characteristics of hazard exposure can result in biased estimates. etween 1979 and 2016, over 418,000 people across 85 coun- In Fig. 1, we illustrate how, from 1996 to 2016, countries with tries and territories have lost their lives in tropical cyclone B more-effective governments had lower mortality from tropical disasters.∗ However, there is substantial variation in the degree of harm. Out of more than 4,000 tropical storms and cyclones recorded between 1979 and 2016, about 20% triggered human- Significance itarian disasters, and less than 5% resulted in more than 100 deaths. As recently as 2008, Cyclone Nargis killed over 138,000 Tropical cyclone disasters frequently result in substantial loss people in Myanmar. Nargis was a powerful category 3 or 4 storm of life. Institutional capacity and economic development are at landfall, but tropical cyclones with similar wind speeds struck believed to play protective roles, but previous efforts have several other countries that year with far fewer fatalities. Under- been unable to disentangle their relative effects. We estab- standing what drives this large variation in impacts may provide lish empirically that stronger national and subnational insti- guidance on how we can prevent mortality from future storms, tutions, independent of income, are associated with lower which will be of increasing importance as countries grapple with tropical cyclone mortality. This suggests that effective insti- complex vulnerabilities to extreme weather events under climate tutions play an important role in the success of disaster risk change (5). reduction strategies. Our approach of accounting for hazard This paper investigates relationships between tropical cyclone intensity, population exposure, and socioeconomic conditions mortality and institutional, economic, and human development at high resolutions can be extended to other hazards and (collectively referred to as “development”). We focus, in par- scales to further examine how institutions moderate risk. ticular, on the role of institutional effectiveness, going beyond previous efforts in two important ways. First, we establish an Author contributions: E.T. and E.A.G. designed research; E.T. performed research; E.T. empirical association between national government effectiveness analyzed data; E.T. wrote the paper; and E.A.G. provided critical feedback.y and tropical cyclone deaths that cannot be explained away by The authors declare no competing interest.y income, health, or education. Second, we present a global analy- Published under the PNAS license.y sis showing that locally elevated infant mortality rates (IMRs) in This article is a PNAS Direct Submission.y the exposure zone are associated with increased tropical cyclone 1 To whom correspondence may be addressed. Email: [email protected] mortality. We interpret this as evidence that tropical cyclones This article contains supporting information online at https://www.pnas.org/lookup/suppl/ are more deadly when they impact areas with weaker public ser- doi:10.1073/pnas.2006213117/-/DCSupplemental.y vices due to limited local institutional capacity or the failure of First published November 3, 2020. national programs to be inclusive of all vulnerable populations. *The statistics presented in this paragraph are the authors’ calculations based on data Natural hazards, including tropical cyclones, result in disas- from refs. 1–3. All datasets and code for this study are publicly available via the ters only when vulnerable human systems are exposed to haz- replication files at https://doi.org/10.6077/89ba-bj79 (4).

28692–28699 | PNAS | November 17, 2020 | vol. 117 | no. 46 www.pnas.org/cgi/doi/10.1073/pnas.2006213117 Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 enn n Gilmore and Tennant hriflecshwcnuietentoa niomn sto fur- is environment and national response, the and conducive preparedness how plays disaster state influences in The ther (23). role change direct climate of risk a and context disaster the and institutions in adaptation of for reduction quality factors the concludes enabling that are Report governance confidence” Assessment high Fifth “very Intergovernmental Change with The Climate cyclones. nat- on tropical from Panel as mortality such reducing hazards, for ural important particularly capital, be social 23). may 19, of (7, institutions levels corre- quality to higher better other theorized or for as rates, been poverty proxy such also lower a deaths, have be disaster that may development reduce effect of relationship GDP aspects observed income The lated the whether 22). drive unclear 21, development is (16, of it facets devel- model, Because other multiple single 20). or include a not 19, in do factors (7, opment studies vulnerabilities cyclone tropical new existing create also are or can individ- risk activities exacerbate on reduce growth-targeting of development that positive; levels economic unambiguously activities of not higher and effects assets the reflect in However, per may risk. investment GDP This collective higher or 18). with ual in 17, countries result strike (16, to they tend capita intensity when physical similar deaths include of that fewer storms adaptation that and observe risk stud- hazard Recent cyclone exposure. tropical and of intensity ies in variations for account to 15). and socioeconomic 14 on refs. (e.g., storms exposure of repeated of impacts areas the in from development arise inciden- could be or could tal, variables socioeconomic and tropical were between exposure countries Correlation cyclone speeds. those wind within dangerous people to exposed more though even cyclones (see data to tropical underlying due from the underestimated in be Exposure tracks may (13). storm basin are codes Ocean missing Indian abbreviations country the alpha-3 in Country 119 occurring 3166 1996–2016. cyclones (exceeding ISO from the winds country on annual strength by based Average cyclone modeled (1). EM-DAT is tropical the km/h) more to from indicate data exposure scores on deaths population higher disaster based WGI; cyclone are tropical the 1996–2016 annual from from Average taken (12). scores are governance effectiveness effective 2016 to government 1996 national from Average 1996–2016. countries, 1. Fig. ntttoa fetvns n nlsvt tmlil scales multiple at inclusivity and effectiveness Institutional able better are hazard of class particular a to restricted Studies oenne otlt n xouefrtoia cyclone-affected tropical for exposure and mortality Governance, IAppendix SI o details). for untoa eeoeete nisiuinlefciees(e.g., characterize effectiveness develop- to institutional conflict, literatures in civil health both heterogeneities the public of from subnational and zone. insights importance exposure economics, and the the data ment in study on conditions draw to local We and able factors are risk national mor- we and Second, factors tality. socioeconomic to between ability our correlated relationships improves be identify so may Doing exposure conditions. cyclone socioeconomic con- that with and possibility precision the storm for increases for trols exposure controlling population 35–38). First, and 33, advantages. intensity 3, multiple (2, provides Project This Precipitation the Unified from Center’s Pre- data diction Climate for Administration rainfall Atmospheric Archive and to Track Oceanic and National Best the (IBTrACS) from Stewardship tracks Climate storm on wind based modeled parametrically profiles to each matched spatially socioeco- for are and data data nomic estimates population socioeconomic gridded meth- and Time-variant spatial meteorological storm. use we match system, to human ods the and of hazard interaction physical the the from we results mortality 2016—where cyclone inclusion. 1996 tropical and to Because capacity from institutional 1979 of first indicators from subnational the government test second dataset, national to this the relationship of analy- the effectiveness—and subsets test Our we two 2016. 2016—where to on to nearly 1979 based of from is dataset disasters sis a both cyclone construct explicitly at tropical We that factors 1,500 exposure. models risk hazard using multiple for level, account subnational of and importance national the the for testing by federal of both. capacity or the institutions, building local on agencies, focus actions, to their target whether to example, underde- makers for policy which of ability at the limited. scales constrains is This vulnerability the cyclone of tropical to understanding contributes velopment our regions. result, affected socioeconomic a in averages and As national quality from differ institutional may do local conditions therefore how and 22), consider 16, dis- (9–11, not hazards level of country climate the (±) other studies to and restricted 30 global are cyclones most and tropical from However, 10 mortality 2). aster between affected (Fig. regions with latitude coastal highly countries, degrees in is within concentrated cyclones also areas tropical but to across Exposure exposure heterogeneous variation physical varied. within-country of sufficiently patterns to when are contribute are mortality cyclone to effects tropical likely their in or and are social inequalities they of these political large, forms the If other institutions, inequality. and local economic groups, of certain quality of uneven marginalization the haz- to cyclone due tropical ard against be protections may national locales from and excluded People vulner- 30–32). of (7, patterns marginalization existing and by ability shaped is investments and policies for 29). institutions and local 28 upon refs. (e.g., depend implementation also integrat- but and effective bureaucracy an plans, central planning—require development evacuation and early risk and disaster risk—including ing shelters cyclone systems, tropical warning managing Government killed 27). for (1, that Myanmar programs Nargis in Cyclone people 2008 Pakistan 138,000 the East approximately and former killed (26) in that Bangladesh) people cyclone 500,000 (now Bhola to 1970 250,000 disas- the estimated cyclone include an tropical These deadliest history. the in of ters have some resources and in govern- will, implicated capacity, contrast, of been In lack the the 24). as own (19, such demonstrated failures hazards its ment repeatedly to have in adaptation societies for important capacity civil function- also and high is states with bilat- countries It ing developed disbursing (25). less and economically aid receiving right; multilateral in the and intermediary when an eral particularly as capac- resources, acts financial Government state complement (24). may adaptation ity individual and collective nti nlss eadestelmttoso rvosefforts previous of limitations the address we analysis, this In reduction risk disaster from benefits who countries, Within PNAS | oebr1,2020 17, November | o.117 vol. | o 46 no. | 28693

SOCIAL SCIENCES Number of tropical cyclone exposures exceeding 63 km/hr

1 5 10 15 20 25 30 35 40 45 50 55 60

Deaths from tropical cyclone disasters

100 to 1000 1001 to 10,000 10,001 to 100,000 More than 100,000

Fig. 2. National tropical cyclone disaster deaths and subnational wind exposure (1979–2016). Total mortality is indicated by the shaded triangles for all countries with at least 100 total deaths from 1979–2016 (1). Areas shaded in gray indicate countries that have not experienced tropical cyclone deaths during this period. The frequency of exposure to sustained winds exceeding 63 km/h is mapped at 2.5-min (∼ 5 km) resolution (author calculations based on data and models by refs. 2, 3, 33, and 34). Exposure from tropical cyclones occurring in the Indian Ocean basin may be underestimated due to missing storm tracks in the underlying data. This region is therefore excluded from the main empirical analysis (see SI Appendix for details).

refs. 39–41). Finally, because we construct hazard and exposure tality in a country-level model that controls for hazard exposure. measures for all recorded tropical cyclones, we can examine the When we include only one of these four development indicators characteristics of storms that were not associated with a recorded at a time, each has a highly statistically significant association with disaster. This is a useful check on potential selection and mea- tropical cyclone deaths (SI Appendix, Table S5). This is consistent surement error issues in this literature and allows us to observe with existing evidence that GDP per capita is a useful proxy for the conditions under which tropical cyclone disaster is avoided. tropical cyclone vulnerability (16, 22); an increase of one log-unit of GDP per capita is predictive of a 66% decrease in deaths in a Results model with no other socioeconomic variables. However, because The effects of institutions, income, and human capital on tropical institutions, income, health, and education are highly correlated, cyclone mortality are estimated via two sets of multivariate neg- the independent effects of these variables cannot be identified by ative binomial regression models. The first set of models tests models with only a single socioeconomic variable. the importance of different national characteristics for cyclone To parse these relationships, we test multiple aspects of deaths, using data from over 900 events across 67 countries national development in combination (SI Appendix, Table S5). between 1996 and 2016. In addition to confirming the correla- This yields evidence of a large and statistically significant asso- tion between several facets of development and disaster deaths ciation between national government effectiveness and lower in the existing literature (9–11, 16, 22), our country-level mod- cyclone mortality. In a model with no other socioeconomic vari- els establish evidence of a robust association between national ables, a 1 SD increase in government effectiveness is associated government effectiveness and mortality from tropical cyclones. with a 71% decrease in deaths. As illustrated in Fig. 3, when Government effectiveness is represented in our models using we add GDP per capita and infant mortality to the model, gov- annual country-level scores, published by the World Governance ernment effectiveness accounts for a 49% decrease in mortality Indicators (WGI) and designed to capture the overall quality per SD, remaining practically and statistically significant. When and independence of public policy and service delivery (12). The we also include education, this reduces the number of observa- second set of models investigates the importance of subnational tions due to missing data, but the effect of governance remains development patterns for disaster mortality, using data from large and statistically significant. The association between gov- tropical cyclone disasters in 59 countries between 1979 and 2016. ernment effectiveness and lower tropical cyclone deaths is robust Socioeconomic conditions in the path of the storm are found to to a range of sensitivity analyses, including ordinary least squares have a large effect on expected mortality. Importantly, we con- (OLS) estimation, as described in SI Appendix, section 2 and trol for hazard exposure in both the national and subnational Tables S6–S14. specifications. In contrast, GDP per capita, health, and education are more sensitive to multivariate specifications. The decrease in mortal- National Government Effectiveness and Socioeconomic Conditions. ity associated with a one log-unit increase in GDP per capita Government effectiveness, real GDP per capita, IMRs, and pri- falls from 66 to 44% when we add government effectiveness mary school enrollment are all good predictors of cyclone mor- to the income-only model. The GDP per capita loses statistical

28694 | www.pnas.org/cgi/doi/10.1073/pnas.2006213117 Tennant and Gilmore Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 nlsv faeso h onr ihwae ntttosor institutions polit- weaker of settlements with and country IMRs select the We of groups. marginalized areas are effec- attributes of by country-level inclusive other afforded and protections governments the national tive whether Field. investigate Wind also Cyclone the We in Conditions Socioeconomic and Institutions hazard. our to to exposed when support disaster capacity further avert higher a to lends have and countries more-developed governance–mortality results that the hypothesis main of bias our direction selection in the that disentan- estimates for indicates account completely result not this cannot (SI does effects, EM-DAT we selection the the While in gle S15). included be hazard Table effective to for more Appendix, likely cyclone with controls less countries and are as wealthier storm governments in well tropical occur as that that indicate capita exposures results per Our effective- GDP exposure. government real includes and EM-DAT, that the regressors ness in of a included vector is of a estimate exposure instance given then cyclone an or that can probability storm We the tropical dataset. a of to model IBTrACS construct regression 1996 the logistical we from on exposures EM-DAT, based country-storm the 2016 all by includes that recorded dataset disasters in empirically. result assessed been previously not literature have the but in 22), discussed study- ref. been by (see have introduced hazards biases versus (e.g., potential disasters hazards The ing natural 22). and other 16, the and 9–11, of cyclones refs. studies tropical is global by in database used posed EM-DAT data risks mortality the of source as primary analysis, the this has This beyond deaths. institutional disaster implications underreport lower may with Second, corruption or mea- countries EM-DAT. capacity of example, possibility the For the error. creates in institutions, surement data well-endowed self-reported represented on and not reliance perhaps the effective are averted, of is events actions disaster hazard the when First, to hazards. due vulnerabil- of natural theories to validate to ity from utilized are reports data when nongovern- EM-DAT concerns on key the other two raises based and This (NGOs). disasters agencies, organizations mental (EM- Nations of Database United database Events governments, Emergency global the a is DAT), analysis per this GDP for with data Exposures. model Hazard versus joint Disasters the effectiveness. also government education in and and capita significance mortality infant statistical education of effects and lose The mortality model. infant the of to addition the with significance infant and capita prediction. per for GDP values mean real at including held shown, are not mortality, Variables CI. 95% the the on based deaths, 3. Fig. enn n Gilmore and Tennant Predicted deaths nodrt opr rpclccoe htd n onot do and do that cyclones tropical compare to order In 100 200 300 400 500 600 700 0 rdce fet fntoa oenetefcieessoeon score effectiveness government national of effects Predicted 2− 2 1 0 −1 −2 h hddae represents area shaded The S5. Table Appendix, SI National government effectiveness score h oreo h mortality the of source The osrcino h idfil aibe sdsrbdin described is variables Methods field and The wind conditions. the socioeconomic of under construction and but country institutional same the local the than in different out- occurred off compare that to worse events us or across allows comes This better metrics. relatively these popu- is by average the field national whether wind the capture in to over (7, lation occur (1979–2016) storms locations decades where vulnerable in four variability physically nearly spatial groups more the marginalized exploit in or We settle 44). them, to to access forced or are in mem- trust group areas less resources, have because fewer bers receive occur groups could excluded This by portions (43). settled to population accountability affected the lack tolls of governments death when that anticipate increase cap- we will specifically marginalization; more of political to effects The the selected them. ture was not groups from that ethnic else benefit of or population exclusion services, the such of of segments provision of all the lack in a capacity reflect insecurity or may food will mortality sani- against infant education, protect Elevated care, that malnutrition. health nets and as safety are provision social such the and and IMRs 42), in tation, (40, role quality level. their services field via the public institutions wind of of for quality the the proxies to at linked as institutions groups of inclusiveness ethnic excluded ically i.4adbsdo h eaiebnma ersinmdlesti- model regression binomial negative in the mated on based and 4 Fig. h hddaesrpeette9%CIs. 95% the represent areas shaded The . S18 Table Appendix , SI winds (sustained zone exposure cyclone tropical of intense more the in exposure ratio aver- storm IM tropical national of the the in winds to ratio (sustained compared IM the zone field of wind effect storm predicted The the (Top) in age. IMR the of ratio the 4. Fig. 1 mh.Peitosaebsdo h siae oespeetdin presented models estimated the on based are Predictions km/h). >119 Predicted deaths Predicted deaths in presented are analysis subnational the of results main The 100 200 300 400 500 100 200 300 400 500 0 0 rdce fet ftewn edI ai ndah.TeI ai is ratio IM The deaths. on ratio IM field wind the of effects Predicted sn oe htcontrols that model a Using S18. Table Appendix, SI n ute lbrtdin elaborated further and PNAS . . . 2.0 1.5 1.0 0.5 2.0 1.5 1.0 0.5 Infant Mortality Ratio(Windfield>119km/hr) Infant Mortality Infant Mortality Ratio(Windfield>63km/hr) Infant Mortality | h rdce feto the of effect predicted The (Bottom ) km/h). >63 oebr1,2020 17, November IAppendix. SI | o.117 vol. | o 46 no. Materials | 28695

SOCIAL SCIENCES for national socioeconomic conditions as well as hazard expo- effectiveness is not publicly available at present, we encourage sure, we find that death tolls are higher when IMRs are elevated future research to test the robustness of these findings using new within the cyclone wind field. For the tropical storm-strength or proprietary data sources. Future work could also investigate wind field (sustained winds of >63 km/h), the model predicts the importance of other facets of governance for disaster mortal- an 11% increase in storm deaths when local IMRs are elevated ity, such as polity and sociopolitical goals (54, 55). Third, our data by 10% above the national average. At higher wind speeds (sus- and research design are not suitable for demonstrating causality. tained winds of >119 km/h), the effect is more pronounced; a The challenges of overcoming multicollinearity in the analysis 10% increase in wind field infant mortality is associated with of observational data and, in particular, disentangling different a 14% increase in storm mortality. The results for these more aspects of governance and the complex processes that underlie intense tropical cyclone wind fields are robust to various permu- the correlation between income and institutions, are well doc- tations of the model and the dataset, while the results for the umented (e.g., refs. 56–59). Our results, however, go beyond weaker tropical storm wind fields lose statistical significance in previous efforts, by demonstrating that the association between some alternative specifications (SI Appendix, Tables S19–S29). national government effectiveness and tropical cyclone mortal- The statistical relationship between elevated IMRs and disas- ity cannot be fully explained by indicators of income, health, or ter deaths may be interpreted in several ways. Infant mortality education. Finally, the trade-off of focusing on a single class of is a measure of public health and has also been employed as hazard is that it limits our ability to generalize these results to a proxy for overall well-being, poverty, or inequality (e.g., refs. other types of natural disaster. However, our approach can be 31, 39, 45, and 46), each of which is plausibly related to disas- adapted to the study of additional hazards, scales, and outcomes ter deaths. However, the importance of within-country variation to gain further insight into the role of institutions and economic in IMRs clearly demonstrates that disaster deaths are not only a development in risk reduction. function of the national context and hazard exposure. Local vul- Our findings are salient to current questions about the inter- nerabilities are important, and particularly so in areas that are section of institutions, sustainable development, and disaster exposed to sustained wind speeds in excess of 119 km/h, the “very risk, questions made more urgent under climate change. The dangerous” threshold for tropical cyclone winds (47). intensity and rainfall of the strongest tropical cyclones are Our analysis of the effects of politically exclusive institutions expected to increase under climate change (37, 60–62), and on disaster mortality is not conclusive and highlights the need for trends in population growth and sea level rise will further further research on this topic. Following the Ethnic Power Rela- contribute to risk in the absence of effective adaptation (22, tions (EPR) classifications, we consider groups to be excluded 37, 63). Many tropical cyclone-affected countries will also face from executive political power if they are powerless, discrimi- increased risk from other climate change impacts, including nated against, or self-excluded (48, 49). By this measure, the extreme weather events such as droughts, floods, and heat effects of exclusion are not precisely estimated (SI Appendix, waves (5). These challenges are amplified by uneven progress Table S18). However, we find that very few tropical cyclones in on eliminating poverty, hunger, disease, illiteracy, environmental our dataset actually impact areas settled by groups that are dis- degradation, and discrimination against women (5, 64). Enhanc- criminated against or self-excluded. Our measure of ethnic group ing institutions may have wide-ranging benefits for disaster risk exclusion therefore primarily captures the effects of powerless reduction as well as climate adaptation and sustainable develop- groups settled in the impact area (SI Appendix, Table S30). Pow- ment. This underscores the value of understanding relationships erless groups, which lack representation, may be less likely to be between institutions and disasters. excluded from national protections compared to groups that are actively discriminated against. Our indicators of exclusion also do Materials and Methods not unpack potentially important heterogeneities in the density Disasters occur when a population is exposed to hazardous conditions and of ethnic group settlements and the de facto and de jure forms of is unable to adapt or cope. Understanding mortality from tropical cyclones political power sharing (50–52). therefore requires information about the spatial intersection of physical hazard and socioeconomic systems. Here, we describe the methods and Discussion data sources used to build our event-based dataset of tropical cyclone dis- Our analysis generates empirical support for the role of govern- asters that extends from 1979 to 2016. This is followed by a description of the econometric methods that underlie our results. The hazard exposure ments and institutions in reducing tropical cyclone risk. First, we variables and the socioeconomic variables are summarized in SI Appendix, show that national government effectiveness is associated with Tables S1 and S2; the source data and methods are also described in further lower mortality from tropical cyclones, independent of GDP per detail in SI Appendix. capita, health, and education. We then demonstrate the importance of within-country heterogeneities in vulnerability through Dataset. Our approach recognizes the importance of accurately accounting global analysis of subnational institutional quality and tropical not only for the intensity of the hazard but also for the number of people cyclone risk. Specifically, we find that death tolls are higher when exposed to hazardous conditions and the local socioeconomic conditions of IMRs, a proxy for the quality and inclusiveness of local insti- the affected population. Basic statistics such as a storm’s maximum wind tutions, are elevated compared to the national average within speed or minimum central pressure are indicators of hazard intensity rather than exposure, and therefore incomplete measures of the severity of the the cyclone wind field. These results lend support for general shock. Many intense storms never pass within striking distance of populated theories of how effective and inclusive institutions can moder- land or weaken sufficiently to pose little threat upon landfall. When intense ate vulnerability and foster resilience to a range of shocks and storms do strike land, minor differences in storm trajectory can have large stressors. implications for the number of people exposed to hazardous conditions. We acknowledge several limitations of this work. First, we rely The speed and longevity of a storm impacts the duration of wind exposure on data that include only the direct, short-term disaster deaths. as well as the cumulative rainfall. Our analysis does not capture how institutions may mediate To translate from hazard to exposure, we develop a method to match longer-term mortality, for example, through their role in mitigat- storm tracks and rainfall to disaster data and then parametrically model ing economic hardship or reestablishing health care and other the intensity and spatial extent of each storm. With the area of exposure spatially delineated, we can then determine the size and socioeconomic con- services in the aftermath of the storm (14, 53). Second, these ditions of the population living there. In brief, this is done by first identifying results may be sensitive to the data sources used to operational- the grid cells that fall within the storm’s wind field, extracting the measures ize the latent concept of institutional capacity. Our analysis relies of interest for each of those grid cells (e.g., population, infant mortality), on the subjective WGI government effectiveness scores. While, and then computing the average conditions in the wind field. Thus, while to our knowledge, a suitable alternative measure of government several variables in this analysis draw on subnational data specific to the

28696 | www.pnas.org/cgi/doi/10.1073/pnas.2006213117 Tennant and Gilmore Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 enn n Gilmore and Tennant Toia yln:1913k/) n C+(rpclCyclone: (Tropical TC2+ fields. wind and km/h), 119–153 Cyclone: land (Tropical over speed Depression: wind (Tropical (C sustained define TD we maximum Finally, resolution. the software 2.5-arc-minute (B) the a in estimate at implemented then (2, (34) we locations model (33), and speed R speeds wind wind parametric 6-hourly a Using the 3). (A) with begin We Madagascar. 5. Fig. See basin. Ocean Indian the or including inclusion the region, to particular robust How- any are period. paper of study this quality exclusion the in during the presented region findings about main this the concerns including for ever, data to specifications the due of for completeness specifications, and main (1996+) our years are from cyclones recent excluded tropical Ocean Indian more scores. effectiveness to government national and 500-km data, a within fall and country the in track. cell storm the grid the over any of rainfall for buffer total storm maximum the 0.5 the of by a duration represented of is at Analysis Rainfall (38). available Gauge-Based present Unified dataset, Global Precipitation based Center is Daily exposure Prediction populations Rainfall Climate the 65). inter- of the (36, are size intensities on different the 4.10) of estimate (Version winds to to fields World exposed wind the Series modeled the of Time with Population Grid acted Gridded Count Earth and Population International Estimates Global for Network’s Center Information the Science from Time-variant, estimates exposure. the population estimate to subnational data population with event, fields wind country-storm the single a for Madagascar. process in this Gafilo thresholds. Cyclone of wind 2004 steps the multiple the for for illustrates performed mapped 5 is is land Fig. This over fields event. 2.5-arc-minute wind country-storm a the at of each rasterized extent spatial then the are and winds resolution, modeled The (33). of details) for version are adapted implemented winds is globally This and a (34). model interpolated using cyclone popu- are tropical parametric gridded data a using with track spatial modeled a data, matching produce socioeconomic for to and suitable order lation winds, In storm. storm the of of representation track central the along intervals storm agencies. as meteorological such and sources, governments additional by consulting published by reports accuracy man- space for were IBTrACS in reviewed and storm EM-DAT a ually closest the between the using matches for the Automated matched looks time. with disaster, and therefore identifier each for were common that, a observations algorithm share spatial The not data. do 3) disaster (2, EM-DAT Project IBTrACS Exposure. the and from Intensity Hazard of Measures (1). in analysis this detailed for each are with disaster associated 2 aggregated for princi- section deaths criteria are disaster our Our data of with event. number these comparison country-storm the storm, for variable: allows the outcome This pal by measures. impacted country-storm country into the of area ABC hsaayi slmtdt h aelt r 17+ fwn n rain- and wind of (1979+) era satellite the to limited is analysis This overlay then can we delineated, spatially been has hazard wind the Once 6-h at georeferenced data pressure and wind of consist data track Best oeigtoia yln idfilsfrCcoeGfio(04 in (2004) Gafilo Cyclone for fields wind cyclone tropical Modeling n olwteE-A,tesuc ftedsse otlt data mortality disaster the of source the EM-DAT, the follow and 3k/) S(rpclSom 318k/) TC1 km/h), 63–118 Storm: (Tropical TS km/h), <63 stormwindmodel rpclccoedt obtained data cyclone Tropical ◦ eouinfo 99to 1979 from resolution h pta xeto the of extent spatial the ) nR(see R in IAppendix, SI > IAppendix SI 5 km/h) 153 SI ,0 etsfo h eaiebnma pcfiain.Teeotirevents outlier These than specifications. more binomial with negative events the exclude from therefore deaths We 5,000 events. outlier large handle to to parameters The variables. control other are estimate and geographic and exposure, in presented regressions is binomial model negative NB2 estimate inconsis- The all therefore analysis. for this be We (74). errors may misspecification standard distributional errors robust of standard cases in distributional model to to tent However, compared robust properties is (73). it useful that several misspecification including has models, model binomial NB2 negative The other 808). a p. Eqs. and 72, in mean represented (ref. the model regression on binomial depends ative variance the that parameter dispersion such assumption this relax significant statistically S20 highly and, and fields, S19 positive Tables wind consistently Appendix, strength (SI remains cyclone of deaths effect tropical estimated the disaster the for on check, only IMR robustness of or elevated a impacts 1999–2000 as indirect locally years 2001–2016 or the years direct exclude the the we to include when due However, country cyclones. the might of tropical mortality infant parts that are in is data concern elevated one IMR be 2000), subnational year underlying the mortality (for the infant invariant that time the Given of country. resolution by the are the varies as dummies data to models, Country field subnational (35). wind all 2000 year in storm the included Sub- the for Global in Project’s Rates Mapping Mortality IMR Poverty Infant the the national from of data ratio on based the mor- IMR, (36, is infant national population The ratio) field. cell (IM wind over-land ratio grid the tality to the restricted by are and therefore weighted mortality and fields infant 65), are the wind variables Both for 5). eth- exclusion Fig. constructed of in political are illustrated exclusion variables (as political intensities these multiple the storm, of each on For data groups. spatial nic and our IMRs subnational in using countries the across there scores that conclude governance to in dataset. us variation allows meaningful This is uncertainty. explicit the the is of are using methodology WGI characterization combined measures the of scores experts perception-based advantage NGO The an While imprecise, and 70). unavoidably model. private, (12, components public, institutions unobserved of formal an surveys by on pub- delivery based of service are quality and the as policies defined lic effectiveness, government national capture y. per 1 country. the by GDP from the lagged The are (67–69). taken and IMRs sources are and other year capita education and the (66) and Indicators on health, Development World income, based of events indicators cyclone National tropical to matched Variables. Socioeconomic storms. Ocean Indian Appendix x principle equidispersion Poisson the violate given simpler data suitable The the is values. because model integer used, E data not nonnegative count is are a model deaths of use storm The that model. regression mial deaths cyclone Tropical or Methods against discriminated than rather cyclones powerless tropical See as the self-excluded. with classified overlap primarily that the are However, settlements self-excluded. group excluded or as ethnic against, groups excluded discriminated classifies powerless, and are rele- power, politically they state on if to data access annual groups’ provides EPR ethnic EPR the The vant from 71). data 49, on (48, based is Family This Dataset constructed. also is group ethnic excluded h hrceitc fec onr-tr event country-storm each of characteristics The and where i [y nld oieooi hrceitc,maue fsomitniyand intensity storm of measures characteristics, socioeconomic include , n rwako h eaiebnma oe sta ti o elsuited well not is it that is model binomial negative the of drawback One proxied are inclusion and quality institutional local countries, Within to WGI the use we 22), and 10 refs. (e.g., work related previous Following h ouainwihe ecnaeo h idfil hti ete yan by settled is that field wind the of percentage population-weighted The i | x i ]= Var o h estvt nlssadadtoa ouetto fthe of documentation additional and analyses sensitivity the for β [y , θ i Prob(Y . PNAS | IAppendix SI x i ] h eaiebnma ersinmdlalw sto us allows model regression binomial negative The . α = | = y y 1/θ oebr1,2020 17, November o event for i onr-ee oieooi aibe are variables socioeconomic Country-level | x eueteNgi N2 omo h neg- the of form (NB2) 2 Negbin the use We . r i λ o icsino h implications. the of discussion a for i ) = i = ). = λ Γ(y exp(x i /(θ i Γ(θ i r oee sn eaiebino- negative a using modeled are + + + i 0 1)Γ(θ β λ y ), i ). i ) | ) i r o.117 vol. ersne ytevector the by represented , i y i (1 − olwn Greene following 2–4, r i ) θ | , o 46 no. | 28697 [2] [4] [3]

SOCIAL SCIENCES are few in number, but catastrophic in their humanitarian impacts.† We Due to copyright, the original EM-DAT data (1) is not included in the repli- therefore estimate comparable OLS models with a transformed dependent cation files; details on how the EM-DAT data can be accessed directly (for variable to accommodate these high-mortality events as a robustness check. noncommercial use) are included. As described in SI Appendix, the main results are robust to OLS estimation with and without the outlier events. ACKNOWLEDGMENTS. We are grateful for the advice and guidance of Anand Patwardhan on this project. We also thank Anna Alberini, Brooke Data Availability. A replication package including the R code and data files Anderson, Florio Arguillas, Christopher Barrett, Molly Elizabeth Brown, generated for and analyzed during the current study has been deposited Christopher Foreman, Matthew Kahn, Richard Moss, Robert Sprinkle, Larry Swatuk, Mathieu Taschereau-Dumouchel, Brian Thiede, Catherine Warsnop, in the Cornell Institute for Social and Economic Research (CISER) Data & and seminar and conference participants at the American Geophysical Reproduction Archive, https://doi.org/10.6077/89ba-bj79 (4). The replication Union, the International Studies Association, Cornell University, New York package includes all publicly available and author generated source data. University, and the University of Maryland School of Public Policy. We acknowledge funding support from the University of Maryland Council of the Environment, the Anne G. Wylie Dissertation Fellowship, the Univer- †Our criteria exclude Thelma (1991) and Haiyan (2013) in the Philippines and Mitch sity of Maryland, and Clark University. This material is based upon work (1998) in Honduras. See SI Appendix for additional Indian Ocean storms that exceed supported, in part, by the US Army Research Laboratory and the US Army 5,000 deaths. Research Office via the Minerva Initiative under Grant W911NF-13-1-0307.

1. D. Guha-Sapir, EM-DAT: The Emergency Events Database (Universite´ Catholique de 25. H. Eakin, M. C. Lemos, Adaptation and the state: Latin America and the challenge of Louvain, Brussels, Belgium, 2018). www.emdat.be. Accessed 16 October 2017. capacity-building under globalization. Glob. Environ. Change 16, 7–18 (2006). 2. K. R. Knapp, M. C. Kruk, D. H. Levinson, H. J. Diamond, C. J. Neumann, The Interna- 26. N. Hossain, The 1970 Bhola cyclone, nationalist politics, and the subsistence crisis tional Best Track Archive for Climate Stewardship (IBTrACS). Bull. Am. Meteorol. Soc. contract in Bangladesh. Disasters 42, 187–203 (2018). 91, 363–376 (2010). 27. B. Howe, G. Bang, Nargis and Haiyan: The politics of natural disaster management in 3. K. R. Knapp, H. J. Diamond, J. P. Kossin, M. C. Kruk, C. J. Schreck, International Best Myanmar and the Philippines. Asian Stud. Rev. 41, 58–78 (2017). Track Archive for Climate Stewardship (IBTrACS) Project, Version 3.10 (NOAA National 28. B. K. Paul, Why relatively fewer people died? The case of Bangladesh’s cyclone sidr. Centers for Environmental Information, 2017). https://doi.org/10.7289/V5NK3BZP. Nat. Hazards 50, 289–304 (2009). Accessed 8 November 2017. 29. S. Veland, R. Howitt, D. Dominey-Howes, Invisible institutions in emergencies: Evac- 4. E. Tennant, E. Gilmore, Replication package for government effectiveness and insti- uating the remote indigenous community of Warruwi, Northern Territory Australia, tutions as determinants of tropical cyclone mortality in Catalog No. 2845, Version 1 from Cyclone Monica. Environ. Hazards 9, 197–214 (2010). (2020). https://doi.org/10.6077/89ba-bj79. Deposited 22 October 2020. 30. W. N. Adger, Vulnerability. Global Environ. Chang. 16, 268–281 (2006). 5. Intergovernmental Panel on Climate Change, Climate Change 2014: Impacts, Adap- 31. E. B. Barbier, J. P. Hochard, The impacts of climate change on the poor in tation, and Vulnerability: Global and Sectoral Aspects, V. R. Barros et al., Eds. disadvantaged regions. Rev. Environ. Econ. Pol. 12, 26–47 (2018). (Cambridge University Press, Cambridge, United Kingdom, 2014). 32. M. Pelling, The political ecology of flood hazard in urban Guyana. Geoforum 30, 6. D. Alexander, Natural disasters: A framework for research and teaching. Disasters 15, 249–261 (1999). 209–226 (1991). 33. B. Anderson et al., Stormwindmodel: Model tropical cyclone wind speeds (2017). 7. P. Blaikie, T. Cannon, I. Davis, B. Wisner, At Risk: Natural Hazards, People’s https://cran.r-project.org/web/packages/stormwindmodel/index.html. Accessed 17 Vulnerability and Disasters (Routledge, 2004). August 2018. 8. United Nations International Strategy for Disaster Reduction, Making Development 34. H. E. Willoughby, R. W. R. Darling, M. E. Rahn, Parametric representation of the pri- Sustainable: The Future of Disaster Risk Management, Global Assessment Report on mary hurricane vortex. Part II: A new family of sectionally continuous profiles. Mon. Disaster Risk Reduction (United Nations, 2015). Weather Rev. 134, 1102–1120 (2006). 9. A. Alberini, A. Chiabai, L. Muehlenbachs, Using expert judgment to assess adaptive 35. Center for International Earth Science Information Network, Poverty Map- capacity to climate change: Evidence from a conjoint choice survey. Glob. Environ. ping Project: Global Subnational Infant Mortality Rates (2005). http://dx.doi.org/ Change 16, 123–144 (2006). 10.7927/H4PZ56R2. Accessed 3 January 2018. 10. N. Brooks, W. N. Adger, P. M. Kelly, The determinants of vulnerability and adaptive 36. Center for International Earth Science Information Network, Gridded Population capacity at the national level and the implications for adaptation. Glob Environ. of the World, Version 4 (GPWv4): Population count adjusted to match 2015 Change 15, 151–163 (2005). revision of UN WPP country totals, revision 10, version 4 (2017). https://doi.org/ 11. M. E. Kahn, The death toll from natural disasters: The role of income, geography, and 10.7927/H4JQ0XZW. Accessed 3 January 2018. institutions. Rev. Econ. Stat. 87, 271–284 (2005). 37. K. J. E. Walsh et al., Tropical cyclones and climate change. Wiley Interdiscip. Rev. Clim. 12. D. Kaufmann, A. Kraay, M. Mastruzzi, Worldwide governance indicators: Methodol- Change 7, 65–89 (2016). ogy and analytical issues (World Bank Policy Research Working Paper No. 5430, World 38. National Oceanic and Atmospheric Administration, CPC Global Unified Precipitation Bank, 2010). https://ssrn.com/abstract=1682130. Accessed 22 October 2018. Data (2018). https://psl.noaa.gov/data/gridded/data.cpc.globalprecip.html. Accessed 13. ISO 3166 Country Codes. https://www.iso.org/iso-3166-country-codes.html. Accessed 1 25 January 2018. April 2016. 39. R. C. Briggs, Poor targeting: A gridded spatial analysis of the degree to which aid 14. J. Anttila-Hughes, S. Hsiang, Destruction, disinvestment, and death: Economic reaches the poor in Africa. World Dev. 103, 133–148 (2018). and human losses following environmental disaster. SSRN:10.2139/ssrn.2220501 (18 40. R. Pongou, B. K. Defo, Z. T. Dimbuene, Excess male infant mortality: The gene- February 2013). institution interactions. Am. Econ. Rev. 107, 541–545 (2017). 15. S. M. Hsiang, A. S. Jina, The causal effect of environmental catastrophe on long- 41. N. von Uexkull, M. Croicu, H. Fjelde, H. Buhaug, Civil conflict sensitivity run economic growth: Evidence from 6,700 cyclones, National Bureau of Economic to growing-season drought. Proc. Natl. Acad. Sci. U.S.A. 113, 12391–12396 Research (2014) Working Paper Series. http://www.nber.org/papers/w20352. Accessed (2016). 18 November 2018. 42. A. Sirag, N. M. Nor, N. M. Raja Abdullah, J. A. Ghani, Public health financing and 16. S. M. Hsiang, D. Narita, Adaptation to cyclone risk: Evidence from the global cross- infant mortality: Does governance quality matter?Publ. Finance Manag. 17, 341–368 section. Clim. Chang. Econ. 3, 1250011 (2012). (2017). 17. P. Peduzzi, H. Dao, C. Herold, Mapping disastrous natural hazards using global 43. J. Busby, N. von Uexkull, Climate Shocks and Humanitarian Crises (Foreign Affairs, datasets. Nat. Hazards 35, 265–289 (2005). 2018). 18. L. A. Bakkensen, R. O. Mendelsohn, Risk and adaptation: Evidence from global 44. J. M. Bretthauer, Conditions for peace and conflict: Applying a fuzzy-set qualitative hurricane damages and fatalities. J. Assoc. Environ. Resour. Econ. 3, 555–587 (2016). comparative analysis to cases of resource scarcity. J. Conflict Resolut. 59, 593–616 19. W. N. Adger, S. Huq, K. Brown, D. Conway, M. Hulme, Adaptation to climate change (2015). in the developing world. Prog. Dev. Stud. 3, 179–195 (2003). 45. P. D. C. Esty et al., State Failure Task Force report: Phase II findings. Environ. Change 20. F. Denton et al., “Climate-resilient pathways: Adaptation, mitigation, and sustain- Secur. Proj. Rep. 5, 49–72 (1999). able development” in Climate Change 2014: Impacts, Adaptation, and Vulnerability: 46. S. Hillesund et al., Horizontal inequality and armed conflict: A comprehensive Global and Sectoral Aspects, C. B. Field et al., Eds. (Cambridge University Press, literature review. Can. J. Dev. Stud. 39, 463–480 (2018). Cambridge, United Kingdom, 2014), pp. 1101–1131. 47. National Oceanic and Atmospheric Administration, Saffir-Simpson hurricane wind 21. S. J. Camargo, S. M. Hsiang, “Tropical cyclones: From the influence of climate to their scale. https://www.nhc.noaa.gov/aboutsshws.php (2020). Accessed 1 January 2020. socioeconomic impacts” in Extreme Events: Observations, Modeling, and Economics, 48. L. E. Cederman, A. Wimmer, B. Min, Why do ethnic groups rebel?: New data and M. Chavez, M. Ghil, J. Urrutia-Fucugauchi, Eds. (American Geophysical Union, 2015), analysis. World Polit. 62, 87–119 (2010). pp. 303–342. 49. M. Vogt et al., Integrating data on ethnicity, geography, and conflict: The 22. P. Peduzzi et al., Global trends in tropical cyclone risk. Nat. Clim. Chang. 2, 289–294 ethnic power relations data set family. J. Conflict Resolut. 59, 1327–1342 (2012). (2015). 23. R. K. Pachauri, L. Mayer.., Eds. Climate Change 2014: Synthesis Report (Intergovern- 50. N. C. Bormann, Power sharing: Institutions, behavior, and peace. Am. J. Polit. Sci. 63, mental Panel on Climate Change, Geneva, Switzerland, 2015). 84–100 (2019). 24. W. N. Adger, Social capital, collective action, and adaptation to climate change. Econ. 51. J. Busby, Taking stock: The field of climate and security. Curr. Clim. Change Rep. 4, Geogr. 79, 387–404 (2003). 338–346 (2018).

28698 | www.pnas.org/cgi/doi/10.1073/pnas.2006213117 Tennant and Gilmore Downloaded by guest on September 30, 2021 Downloaded by guest on September 30, 2021 1 negvrmna ae nCiaeChange, Climate on Panel Intergovernmental 61. government. of quality The Vishny, R. Shleifer, A. Lopez-de-Silanes, F. Porta, L. R. 58. 0 .H Christensen H. J. 60. Putnam, D. R. 59. system. international the and development, Democracy, Boix, democracy. C. and Income Yared, 57. P. Robinson, A. J. Johnson, S. Acemoglu, D. change 56. climate to nations of Lesnikowski capacity A. 55. adaptive the Ranking Haddad, M. B. 54. 2 .W tø,S ae,B .Gaa,H tad nlso,dseso,adcon- and dispersion, Inclusion, Strand, Kishore H. N. Graham, 53. A. B. Gates, S. Strøm, W. K. 52. enn n Gilmore and Tennant neBss otiuino okn ru oteFfhAssmn eoto the of 2013). Report Stocker Kingdom, F. Assessment United T. Fifth Cambridge, Change, Press, the University Climate to on I Group Panel Working Intergovernmental of Contribution Basis: ence Panel 1217–1308. Intergovernmental pp. the 2013), of Kingdom, Report United Assessment Fifth Change the Climate to on I Group Working of in change” climate Organ. Econ. Law (2011). 809–828 105, Rev. nvriyPes rneo,N,1994). NJ, Princeton, Press, University change. climate of impacts (2013). health to tion explicit. are goals (2005). socio-political when (2018). 162–170 tan:Pwrhrn ntewrdssae,1975–2010. states, world’s (2017). the in Powersharing straint: 0–4 (2008). 808–842 98, otlt nPet ioatrHriaeMaria. Hurricane after Rico Puerto in Mortality al., et aigDmcayWr,CvcTaiin nMdr Italy Modern in Traditions Civic Work, Democracy Making ainllvlfcosafcigpand ulcadapta- public planned, affecting factors National-level al., et Ciaepeoeaadterrlvnefrftr regional future for relevance their and phenomena “Climate al., et 2–7 (1999). 222–279 15, .F Stocker F. T. , lmt hne21:TePyia cec ai.Contribution Basis. Science Physical The 2013: Change Climate d.(abig nvriyPes Cambridge, Press, University (Cambridge Eds. al., et lb nio.Change Environ. Glob. lb nio.Change Environ. Glob. lmt hne21:TePyia Sci- Physical The 2013: Change Climate r .Plt Sci. Polit. J. Br. d.(Cambridge Eds. al., et .Eg.J Med. J. Engl. N. m oi.Si Rev. Sci. Polit. Am. 1153–1163 23, 165–176 15, 165–185 47, (Princeton m Econ. Am. 379, J. 4 .M Hilbe, M. J. 74. Trivedi, K. P. Cameron, C. A. 73. Greene, Politi- H. W. Wimmer, A. 72. Cederman, E. L. Girardin, L. Weidmann, B. N. https://datacatalog.worldbank.org/dataset/ Wucherpfennig, J. (2020). 71. Indicators of (Ministry Governance indicators World welfare and 70. health 2016 2014 Statistics: Welfare, of and Digest 2020). Health Annual 9.1, of Kong (Version Ministry Table Hong 69. World Department, Penn Statistics Timmer, and P. M. Census Inklaar, 68. I. R. Feenstra, C. R. https://datacatalog.worldbank.org/dataset/ 67. (2020). Indicators Development World 66. Count Population Global Network, Information Science Earth International for Center 65. climate of impact The Nations Bakkensen, United L. 64. Chonabayashi, tropi- S. major Emanuel, K. in Mendelsohn, increase R. Global Velden, 63. S. C. Olander, L. T. Knapp, R. K. Kossin, P. J. 62. rs,e.2 e ok Y 2013). NY, York, New 2, ed. Press, Dataset1. GeoEPR the Sci. Introducing Peace time: Manag. and Conﬂ. space across groups ethnic relevant cally 2020. May 13 Accessed worldwide-governance-indicators. https://www.mohw.gov.tw/lp-134-2.html. 2017. 2017). August (Taiwan), 18 Accessed ROC Welfare and Health 2016). Region, Administrative Special Kong (Hong Edition 2020. May 12 Accessed https://doi.org/10.15141/S50T0R. 2020. May 13 Accessed 3 world-development-indicators. Accessed https://doi.org/10.7927/H4CC0XNV. (2017). Estimates 2018. January Series Time Grid 2015). damage. cyclone tropical global on change decades. four past the over U.S.A. probability exceedance cyclone cal 17–18 (2020). 11975–11980 117, oeigCutData Count Modeling cnmti Analysis Econometric h ilnimDvlpetGasRpr 2015 Report Goals Development Millennium The PNAS | 2–3 (2011). 423–437 28, oebr1,2020 17, November ersinAayi fCutData Count of Analysis Regression CmrdeUiest rs,NwYr,N,2014). NY, York, New Press, University (Cambridge Petc al d ,Bso,M,2012). MA, Boston, 7, ed. Hall, (Prentice a.Ci.Change Clim. Nat. | o.117 vol. 0–0 (2012). 205–209 2, CmrdeUniversity (Cambridge | rc al cd Sci. Acad. Natl. Proc. o 46 no. Uie Nations, (United | 28699

SOCIAL SCIENCES