American Economic Review: Papers & Proceedings 2017, 107(5): 446–450 https://doi.org/10.1257/aer.p20171053 Heat Exposure and Youth Migration in Central America and the Caribbean† By Javier Baez, German Caruso, Valerie Mueller, and Chiyu Niu* Emerging evidence demonstrates migration may render greater consequences on the mobil- may be used by individuals in the southern ity of individuals than natural disasters. Social hemisphere to adapt to environmental changes. protection programs often target populations Few studies measure the environmental drivers affected by natural disasters reducing their vul- of migration in the Latin America and Caribbean nerability, but individuals rarely receive social region as a whole. Furthermore, while it is assistance due to being exposed to a heat wave. well-documented that the liquidity-constrained In addition, widespread disasters can compro- often move temporarily or over short distances mise employment opportunities and inflate in response to a localized, negative shock transportation costs; both reducing net migration (Fussell, Hunter, and Gray 2014), the permanent benefits. migration impact of prolonged or repeated expo- The growth literature has pointed to numer- sure to changes in the environment is poorly ous examples in which environmental migra- understood. First, ambiguities about risk caused tion contributes to urbanization (Henderson, by gradual environmental degradation may Storeygard, and Deichmann 2017) without result in minimal behavioral change (Lee et al. growth (Poelhekke 2011). We therefore addition- 2015). Second, low-income households may not ally predict how gradual changes in temperature be able to afford transportation between loca- influence the skill set of who migrates and their tions or have sufficient social capital to secure location choice. We further confirm using two employment in urban areas (Bryan, Chowdhury, separate datasets whether the observed mobility and Mobarak 2014). These features challenge patterns correspond to changes in income levels, the forced migration rhetoric, instead offering by examining the GDP losses at the district and a scenario in which populations are trapped in national levels. The former may be indicative of place (Black et al. 2011). local vulnerability, while national vulnerability We build upon recent work in the region from prolonged heat exposure can narrow the using a similar triple difference-in-difference set of employment opportunities at destina- quasi-experimental design and dataset to test tions reducing the desirability of environmental whether repeated and/or prolonged heat expo- migration. sure affect inter-province migration (Baez et al. 2017). We previously illustrated that youth are I. Data more likely to move in response to hurricanes and droughts. Exposure to temperature extremes Migration.—Migration data are taken from censuses conducted in the following countries * Baez: World Bank, 1818 H Street, NW, Washington, (Minnesota Population Center 2015): Costa DC 20433 (e-mail: jbaez@worldbank.org); Caruso: World Rica (2000, 2011), Dominican Republic (2002, Bank, 1818 H Street, NW, Washington, DC 20433 (e-mail: 2010), El Salvador (1992, 2007), Haiti (1982, gcaruso@worldbank.org); Mueller: International Food 2003), Mexico (2000, 2010), Nicaragua (1995, 2005), and Panama (2000, 2010). The infor- Policy Research Institute, 2033 K Street, NW, Washington, DC 20006 (e-mail: v.mueller@cgiar.org); Niu: Department of Agricultural and Consumer Economics, University mation allows for measurement of individual of Illinois, 326 Mumford Hall, 1301 W. Gregory Drive, inter-province migration over the five years Urbana, IL 61801 (e-mail: cniu3@illinois.edu). We thank prior to the census. We create three migration Solomon Hsiang for his constructive comments. †  Go to https://doi.org/10.1257/aer.p20171053 to visit the variables based on the province destination of article page for additional materials and author disclosure each individual, whether the person migrated statement(s). across provinces, whether the person migrated 446 VOL. 107 NO. 5 Heat Exposure and Youth Migration 447 across provinces to the national capital, and ues less than zero are replaced by the value zero. whether the person migrated across provinces “affected” areas are defined by provinces having to a provincial capital. Individual age, gender, positive z-score values, allowing heat exposure education, and province origin are used to for- to vary in intensity by maintaining the continu- mulate the explanatory variables included in the ous z-score values.5 regressions.1 Precipitation data are taken from the Climate Research Unit’s Time Series of the University Climate.—Daily temperature are extracted of East Anglia (1983–2010). Five-year average from the Global Land Data Assimilation System precipitation and precipitation squared variables Version 2 (1983–2010) (Rodell and Beaudoing are created to reduce bias from unobserved, cor- NASA/GSFC/HSL 2015).2 We focus on heat related weather characteristics in regressions exposure over the migration period.3 To formu- (Auffhammer et al. 2013). All temperature and late a standardized measure of heat exposure, we precipitation variables are then merged by the perform the following steps. First, we create a origin province and survey year of each individ- binary indicator for whether the daily minimum ual in the censuses. temperature is in the ninetieth percentile of the province distribution.4 Second, we create vari- Gross Domestic Product.—To examine ables reflecting the number of days that the daily whether heat-induced income fluctuations drive minimum temperature exceeds the ninetieth per- migration patterns, we use two sources of GDP centile over moving five-year averages, starting data adjusted using the purchasing power par- with 1983 and ending with 2010. Third, we use ity rates in 2011 international dollars. The first the distribution (24 points) of the number of source comes from statistical models produced “extremely hot” days over periods of five years by the World Bank to predict province-level to construct the means and standard deviations GDP.6 One shortcoming of the detailed data is of each province distribution of heat. its limited spatial (e.g., Costa Rica, El Salvador, Next, we use the above data to create a heat Nicaragua, and Panama) and temporal (2000, exposure z-score to facilitate the interpretation 2005, 2010) availability. We therefore also of results: the number of days exceeding the exploit the World Development Indicators ninetieth percentile (over the five-year migra- (WDI) database which concentrates on the tion period in the census follow-up round), and national GDP for all seven countries over the the means and standard deviations of each ori- time frame of our censuses (1987–2010). gin province distribution of heat. Areas with heat exposure values of zero or below in the II. Identification follow-up round are considered “unaffected” for the purpose of the experiment. The z-score val- We evaluate the effect of repeated and prolonged heat exposure on inter-prov- 1  ince migration by building on the triple ­difference-in-difference (DID) design in Baez et Summary statistics of the explanatory variables included in the empirical model can be found in the online Appendix. 2  We use this dataset over others that measure surface al. (2017). Double DID models are commonly temperature as it offers daily (rather than monthly) values at used to identify the impact of shocks on out- a fine scale. It should be noted the data come from a surface comes in the absence of panel data, comparing hydrology model that takes temperature as an input from the Princeton Global Meteorological Forcing Dataset. outcomes before and after exposure in affected 3  Typically, the literature focuses on contemporaneous and unaffected provinces. One drawback of the (or lagged) exposure during the year of move. We create a double DID approach is it assumes there are no five-year measure of exposure, since censuses ask whether other micro-level shocks in affected areas at the the individual is located in a different province than he/she time of exposure. Using the same data, Baez et al. (2017) exploit an additional dimension, resided in five years prior to the interview. 4  We use minimum (rather than average or maximum) temperature values as a conservative measure of heat. ­ ecomposing timing of birth, for identification d Transforming temperature into percentile indicators stan- dardizes the definition of extreme heat by accounting for local frames of reference. Migration responses to extreme 5  heat have been shown to be robust to the use of percentile A figure illustrating the distribution of the heat exposure and average temperature measures in other settings (Mueller, variable is provided in the online Appendix. Gray, and Kosec 2014). 6  For further details, refer to the online Appendix. 448 AEA PAPERS AND PROCEEDINGS MAY 2017 the sample by high-mobility (ages 15–25, more likely affected in the follow-up censuses 26–35) and low-mobility (ages 36–45, 46–55, (online Appendix). Second, we assume differ- and 56–65) groups. This additional dimension ences across birth cohorts in migration rates circumvents the bias caused by time variant would be similar across “affected” and “non-af- shocks by further drawing comparisons of the fected” districts in the absence of the shock. We changes in outcomes by heat exposure across validate that the changes in the baseline charac- age groups. teristic averages across age groups and exposure A linear probability model is used to quantify categories are statistically insignificant (online the effects of heat exposure on the migration of Appendix). men and women: III. Results ​ ​​  = ​ (1)  ​​Mijkat β1​ ​​  (Tem​ pk​ ​​  × Ag​e​a​​ × After) Table 1 provides the parameter estimates and   β​ + ​ 2​​  (Tem​ pk​ ​​  × Ag​e​a​​  ) standard errors from model (1) using the three migration outcomes disaggregated by the gender   β3 + ​ ​ ​​  (Tem​ pk​ ​​  × After) of the individual. We only observe a positive and statistically significant effect on the migration of   β​ + ​ ​ ​​ × After) + θ ​X​ 4​​  (Ag​ea ijkt​​ women to a provincial capital. The tendency of women to migrate in response to temperature   α​ + ​ δ​ j​​  + ​ γ​ t​​  + ​ a​​  + ​ ϵ​ijkat​​  , ​ variability is consistent with recent evidence in South America (Thiede, Gray, and Mueller where ​​Mijkat ​ ​​​is a binary variable for whether 2016). Women may be used to diversify risk, if individual ​ i​at destination province j ​ ​from origin the opportunity cost of having an absent young province ​ k​in age group ​ a​at time ​ t​migrated in male member of the household exceeds that of the last five years; ​Ag​e​a​​​ is a vector of 10-year a young female member. Alternatively, income age indicators (15–25, 26–35, 36–45, 46–55, losses caused by fluctuations in temperature 56–65 omitted); ​ Afte​rt​​​​ signifies the follow-up may affect the demand for goods and services census for the country; ​​ X​ijkt​​​is a vector of pre- provided by female-dominated industries (e.g., shock variables (indicators for being male and seamstresses). having completed primary school, and five-year Table 2 presents the migration impacts of average precipitation and precipitation squared heat exposure among the unskilled (has not at origin). Location, generational and temporal completed a primary education).8 A 1 standard factors that influence migration patterns are con- deviation increase in heat exposure more than trolled for by including origin ​​ αj​​​​  , age ​​γa​ ​​​  , and doubles (quadruples) the probability of a young year ​​δt​​​​ fixed effects.7 Standard errors are clus- (ages 15–25), unskilled woman migrating to a tered by origin province and birth year. provincial (national) capital. Restricting the Identification rests on the following assump- focus to the unskilled also introduces a positive tions. First, our methodology assumes disasters and significant effect on the provincial capital would affect the migration status of an individual migration of women ages 26–35, and a positive in the 15–35 age range significantly more than and significant effect on the provincial capital individuals outside of that age range (36–65). migration of young men. Difference-in-difference regression results The GDP regressions qualitatively suggest which compare the change in the mobility pat- that the observed migration patterns do not coin- terns across age groups confirm that younger cide with any broad economic losses (online men and women are more likely to move and the Appendix). Young women may be used by migration patterns of young men and women are households to obtain auxiliary income to smooth consumption. Consequently, they travel to pro- vincial and national capitals because these areas 7  Controlling for pull factors through the inclusion of destination province fixed effects does not change infer- ences on the parameters of interest. We focus on the abbre- viated empirical model given concerns over the potential for 8  biased estimates generated from the endogeneity of location There are no significant impacts on the skilled sample decisions. (online Appendix). VOL. 107 NO. 5 Heat Exposure and Youth Migration 449 Table 1—Impact of Heat Exposure on Migration Sample: Women Men Destination: National Provincial National Provincial Any capital capital Any capital capital Temp ​×​ after ​×​ age 15–25 −0.00011 −0.00106 0.00300 −0.00018 0.00012 0.00161 (0.0058) (0.00110) (0.00178) (0.00562) (0.00081) (0.00168) Temp ​×​ after ​×​ age 26–35 0.00282 0.00027 0.00169 0.00053 −0.00003 0.00062 (0.00654) (0.00107) (0.00188) (0.00704) (0.00099) (0.00190) Temp ​×​ after ​×​ age 36–45 0.00193 0.00026 0.00864 0.00008 0.00040 0.00001 (0.00546) (0.00009) (0.00187) (0.00615) (0.00083) (0.00183) Temp ​×​ after ​×​ age 46–55 0.00284 0.00061 0.00069 −0.00074 0.00046 −0.00088 (0.00590) (0.00109) (0.00222) (0.00649) (0.00093) (0.00214) Constant −0.00502 −0.01988 −0.02864 −0.01062 −0.01573 −0.02328 (0.01321) (0.00342) (0.00468) (0.01421) (0.00253) (0.00440) Mean temp 0.751 0.752 0.749 0.737 0.738 0.735 Mean control migration rate 0.023 0.001 0.008 0.024 0.001 0.008 R2 0.073 0.242 0.113 0.061 0.259 0.098 Observations 8,599,759 8,283,727 8,310,807 7,883,317 7,589,007 7,618,804 Notes: Temp represents the standardized number of excessive heat days experienced over a five-year period. Origin-province by birth year clustered standard errors in parentheses. Table 2—Heat-Induced Migration Patterns of the Unskilled, Without Primary Education Sample: Women Men Destination: National Provincial National Provincial Any capital capital Any capital capital Temp ​×​ after ​×​ age 15–25 0.00571 0.00443 0.00760 0.00201 0.00111 0.00212 (0.00550) (0.00185) (0.00201) (0.00447) (0.00082) (0.01263) Temp ​×​ after ​×​ age 26–35 0.00637 0.00109 0.00271 0.00010 0.00082 0.00012 (0.00533) (0.00096) (0.00141) (0.00491) (0.00069) (0.00121) Temp ​×​ after ​×​ age 36–45 0.00354 0.00160 0.00190 −0.00136 0.00120 0.00003 (0.00455) (0.00098) (0.00143) (0.00452) (0.00058) (0.00106) Temp ​×​ after ​×​ age 46–55 0.00256 0.00115 0.00062 −0.00027 0.00025 −0.00104 (0.00463) (0.00084) (0.00142) (0.00444) (0.00058) (0.00118) Constant 0.06011 0.01794 0.00935 0.01811 0.00583 −0.00985 (0.01502) (0.00703) (0.00849) (0.01392) (0.00347) (0.00570) Mean temp 0.782 0.784 0.782 0.737 0.739 0.737 Mean control migration rate 0.020 0.001 0.005 0.019 0.001 0.005 R2 0.117 0.324 0.212 0.080 0.318 0.145 Observations 2,688,088 2,612,545 2,619,639 2,246,249 2,184,378 2,191,324 Notes: Temp represents the standardized number of excessive heat days experienced over a five-year period. Origin-province by birth year clustered standard errors in parentheses. are at least perceived to offer a greater range of young women9 and 1,578 young men. The employment opportunities given their skills. total effect is smaller than that observed under IV. Discussion 9  This calculation comes from multiplying the total num- ber of unskilled workers (2,619,639), the percentage of Our results imply that a 1 standard deviation individuals aged 15–25 (0.35), and the coefficient in Table 2 increase in heat would affect the lives of 7,314 (0.008). 450 AEA PAPERS AND PROCEEDINGS MAY 2017 droughts (41,559 people) or hurricanes (13,931 Dimensions of Human Migration: The people), but could increase with climate change. Demographer’s Toolkit.” Global Environmen- In all instances, we are likely to omit a signifi- tal Change 28: 182–91. cant fraction of short-distance moves due to data Henderson, J. Vernon, Adam Storeygard, and limitations. What is most disconcerting is that Uwe Deichmann. 2017. “Has Climate Change youth are more likely to wind up in urban centers Driven Urbanization in Africa?” Journal of when exposed to heat than the common disasters Development Economics 124: 60–82. experienced in the region. Additional research is Lee, Tien Ming, Ezra M. Markowitz, Peter D. warranted over the welfare implications of these Howe, Chia-Ying Ko, and Anthony A. Lei- choices in the long term and the interventions serowitz. 2015. “Predictors of Public Cli- available to minimize distress migration. mate Change Awareness and Risk Perception around the World.” Nature Climate Change 5: 1014–20. References Minnesota Population Center. 2015. “Inte- grated Public Use Microdata Series, Interna- Auffhammer, Maximilian, Solomon M. Hsiang, tional: Version 6.4.” University of Minnesota Wolfram Schlenker, and Adam Sobel. 2013. (accessed June 15, 2015). “Using Weather Data and Climate Model Out- Mueller, V., C. Gray, and K. Kosec. 2014. “Heat put in Economic Analyses of Climate Change.” Stress Increases Long-Term Human Migration Review of Environmental Economics and Pol- in Rural Pakistan.” Nature Climate Change 4: icy 7 (2): 181–98. 182–85. Baez, Javier, German Caruso, Valerie Mueller, and Poelhekke, Steven. 2011. “Urban Growth and Chiyu Niu. 2  017. “Droughts Augment Youth Uninsured Rural Risk: Booming Towns in Migration in Northern Latin America and the Burst Times.” Journal of Development Eco- Caribbean.” Climatic Change 140 (30): 423– nomics 96 (2): 461–75. 35. Rodell, Matthew, and Hiroko Kato Beaudoing. Black, Richard, Stephen R. G. Bennett, Sandy M. 2015. “GLDAS Noah Land Surface Model Thomas, and John R. Beddington. 2011. “Cli- L4 monthly 0.25 x 0.25 degree Version 2.0.” mate Change: Migration as Adaption.” Nature Goddard Earth Sciences Data and Informa- 478: 447–49. tion Services Center, Greenbelt, MD. https:// Bryan, Gharad, Shyamal Chowdhury, and Ahmed disc.sci.gsfc.nasa.gov/uui/datasets/GLDAS_ Mushfiq Mobarak. 2014.  “Under-Investment NOAH025_3H_2.0/summary#citation in a Profitable Technology: The Case of Sea- (accessed August 2016). sonal Migration in Bangladesh.” Econometrica Thiede, Brian, Clark Gray, and Valerie Mueller. 82 (5): 1671–1748. 2016. “Climate Variability and Inter-provin- Fussell, Elizabeth, Lori M. Hunter, and Clark L. cial Migration in South America, 1970–2011.” Gray. 2014. “Measuring the ­ Environmental Global Environmental Change 41: 228–40.