Climate Change as a Migration Driver in Mexico,

Transcription

1 University of Colorado, Boulder CU Scholar Sociology Graduate Theses & Dissertations Sociology Spring Climate Change as a Migration Driver in Mexico, Raphael J. Nawrotzki University of Colorado at Boulder, raphael.nawrotzki@colorado.edu Follow this and additional works at: Part of the Climate Commons, Demography, Population, and Ecology Commons, and the Natural Resource Economics Commons Recommended Citation Nawrotzki, Raphael J., "Climate Change as a Migration Driver in Mexico, " (2014). Sociology Graduate Theses & Dissertations This Dissertation is brought to you for free and open access by Sociology at CU Scholar. It has been accepted for inclusion in Sociology Graduate Theses & Dissertations by an authorized administrator of CU Scholar. For more information, please contact cuscholaradmin@colorado.edu.

2 CLIMATE CHANGE AS A MIGRATION DRIVER IN MEXICO, by RAPHAEL J. NAWROTZKI B.S., University of Applied Sciences Darmstadt, 2007 M.S.A., Andrews University, 2009 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirement for the degree of Doctor of Philosophy Department of Sociology 2014

3 This thesis entitled: Climate Change as a Migration Driver in Mexico, written by Raphael J. Nawrotzki has been approved for the Department of Sociology Dr. Lori M. Hunter Dr. Fernando Riosmena Dr. Richard G. Rogers Dr. Fred C. Pampel Dr. Daniel M. Runfola Date: The final copy of this thesis has been examined by the signatories, and we find that both the content and the form meet acceptable presentation standards of scholarly work in the above mentioned discipline.

4 ABSTRACT Nawrotzki, Raphael J. (Ph.D., Sociology) Climate Change as a Migration Driver in Mexico, Thesis directed by Professor Lori M. Hunter Grounded in the Sustainable Livelihoods framework, this study investigates the impact of climate change on various migration streams in Mexico, comparing migration from rural versus urban areas, international versus domestic moves, and first versus last moves within households. To measure the effects of climate change on these migration streams, a set of 17 climate change indices, proposed by the Expert Team on Climate Change Detection and Indices (ETCCDI), was generated using daily temperature and precipitation data for 214 Mexican weather stations obtained from the Global Historical Climate Network-Daily (GHCN-D) data set. Cokriging as a method of spatial interpolation was employed to assign climate change index values to 111 Mexican municipalities, for which detailed migration histories and relevant sociodemographic characteristics were obtained from the Mexican Migration Project (MMP). Multi-level event history models were employed to estimate the impact of various climate change indices on household-level migration patterns from 1986 to The results indicate climate change more strongly impacts migration from rural compared to urban areas and international moves more so than domestic moves. In addition, within households, first moves are more sensitive to climate change than later moves. First international moves from rural areas are the most sensitive streams and are predominantly influenced by warming temperatures and decreasing rainfall. Significant socio-climatic interactions demonstrated that the effect of climate change on iii

5 migration varies by characteristics such as the proportion of the male labor force employed in the agricultural sector, municipality-level wealth, and access to migrant networks. Investigating the timing of migration suggests a direct response pattern -- the climate signal in the preceding year leads to the strongest migration response and declines in importance thereafter. As compared to other predictors, the strength of the climatic effect is about a tenth of the size of the strongest sociodemographic factors (e.g., social networks). The findings suggest that projected increases in temperature and declines in precipitation over the 21 st century may increase Mexico-U.S. migration. Climate change adaptation programs intended to reduce international migration may be most effective when targeted to improve the livelihoods of agricultural-dependent households in rural Mexico. iv

6 ACKNOWLEDGMENTS I would not have been able to complete this dissertation research without the incredible help of various outstanding individuals. First of all, I am grateful to all my committee members, Lori Hunter, Fernando Riosmena, Rick Rogers, Fred Pampel, and Dan Runfola, who have been always helpful and patient with my questions in various areas of theory, methodology, and analysis. Foremost, I am incredibly thankful to Lori as my advisor who has been a great mentor and friend throughout my time as graduate student and whose tireless help, support, and guidance during the entire process of conducting this dissertation research has been invaluable. Special thanks also to Fernando as Mexican migration expert who has taught me much about migration dynamics in Latin America and who has always been a great resource to discuss substantive and methodological issues. I would also like to express my gratitude to Rick whose population class has laid the foundation for my academic development in the field of demography and whose valuable insights and feedback have had a major impact on the quality and development of this dissertation project. I would like to express my sincere gratitude to Fred from whom I have learned almost everything I know about statistics and research methods and whose friendly help, support, and advice with various statistical issues has been invaluable for the completion of this project. I would also like to thank Dan who has been of great help in solving various spatial problems that emerged during this research, and whose profound insights, questions, and comments have vastly improved the quality of this study. I am also thankful for the emotional support of my parents Dagmar and Andreas and my siblings Tirza and Simon, who have always encouraged me to give my best but also to take some breaks when needed. A special thanks also to my grandparents who have financially supported my academic ventures on a monthly basis for more than 11 years. v

7 And last but not least, I acknowledge the source of all energy, intelligence, and knowledge, which I believe is not of this word but resides with the great architect and sustainer of all things who alone is immortal and who lives in unapproachable light, whom no one has seen or can see (1 Timothy 6:16, The Bible). vi

8 TABLE OF CONTENTS CHAPTER I INTRODUCTION...1 II CONCEPTUAL FRAMEWORK AND LITERATURE REVIEW...6 Conceptual Framework...6 The Sustainable Livelihoods (SL) Framework...6 Causal Chains Connecting Climate Change and Migration...12 Migration as Self Insurance Mechanism: New Economics of Labor Migration...13 Prior Studies Investigating the Climate Change-Migration Association...14 The Study Region: Mexico...19 Mexican Migration: Historical Trends and Current Patterns...19 Climate Change Context and Vulnerability...22 Scope of the Study...28 III DATA AND METHODS...32 Data...32 Demographic Data...32 Climate Data...33 Non-Climatologic Spatial Data...34 Demographic Unit of Analysis...34 Spatial Unit of Analysis...35 Time Frame...36 Variable Construction...37 Outcome Variables...37 Primary Predictor Variables...42 Secondary Predictors (Controls)...58 vii

9 Estimation Strategy...65 IV RESULTS: CLIMATE CHANGE AS A MIGRATION DRIVER: DIFFERENCES BETWEEN RURAL AND URBAN AREAS...71 Hypotheses...71 Climate Change Effects: Temperature and Precipitation...71 Climate Change Effects: Rural vs. Urban...74 Hypothesis Testing Conventions...75 Results and Discussion...77 Model Building...77 Climate Change Effects: Temperature and Precipitation...83 Climate Change Effects: Rural vs. Urban...90 Headship, Documentation Status, and Historical Regions...95 Socio-climatic Interactions...97 V RESULTS: CLIMATE CHANGE AS A MIGRATION DRIVER: DIFFERENCES BETWEEN INTERNATIONAL AND DOMESTIC MIGRATION Hypotheses Results and Discussion Model Building Climate Change Effects: International vs. Domestic Moves Strength of the Observed Effects Socio-climatic Interactions VI RESULTS: CLIMATE CHANGE AS A MIGRATION DRIVER: DIFFERENCES BETWEEN FIRST AND LAST MOVES Hypotheses Results and Discussion Model Building viii

10 Climate Change Effects: First vs. the Last Move Timing of Climate Change Related Migration Socio-climatic Interactions VII CONCLUSIONS Theoretical Contributions: Expanding the Environmental Dimensions of the Sustainable Livelihoods Framework Limitations Policy Implications REFERENCES APPENDIX A B TECHNICAL DEFINITION OF THE ETCCDI CORE CLIMATE CHANGE INDICES MULTIPLE IMPUTATION (MI) C SPATIAL INTERPOLATION USING KRIGING D ROBUSTNESS TESTS ix

11 TABLES Table 3.1 Life table describing the event of different migration types during the period for households of all MMP communities (rural and urban) List of ETCCDI climate change indices measuring temperature extremes List of ETCCDI climate change indices measuring precipitation extremes Summary statistics of climate change indices for selected periods Secondary predictors (controls) for the analysis of climate change impacts on migration in Mexico, Odds of international migration from households in Mexico, Odds of international migration from households in rural and urban Mexico, Estimates of the effect of climate change on international migration from rural and urban Mexico, Estimates of the effect of climate change on international migration from rural Mexico, , stratified by headship status, documentation status, and location characteristics Estimates of the effect of climate change on international migration from urban Mexico, , stratified by headship status, documentation status, and location characteristics Interaction between climate change indices and the male labor force employed in the agricultural sector predicting the odds of international migration from rural Mexico, Interaction between climate change indices and the municipality-level wealth index predicting the odds international migration from rural Mexico, Interaction between climate change indices and international migrant prevalence in the community predicting the odds international migration from rural Mexico, x

12 4.9 Interaction between climate change indices and the municipality-level wealth Index predicting the odds of international migration from urban Mexico, Interaction between climate change indices and occupation predicting the odds of international migration from urban Mexico, Odds of international and domestic migration from households in rural Mexico, Estimates of the effect of climate change indices on international and domestic migration from rural Mexico, Comparison of the size of the climate change effect in relation to sociodemographic factors, predicting international migration from rural Mexico, Interaction between climate change indices and occupational status predicting the odds of domestic migration from rural Mexico, Odds of a first and last international move from households in rural Mexico, Estimates of the effect of climate change on the first and the last international migration from rural Mexico, Estimates of the timing of the migration response in relation to the climate signal for the first international move from rural Mexico, Interaction between climate change indices and municipality domestic migrant prevalence predicting odds of the last international move from rural Mexico, Interaction between climate change indices and the proportion of irrigated farmland, predicting the odds of the last international move from rural Mexico, Summary of the hypothesis testing results to investigate the climate change migration association in Mexico, D.1 Jack-knife type procedure to investigate the robustness of the observed climate change effects for international migration from rural and urban Mexico, xi

13 D.2 Jack-knife type procedure to investigate the robustness of the observed climate change effects comparing international to domestic migration from rural Mexico, D.3 Jack-knife type procedure to investigate the robustness of the observed climate change effects comparing first to last international migration from rural Mexico, xii

14 FIGURES Figure 2.1 Sustainable livelihood framework with added emphasis on climate change and migration Climatic zones across Mexico derived from a Koeppen classification Geographic location of MMP municipalities and spatial distribution of weather stations across Mexico Hazard functions for international and domestic migration from MMP Communities during the years Interpolated surface of values for maximum length of dry spell (cdd) Root Mean Squared Error (RMSE) values across weather stations for average wet-day precipitation (sdii) across Mexico Root Mean Squared Error (RMSE) values across weather stations for % warm nights (tn90p) across Mexico Difference between current interpolations and values based on HadEX2 grids for sdii (a) and tn90p (b) for the year Confidence matrix used to evaluate the hypothesized effect of climate change on migration based on agreement and evidence Interaction between the warm spell duration (a) and number of days of very heavy precipitation (b) and the male labor force employed in the agricultural sector in predicting the odds of international migration from rural Mexico, Interaction between the warm spell duration (a) and number of days of very heavy precipitation (b) and the municipality-level wealth index in predicting the odds of international migration from rural Mexico, Interaction between warm spell duration and migrant networks in predicting the odds of international migration from rural Mexico, Interaction between warm spell duration and the municipality-level wealth index, predicting the odds of international migration from urban Mexico, Interaction between warm spell duration (a) and the number of days of heavy precipitation (b) and occupational status, predicting the odds of international migration from urban Mexico, xiii

15 5.1 Interaction between the coldest day (a) and the number of days with very heavy precipitation (b) and occupational status predicting the odds of domestic migration from rural Mexico, Interaction between the number of frost days (a) and the number of days of very heavy precipitation (b) and the percentage of domestic migrant households in predicting the odds of the last international move from rural Mexico, Interaction between the number of summer days (a) and the precipitation on extremely wet days (b) and the proportion of the farmland irrigated in predicting the odds of the last international move from rural Mexico, Refined Sustainable Livelihoods framework with added information on the effect of climate change on migration dynamics for Mexico, C.1 Semivariogram used to obtain spatial weights for the interpolation of values for unknown points xiv

16 CHAPTER I INTRODUCTION Climate change has become a publically recognized problem of global magnitude. The award winning work of the Intergovernmental Panel on Climate Change (IPCC) first brought public-wide attention to the impact of industrialization and anthropogenic greenhouse gas (GHG) emission on changes within the climatic system (IPCC, 2007, 2013). Furthermore, a general consensus exists that particularly poor, less developed countries (LDCs) will suffer the most from climate change impacts including increased frequency and intensity of storms, floods and droughts as well as more gradual changes involving sea level rise and desertification (Adamo and de Sherbinin, 2011; Huq et al., 2003; Roberts and Park, 2006; UNHR, 2007). The differential impact on LDCs is a result of relatively higher dependence on agriculture and local natural resources, as well as a lack of financial capital to employ technological barriers as protection against adverse climatic shocks (Gutmann and Field, 2010). In the face of environmental strain related to, for example, increasing temperature or decreasing rainfall, rural populations may employ in situ (in place) adaptation strategies (Bardsley and Hugo, 2010; Davis and Lopez-Carr, 2010). As examples, in situ adaptation strategies may include selling assets, intensifying or adopting new livelihood activities, using formal and informal credit, reducing nonessential expenditures, and/or drawing on social networks and public programs for assistance (Gray and Mueller, 2012a). In addition to in situ strategies, a household may send a member elsewhere to seek employment and send remittances, or a household may locate entirely (McLeman and Smit, 2006; Warner et al., 2010). Migration as an ex situ livelihood strategy may occur simultaneously during the implementation of in situ strategies, as well as after those strategies have failed (Ezra and Kiros, 2001; Meze-Hausken, 2000). A number of studies have investigated the impact of rainfall decline, and droughts, on migration (e.g., Henry, Schoumaker, and Beauchemin, 2004; Hunter, Murray, and Riosmena, 2013), but migration research employing direct measures of climate change is rare -- or non-existent. Moreover, households in particular settings likely engage some forms of migration more than others (e.g., international vs. 1

17 domestic), yet little scholarship has examined these distinctions (see Gray and Bilsborrow, 2013, and Henry, Schoumaker, and Beauchemin, 2004, for exceptions). The research presented here offers a unique contribution to the migration-environment literature through its inclusion of multiple measures of climate change, combined with its focus on differential climate impacts on distinct migration flows including international/domestic, urban/rural, and as related to timing of the move (first/last move within households). This study is grounded in, and builds on, current migration theory, particularly the New Economics of Labor Migration perspective (Stark and Bloom, 1985). This perspective considers migration as a strategy through which households in developing countries overcome various livelihood constraints in the absence of functioning insurance and capital markets (Massey et al., 1993). A household may send a migrant to an international destination to overcome liquidity constraints in order to accumulate the financial means to buy property, build a house, start a business, or invest in production equipment such as irrigation systems (Massey et al., 1993; Massey, 1987b; Massey and Parrado, 1998). More important for the study of the environment - migration relationship, households employ migration as a self-insurance mechanism in the absence of functioning insurance markets. For example, in the absence of crop insurance systems a household may place a member strategically in a location where the agricultural market conditions are independent of those in the origin (Stark and Bloom, 1985). When local conditions in the agricultural sector deteriorate due to climate change impacts, the household can rely on migrant remittances for support (Massey et al., 1993). I also make use of the Sustainable Livelihoods Framework (Ashley and Carney, 1999; Chambers and Conway, 1991) to better understand migration decisions within the complex socioeconomic and sociodemographic contexts within which households are embedded. This framework proposes that access to various livelihood capitals, including natural, social, human, physical and financial capital determine the type and variety of livelihood strategies employed by a household (Morse et al., 2009). Of key importance to this research project is the relationship between natural capital and migration as a livelihood strategy. Natural capital encompasses the natural resource stocks (soil, water, weather) and 2

18 environmental services (e.g., hydrological cycle, carbon cycle) from which resources flow that offer livelihood options (Scoones, 1999). Climate change directly impacts natural capital, which in turn may lead to changes in livelihood strategies such as farming patterns and migration responses. By drawing on these two frameworks, this research adds to a growing body of literature expanding present knowledge regarding the complexities of environment - population dynamics (e.g., Black et al., 2011). This dissertation project offers an inductive approach to improve our understanding as to which features of the changing climate system (e.g., high temperature extremes, maximum wet day precipitation) are associated with particular streams of out-migration. Further, the study clarifies whether climate change impacts migration largely in rural areas, likely through impacts on the agricultural sector (e.g., Feng and Oppenheimer, 2012), or if climate effects also manifest in migration from urban areas (e.g., Adamo, 2010). The investigation of the climate change - migration association has important implications for policymakers. Based on the findings of this study, policymakers will be better positioned to design policies and livelihood based programs to target climate change sensitive migration flows and better assist vulnerable subpopulations. For an investigation of climate change impacts on migration patterns, it is necessary to first define what is meant by climate change. Climate typically refers to the weather in a location averaged over long periods of time, whereas weather refers to temperature and precipitation at a specific time and place (Deschenes and Greenstone, 2007, p. 354). In the same vein, the glossary of the U.S. Environmental Protection Agency (EPA) defines climate change as follows: Climate change refers to any significant change in the measures of climate lasting for an extended period of time. In other words, climate change includes major changes in temperature, precipitation, or wind patterns, among others, that occur over several decades or longer (EPA, 2012). This definition closely resembles those proposed by the Intergovernmental Panel of Climate Change (IPCC, 2014) and the United Nations Framework Convention on Climate Change (UNFCCC, 3

19 1992). However, all three sources are vague regarding the specific timeframe over which a change must occur to be labeled climate change. For example, the EPA defines this timeframe as several decades (EPA, 2012; Baede, 2001) while the United Nations Framework Convention on Climate Change uses the phrase comparable time periods (UNFCCC, 1992). A commonly used practice to approximate climate change is to relate weather information to a reference period of 30 years, in most cases 1961 to 1990, which is often referred to as climate normal (c.f., Conde, Ferrer, and Crozco, 2006; Frich et al., 2002; Klein Tank et al., 2006; Poston et al., 2009). Although some experts argue that climate change is actually a gradual development over much longer time spans (centuries or more), changes across shorter time periods (e.g., 30 to 40 years) are more likely to be consciously experienced by humans and thus to induce behavioral responses such as migration. In the current study, migration is associated with changes in climate extremes across a period of 39 years (1961 to 1999) -- a period longer than the climate normal but shorter than centuries. Although some researchers would prefer the term weather extremes, changes such as those measured have been associated with climate change (Alexander et al., 2006; Peterson and Manton, 2008). Therefore, the label climate change is used throughout the text. Natural and anthropogenic substances and processes that alter the earth s energy budget drive climate change. Increasingly, evidence suggests that much of the observed climate change is anthropogenic in nature. The recent IPCC report (IPCC, 2013, p. 17) states that It is extremely likely that human influence has been the dominant cause of the observed warming since the mid-20 th century. For example, experiments show that human influence has more than doubled the probability of occurrence of temperature extremes in some locations (Bindoff et al., 2013). Such attributions have important political and moral implications regarding liability and compensation for costs related to climate change adaptation and response measures (Nawrotzki, 2014). Even so, this dissertation project is not concerned with attribution of the causes of climate change. Rather, the specific focus is on the climate change - migration association for Mexico during the period of with the primary aims of empirically understanding this association and contributing to migration and livelihood theory. 4

20 The manuscript is structured as follows. Following this brief introduction (Chapter I), the next chapter (Chapter II) introduces the sustainable livelihood framework as the theoretical construct in which this study is grounded. This chapter also reviews the published literature on the climate - migration association and presents Mexico as the national focus. Mexico provides a unique case because of an established history of migration as well as the sensitivity of various urban and rural sectors to climate change. Chapter III presents the data used, the methods employed to construct the relevant variables, and the estimation strategy. The following three chapters present key results of the investigation of the impact of climate change on migration from rural versus urban areas (Chapter IV), differences in international versus domestic migration (Chapter V), and a comparison between first and last moves within particular households (Chapter VI). In the concluding chapter (Chapter VII), the central findings are used to expand the theoretical framework, followed by a discussion of limitations and policy implications. 5

21 CHAPTER II CONCEPTUAL FRAMEWORK AND LITERATURE REVIEW Conceptual Framework None of the classical migration frameworks have been explicitly designed to address climate change-related migration (see Massey et al., 1993 for an overview). However, a few frameworks implicitly include environmental factors as contextual features operating either as migration push (disamenities, hazards, etc.) or pull (amenities) factors (Hunter, 2005). Black et al. (2011) have recently developed a framework that reflects migration as being impacted by macro-level factors categorized as political, social, economic, demographic, and environmental drivers. These drivers interact to impact migration decisions and are themselves influenced by environmental change. Black et al. s framework also reflects understanding of migration as influenced by household (micro-level) and community (meso-level) characteristics. However, the framework provides little detail on these micro and meso scales. Since the present study takes a micro-level focus to explore the impact of climate change on household-level migration decisions, the sustainable livelihoods framework (Carney et al., 1999; Scoones, 1999) was deemed a particularly useful conceptual tool. The Sustainable Livelihoods (SL) Framework The sustainable livelihoods framework was originally designed as a tool to understand the complex context that determines household vulnerability and resilience to external shocks such as harvest failures or food crises (Ashley and Carney, 1999). The framework has been widely used by researchers (e.g., Gray, 2009; Massey, Axinn, and Ghimire, 2010; Nawrotzki, Hunter, and Dickinson, 2012), policymakers, and practitioners as a tool for livelihoods analysis 6

22 with the goal of enhancing resilience and to foster sustainable development (Knutsson, 2006). Sustainable livelihoods have been defined by the architects of this framework as follows (Chambers and Conway, 1991, p.6): A livelihood comprises the capabilities, assets (stores, resources, claims and access) and activities required for a means of living: a livelihood is sustainable which can cope with and recover from stress and shocks, maintain or enhance its capabilities and assets, and provide sustainable livelihood opportunities for the next generation; and which contributes net benefits to other livelihoods at the local and global levels and in the short and long term. The SL framework uses a multiple-capital approach to investigate the livelihood sustainability (Morse et al., 2009). A household s overall access to, and possession of livelihood assets including financial, physical, human, social, and natural capital allows a household to engage in various livelihood strategies (Carney et al., 1999). The more diversified a household's livelihood strategies, the more resilient the household is in the face of shocks (Meze-Hausken, 2000; Scoones, 1999). Climate shocks may be particularly detrimental for agricultural dependent rural households (McLeman and Smit, 2006) as well as for the urban poor (Satterthwaite et al., 2007). Each group may lack the resources necessary to guard against climate shocks through technological adjustments (e.g. irrigation) or insurance. Households respond to livelihood strain through a variety of strategies aimed to increase resilience, with migration representing one such strategy (Chambers and Conway, 1991; Scoones, 1999). The following sections review the complex ways in which the five livelihood capitals are linked to migration as a livelihood strategy. Financial capital represents the financial resources available to a household (e.g., income, savings, and remittances) and may influence the propensity to migrate in contrasting ways. On one hand, capital is necessary to finance a move, especially to an international destination (Brown and Bean, 2006). On the other hand, relative financial deprivation and liquidity 7

23 constraints can motivate a move (Katz and Stark, 1986). In addition, financial capital may be used to purchase insurance, fortify buildings to withstand storms and flooding, purchase drought resistant seeds, or install irrigation systems. All of these strategies may increase climatic resilience and reduce the necessity of livelihood diversification through migration (Warner and van der Geest, 2013). Particularly important for Mexican migration, social capital includes the social resources, such as networks, group membership and trust relationships, upon which individuals and households draw (Woolcock and Narayan, 2000). The migration literature provides evidence that access to migrant networks strongly increases the likelihood of a move through cumulative causation (de Haas, 2010; Fussell and Massey, 2004). Connecting places of origin and destination, these networks reduce the costs and increase the benefits of migration (Massey, 1990), and frequently determine the destination of a move (Juelich, 2011). In the presence of well-established migration corridors, the sensitivity of migratory responses towards climate triggers may be strongly amplified (Bardsley and Hugo, 2010; Adamo and de Sherbinin, 2011). At the individual level, human capital includes labor skills, knowledge, and ability, which are central to various livelihood pursuits. Whether human capital increases or decreases outmigration depends on its economic returns at the origin and destination (Taylor et al., 1996). For example, education has been shown to increase domestic rural-urban migration, but not international migration, especially moves from developing to more industrialized countries (Henry, Schoumaker, and Beauchemin, 2004; Massey, 1987b; Stark and Taylor, 1991). At the household level, human capital includes the availability of household members to be employed in a family business or engage in other employment. Human capital is also represented by the availability of household members for migration. When climate change 8

24 adversely impacts local employment, unemployed laborers become potential migrants. Climate change may impact the employment situation in rural areas through the agricultural sector (Munshi, 2003) and in urban areas through the weather dependent service sectors, such as tourism (Amelung, Nicholls, and Viner, 2007; Lise and Tol, 2002). Physical capital represents the basic infrastructure and production equipment enabling the pursuit of various livelihood strategies. Irrigation systems, for instance, may enable production even under adverse climatic conditions and reduce the necessity of sending a household member elsewhere (Eakin, 2005). In contrast, climate change may increase the frequency and severity of extreme weather events such as storms and floods (IPCC, 2013, 2014), which may lead to the destruction of physical capital such as houses, roads, factories and production equipment. This may, in turn spur out-migration (Fussell, Sastry, and Van Landingham, 2010; Fussell, Curtis, and DeWaard, 2014). Extreme weather events may be particularly destructive in urban areas due to high population density and the fact that many of the large urban centers are situated in environmentally vulnerable locations such as coastal areas and dry-lands (Adamo, 2010). Natural capital includes access to land, water, and natural resources, which enable rural households to engage in agricultural pursuits and/or natural resource collection for both sustenance and income generation (Scoones, 1999). Climate change, through shifting environmental conditions and an increase in the frequency of weather extremes (e.g., droughts, floods, storms), has the potential to substantially alter the productivity of the agricultural sector (Eakin and Appendini, 2008; Nawrotzki and Akeyo, 2009). In some cases, these alterations may result in livelihood insecurity, which might be followed by migration (Meze-Hausken, 2000). Natural capital is linked to the other livelihood capitals (Nawrotzki, Hunter, and Dickinson, 2012). For example, in rural areas adverse climate conditions can negatively impact 9

25 crop yields (natural capital), which may reduce agricultural employment opportunities and associated income (financial capital), influence migration (loss of household human capital, gain in social capital), or require the selling of assets (physical capital) to recover costs associated with a harvest failure (Warner and van der Geest, 2013). Climate change impacts on natural capital are likely to be felt in urban centers as well. For example, droughts may lead to challenges with regard to domestic water supplies (Revi, 2008) as well as impacting manufacturing and production processes that require large amounts of water as inputs (e.g., cement and paper production) or for cooling. In addition, climate change may lead to increases in food prices through adverse impacts on urban and peri-urban agriculture (Satterthwaite et al., 2007) and may also increase energy costs if production is based on hydropower generation or biomass combustion (Schaeffer et al., 2012). Whether or not certain livelihood strategies can be pursued and successfully implemented depends largely on the sociopolitical context, which encompasses laws, policies, institutions, and governance (Eakin, 2000; Mahdi, Shivakoti, and Schmidt-Vogt, 2009), as well as the sociocultural context which includes gender roles (Juelich, 2011), value of children, and spiritual connection to the land (de Sherbinin et al., 2008). Other contextual factors include household and community demographic composition, as well as the macro-level economic context (Fussell and Massey, 2004; Black et al., 2011). For example, border protection policies shape the ease with which an international move can be undertaken. Other policies and programs may help to protect certain livelihood assets. For instance, crop insurance systems may help offset climate-related damages to the harvest (Juelich, 2011). In addition, publically-funded agricultural extension services may provide information on the use of advanced irrigation technologies (c.f., Bourzac, 2013) and may make seeds of drought resistant crop varieties available at reduced prices (c.f., 10

26 Eisenstein, 2013). Often non-governmental and private agencies provide additional support for livelihoods through the implementation of aid and development programs (Morse et al., 2009). A visual depiction of the employed SL framework is provided in Figure 2.1. Figure 2.1: Sustainable Livelihood framework with added emphasis on climate change and migration Note: Thickness of the climate change impact arrows represent anticipated strength of the climate change effects on various sectors based on Boyd and Ibarraran (2009) (not measured in this study). The interconnectedness between various livelihood capitals is illustrated by dashed lines. 11

27 The visual representation of the SL framework illustrates the causal pathways of the influence of climate change on various economic sectors, leading to effects on capitals, which in turn results in migration as a household-level livelihood strategy. Feedback likely occurs when climate change induced migration leads to a change in capital composition. For example, a household may sent a migrant elsewhere in response to adverse impacts of droughts on agricultural production. The migrant remittances may be used to purchase drought resistant crops and install irrigation systems. The resulting increased level of climate resilience will then decline future migration probabilities. However, since such feedback loops cannot be adequately represented in a regression framework, they were not included in the framework representation. Causal Chains Connecting Climate Change and Migration The modified sustainable livelihoods framework helps to illustrate the causal chains connecting climate change with out-migration, and the ways in which they may differ for rural and urban areas: For rural areas, climate change (external shock) directly impacts agricultural production and crop yields (natural capital), which in turn affects livelihoods and may influence migration probabilities (livelihood strategy) (Feng, Krueger, and Oppenheimer, 2010; Feng and Oppenheimer, 2012). In urban areas, climate change may impact employment opportunities in the manufacturing sector when weather extremes destroy equipment, buildings, and infrastructure (Revi et al., 2014), and increase production costs through the impact on power generation (e.g., hydro power) (Pereira de Lucena et al., 2009). In addition, weather dependent service sectors such as tourism may be impacted by climate change (Amelung, Nicholls, and Viner, 2007; Lise and Tol, 2002). In this way, climate change may also impact urban livelihoods and, therefore, shape migration as a potential livelihood strategy (Adamo, 2010). 12

28 However, rural and urban areas do not exist in isolation. Rather, they are connected economically and socially. Residents in urban centers depend on rural ecological services (e.g., food production) while rural populations often require urban goods and services (e.g., access to markets to sell goods, purchase of factory made fertilizers or machinery) (Satterthwaite et al., 2007). Due to these connections and interdependencies, the effects of climate change on the rural agricultural sector will have indirect impacts on urban economies. Using general equilibrium models, Boyd and Ibarraran (2009: p. 388) confirm this ripple effect of rural agriculture on urban non-agriculture production sectors. Their models reveal that drought conditions can result in food and electricity price increases, followed by productivity slowdowns in in the urban manufacturing, chemicals, and refining sectors. Migration as Self Insurance Mechanism: New Economics of Labor Migration While the SL framework constitutes an excellent tool for understanding the sociodemographic context in which livelihoods are embedded, it provides only limited information on migration as the livelihood strategy of particular interest in this study. The New Economics of Labor Migration (NELM) theory is well suited to expand the conceptual understanding of the climate change migration association begun with Sustainable Livelihoods. NELM assumes that migration decisions are made jointly by the family unit (Massey et al., 1993; Taylor, 1999), while the household shares costs and benefits of migration, according to an implicit contract between the migrant and the family members left behind (Sana and Massey, 2005; Stark and Bloom, 1985). All household members pool resources, and sometimes borrow from neighbors to finance an expensive move. In turn, the family members remaining in the household of origin expect that migrants will remit income. The migrating household member is 13

29 often strategically placed in a destination where market conditions are only weakly correlated or uncorrelated to the conditions at home (Massey et al., 1993; Stark and Bloom, 1985). International destinations are ideal because of the frequently substantial difference in politicaleconomic and climatic conditions in distant locations. In this way, remittance income is largely independent of local market conditions (Rosenzweig and Stark, 1989). Therefore, sending a migrant functions as a household-based informal insurance mechanism against market failures that may result from the impacts of adverse climate change on the agricultural sector (Liverman, 1990, Adamo and de Sherbinin, 2011). This risk management strategy is of particular importance for poor families residing in agricultural-dependent regions with limited access to formal financial and insurance markets (Lucas and Stark, 1985; Stark and Levhari, 1982). Furthermore, the ability of households to self-insure through migration lowers barriers to investment in agricultural production (Lindstrom and Lauster, 2001) and thus may foster technological adaptation to potential environmental changes. For example, migrant remittances may be used to purchase irrigation systems or drought resistant seeds that increase climate change resilience (c.f., Massey et al., 1993). The insurance aspect of having a migrant family member is evident from an increased level of remittances in response to crop income shocks among households in the Kayes areas of Mali (Gubert, 2002) and in the Philippines (Yang and Choi, 2007). These case studies demonstrate that migrants remain connected to the household and, in crisis situations, can function as guarantor of a more stable livelihood through remittances. Prior Studies Investigating the Climate Change-Migration Association 14

30 Research from all major continents shows that climatic variability and weather extremes are associated with outmigration, although the observed relationships tend to vary across time and place. Indeed, diverse archaeological, palaeo-ecological, and palaeo-anthropological evidence suggests that environmental and climatic changes have impacted human settlement and migration patterns in various parts of the globe throughout history (McLeman and Smit, 2006). For example, climate change around 2550 BP caused a rise in the water table and inundation in the Netherlands, resulting in loss of cultivated land and subsequent migration and colonization of the salt marshes in the northern part of the country (van Geel, Buurman and Waterbolk, 1996). Changes in rainfall patterns were a significant factor in promoting southward migration of Sotho- Tswana speaking people from equatorial East Africa during the first few centuries of the last millennium (Tyson et al., 2002). A shift in the monsoon rainbelt impacted migration patterns, allowing them to either penetrate deeper into the Sahara in times of abundance or forcing retreat in search of water and pasture during arid periods (Brooks et al., 2005). As another example, in ancient China, both livestock losses and crop failure during cold and/or dry climatic periods were associated with southward and eastward migration of people in the southern Mongolian grasslands and eastern central Asia (Fang and Liu, 1992). Finally, evidence from the American continent suggests that climatic factors were important determinants of seasonal settlement patterns of the Lakota Indians of the North American Plains prior to European settlement (Fixico, 2003). Similar evidence for the impact of climate and weather on migration patterns can be found in the literature in more recent years. The literature review below is limited to developing settings, given the analytical focus of this study. It is further organized by geographic region. 15

31 Africa. In one of the earliest quantitative studies of the migration-environment connection, Findley (1994) conducted research in Mali and demonstrated a large increase in the migration of women and children during the severe drought. Much of this migration was circular and short-term migration within Mali. On the other hand, international long-term migration (mostly to France) declined, perhaps due to a lack of resources and the challenges of a long planning horizon required for an international move. A decade later, Henry, Schoumaker, and Beauchemin (2004) observed similar migration dynamics for Burkina Faso. Using largescale survey data, they found that, among males, a long-term decline in rainfall significantly increased rural-rural migration but deterred international moves. In contrast, female mobility, both domestic and international, was unrelated to weather patterns. More recently, Gray and Mueller (2012a) investigated the effects of droughts on population mobility for rural Ethiopian highlands. They found that an increase in drought conditions propelled men s long-distance mobility and labor migration. Landless households were especially likely to respond to drought through migration. Overall, these studies suggest that rainfall shortages, including droughts, tend to be associated with local/domestic migration but not international moves. Moreover, the climate related migration response differs across both individual and household socioeconomic characteristics including gender and wealth. However, a recent examination of two rural communities in Ghana s forest-savannah transition zone found no significant relationship between climate-related factors (rainfall irregularity and bush fires) and the intention to migrate, although households considered these climatic factors as the most pressing stressors (Abu, Codjoe and Sward, 2014). 16

32 Asia: Only a few studies investigate the climate-migration association in Asia, most of which were conducted within the European Commission co-sponsored Environmental Change and Forced Migration Scenarios (EACH-FOR) project (Warner, 2011). Drawing mostly on qualitative research these studies demonstrate that temporary migration is a vital coping strategy under drought conditions in Orissa, India (Juelich, 2011) as well as under intensified flooding in the Mekong Delta, Vietnam (Dun, 2011). A recent large-scale quantitative investigation demonstrated that drought-related crop failure was significantly associated with an increase in both local and long-distance mobility from rural Bangladesh (Gray and Mueller, 2012b). Overall, these studies suggest that both flooding and drought may increase migration both locally and across long-distances. Latin America: Few published studies have explored the impact of climatic variability on migration from Mexico to the U.S. Using rainfall in the origin community as an instrument for the size of the network at the U.S. destination, Munshi (2003) found that rainfall deficits increased international outmigration from Mexico. This association was confirmed at the statelevel by Feng, Krueger, and Oppenheimer (2010) who found climate-driven changes in crop yields related to a significant increase in emigration during the period of 1995 and 2005 in 16 rural Mexican states (Feng and Oppenheimer, 2012). More recently, Hunter, Murray and Riosmena (2013) employed a categorical measure for long-term changes in rainfall patterns and found that households subjected to drought conditions were far more likely to send a migrant compared to those subjected to wet conditions, but only in regions with strong migration histories, especially in households with prior migration experience. Another study used a 12-year period to investigate the impact of severe drought conditions on migration (Nawrotzki, Riosmena, and Hunter, 2013) during the period and confirmed the general pattern of 17

33 an increase in out-migration under conditions of rainfall decline, but further noted that this impact is strongest for the arid northern states during the crop-growth period of corn. In addition, research using migration data from the latest census round (2010) stressed the importance of community-based migrant networks in facilitating international moves in response to climate variability (Riosmena, Nawrotzki, and Hunter, 2014). Similar patterns emerge for other Latin American countries such as Ecuador (Gray, 2009, 2010) and El Salvador (Halliday, 2006). For example, in a recent study of rural Ecuador, Gray and Bilsborrow (2013) found that a decline in rainfall, relative to a 25-year long-term average, increased international outmigration. The rainfall effect on international migration was consistent across subpopulations, disaggregated by gender and farm size. However, for domestic migration the results were mixed. A decline in precipitation increased short distance but decreased long distance domestic moves. Overall, these studies suggest that in countries of Latin America international migration is a commonly employed strategy in response to changes in climatic conditions. Most likely the commonly observed association between rainfall decline and migration is indirect in nature (Middleton and Thomas, 1997). Many rural households depend on local land and water resources to subsist -- both as related to agricultural output for consumption and for income generation. Climate change associated shifts in rainfall and temperature patterns may reduce productivity, yielding adverse impacts on rural sustenance and incomes. The inability to farm or subsist on the land often results in rural unemployment and general poverty, leading to migration (de Janvry et al., 1997; Middleton and Thomas, 1997). Overall, the above literature review has shown that the relationship between climatic factors and human mobility has a millennia long history. Moreover, evidence from all continents 18

34 demonstrates that climatic conditions continue to be associated with migration patterns in recent years. However, the relationship between specific environmental challenges and change and migration is different across space and time. While on the African continent climate factors appear to motivate largely short-distance circular moves, international mobility is frequently found in countries of Latin America. The Study Region: Mexico Mexico is uniquely positioned for the study of the association between climate change and migration due to Mexico s established history of high levels of migration, and the vulnerability of various economic sectors and related livelihood strategies to climate change. Mexican Migration: Historical Trends and Current Patterns International Migration. International migration from Mexico to the U.S. is a centurylong tradition. The first substantial flows of Mexican migration to the U.S. began in the early 1900s mostly from rural areas of western Mexico when labor recruiters sought railroad and agricultural labor (Fussell, 2004). Migration flows continued to grow early in the twentieth century in response to a mix of push and pull factors associated with U.S employer recruitment efforts and the Mexican Revolution (Cardoso, 1980). This initial surge in migration flows abated during the Great Depression (Hoffman, 1974) but regained momentum in 1942 in response to the Bracero Program, a bi-national labor accord aimed at providing Mexican farm labor to the U.S. during World War II (Calavita, 1992). The Bracero Program was discontinued in 1964 as part of broader civil rights and immigration reforms (Calavita, 1992). Despite the lack of official program support, both documented and especially undocumented migration continued 19

35 (Cornelius, 1992; Massey, Durand, and Malone, 2002). Even under conditions of increased border militarization, the number of Mexicans in the U.S. increased by 450 % between 1980 and 2000 and the share of the immigrant population continued to expand (Massey and Capoferro, 2004). A factor that contributed to this massive increase in migration flows was the establishment of the North American Free Trade Agreement (NAFTA) of 1994 (Sanchez Cohen et al., 2012). NAFTA contributed to a decline in job opportunities in the agriculture sector since small-scale Mexican farmers were unable to compete with U.S. and Canadian agricultural imports and frequently reverted to undocumented migration to find employment in the U.S. (Fussell, 2004). In contrast, an increase in documented migration resulted from the 1986 Immigration Reform and Control Act (IRCA), which provided a relatively easy path to U.S. citizenship for large numbers of undocumented migrants already residing within the U.S. (LoBreglio, 2004). The Special Agriculture Worker (SAW) and Replenishment Agricultural Workers (RAW) sub-components of IRCA have likely led to a large increase in the number of migrants due to a program design that encouraged fraudulent claims to obtain legal permanent resident status (Martin, 1990). IRCA brought about some other changes in immigration law intended to deter immigration from Mexico, including the penalization of employing undocumented migrants and a substantial increase in the border patrol budget and programs (Orrenius and Zavodny, 2003). Although the majority of Mexican migrants to the U.S. have traditionally come from rural regions, urban areas (especially small cities) have increasingly contributed to the migration flow (Fussell and Massey, 2004). Through this unique century-long history of migration, a dense migrant network has been developed. These networks operate as migration corridors (Bardsley 20

36 and Hugo, 2010, p. 249), which may facilitate future migration under conditions of declining livelihoods due to factors such as climate change (Adamo and de Sherbinin, 2011). Domestic migration. During most of Mexico s contemporary history there was little migration of population from one state to another or even between communities (Whetten and Burnight, 1956). 1 However, beginning with the Mexican revolution of 1910, migration increased. Civil war and political turmoil motivated migration and the final abolition of the haciendas (a colonial system of large land holdings) released large numbers of workers to seek livelihoods elsewhere (Whetten and Burnight, 1956). Over the following decades two distinct domestic migration patterns developed rural-urban migration and step migration. As an example of rural-urban migration flows, the rural population of Oaxaca became heavily engaged in migration to Mexico City during the 1990s. Men relocated for the prospect of finding jobs in the construction and service industry while women frequently sought employment as domestic helpers and maids (Cohen, 2004). This migration stream dates back to the early 1960s fueled by government-sponsored industrial development in urban centers such as Mexico City, Guadalajara, and Monterrey (see Fussell, 2004, and references therein). In recent decades, domestic migration streams have become more diverse and include moves to coastal cities for employment in the tourist industry (Fussell, 2004) and migration to urban centers to obtain higher education at colleges and universities (Cohen, 2004). 1 Three reasons have been proposed by Whetten and Burnight (1956) for this lack of mobility: First, until 1910, the hacienda system (colonial system of large land grants) tended to dominate the social and economic life of the nation and in most cases the hacienda population was virtually bound to the soil as indentured labor. Second, the Mexican economy operated near the subsistence level with few employment opportunities that would have attracted workers from one area to another. Third, channels of communication were underdeveloped, which prevented the spread of information regarding potential opportunities elsewhere. In addition, the lack of infrastructure (e.g., paved roads and highways) and transportation technology (e.g., cars and bus services) posed a major obstacle to traveling beyond one s community (Cohen, 2004). If migration occurred, it was marked by short-term seasonal moves for work in plantations, fields, and road and construction projects (Cohen, 2004). 21

37 With regard to step migration, research has demonstrated that domestic migration is frequently a means of entering the broader international migration stream between Mexico and the U.S. (Davis, Stecklov, and Winters, 2002). Individuals may migrate to border states to gain access to migration-related information and contacts, such as migrant smugglers (coyotes) or career migrants (Orrenius, 2001). It has been shown that domestic migrants to the border from interior communities are 67% more likely to engage in undocumented migration to the U.S. compared to those who have never made a trip to the border (Fussell, 2004). Overall studies from various parts of Mexico demonstrate that international migration to the U.S. as well as domestic migration are common phenomena (Lindstrom and Lauster, 2001; Cohen, 2004). However, both streams show differences in demographic make-up. For example, while lower educated individuals are more likely to migrate internationally, higher educated Mexicans more frequently migrate domestically (Taylor et al., 1996). While the international migration stream is male dominated, domestic migration is much more gender balanced (Cohen, 2004). 2 Climate Change Context and Vulnerability For an informed investigation of the impact of climate change on migration patterns, it is important to understand the climatic context of Mexico and the vulnerability of various economic sectors and associated livelihood strategies to changes in weather patterns. The Climatic Context of Mexico. The country of Mexico covers an area of around 2 million square kilometers with varying climatic zones. Overall, two distinct seasons can be identified: a warm, rainy season (May to October) and a cold, dry season (November to April) 2 A detailed description of the sociodemographic characteristics of Mexican migrants is provided in the variable construction section of the Data and Methods chapter. 22

38 (Schwartz, 1977; Pearce and Smith, 1990). The latter part of the warm, rainy season is governed by the Mexican monsoon, but there is a relatively drier period between July and August known as the sequia intraestival (mid-season drought) or canicula (August drought) that varies in intensity and timing from year to year (Endfield, 2007; Marty, 1992). However, there are large geographical variations in climatic conditions. The northern states from the U.S. border to Mexico City are characterized by a semiarid dry to very dry climate throughout the year. Instead, the central-western part of the country experiences a temperate sub humid climate, while the southern coasts and the Yucatan Peninsula are warm and sub humid. Warm humid conditions prevail in the area west of the Yucatan Peninsula to the south central interior (Boyd and Ibarraran, 2009; Marty, 1992). The trade winds deliver summer rainfall to the central and southern regions of the country when the Intertropical Convergence Zone (ITCZ) shifts northward, and the Sub-Tropical High Pressure belt brings stable dry conditions to the country during the winter when the ITCZ is displaced in the direction of the equator. However, the northern part of the country is little affected by the summer trade winds (Endfield, 2007). Inter-annual climate variability is frequently due to El Niño Southern Oscillation (ENSO) events. During El Niño winters, precipitation increases over northwestern Mexico, while it decreases in the central region around the Isthmus of Tehuantepec (e.g., southeastern parts of Veracruz and Oaxaca). In contrast, during El Niño summers, negative precipitation anomalies dominate over most of Mexico, as experienced during the years of (Magana et al., 2003). The various climatological mechanisms associated with El Niño precipitation decline include a shift in the ITZC, more intense trade winds, a decreased number of tropical cyclones over the Intra Americas Seas (IAS), and reduced relative humidity (Magana et al., 2003). 23

39 Knowledge of these climatic patterns is important for the present analysis and provides relevant information for the investigation of temporal and spatial variation in the weather-migration association. Figure 2.2 below shows climatic zones derived from a modified Koeppen classification for Mexico. Figure 2.2: Climatic zones across Mexico derived from a modified Koeppen classification Note: Panel (a) displays humidity classification while panel (b) shows temperature classifications. Source: data were obtained from INEGI (2000). Topographic features also impact Mexico s climate (Schwartz, 1977). The Mexican plateau (Altiplano), with an average elevation of 1,500 meters above sea-level, is flanked on the western edge by the Sierra Madre Occidental and on the east by the Sierra Madre Oriental. These two mountain ranges (Cordilleras) can trap stable air masses, deflect winds, and lead to distinct temperature and precipitation patterns (Endfield, 2007; Marty, 1992). For example, adiabatic cooling of air as it ascends over the mountain ranges leads to the release of moisture in the form of heavy rains over the mountain slopes. In addition, frosts are relatively common at high 24

40 elevations, particularly between October and April and can threaten crop yields even during the summer months (Eakin, 2005, p. 1926). Since topography has important influences on weather patterns, data on elevation are used in the construction of the climate change measures. Projected Climate Change. The newly released Fifth Assessment Report of the Intergovernmental Panel on Climate Change suggests that anthropogenic climate change will have major impacts on temperature and precipitation regimes for North and Central America (IPCC, 2013). Results from an ensemble run of 39 climate models within the Coupled Model Intercomparison Project (CMIP5), using a relative optimistic Representative Concentration Pathway with a radiative forcing of 4.5 W m -2 (RCP 4.5), suggest that over the 21st century precipitation will decline across Mexico (Christensen et al., 2013) while temperatures will increase (Collins et al., 2013). In an earlier study, Wehner et al. (2011) have used an ensemble of 19 models (CMIP3) to study the effect of climate change on droughts in Mexico and the continental United States under the A1B emission scenario. They also found an overall increase in future drought frequency and severity, and the drought of might foreshadow these changes (Stahle et al., 2009). Thus, the investigation of the climate change migration association has particular relevance for Mexico. Climate Change in Rural Areas. Although most rural Mexican households do not rely entirely on agriculture, income from farming activities contributes in important ways to sustenance and livelihood portfolios (Wiggins et al., 2002; Winters, Davis, and Corral, 2002; Conde, Ferrer, and Orozco, 2006). For rural Mexicans, agriculture contributes between 23% and 67% of household income depending on the size of land holdings (de Janvry and Sadoulet, 2001). This reliance on agriculture makes rural Mexicans vulnerable to climatic shifts and resulting adverse impacts on crop production (Eakin, 2000, 2005; Eakin and Appendini, 2008; 25

41 Endfield, 2007; Saldana-Zorrilla and Sanberg, 2009; Schroth et al., 2009; Vasquez-Leon, West, and Finan, 2003). Climate/weather events were responsible for approximately 80% of economic losses in Mexico between 1980 and 2005, mostly due to adverse impacts on the agricultural sector (Saldana-Zorrilla and Sanberg, 2009). The sensitivity of the agriculture sector to climate variability can be partially attributed to lower levels of infrastructure designed to mitigate the impacts of environmental stresses (Endfield, 2007). For example, only 23.15% of the arable and permanently-cropped land was irrigated in 2001 (Carr, Lopez, and Bilsborrow, 2009). Given this sensitivity of the agricultural sector and the dependence of rural households on agricultural production for sustenance and income generation, climate change effects on crop production will likely impact rural livelihoods. Climate Change in Urban Areas. Due to a limited amount of publications on the topic, the below discussion of climate impacts on Mexican cities also incorporates studies from other countries. Overall, climate change will impact urban areas largely through extreme weather events such as storms, floods and droughts (IPCC, 2014; Revi et al., 2014). Climate change related-droughts may influence urban livelihoods in multiple ways such as intensified water shortages as well as food insecurity due to reduced supplies from the surrounding rural areas (Revi et al., 2014). For example, Mexico City has already had to import around a third of its raw water from neighboring river basins (Connolly, 1999) and prolonged droughts can severely limit the access to these water sources with adverse impacts on residents and businesses alike (Satterthwaite et al., 2007; Romero-Lankao, 2010). In addition, droughts may reduce the energy supply and increase energy prices if hydro power is used as an energy source (Pereira de Lucena et al., 2009; Schaeffer et al., 2012). 26

42 With regard to temperature effects, changes in mean temperatures and extremes may alter patterns of urban energy consumption for cooling and heating (see Mideksa and Kallbekken, 2010, for a review), making industrial production more expensive and potentially impacting job availabilities. In addition, an increase in temperature may adversely impact urban livelihoods through various negative health outcomes (Burkart et al., 2011; WHO and WMO, 2012). Warming may worsen urban air quality through an increase in ozone concentration (Jacob and Winner, 2009; Weaver et al., 2009) and may disrupt or limit income-earning opportunities (Kovats and Akhtar, 2008). These adverse impacts are intensified by the urban heat island effect, which often exacerbate warming conditions (Wilby, 2007; Adachi et al., 2012). The urban poor are strongest impacted by adverse climate change due to a lack of resources to successful implement adaptation and coping strategies(satterthwaite et al., 2007; Kovats and Akhtar, 2008). A climate related increase in storm surges and heavy precipitation will increase the likelihood of flooding in urban areas (IPCC, 2014), which will affect urban infrastructure such as buildings, bridges, roads, pipelines, wire networks, and telecommunication systems on which many business and economic activities depend (Hallegatte et al., 2010; Fussell, Sastry, and Van Landingham, 2010). Flooding is particularly problematic when drainage systems are underdeveloped and sewer and wastewater systems are impacted, leading to contamination of freshwater sources (Romero-Lankao, 2010; Revi, 2008; Willems et al., 2012). In addition, floods and storms have profound impacts on the transport sector with problematic results for various dependent economic activities (Gasper, Blohm and Ruth, 2011; Koetse and Rietveld, 2009). As an example, a major flood in 2007 seriously impacted the city of Villahermosa (located in the Mexican state of Tabasco), leading to water related losses and damages to infrastructure and various economic activities that amounted to 30% of Tabasco s annual GDP (CEPAL, 2008). 27

43 Finally, the urban tourism sector is sensitive to climate change and income and employment opportunities will decline if droughts and storms alter the attractiveness of destinations (Amelung, Nicholls, and Viner, 2007; Lise and Tol, 2002). In summary, climate change will likely impact rural livelihoods through the agricultural sector and will impact urban livelihoods through a broad spectrum of city functions, infrastructure, services, and related employment opportunities. Resulting livelihood insecurities may lead to migration responses both in rural (e.g., Feng and Opppenheimer, 2012) and urban areas (Adamo, 2010). However, it is important to stress that urban and rural livelihoods are connected and climate impacts on rural sites of natural resource extraction and agricultural production may indirectly influence well-being in urban centers (Boyd and Ibarraran, 2009; Wackernagel et al., 2006). The preceding discussion illustrates the uniqueness of Mexico as a case for the study of the climate change-migration association due to the relative sensitivity of both rural and urban areas to climate change and a long-standing history of mobility and established social networks that may make relocations more sensitive to environmental factors. Scope of the Study The collection of results from available studies suggests that climate and weather events indeed impact migration patterns. However, many questions remain unanswered and this study aims to contribute new insights on the climate change-migration connection with a focus on both the driver (climate change) and outcome (migration). Driver. As to climate change as a migration driver, most existing research has used rather crude measures of climate and weather in terms of spatial and temporal resolution. For example, many measures reflect conditions at a broad regional or state scale, while also reflecting average 28

44 conditions over the course of a year or longer periods of time (e.g., Hunter, Murray, and Riosmena, 2013). Moreover, none of these studies have employed direct measures of climate change, impeding the interdisciplinary dialog between the demographic and climate change community (Hunter and O Neill, 2014). Specifically, existing research tends to use singular measures of temperature or rainfall but lack the detail and nuance allowing for linkages to broader shifts in climate patterns. Overall, rainfall has been the main focus of most research on migration-environment connections, while temperature effects have been largely omitted (see Poston et al., 2009, for an exception). This is unfortunate, since compared to precipitation, temperature effects have been shown to stronger impact yields of certain crop types (e.g., maize, wheat) (Lobell and Field, 2007), which are highly important staple crops in many countries including Mexico. In response to these shortcomings, I employ a suite of 27 climate change indices (Alexander et al., 2006) that are widely used by the climate change research community (c.f., Bindoff et al., 2013). These indices capture nuanced differences in long term climate patterns at the finest spatial scale currently possible (daily temperature and precipitation data at the municipality level). Such details allow exploration of questions such as: Which types of climate change (e.g., dry spell duration vs. number of summer days) are most strongly associated with migration? What are the directionalities of the temperature, precipitation, migration associations? Outcome. As to the outcome -- migration -- most of the published research has focused on rural communities (e.g., Hunter, Murray, and Riosmena, 2013). Based on the assumption that the impact of climate change on migration largely operates through adverse impacts on the agricultural sector, this assumption is quite plausible. However, non-agricultural sectors such as manufacturing and the service sector may also be directly or indirectly impacted by climate 29

45 change (e.g., Boyd and Ibarraran, 2009), influencing livelihood security of residents in urban areas (Adamo, 2010). In urban areas the most important climate change impacts will likely constitute weather extreme events such as droughts, storms, and floods (Revi et al., 2014). As such, this project is also designed to answer the question: Does climate change differentially impact migration from urban and rural areas? Yet another unsolved puzzle in the available literature is the fact that for some regions (e.g., Africa) environmental factors appear to differentially drive domestic migration (Henry, Schoumaker, and Beauchemin, 2004), while for other regions (e.g., Latin America) international migration demonstrates a stronger association with environmental factors (Feng and Oppenheimer, 2012). Focusing on the case of Mexico, no published research has attempted to contrast domestic and international migration probabilities in response to climate change. 3 In this study, I therefore ask: Does climate change differentially impact domestic and international migration? Finally, a few studies have stressed the importance of social capital and networks for the climate change migration association (e.g., Bardsley and Hugo, 2010; Riosmena, Nawrotzki, and Hunter, 2014). After the first move, households possess a personalized type of social capital, developed through the migration experience and the connections forged in destination areas (e.g., Fussell, 2004). However, it remains to be explored whether this internal social capital, available during later moves, facilitates climate change related migration. This study examines this possibility through an analytical focus contrasting first and last moves from a household -- with the last move presumably occurring during a period in which the household has access to social 3 Some ongoing research presently under review (Runfola et al., 2014) employed municipality-level precipitation data in combination with migration information from the 2010 census round to investigate the impact of climate variability on domestic and international migration. This study finds that a decline in in rainfall is associated with elevated levels of domestic and international migration particularly in municipalities with high dependence on rain-fed agriculture (<50% crop land irrigated). 30

46 capital related to prior migration(s). Therefore, the final research question explored in this study is: Does climate change differentially impact first and last moves? In sum, previous research suggests that a rainfall shortage and other forms of environmental strain may influence migration patterns. This project advances knowledge due to its use of detailed climate change information to investigate the impact of climate change on difference types of migration by contrasting moves from rural vs. urban origin areas, international vs. domestic migration, and first vs. last move. The results from this investigation will be used to refine the theoretical framework by explicitly linking certain forms of climate change to different migration streams. 31

47 CHAPTER III DATA AND METHODS Data To investigate the climate change-migration association, I combined sociodemographic data from the Mexican Migration Project (MMP) 4 with climate information obtained from the Global Historical Climatology Network Daily data set (Menne et al., 2012). The analysis was conducted with households residing in rural and urban areas of Mexico during the period Multilevel event history models allow estimating the impact of climate change on four distinct migration flows (first international move, first domestic move, last international move, last domestic move) at the household level. Demographic Data Some studies on the association between climate change and migration have used census data (Feng, Kruger, and Oppenheimer, 2010; Nawrotzki, Riosmena, and Hunter, 2013). Drawbacks of currently available census data are that they do not allow for the differentiation between migration types such as first and last migration, are cross-sectional in nature, and have coarse temporal resolution. The ethnosurvey design, developed by Massey and others (Massey, 1987a), has been proposed as a viable alternative to traditional censuses. The ethnosurvey blends anthropological and survey research methods to conduct in-depth surveys of a random sample of households within specific communities non-randomly chosen to represent an array of sizes and patterns of social and economic origins. A key component of the ethnosurvey is the gathering of 4 The Mexican Migration Project (MMP) is a collaborative research project based at Princeton University and the University of Guadalajara. The MMP provides high quality public data at 32

48 detailed life histories, which allow for the construction of retrospective event histories that can be used in longitudinal analysis (Massey and Capoferro, 2004). Mexican Migration Project. The Mexican Migration Project (MMP) constitutes the first ethnosurvey and began data collection in 1982 (Massey and Capoferro, 2004). Every year, the MMP selects between 4 and 6 Mexican communities and interviews a simple random sample of approximately 200 households in each community. To date, the MMP has surveyed 143 Mexican communities in 24 states. It is important to stress that the MMP does not yield a probability sample of Mexico because communities are not randomly selected. However, Massey and Zenteno (2000) and Massey and Capoferro (2004) used data from Mexico s National Survey of Population Dynamics (ENADID by its Spanish acronym) to validate the accuracy of the MMP and found that the MMP very accurately captured the characteristics (e.g., gender, age, marital status, education) and behavior (e.g., trip duration) of international migrants. Although for domestic migrants such a validation exercise has not been performed, the MMP has been used for research on domestic migration that has been published in widely respected journals such as Demography (Curran and Rivero-Fuentes, 2003) and Social Forces (Fussell, 2004). Due to challenges with obtaining comparable data, validation was beyond the scope of this project. Climate Data Climate data were selected from a dataset known as Global Historical Climatology Network Daily (GHCN-D) (version number: 2.93-upd ), compiled and made publically available by the National Oceanic and Atmospheric Administration (NOAA). The data are provided pre-compiled at a daily time resolution, which allows the assessment of climate 33

49 change-associated phenomena such as the frequency of heavy rainfall or heat wave durations based on measures of maximum and minimum daily temperature and total daily precipitation (Menne et al., 2012). NOAA routinely applies rigorous multi-tiered quality assurance checks to the full dataset to guarantee the highest possible levels of data integrity (see Menne et al., 2012 for a detailed description). GHCN-D has been used for climate monitoring and change assessments in a wide range of prior published work (e.g., Alexander et al., 2006; Caesar, Alexander, and Vose, 2006). Non-Climatologic Spatial Data In addition to the two core data sources described above, a number of spatial data layers are integrated to construct important control variables, such as distance to the U.S.-Mexico border, or the density of road networks. These data layers are described in depth in the variable construction section below. Demographic Unit of Analysis Although migration ultimately occurs at the individual or person level, in developing countries the decision to stay or go is typically reached within some larger family or household unit, a key assumption of various migration theories including the New Economics of Labor Migration theory (Massey et al., 1993; Taylor, 1999). Research suggests that within the Mexican culture, the household is the fundamental unit through which individuals create a sense of identity and belonging within communities and through which status and prestige are obtained (Cohen, 2004). It is against this backdrop that migration is considered embedded in a household s needs, desires, and aspirations, rather than as a strategy employed by an individual 34

50 removed from his/her social context (Kanaiaupuni, 2000). As has been described for the case of rural Oaxaca, Mexico, households pool resources and frequently borrow money from extended family members or friends to finance the move of a household member (Cohen, 2004, Chapter 2). In this way, migration is a strategy that allows the household to access new work opportunities to supplement other livelihood activities (Cohen, 2004). Based on these considerations, the analysis is conducted at the household level in line with prior studies of Mexican migration (de Janvry et al., 1997; Hunter, Murray, and Riosmena, 2013; Kanaiaupuni, 2000). Spatial Unit of Analysis Within the MMP data, households are nested within communities. Communities in turn are located in municipalities. With special permission, the MMP administration provided information on the location of communities in specific municipalities. Spatial location of municipalities is available from a shapefile provided by INEGI (INEGI, 2012), which enabled me to link the spatial climate data and spatial controls to the sociodemographic MMP data at the municipality scale, making the municipality level the spatial unit of analysis. In most cases communities are located in separate municipalities but in some cases more than one community is located in a given municipality. Figure 3.1 displays the geographic location of the 111 municipalities that contain the 138 MMP communities for which sociodemographic data are available. Of these municipalities, 62 contain rural communities and 43 contain urban communities, while an additional six municipalities contain both urban and rural communities. 35

51 Figure 3.1: Geographic location of MMP municipalities and spatial distribution of weather stations across Mexico Note: Total municipalities: n=111; rural municipalities: n=62; urban municipalities: n=43; municipalities containing both rural and urban municipalities n=6 Time Frame The time period under investigation is This time frame was chosen for the following methodological and theoretical reasons: (1) As outlined above, the 1986 Immigration Reform and Control Act (IRCA) had major impacts on the policy context of Mexico-U.S. migration, and thus only the post-irca period will be investigated. (2) The number of weather stations available in the GHCN-D data set drops from an average of n=182 for the years 1961 to 1998 to n=15 after 1998, due to the historical focus of this data set, rendering interpolation methods (see details below) for years after 1998 unreliable. Since climate change in the year t-1 36

52 (lagged by one year) is used in the analysis to predict migration in the year t, climatic data of 1998 can be used to predict migration in (3) A change in macro-level factors has substantially altered the milieu of migration since the early 2000s. Several factors taken together, such as a growing anti-immigrant sentiment (Varsanyi, 2011), increasingly strict federal, state, and local immigration enforcement policies (Hanson, 2009), a much more robust border enforcement effort (Orrenius, 2004), 5 and the worsening U.S. economic climate (Papademetrious and Terrazas, 2009) have substantially changed the nature of Mexico-U.S. migration dynamics. Due to these reasons, I investigate the impact of climate change on migration patterns for the years , which restricts the generalizability of the findings to this period. Variable Construction Outcome Variables The Mexican Migration Project (MMP) defines migration as a move that involved a change in usual residence, excluding short trips for vacation, shopping, visits, and commuting (Fussell, 2004). This research investigates four distinct types of migration: first international move, first domestic move, last international move, and last domestic move. All four outcome variables were coded either 0 if no move occurred or 1 if the household experienced a move. For 5 Border enforcement has discouraged crossing into more heavily enforced states, and has resulted in increased likelihood of migrant injury and death during crossing attempts, and a spatial redistribution of illegal entry attempts between and within states (Orrenius, 2004). Some research suggests that border enforcement efforts impact the circularity of undocumented migration rather than the overall flow size (Massey, Durand, and Malone, 2002; Massey and Riosmena, 2010). If migration costs at some point outweigh the benefits, a decline of migration or a spatial diversion to other destinations can be expected (de Haas, 2010). Hence, border enforcement reduces the number of migrants entering the U.S. but also leads to longer stays to recoup the higher entry costs (Angelucci, 2012). However, border enforcement has more complex effects on migration dynamics, and it has been shown that (1) migrants from traditional sending states are more sensitive to enforcement than migrants from new emigration states, (2) that as the level of border controls increases, its marginal deterrent effect becomes higher, and (3) that under conditions of higher enforcement, migrant selectivity increases, and the better-educated are more likely to make a move (Angelucci, 2012). Combined, these evidences suggest a changed nature in migration dynamics in the post-2000 era. 37

53 the first migration the earliest year was selected for a given household when different years for the first move of various household members were reported. When different years for a last move were reported for various household members, the latest year was chosen to represent the last move for the household unit. For the last move, only moves were considered when another move had occurred from the household in an earlier year. This treatment clearly distinguishes first from last moves and allows evaluating the effect of internal social capital on the odds of migration. 6 This analytical focus is logical given this project's focus, yet, the MMP only reports the first and the last move and it remains unclear how climate change may differently impact intermediate (between first and last) moves. For domestic migration, only moves to a municipality other than the present residence (out-migration) were used, to purge the measure of in-migrants to the present municipality. For the discrete event history models (see details in Estimation Strategy) data must be structured as a household-period file. In this format, each row represents a time period during the observation window ( ) in which the household either experienced or did not experience a migration event. Censoring determines the structure and composition of the data sets used for this analysis. I employ data for all communities that were surveyed after 1986 in the sample. Right censoring (Allison, 1984) occurred for communities surveyed before the year However, it is not recommended to remove right censored cases from the analysis because this might introduce 6 The first and the last move may be differentially impacted by climate change. Through the first move a household gains internal social (networks at destination) and human (knowledge how to cross the border) capital (Fussell, 2004). Some research suggests that for households with a prior migration experience, internal social and human capital becomes the primary predictor of a move and the influence of external conditions declines (de Janvry et al., 1997). However, the established internal social capital may also fulfill the function of migrant corridors (Bardsley and Hugo, 2010) and thus increase the sensitivity of migration to external factors (c.f., Hunter, Murray, and Riosmena, 2013). It is therefore unclear how internal social and human capital impacts the climate-change migration association, a topic investigated in Chapter VII. 38

54 substantial bias (Box-Steffensmeier and Jones, 2004; Steele, 2005), particularly because the MMP communities were not selected randomly. I retain right censored cases in the analysis based on the assumption that the censoring is non-informative, meaning that the time of migration is independent of the time a particular community was surveyed (Allison, 1984; Steele, 2005). Domestic and international migration are not mutually exclusive events and leftcensoring requires the construction of separate data sets. Left censoring arises when a household experiences the event before the start of the observation period and a common practice is to remove these households from the analysis (Allison, 1984; Steele, 2005). For example, if a hypothetical household experienced a first international migration event in 1974, then this household needs to be omitted from the data set when investigating the odds of international migration because it was not at risk for a first international move during the study window (left censoring). However, the same household was at risk of domestic migration until 1995 when the household experienced its first domestic move and should therefore be included in the data set of investigating the determinants of domestic migration. Having slightly different samples is not ideal but a situation that commonly emerges when investigating different migration types (c.f., Curran and Rivero-Fuentes, 2003). Households that are highly responsive to climate change may have already conducted a move prior to the study window and are therefore excluded from the sample. Due to this exclusion, the sample contains individuals less prone to migrate, resulting in more stringent statistical tests that may underestimate the true population effects of climate change on migration. To reduce the number of records in the household-period file to a computationally manageable size, I employ a common practice and extended the observation intervals from one- 39

55 year to two-year periods (Allison, 1984; Steele, Goldstein, and Brown, 2004). Table 3.1 shows the households at risk of migration (rt) and the number of observed moves (migt) that occurred during the particular time interval for both rural and urban municipalities. From these numbers, I t computed the hazard rate ( h mig t r ) and the survival rate ( t 1 s t (1 h j ) ) (Steele, 2005). j 1 t Table 3.1: Life table describing the event of different migration types during the period for households of all MMP communities (rural and urban) Migration Type: First international First domestic Years rt migt ht st rt migt ht st 2-year periods Total year period Migration Type: Last international Last domestic Years rt migt ht st rt migt ht st 2-year periods Total year period Note: rt = households at risk during interval; migt = number of migrations during interval; ht = hazard of migration during interval; st = probability of survival up to the start of interval. 40

56 Table 3.1 shows that among first moves, there were more international (14.2%) compared to domestic (7.9%) migrants during the study period, Similarly, among last moves more households engaged in international (5.7%) compared to a domestic (2.4%) migration. The detailed life table shows adequately large migrant counts across all periods for first and last international moves as well as for the first domestic move, suggesting that the statistical tests for these migration streams are sufficiently powered. However, for last domestic moves only small numbers of migrant counts are available, which are further decreased when conducting analysis for rural and urban sub-samples. A preliminary analysis showed that these small numbers lead to unstable estimates and the last domestic move was therefore not further investigated. Figure 3.2 visually depicts the trajectory of the hazard of different migration types. The curves for first and last international migration show a dip in the migration hazard during the period of Lower levels of outmigration may be associated with good economic conditions in origin regions while high levels of outmigration may result from economic downturns. And indeed, the high levels of migration coincide with major economic crises in Mexico during the mid-1980s (Lustig, 1990) and during (McKenzie, 2006), which reportedly fueled out-migration. In addition to the macro-economic effect, U.S. immigration policy reforms likely impacted the number of migrants in the late 1980s. The Special Agriculture Worker (SAW) and Replenishment Agricultural Workers (RAW) subcomponents of IRCA strongly encouraged Mexico-U.S. migration (Martin, 1990). 41

57 Figure 3.2: Hazard functions for international and domestic migration from MMP communities during the years Note: int= international migration; dom= domestic migration In contrast, first domestic migration rates declined across the study period. Domestic migration may be less responsive to macro-economic conditions and U.S.-Mexico border policies. Perhaps the decline in domestic migration rates reflects the more rural character of the MMP sample and a deceleration of rural to urban migration due to industrial downturn and urban wage declines (c.f., Perz, 2000). Under conditions of economic crisis and economic restructuring (e.g., Sanchez Cohen et al., 2012), the incentive to migrate internationally may increase concurrently with a decline in the incentive to move to urban areas within Mexico. Primary Predictor Variables 42

58 The primary predictors in this research project are a set of 27 climate change indices, suggested by the Expert Team on Climate Change Detection and Indices (ETCCDI) 7 (Alexander et al., 2006; Peterson, 2005). ETCCDI developed the general formal representation of the indices, allowing investigators to construct these measures based on specific research project needs regarding time period and geographical focus. These indices were introduced in the IPCC Third Assessment Report (TAR) and focus primarily on climate extremes (Peterson et al., 2001). The indices have been widely employed in the climatological research community to investigate long term climatic trends (e.g., Alexander et al., 2006; Bindoff et al., 2013; Frich et al., 2002; Klein Tank et al., 2006). However, these indices have not been used for demographic studies of migration behavior, perhaps due to a lack of collaboration between the demographic and climate change communities (Hunter and O Neill, 2014), and because of the relative complicated way of constructing these measures, requiring daily temperature and precipitation data as well as an understanding of spatial modeling. The climate change indices can be grouped into measures of high and low temperature extremes (Table 3.2) and measures of high and low precipitation extremes (Table 3.3). A few measures that did not fit into this classification are listed as temperature & precipitation (other) in Table 3.3 and include also a measure of the average daily temperature and average precipitation for a particular year. The average daily precipitation measure is comparable to the average precipitation measures used in prior studies (Hunter, Murray, and Riosmena, 2013; Nawrotzki, Riosmena, and Hunter, 2013) but has higher spatial resolution, operating at the municipality level. 7 The Expert Team is part of the Climate Variability and Predictability (CLIVAR) project, which is jointly sponsored by the World Meteorological Organization s Commission for Climatology (CCl) and the World Climate Research Programme (WCRP). 43

59 44 Table 3.2: List of ETCCDI climate change indices measuring temperature extremes Indicator Name ID Indicator definition Unit Temperature (high) No. summer days su Annual count when daily max temperature > 25 C days Tropical nights tr Annual count when daily min temperature > 20 C days Warm spell duration wsdi Annual count when at least six consecutive days of max temperature > 90th percentile days Warmest day txx Annual max value of daily max temperature C Warmest night tnx Annual max value of daily min temperature C % warm nights tn90p Percentage of days per year when daily min temperature > 90th percentile % % warm days tx90p Percentage of days per year when daily max temperature > 90th percentile % Temperature (low) No. frost days fd Annual count when daily min temperature < 0 C days Ice days id Annual count when daily max temperature < 0 C days Cold spell duration csdi Annual count when at least six consecutive days of min temperature < 10th percentile days Coldest day txn Annual min value of daily max temperature C Coldest night tnn Annual min value of daily min temperature C % cool nights tn10p Percentage of days per year when daily min temperature < 10th percentile % % cool days tx10p Percentage of days per year when daily max temperature < 10th percentile % Note: Table adjusted based on Donat et al. (2013).

60 45 Table 3.3: List of ETCCDI climate change indices measuring precipitation extremes Indicator Name ID Indicator definition Unit Precipitation (high) No. days heavy precip r10mm Annual count of days when precip > 10mm days No. days very heavy precip r20mm Annual count of days when precip > 20mm days Wet spell duration cwd Max number of consecutive days with precip > 1mm days Max 1-day precip rx1day Annual max 1-day precip amount mm Max 5-day precip rx5day Annual max consecutive 5-day precip amount mm Precip very wet days r95ptot Annual total precip from days when precip > 95th percentile mm Precip extremely wet days r99ptot Annual total precip from days when precip > 99th percentile mm Total wet-day precip prcptot Annual total precip from days when precip > 1 mm mm Precip intensity index sdii The ratio of annual total precip to the number of wet-days (precip > 1mm) mm/day Precipitation (low) Dry spell duration cdd Max number of consecutive days when precip < 1mm days Temperature & Precipitation (other) Average precip aprec Average daily precipitation mm/day Average temperature atemp Average daily temperature C Temperature range dtr Annual mean difference between daily max and min temperature C Growing season length gsl Count between six day periods with daily mean temperature > 5 C and < 5 C days Note: The 27 ETCCDI climate change indices contain a precipitation index that is constructed similarly to r10mm and r20mm but lets the user define the precipitation threshold. This measure was not employed and only the remaining 26 indices were considered for the present analysis. Table adjusted based on Donat et al. (2013).

61 Using this set of climate change indices instead of relatively crude measures of average rainfall has many benefits. First of all these measures allow for the investigation of nuanced differences in climatic variability and change. Changes in climatic extremes such as the number of days with very heavy precipitation (r20mm) are likely to stronger impact the agricultural sector and resulting migration, as compared to the average annual precipitation. Second, use of these climate change indices allows bridging the divide between the demographic and climate change research communities and will allow more informed policy recommendations as to the impact of climate change on migration dynamics. However, although daily temperature and precipitation data are required to construct the ETCCDI indices, they represent annual aggregates -- meaning they are bounded by calendar years. The agricultural sector will be most sensitive to changes during certain months of the year (growing season) (Steduto et al., 2012) and, as such, more nuanced temporal effects are not captured. A seasonal analysis of climatic impacts is beyond the scope of this research project and would require detailed knowledge about sensitive growing periods and crop type planted, which vary across Mexico. Constructing the climate change indices based on the formal definition suggested by ETCCDI and daily temperature and precipitation raw data obtained from the GHCN-D data set was a four step approach involving missing data imputation, climate change index computation, spatial interpolation, and computation of relative change measures. Missing Data Imputation. Unfortunately, the 38-year time series ( ) of daily temperature and precipitation readings for the 200+ Mexican weather stations were not complete and about 20.6% of the records were missing. The GHCN-D data set undergoes rigorous quality checks (see Menne et al., 2012 for a detailed description) and daily temperature and precipitation readings that fail to meet these standards, often due to some type of instrumentation error, are not 46

62 included (coded missing). To maximize the use of the available climate information, missing data were imputed using Multiple Imputation (MI) (Rubin, 1987). MI is the most sophisticated method of missing data imputation and generally outperform more simplistic imputation approaches because it accounts for the uncertainty in the imputation process through explicitly adding random variation to the estimates (Allison, 2002; Little and Rubin, 2002; Honaker and King, 2010). The R package Amelia (Honaker, King, and Blackwell, 2011) was used to implement MI because of its capability to impute time-series data. To account for the temporal dependence (e.g., the temperature on a prior day is more similar to the missing value compared to the temperature 70 days ago), a second order polynomial for time was included as a covariate in the imputation model. Appendix B provides a more detailed description of the MI approach. The accuracy of the imputation was checked by inspecting density, overimputation, and overdispersion plots, suggesting that the imputation model performed reasonably well (Honaker, King, and Blackwell, 2011). Climate Change Index Computation. I used the R package climdex.pcic, managed and released by the Pacific Climate Impacts Consortium (Bronaugh, 2014) to construct the climate change indices. The climate change indices are computed from daily observations but reflect specific climatic conditions during a given year for a particular weather station. A detailed formal description of all indices is provided in Appendix A. I manually computed a selection of climate change indices based on the raw input data to investigate whether the climdex.pcic package accurately computed the indices. I was able to replicate the climdex.pcic package output, establishing confidence in the climdex.pcic outputs. Spatial Interpolation. The climate change indices are computed for the 200+ weather stations that are not necessarily situated at the particular location of a MMP municipality but are 47

63 scattered across Mexico. I use cokriging to interpolate values of the climate change indices for the 111 MMP municipalities. Other researchers have generated data products at a more coarse resolution, using similar interpolation techniques to generate gridded surfaces of the ETCCDI climate change indices, which have been used to investigate historical climate trends (Donat et al., 2013). To perform the cokriging interpolation, I used the ArcGIS Geostatistical Analyst (ESRI, 2004) implemented in the programming language Python (van Rossum and Drake, 2011) by means of the arcpy module, to allow for batch processing. To improve the interpolation, a measure of altitude (digital elevation model, DEM) was included as a covariate in the cokriging model. The DEM is based on remotely sensed images from the Shuttle Radar Topography Mission (SRTM) with a 1 kilometer (30 arc-seconds) grid cell resolution, created and released by the U.S. Geological Survey (USGS) and the National Geospatial-Intelligence Agency (NGA) (Danielson and Gesch, 2011). Appendix C provides a detailed description of the employed cokriging interpolation method. Figure 3.2 shows the weather stations and resulting interpolation surface for one climate change index (dry spell duration, cdd) as an example. From the smoothly interpolated surface, only values for MMP municipalities are relevant for the subsequent data analysis. The spatial location and shape of the municipalities was identified using a polygon layer provided by INEGI (2012). 8 Using a lattice of points with a mesh size of 700 meters, I extracted the interpolated values for each municipality and computed the average climate change index. Because of limited variation in the data, the interpolation 8 INEGI provides municipality layers for the years 1995, 2000, and 2005, and For the variable construction the 2010 polygon layer was used. Visual investigation showed only negligible changes in administrative boundaries across years and the use of the 2010 shape file was deemed appropriate. 48

64 model for the climate change indices number of ice days (ID) and number of tropical nights per year (TR) did not converge and these indices were omitted from the analysis. 9 Figure 3.3: Interpolated surface of values for maximum length of dry spell (cdd) Note: Small circles represent location of weather stations for which climate change index data were available. The color of the circles as well as the color of the resulting interpolation surface reflects the length of dry spell duration in days. 9 Some summary statistics on the climatic raw data help to explain the challenge. The mean value of the daily maximum temperature was C (min=-5.00 C, max= C). The mean value of the daily minimum temperature was C (min= C, max=40.00 C). Ice days are defined as days when the daily maximum temperature is below 0 C, which happened for only 0.004% of the days across the 38 year period. Tropical nights are defined as days when the daily minimum temperature is greater than 20 C, which happened for 19.10% of the days across the 38 year period, indicating that on 80.90% of days a 0 value on the tropical nights index was assigned. The large number of zero values prevented the interpolation model to converge successfully. 49

65 I assessed the quality of the interpolations using a split-sample approach and by comparing my estimates to those produced by other researchers (Donat et al., 2013). For the split-sample approach, the full sample (100% of non-missing values) was randomly split into subset 1 (90% of points) and subset 2 (10% of points). The interpolation was then performed with the subset of 90% of the points. For subset 2 (10% of points) the interpolated value was obtained and the difference between the predicted and the observed values allowed the computation of station-specific error values. Using a bootstrapping approach, the split-sample procedure was repeated until a minimum of 10 error values (a compromise between accuracy and computation time) was available for each individual weather station. I then computed the root mean squared error (RMSE) statistic (Bolstad, 2012). Surfaces of RMSE values by weather station are displayed in Figure 3.4 and 3.5 for an example of a temperature (tn90p) and precipitation (sdii) index across two years. Overall the spatial distribution of RMSE values reveals no systematic under or overprediction. However, the interpolation model performs less well for the temperature measure (tn90p) in the later year (1990). This effect might stem from varying data quality, missing temperature data for the particular year, or might be attributed to an increase in local extremes (see Table 3.4 below) that are more difficult to predict. Local climatic extremes are potentially indicative of climate change which may add spatio-temporal uncertainty across time. This uncertainty may be quantified in future research to derive confidence bands around the obtained estimates. 50

66 Figure 3.4: Root Mean Squared Error (RMSE) values across weather stations for average wetday precipitation (sdii) across Mexico Note: Panel (a) RMSE for the interpolation of sdii in 1975; Panel (b) RMSE for the interpolation of sdii in Green polygons represent MMP municipalities. 51

67 Figure 3.5: Root Mean Squared Error (RMSE) values across weather stations for % warm nights (tn90p) across Mexico Note: Panel (a) RMSE for the interpolation of tn90p in 1975; Panel (b) RMSE for the interpolation of tn90p in Green polygons represent MMP municipalities. 52

68 In order to gauge whether the values for the climate change indices are in a reasonable range, I compared my estimates to the climate index data available within the HadEX2 data set (Donat et al., 2013). HadEX2 is a global data set of the ETCCDI climate change indices for the period 1901 to 2010, but at a very coarse resolution (3.75 x 2.5 longitude-latitude grids). The HadEX2 data set employs a collection of climate data compiled from various sources and computes climate change indices for each weather station. The values are interpolated using a modified version of Shepard s angular distance weighting algorithm, another common method of interpolation, and subsequently gridded (Donat et al., 2013). I computed for each municipality the difference between the cokriging interpolations (constructed for this research project) and the HadEX2 data and display the values for sdii and tn90p in Figures 3.6. Overall the computation of difference values demonstrates that for the majority of municipalities, the values for sdii and tn90p computed for this research project, fall in the same range as the values in the HadEX2 data set. However, some differences occur, which may be attributed to dissimilarities in raw data, the employed interpolation method, and/or the level of aggregation (very coarse spatial resolution of the HadEX2 data). 53

69 Figure 3.6: Difference between current interpolations and values based on HadEX2 grids for sdii (a) and tn90p (b) for the year 1990 Note: Difference measures were computed by subtracting current (this research project) interpolations from the respective HadEX2 values. Positive values indicate higher values in current interpolations while negative values indicate lower values in current interpolations compared to HadEX2. 54

70 Computation of Relative Change Measures. I express all climate change measures as standardized variables (sometimes called a z-score or a standard score) by subtracting the mean of the baseline period ( ) from the value for each case (c.f., Gray and Bilsborrow, 2013), and then dividing the resulting difference score by the standard deviation of the original variable (across all years and municipalities to reflect complete variation) (Equation 1). Equation 1: x ' ( x x bp ) / sd x In Equation 1, x ' represents the standardized version of variable x, x bp represents the baseline mean value of x, and sd x represents the standard deviation of x. Each value indicates the difference from the mean of the original variable in standard deviation units, which allows a comparison across variables. To guard against the impact of short term fluctuations, x was computed as three-year average, prior to the standardization (c.f., Henry, Schoumaker, and Beauchemin, 2004; Hunter, Murray, and Riosmena, 2013). Table 3.4 displays period specific mean values for each climate change index. 55

71 Table 3.4: Summary statistics of climate change indices for selected periods Indicator Name ID SD Min Max Temperature (high) No. summer days su Warm spell duration wsdi Warmest day txx Warmest night tnx % warm nights tn90p % warm days** tx90p Temperature (low) No. frost days fd Cold spell duration csdi Coldest day txn Coldest night** tnn % cool nights tn10p % cool days tx10p Precipitation (high) No. days heavy precip r10mm No. days very heavy precip r20mm Wet spell duration cwd Max 1-day precip** rx1day Max 5-day precip rx5day Precip very wet days** r95ptot Precip extremely wet days r99ptot Total wet-day precip prcptot Average wet-day precip** sdii Precipitation (low) Dry spell duration cdd Temperature & Precipitation (other) Average precip aprec Average temperature atemp Temperature range dtr Note: Columns report mean values by period; Standard Deviation (SD), Min, and Max values are reported across all periods; **climate change indices (tx90p, tnn, rx1day, r95ptot, sdii) were dropped from the analysis due to high correlation with other indices; municipality periods: n=

72 Table 3.4 reveals trends in the climate change indices that are partially in line with global long term climatic trends (IPCC, 2013). For most of the high temperature measures an increase in temperature extremes can be discerned while the temperature range widened. For example, the value for the number of summer days increased from 0.00 ( ) to 0.14 ( ) to 0.25 ( ). A value of 0.25 for the period indicates that the number of summer days increased by 0.25 standard deviation units, comparing the period of to the 30-year reference period ( ). However, it is not possible to discern a clear trend in the precipitation measures. The growing season length (GSL) index showed little to no variation and was omitted from the analysis due to resulting unstable estimates in the regression models. Some of the climate change measures only differ in the level of threshold and/or are highly correlated (r > 0.70). Because it is the goal of this study to compare substantively different types of climate change, highly correlated variables were dropped from the analysis. 10 Climate change indices that were not considered further are marked with asterisks in Tables 3.4. Within the reduced subgroups, all correlations among variables remain below r = The selection process between highly correlated variables necessarily involves some subjectivity. Among pairs of highly correlated variables, I selected the index that captures the most extreme climatic conditions. Among high temperature indices, the warm spell duration (wsdi) is highly correlated with the percentage of warm days (tx90p) (r = 0.91). For the analysis, I chose the warm spell duration because it appears to reflect drought conditions better than the more general measure of percentage of warm days, which does not account for the connection (spell) of warm days. Among low temperature indices, the coldest temperature during day time (txn) and during night time (tnn) are highly correlated (r = 0.73) and I selected txn. This choice was rather arbitrary since both measures likely capture similar trends in cooling. Among high precipitation variables, indices come frequently in pairs that differ only in the threshold value. The indices for the number of days of heavy (r10mm) and very heavy (r20mm) precipitation are only moderately correlated (r = 0.51), and therefore I retain both variables in the analysis. In contrast, the indices for maximum 1-day precipitation (rx1day) and maximum 5-day precipitation (rx5day) are highly correlated (r = 0.88) and I selected rx5day as the measure reflecting more extreme event occurrence. However, rx5day is also highly correlated with the average wet-day precipitation (sdii) (r = 0.83) and I selected rx5day since rx5day appears to better capture precipitation extremes compared to the more general measure of average wet-day precipitation. 57

73 Secondary Predictors (Controls) Informed by the SL framework, variables representing the economic environment and livelihood capitals (e.g., financial, physical, human, social, and natural) were included as controls. These variables operate at the household and municipality levels and were constructed as time-varying whenever longitudinal information was available and as time-invariant predictors when only cross-sectional data for one time point could be obtained. Table 3.5 presents summary statistics for all secondary predictors. Household Social Capital. Migration is a gendered phenomenon in the relatively patriarchal Mexican society with very low rates of female migration (Lindstrom and Lauster, 2001), and distinct migration patterns among female heads of households (Cerrutti and Massey, 2001). This is likely the result of culturally embedded social norms that determine access to social capital (Kanaiaupuni, 2000). Due to these considerations, I included a dummy variable for the gender of the household head (coded 1=male and 0=female) in all models. In addition, the marital status of the household head may reflect access to social capital through an extension of the kinship network as well as through the ability to share responsibilities and resources among partners (Sanders and Nee, 1996). Kinship networks may provide resources that help to finance a move but may also provide local employment opportunities and therefore reduce the probability to migrate. These contrasting effects may be the reason why research in Mexico often finds only a weak association between marital status and migration (Massey and Riosmena, 2010; Riosmena, 2009). Making use of the detailed marital histories in the MMP data set, I included marital status (1=married, 0=not married) as a time-varying predictor without any clear expectations regarding the directionality of the effect. 58

74 Table 3.5: Secondary predictors (controls) for the analysis of climate change impacts on migration in Mexico, Variable Unit Time-Varying Source Mean SD Household level Social capital Female 1 0 No MMP Married 1 0 Yes MMP Human capital No. of children Count Yes MMP Education Years Yes MMP Working experience Years Yes MMP Occupation: NLF 1 0 Yes MMP Occupation: Blue collar 1 0 Yes MMP Occupation: White collar 1 0 Yes MMP Physical capital Owns property 1 0 Yes MMP Owns business 1 0 Yes MMP Municipality level Social capital International migrants % Yes MMP/COMMUN Domestic migrants % Yes IPUMS-I Physical capital Road network km/10sqkm No groads Distance border 100km No ESRI Distance city 100km No ESRI Rural 1 0 No MMP/COMMUN Financial capital Wealth index SD scale Yes IPUMS-I Natural capital Land area planted % No INEGI Farmland irrigated % No INEGI Base precip mm/day No GHCN-D Base temp Deg. C No GHCN-D Economic environment Male labor in Ag. % Yes MMP/COMMUN Note: precip = precipitation; temp = temperature; km = kilometer; NLF = unemployed/not in labor force; Household periods: n= 124,478; Municipality periods: n=

75 Household Human Capital. The presence of young children ties human capital in terms of labor capacity to nurturing and household tasks. As such, less human capital is available for external activities such as labor migration. This might explain why for the Mexican context the presence of young children usually deters migration (Nawrotzki, Riosmena, and Hunter, 2013; Riosmena, 2009; Massey and Riosmena, 2010). To capture this effect, I made use of the fertility schedules within the MMP data set and constructed a time-varying predictor, indicating the number of young children (age < 5 years) within the household during each observation year. To capture the effect of other forms of human capital, I constructed time-varying predictors for the number of years of schooling (Fussell, 2004; Massey and Riosmena, 2010), as well as cumulative work experience in years (Riosmena, 2009) of the household head from the available labor histories in the MMP data set. Moreover, a set of time-varying dummy variables captures the occupational status of the household head (unemployed/not in labor force [NLF], blue-collar, white-collar) (Massey and Riosmena, 2010). Prior research finds that undocumented migrants to the U.S. are largely young men with little education or employment experience (Massey et al., 1987; Fussell, 2004). Household Physical Capital. Research from around the world has documented that a primary motivation for international migration is to finance the acquisition or construction of a home in the absence of accessible mortgage markets (Taylor et al., 1996; Massey, 1987b). I constructed a time-varying measure of whether the household owns a house or lot (1=property owner, 0=not property owner) (Massey and Riosmena, 2010). Also, to obtain the necessary capital for business formation, a household might employ migration as a tool to access funds through remittances (Woodruff and Zenteno, 2007). To account for this relationship, a timevarying measure reflects the ownership of a business (1=business owner, 0=not business owner). 60

76 Based on the causal relationship, it can be expect that households that do not own property or a business are more likely to migrate compared to property and business owners. Municipality Social Capital. Access to migrant networks serves as a municipality-level predictor that represents the social capital available within the larger community (de Janvry et al., 1997; de Haas, 2010). Networks are specific to certain types of migration and I constructed a separate measure for international and domestic migration prevalence. The COMMUN supplementary data file of the MMP provides a measure of the percentage of adults residing in the community with migration experience. This measure is available for the years 1980, 1990, and 2000, and values for unmeasured years were linearly interpolated to construct a time varying predictor. 11 To capture the presence of domestic migration networks, I made use of data from the Mexican census, obtained via IPUMS-I (Ruggles et al., 2003; MPC, 2013) for the years 1990 and I constructed a municipality-level variable of the percentage of households that had at least one member living in another municipality or state five years before the census round. I employed linear interpolation to construct a time varying predictor. Unfortunately, no census data is available for 1980 and in a few cases the linear interpolation for years earlier than Assuming a linear trend constitutes a simplification of the complex reality and therefore tends to hide variations. In addition, linearly interpolating multiple variables may increase the correlation amongst these predictors, potentially leading to issues of multi-collinearity. However, linear interpolation was employed as a common method in event history analysis to obtain time-varying predictors from a limited number of observations (Allison, 1984). Computing the variance inflation factor demonstrated that multi-collinearity does not impact the estimates. Moreover, preliminary model runs were conducted with a linearly interpolated version and a semi timeinvariant version of the same control variables. The comparison revealed similar coefficients and standard errors, suggesting that linear interpolation has only little effect on the estimates. 12 There was a mismatch between the rural/urban classification in the IPUMS data (variable = urban ) and the MMP data set (variable = metrocat ) and a few municipalities designated as urban (n=7) or rural (n=1) in the MMP did not have urban or rural populations, respectively, in the IPUMS data. In such a case I employed the measures from the available population as proxy, irrespective of rural/urban classification. Although this treatment increases the uncertainty in this measure, it appears to be preferable over omitting the variable from the analysis. This note similarly applies to the wealth index described below. 61

77 results in negative values. To guard against unreasonable negative migration prevalence values, I capped the migration measure at 0. For municipalities that have observed data only for one year, this measure was used as time invariant across all years, assuming no change in domestic migration prevalence. Using only the available data was considered more conservative compared to employing the average slope, which would require additional assumptions on the trend. Municipality Physical Capital. For some regions and countries, access to roads impacts the likelihood of migration (Barbieri and Carr, 2005). For example, Gray (2009, 2010) finds that proximity to highways was positively associated with local moves, but negatively associated with international moves in Ecuador. I constructed a time invariant measure of access to roads by computing the combined length of all road segments within each municipality (in kilometers) and relating this metric to the size of the municipality (in 10 square kilometers). The employed road layer was provided by the Global Roads Open Access Data Set (groads), has an accuracy of meters, and uses the best available public domain roads data (CIESIN and ITOS, 2013). However, the data represents road conditions around 1995 and does not allow capturing changes in the available infrastructure. In addition, I computed the average Euclidean distance (in 100 kilometers) from each municipality to the U.S.-Mexico border as well as to the closest urban center (> 75,000 inhabitants, n=35) as time invariant measures. 13 While proximity to the U.S.-Mexico border should encourage international migration, proximity to major urban centers likely makes domestic moves more attractive. I used the World Administrative Divisions polygon layer and 13 To calculate the distance to the U.S.-Mexico border, country-level polygons for Mexico and the U.S. were buffered with a 5 meter buffer, and an intersection operation isolated the U.S.-Mexico border. After this step, a raster data layer (700m x 700m grid-cells) with values for the Euclidean distance to the U.S.-Mexico border was created. Then, zonal statistics were computed to obtain the mean distance to the U.S.-Mexico border for each of the MMP municipalities. A similar computation was performed to obtain the Euclidian distance to the nearest major city. 62

78 the World Cities polypoint layer from the ESRI s spatial data library (ArcGIS Online), for the computation (ESRI, 2012). Finally, to account for differences in access to infrastructure, services, and amenities more generally, I included a dummy variable that identifies the urbanization status of each community (rural=1; urban=0). This variable comes from the COMMUN supplement of the MMP. Communities located in a metropolitan area (a state s capital city or other large city) or a smaller urban area (10, ,000 inhabitants) were considered to be urban while communities located in a town (2,500 10,000 inhabitants) or a rancho (< 2,500 inhabitants) were considered to be rural. Different ways of categorizing rural vs. urban are possible and some researchers have focused on large metropolitan areas as representation of a truly urban environment (e.g., Fussell and Massey, 2004). However, the employed classification was based on the assumption that locations with population of less than 10,000 inhabitants can be considered rural, and likely depend heavier on the agricultural sector than more urbanized locations. An added benefit of this classification is the relatively balanced distribution of cases to both categories (59% rural, 41% urban) which provides greater statistical power than narrowing the groups to more extreme cases of urban and rural municipalities. Municipality Financial Capital. To account for region-specific differences in wealth status, I constructed a standardized wealth index using data from the Mexican Census for the years 1990 and 2000 (Ruggles et al., 2003; MPC, 2013). The wealth index includes 11 variables measuring the housing quality (building materials used for wall, floor, roof, the number of rooms and bedrooms, type of kitchen, type of toilet), and the quality of services (electricity, water supply, sewage collection, fuel source), and demonstrates a high level of reliability (Cronbach s alpha: ). The wealth index, computed at the household level, was aggregated to the municipality level as a measure of overall wealth status and development (Nawrotzki, Hunter, 63

79 and Dickinson, 2012). Values for unmeasured years during the observation period were linearly interpolated to obtain a time-varying covariate (Allison, 1984). For municipalities that have observed data only for one year, this measure was used as time invariant measure across all years. Municipality Natural Capital. Communities that largely depend on agricultural production are more vulnerable to the adverse impacts of climate change. To account for this effect, I constructed a measure of agricultural dependence by computing the proportion of the surface area planted in each municipality. This time-invariant measure uses the average hectares of planted surface area (major crops only) during the years (INEGI, 2012). However, households will be more resilient to climate shocks if technology (e.g., irrigation systems) makes the agricultural production largely independent of weather irregularities (Eakin, 2005). To capture community differences, I computed a time-invariant measure of the proportion of irrigated farmland for the year 2003 (INEGI, 2012). Data availability restricted the computation of the agricultural dependence and irrigation measures to years outside the study window. Although, time-varying information would be preferable, I assume that these variables changed only little during the study period. Finally, the general climatic conditions in a particular municipality may impact the vulnerability to climatic extremes (Nawrotzki, Riosmena, and Hunter, 2013). For example, an increase in extreme temperatures might be particularly detrimental for the agricultural sector in a dry hot climate while it is less problematic in a cool, humid environment. To account for the general climatic conditions in each municipality, I included a variable for the average daily temperature ( C) and the average daily precipitation (mm) during the 30 years baseline period ( ) as time-invariant controls in the models. 64

80 Municipality Economic Environment. To approximate job availability in climate sensitive sectors, I constructed a measure for the proportion male labor force employed in the agricultural sector. This measure comes from the COMMUN supplement of the MMP data set and is available for the years 1980, 1990, and Values for unobserved years were linearly interpolated to obtain a time varying covariate (Allison, 1984). Estimation Strategy To investigate the climate change-migration association, I employed a discrete-time event history analysis within a hierarchical framework. The event-history analysis follows each household through the study period, noting changes in sociodemographic characteristics, as well as in the environment and evaluates how these characteristics affect the odds of a move in a given 2-year period. The analysis used each period in which the household head was 15 years or older, beginning in 1986 (post-irca area), up until the time a migrant was sent to an international or domestic location, or the end of the study period (1999) if no move occurred. The logistic event history models take the general form suggested in Equation 2 (Allison, 1984; Goldstein, 2011; Singer and Willett, 2003; Steele, 2005). Essentially, the discrete-time event history model constitutes a logistic regression model on a household-period data structure, containing a record for each period that a household was at risk of experiencing the migration event (Allison, 1984; Steele, 2005). In order to reduce the possibility of endogeneity all predictor variables were lagged by one year (Gray, 2009). Event history models have been frequently applied in migration research to investigate socioeconomic as well as environmental determinants of a move (e.g., Fussell, Hunter, and Gray, 2014; Gray 2009, 2010; Gray and 65

81 Mueller, 2012a; Hunter, Murray, and Riosmena, 2013; Lindstrom and Lauster, 2001; Massey and Riosmena, 2010). Equation 2: D j Pij logit ( h ij ) log 1 Pij D j D j The discrete-time hazard h ij for interval i is the probability P that a household j experiences a migration event during the particular interval, given that no migration event has occurred in any previous interval. The parameter represents the baseline hazard of migration and was included as a set of dummy variables D, one dummy variable for each period, to invoke the most flexible representation for time (Allison, 1984; Singer and Willett, 2003). The model contains no single stand-alone intercept (e.g., β 0 ). Instead the parameters through act like multiple intercepts, one per time period, representing the log odds of event occurrence in that particular time period (Singer and Willett, 2003). This multi-part intercept captures the time trend in migration and helps account for variations in economic conditions and border enforcement policies, which might have led to a process of self-selection of migrants with certain observed and unobservable characteristics (Angelucci, 2012). The odds of a household sending a migrant are affected by household-level characteristics but also by municipality-level factors. To appropriately account for the nested data structure, I use a multilevel version of the event history model (Equation 3) (Courgeau, 2007; Goldstein, 2011; Steele, Diamond, and Amin, 1996). The models use a two-level structure 66

82 in which households and time (level-1) are nested within municipalities (level-2). As Singer and Willett (2003) point out, it is not necessary to explicitly account for the nonindependence of multiple records by nesting time within households. The hazard function describes the conditional probability of event occurrence, where the conditioning depends upon the household surviving until each particular time period and the households values of the substantive predictors in each time period. As such all records in the household-period data set are conditionally independent (Singer and Willett, 2003, p. 384). 14 Equation 3: logit ( hijk ) u0k The odds of migration are predicted for a given period i for a household j located in municipality k. The variance component u 0 k indicates that the odds of migration are allowed to differ across municipalities (level-2). To this basic model, a number of control variables are added (equation 4). Equation 4: logit ( y hijk ) n( xnz) u0k n 1 14 The conditional independence was empirically confirmed. When a three-level model structure was used (level 1: time; level 2: households; level 3: municipalities), the variance of the household-level random effects dropped to zero in the full model (all covariates included). When the household random effects variance is zero a three-level model produces the same results as a two-level model but is more parsimonious and results in a better model fit as judged by the BIC statistic. This observation is in line with other published studies that do not cluster time in individuals when employing discrete-time event history models (e.g., Gray and Bilsborrow, 2013; Steele, Diamond, and Amin, 1996; Steele, Goldstein, and Browne, 2004). 67

83 The coefficients, represent the effects of various secondary predictors x ). These n predictors may operate at different levels as indicated by the generic subscript z, which can take the form ijk (all time-varying predictors and time-invariant household-level variables), or k (time-invariant municipality-level predictors). In the next step of the modeling exercise, one climate change index at a time is entered into the model. ( nz Equation 5: logit ( y hijk ) 1 ( ciik ) n ( xnz) u0k n 2 In equation 5, the coefficient 1 shows the effect of a particular climate change index ( ci ik ) on the odds of outmigration. Although the climate change indices are municipality-level variables, they vary across time and therefore operate at level-1 as indicated by the ik subscript. It has been shown that time-varying context-level variables, such as the climate change indices, can be effectively modeled using a 2-level structure (Barber et al., 2000). The final analytical step involves using interaction models to explore whether particular subgroups of the study population are especially responsive to climate change through migration. These interactions can be uni-level interactions, when both variables operate at the same level (Equation 6a), or cross-level, when the two variables involved operate at different levels (Equation 6b). Equation 6a: logit ( y hijk ) 1 ( ciik ) 2( edu ijk ) 3( ciik * edu ijk ) n ( xnz) u0k n 4 68

84 Equation 6b: logit ( hijk ) ( ciik ) 2( btemp k ) 3( ciik * btemp k ) n ( xnz) u0k u1 ( ci 1 k ik n 4 y ) In Equation 6a, the coefficient 3 represents the effect of the uni-level interaction between the climate change index ( ci ik ) and education ( edu ijk ). In contrast, in Equation 6b the same coefficient 3 represents a cross-level interaction between the climate change index ( ci ik ) and the time invariant municipality predictor for the baseline temperature ( btemp k ). I follow a conventional approach in multilevel modeling and allow the lower level variable to vary randomly across higher level units when estimating cross-level interactions (Subramanian et al., 2009; Dedrick et al., 2009). As such a second variance component u ( ci ) is added to Equation 1k ik 6b to estimate a different slope of the climate change index for each municipality. The two 2 random effects are assumed to be normally distributed with a mean of zero and a variance of u. Assuming a joint multivariate normal distribution, the random effects have the following variance-covariance structure. Equation 7: u u 0k 1k ~ N(0, u ) : u 2 u 2 u u1 For ease of interpretation, the log odds are transformed to odd ratios using the expression: exp. The odd ratios can be interpreted as the multiplicative effects of a unit increase in OR 69

85 the predictor on the odds of migration. I fit the models using the multilevel package lme4 (Bates, 2010; Bates et al., 2014) within the R statistical environment version 3.1 (R Core Team, 2014). 15 In summary, this chapter has outlined the data and methods employed in this research project to investigate the climate change-migration association in Mexico. I construct a suite of ETCCDI climate change indices based on daily temperature and precipitation data derived from the Global Historical Climate Network Daily data set. Using spatial interpolation methods these climate change indices were linked at the municipality level to detailed migration histories obtained from the Mexican Migration Project. Controlling for various sociodemographic and economic factors, multilevel event history models were employed to investigate the effect of climate change on the odds of migration. The following three chapters present results. Chapter IV explores the influence of various types of climate change (high, low temperature; high, low precipitation) on international migration, as well as differences in climate change impacts on migration from rural and urban areas in Mexico. Chapter V investigates whether climate change more strongly impacts international or domestic migration. Finally, Chapter VI explores the influence of household internal social and human capital on the climate change-migration relationship by contrasting first and last international moves. 15 All models were fitted using the optimizer bobyqa. I used the integer scalar setting of nagq=0. With this setting the random effects and the fixed-effects coefficients are optimized in the penalized iteratively reweighted least squares step (Bates et al., 2014). Compared to nagq=1, the nagq=0 setting shows better convergence properties (more robust) and improved speed at the cost of precision. A comparison of model runs with nagq=0 to models runs with nagq=1 showed that both methods produce identical fixed-effects coefficients (as the analytical focus of this study) but slightly different random effects estimates. 70

86 CHAPTER IV RESULTS: CLIMATE CHANGE AS A MIGRATION DRIVER: DIFFERENCES BETWEEN RURAL AND URBAN AREAS Hypotheses Climate change impacts both rural and urban livelihoods (IPCC, 2014) although as a multi-faceted phenomenon, the impacts of climate change likely vary spatially. This chapter examines the association of urban and rural migration with five groups of climate change indices: (1) high temperature, (2) low temperature, (3) high precipitation, (4) low precipitation, and (5) other temperature and precipitation measures. Measures in the last category were included largely to allow for a comparison with prior work and are not used for the hypotheses testing. Not only are there different types of climate change that drive migration, different types of migration may be differently impacted by climate factors. For Mexico, most research has investigated the impacts of environmental drivers on international moves to the U.S. (e.g., Feng and Oppenheimer, 2012). In this chapter, I similarly focus on international migration, more precisely a first international move, and postpone an analysis of other migration types to the following two chapters (Chapter V: international vs. domestic migration; Chapter VI: first vs. last move). Climate Change Effects: Temperature and Precipitation In rural areas increases in temperature extremes as well as increases in mean temperature will likely lead to adverse effects on crop yields (Challinor et al., 2007). Such changes may, in turn, lead to increased levels of outmigration where livelihoods heavily depend on agricultural 71

87 production (Feng, Kruger, and Oppenheimer, 2010; Feng and Oppenheimer, 2012). Reference to the plant physiology and growth cycle helps to explain this temperature effect: Changes in mean temperature impact the evaporative and transpirative demand of plants (Priestley and Taylor, 1972; Stone, 2000) and can change crop-growth and maturity duration (e.g., Roberts and Summerfield, 1987). Increased temperature shortens the length of the growing cycle, decreasing the opportunity for CO 2 assimilation and photosynthesis, which in turn leads to a reduction in total biomass and grain yield (Bassu et al., 2014). In addition, episodes of high temperature at critical states in the crop development cycle can impact yield independently of any substantial changes in mean temperature (e.g., Wheeler et al., 2000; McKeown et al., 2005). On the other hand, a cooling or transition to more moderate temperatures could be considered beneficial in a warm, arid country such as in Mexico where temperatures frequently surpass those optimal for plant growth (c.f., Bassu et al., 2014). In urban areas, changes in temperature may impact the tourism sector (Amelung, Nicholls, and Viner, 2007) as well as the manufacturing sector through changes in energy consumption for cooling and heating (Mideksa and Kallbekken, 2010). For the tourism sector a moderate increase in temperature may be beneficial, especially for sand and sun destinations. In contrast, the relationship for the manufacturing sector is unclear since an increase in temperature would require more cooling while a decline in temperature may require more heating and in both cases the energy demand would rise. However, temperature effects on urban areas are likely to be smaller than on rural areas (Boyed and Ibarraran, 2009). As such I hypothesize that an increase in high temperature extremes (warming) increases international 72

88 migration from Mexico (H4.1) and that an increase in low temperature extremes (cooling) decreases international migration from Mexico (H4.2). 16 Most studies in rural Mexico have found a decline in the average level of precipitation to be associated with an increase in outmigration (Munshi, 2003; Hunter, Murray, and Riosmena, 2013; Nawrotzki, Riosmena, and Hunter; 2013). Although not modeled per se, it is likely that an increase in dry spells may impact migration through a negative effect on the agricultural sector. In contrast, increases in precipitation likely improve overall growing conditions (Lobell and Field, 2007). Since precipitation extremes often capture a general trend of an increase in average precipitation, it can be assumed that such changes improve agricultural production, potentially leading to a decline in rural outmigration. Similarly, droughts and precipitation declines will negatively impact urban areas, and may induce migration (Adamo, 2010). Droughts may reduce fresh water availability and increase energy prices (hydro-power) (Schaeffer et al., 2012). In addition, rural and urban areas are connected (Wackernagel et al., 2006) and drought conditions that decline agricultural production in rural areas will likely lead to an increase in food prices in urban areas (Revi et al., 2014). Hence, the effect of changes in precipitation on rural and urban sectors suggest similar directionalities with livelihood based migration responses a decline in rainfall and increase in drought conditions should encourage outmigration. However, the effect of precipitation is likely bound to thresholds, and too much rainfall may lead to flooding and damage crops in rural areas (Rosenzweig et al., 2002) and adversely impact infrastructure and various economic activities in urban areas (Revi, 2008; Fussell, Sastry, and Van Landingham, 2010). Extreme weather events such as storm surges and flooding will likely lead to an increase in migration when livelihoods are negatively impacted (Dun, 2011; 16 In this and the following chapters, international migration is used to refer to international migration to the U.S. The MMP focusses on Mexico-U.S. migration and does not contain data on other international moves. 73

89 Eakin and Appendini, 2008). However, in the mostly arid climate of Mexico it is unlikely that these precipitation thresholds were surpassed, especially given that the 1990s were a particularly dry period (Stahle et al., 2009). As such, I hypothesize that an increase in high precipitation extremes (more wet) decreases international migration from Mexico (H4.3) and that an increase in low precipitation extremes (more dry) increases international migration from Mexico (H4.4). Climate Change Effects: Rural vs. Urban Although historically the largest fraction of international migrants originated in rural areas, the number of international migrants from urban areas has increased over the past few decades (Durand, Massey and Zenteno, 2001; Marcelli and Cornelius, 2001; Riosmena and Massey, 2012). It is unclear if migration from urban and rural areas is similarly responsive to climate changes. It may be logical to assume greater rural sensitivity due to higher levels of dependence on farming and the agricultural sector (Conde, Ferrer, and Orozco, 2006). For that reason, studies of the migration environment association in Mexico have been frequently limited to rural areas (Hunter, Murray, and Riosmena, 2013; Nawrotzki, Riosmena, and Hunter, 2013). However, climate change may impact various non-agricultural sectors of the economy such as the manufacturing sectors (i.e., manufacturing, chemicals, refining, electricity) (Boyd and Ibarraran, 2009) and service sector (i.e., tourism) (Lise and Tol, 2002) and may therefore induce migration from urban areas as well. However, economic studies found the impact of climate change to be strongest for the rural-based agricultural sector with smaller impacts for urban-based manufacturing and service sectors (Boyd and Ibarraran, 2009), leading me to hypothesize that climate change more strongly impacts international migration from rural compared to urban areas in Mexico (H4.5). 74

90 Agreement: % significant coefficients in anticipated direction Hypothesis Testing Conventions Because each individual hypothesis actually relates to a group of several climate change indices, evaluative criteria must be defined allowing for the confirmation or rejection of a particular hypothesis based on multiple variables. Drawing on the climate change literature, I adopt a modified version of an evidence evaluation scheme developed for the IPCC Fifth Assessment Report (Mastrandrea et al., 2010), presented in Figure 4.1 below. Figure 4.1: Confidence matrix used to evaluate the hypothesized effect of climate change on migration based on agreement and evidence Evidence: % coefficients significant 0-33% 34-66% % % C3 High agreement Limited evidence C4 High agreement Medium evidence C5 High agreement Robust evidence 34-66% C2 Medium agreement Limited evidence C3 Medium agreement Medium evidence C4 Medium agreement Robust evidence 0-33% C1 Low agreement Limited evidence C2 Low agreement Medium evidence C3 Low agreement Robust evidence Note: Shades of grey as well as class IDs C1 to C5 reflect the confidence in a particular finding. The five confidence classes are: C1=very low; C2=low; C3=medium; C4=high; C5=very high. Source: Adjusted from Mastrandrea et al. (2010). 75

91 Using this evaluation scheme, the evidence for a climate change effect is first evaluated, regardless of relationship direction. When for example, 50% of the variables in a certain climate change category (e.g., high temperature extremes) are significant, there is medium evidence for a climate change effect. In the next step, the directionality is evaluated. For example, when considering the above 50% significant coefficients, there is high agreement if the sign of 75% of those coefficients is in the hypothesized direction. Percentage points are rounded to the closest integer. A cell with high agreement and medium evidence is assigned a high level of confidence (C4). This confidence is distinct from statistical confidence since it represents a qualitative assessment of the evidence and agreement in directionality of a certain climate change effect, while statistical confidence reflects the percentage of times (e.g., 95%) that a population parameter falls in a certain range of values if independent samples are drawn repeatedly from the same population. When no coefficient in a given climate change group reaches statistical significance, a very low confidence (C1) is assigned. A hypothesis is confirmed when there is high (C4), or very high (C5) confidence, and rejected when there is low (C2), or very low (C1) confidence. For medium evidence (C3), the ambiguous evidence results in the hypothesis being neither confirmed nor rejected. For hypotheses that compare the strength of an overall climate change effect (for example, hypothesis 4.5), I employ only the evidence scale (e.g., significance of a coefficient) since directionality is not important in these cases. Rejecting or accepting a hypothesis by evaluating the percentage of significant climate change indices is based on the assumption that each climate change index contributes unique information and captures a different aspect of the multi-dimensional construct of climate change. To guarantee the unique contribution of each 76

92 climate change index within a particular group, only substantively different indices were retained so that the correlation between all possible pairs of indices remains below a value of r = 0.70 (see details in Methods chapter). Results and Discussion Model Building As the first step in investigating climate change effects on migration, I develop a reliable multivariate base model to predict international migration from Mexico. Table 4.1 shows the model building for the complete sample (rural and urban), including the multi-part intercept for the baseline hazard (Model 1), then with the addition of household covariates (Model 2), and ultimately including community covariates (Model 3). The results shown in Table 4.1 are largely in the anticipated direction, lending credibility to the base model. The results confirm that migration in Mexico is a gendered phenomenon with lower odds of migration when the household head is female (Lindstrom and Lauster, 2001). A cross tabulation shows that about two third (66%) of these female heads are unmarried. The number of young (age < 5 years) children also reduces the odds of sending a migrant to an international destination (Nawrotzki, Riosmena, and Hunter, 2013; Massey and Riosmena, 2010). A young child requires attention and care so that less human capital is available for external ventures such as an international move. 77

93 Table 4.1: Odds of international migration from households in Mexico, Model 1 Model 2 Model 3 b sig. b sig. b sig. Period *** 0.06 *** 0.07 *** Period *** 0.07 *** 0.08 *** Period *** 0.08 *** 0.08 *** Period *** 0.06 *** 0.06 *** Period *** 0.08 *** 0.09 *** Period *** 0.09 *** 0.09 *** Period *** 0.12 *** 0.12 *** Household Level Female 0.55 *** 0.55 *** Married No. of children 0.86 *** 0.86 *** Education Working experience 0.72 *** 0.71 *** Occupation: NLF Occupation: White collar 0.36 *** 0.36 *** Owns property Owns business 0.81 ** 0.81 ** Municipality Level International migrants 1.36 *** Domestic migrants 0.93 Road network 0.87 Distance city 1.09 Wealth index 1.26 Land area planted 1.00 Farmland irrigated 1.17 Base precip 0.97 Base temp 0.97 Male labor in Ag * Rural 1.17 Model statistics Var. Intercept (Mun) BIC N (HH-period) N (Mun) Note: Coefficients reflect odd ratios; baseline hazard is captured by the following period dummies: Period 1 ( ), Period 2 ( ), Period 3 ( ), Period 4 ( ), Period 5 ( ), Period 6 ( ), Period 7 ( ); Occupation: Blue collar used as reference; NLF = unemployed/not in labor force; all predictors were lagged by one year; * p<0.05; ** p<0.01; *** p<

94 An increase in working experience, defined as the cumulative number of years employed, reduces household migration. In addition, blue collar workers are much more likely to migrate internationally compared to white collar workers. Prior research has demonstrated that Mexican migrants coming to the U.S. are mostly young, uneducated males (Massey et al., 1987; Fussell, 2004) and the observed effects confirm this individual-level pattern at the household level. Households that own a business are less likely to send a migrant internationally. Operating a business requires human capital, which is then unavailable for other livelihood strategies such as an international move. In addition, international migration frequently serves the purpose of obtaining the necessary capital to start a business (Massey and Parrado, 1998; Woodruff and Zenteno, 2007) and once this goal has been achieved a further move may become less necessary. Only a few of the municipality-level variables are significant predictors of international migration. Moreover, the BIC statistic demonstrates a worse model fit when the municipalitylevel variables are included. However, recall that the inclusion of these variables was motivated by their potential indirect effect on the climate change-migration relationship. The moderation effect will be explored using interactions in a subsequent step of the analysis. Among the municipality predictors, the presence of adults with migration experience has a strong, positive impact on the odds of international outmigration. This finding is in line with prior research, demonstrating the importance of migrant networks for an international move (Fussell and Massey, 2004). In addition, the percentage of males in the labor force employed in the agricultural sector is positively associated with the odds of an international move. This may reflect the agricultural sector s sensitivity to climatic stressors (Boyed and Ibarraran, 2009). In the next analytical step, separate models for rural and urban populations were estimated (see Table 4.2). 79

95 Table 4.2: Odds of international migration from households in rural and urban Mexico, All Rural Urban b sig. b sig. b sig. Period *** 0.13 ** 0.05 *** Period *** 0.16 * 0.05 *** Period *** 0.15 * 0.06 *** Period *** 0.12 ** 0.05 *** Period *** 0.17 * 0.06 *** Period *** 0.17 * 0.07 *** Period *** ** Household Level Female 0.55 *** 0.48 *** 0.63 *** Married No. of children 0.86 *** 0.89 *** 0.81 *** Education * Working experience 0.71 *** 0.72 *** 0.70 *** Occupation: NLF Occupation: White collar 0.36 *** 0.31 *** 0.43 *** Owns property Owns business 0.81 ** 0.78 * 0.84 Municipality Level International migrants 1.36 *** 1.61 *** 1.22 *** Domestic migrants Road network *** Distance city Wealth index Land area planted Farmland irrigated Base precip Base temp * 1.03 Male labor in Ag * Rural 1.17 Model statistics Var. Intercept (Mun) BIC N (HH-period) N (Mun) Note: Coefficients reflect odd ratios; baseline hazard is captured by the following period dummies: Period 1 ( ), Period 2 ( ), Period 3 ( ), Period 4 ( ), Period 5 ( ), Period 6 ( ), Period 7 ( ); Occupation: Blue collar used as reference; NLF = unemployed/not in labor force; all predictors were lagged by one year; * p<0.05; ** p<0.01; *** p<

96 All models were checked for multi-collinearity and distance to border was highly correlated with baseline temperature (r = 0.658) and baseline precipitation (r = 0.453). The distance measure was, therefore, removed from the set of predictors. With this adjustment the variance inflation factor (VIF) remains below 3.0 for all substantive predictors, suggesting that multi-collinearity is not influencing the estimates. The separate models for rural and urban areas reveal many commonalities with regard to drivers of international moves. The few differences include education, business ownership, road network density, and baseline temperature. Education impacts the odds of an international move only in urban areas where households with a more educated household head are less likely to send a migrant to the U.S. (Fussell, 2004). Heads of migrant households in rural areas are less educated (µ = mean = 5.99 years schooling) than migrant households in urban areas (µ = 7.16 years schooling). However, compared to the non-migrant population (rural: µ = 5.56 years schooling; urban: µ = 7.33 years schooling), rural migrants are more educated while urban migrants are less educated. As such the observation that international migrants to the U.S. are largely uneducated young adult males (e.g., Fussell, 2004) depends on the comparison group and may reflect the situation in urban but not necessarily in rural municipalities. Road networks significantly predict international migration only from urban areas. The denser the road network, the lower the probability that the household will sent a migrant. Road networks may facilitate access to local employment opportunities or may capture different levels of industrialization within urban areas (c.f., Gray, 2009; Barbieri and Carr, 2005). In contrast, international migration from rural areas is more significantly shaped by business ownership and baseline temperature. The negative coefficient for business ownership suggests that moving abroad to remit funds to start a business is largely a rural phenomenon (c.f., Massey and Parrado, 81

97 1998). In addition, the effect of the baseline temperature on an international move is significant in rural areas only, indicating a stronger association between weather conditions and migration patterns in rural compared to urban areas. In the next analytical step, climate change indices are added to the full model. To investigate the effect of each individual climate change measure, I included one measure at a time and estimated the full model (including all control variables). The coefficients and significance levels are presented in Table 4.3. I perform a jack-knife type procedure (Nawrotzki, 2012; Ruiter and de Graaf, 2006) to investigate whether the significance of the effects is sensitive towards the composition of MMP municipalities involved. To check for influential municipalities, the models were estimated by leaving out one municipality at a time from the sample set. The results are reported in appendix Table D.1, demonstrating a high degree of robustness of the estimates. To test the hypothesis regarding the effect of climate change on international outmigration, I employ results from the complete sample including both rural and urban households (Table 4.2, Model: All). However, I will base the explanations of the observed effects on the region for which the effects were significant. For example, if an effect in the combined model was significant but the regional models (rural or urban sub-models) demonstrate that the effect only applies to rural areas, I will make references to the rural agricultural sector instead of discussing impacts on the urban manufacturing and service sectors. For a more detailed discussion of the impact of climate change on the agricultural sector, I will use maize as example of an important staple crop in Mexico (Keleman, Hellin, and Bellon, 2009). Research has shown that maize together, with wheat, is particularly negatively impacted by climate change (IPCC, 2014). 82

98 Table 4.3: Estimates of the effect of climate change on international migration from rural and urban Mexico, All Rural Urban Indicator Name ID b sig. b sig. b sig. CoefDif Temperature (high) No. summer days su 0.83 ** 0.73 *** 0.99 Yes Warm spell duration wsdi 1.14 ** 1.13 * 1.12 No Warmest day txx No Warmest night tnx No % warm nights tn90p 1.15 *** 1.19 *** 1.11 No Temperature (low) No. frost days fd 0.81 * 0.65 ** 0.89 No Cold spell duration csdi No Coldest day txn 1.44 *** 1.88 *** 1.01 Yes % cool nights tn10p No % cool days tx10p * 1.03 Yes Precipitation (high) No. days heavy precip r10mm 0.86 * 0.83 * 0.91 No No. days very heavy precip r20mm 0.75 *** 0.74 ** 0.78 No Wet spell duration cwd No Max 5-day precip rx5day 0.82 *** 0.80 ** 0.84 No Precip extremely wet days r99ptot 0.84 *** 0.82 *** 0.83 * No Total wet-day precip prcptot 0.58 *** 0.61 ** 0.45 ** No Precipitation (low) Dry spell duration cdd 0.84 ** 0.67 *** 1.10 Yes Temperature & Precipitation (other) Average precip aprec 0.69 ** ** No Average temperature atemp No Temperature range dtr No Note: Coefficients reported in odd ratios; CoefDif = indicates whether coefficients of the rural and urban model are significantly different based on Paternoster et al. (1998); All = models use all available cases from rural and urban areas; each coefficient was estimated using the complete set of household and municipality control variables; * p<0.05; ** p<0.01; *** p<0.001 Climate Change Effects: Temperature and Precipitation High Temperature. Table 4.3 shows that three of the five coefficients (60%) in the high temperature group are significant, indicating medium evidence according to the confidence matrix (Figure 4.1). Of these three significant coefficients, two (67%) demonstrate a positive 83

99 effect of high temperatures on international migration, which leads to the assignment of a high agreement in the confidence matrix. Medium evidence and high agreement result in a confidence class C4 (high confidence). Therefore, the results allow confirmation of hypothesis H4.1 because there is high confidence that an increase in high temperature extremes (warming) increases international migration from Mexico. A similar relationship has been observed for the U.S. (Poston et al., 2009) and Pakistan (Mueller, Gray, and Kosec, 2014), where higher temperatures were associated with an increase in outmigration. With regard to specific climate change indices, an increase in the warm spell duration as well as an increase in the percent of warm nights, relative to the 30-years baseline period ( ), increase international migration. Particularly warm spells may serve as indicators of droughts and have likely negative impacts on the crop yield and the agricultural sector (Turner et al., 2011). Under identical precipitation, higher temperatures lead to the drying of the soil because of an increase in evapotranspiration (Mendelsohn, 2007). However, when sufficient precipitation is available, summer days (temperature exceeding 25 C) are not necessarily bad if they are evenly spread across the growing season or occur during the harvest season. For example, an increase in summer days may extend the growing season, increase plant metabolism, and facilitate the drying of the crop prior to harvest (Turner et al., 2011; Challinor et al., 2007; Mendelsohn, 2007). However, in agricultural production plant growth and temperature are non-linearly related (Tollenaar, Daynard, and Hunter, 1979). For maize, an increase in temperature increases plant growth until the optimal temperature of 31 C is reached (Sanchez et al., 2014). Higher temperatures lead to a decline in plant growth and the maximum growth temperature is reached at 42 C after which the plant dies (Sanchez et al., 2014). However, crop yield is sensitive to 84

100 different stages in the plant growth cycle. For example, in maize the period of tassel initiation is important for crop yield because during this stage the number of kernels is defined (Tollenaar and Bruulsema, 1988). In addition, maize is particularly sensitive to high and extreme temperatures during anthesis (flowering) and pollination. For example, temperatures over 32 C can reduce the percentage of non-germinating pollen by up to 51% (Schoper, Lambert, and Vasilas, 1987). Finally, maize kernel yield is affected by high temperatures, which shorten the kern filling period and can lead to lightweight grain or kernel abortion (Sanchez et al., 2014). This reference to the heat sensitivity of maize might explain why the percentage of warm nights (percentage of days where the daily minimum temperature exceeds the 90 th percentile), as a measure of more extreme high temperature events is positively associated with international migration. Overall these observations are in line with research that has demonstrated at the global level that warming temperature trends between 1980 and 2000 negatively impacted crop yield of wheat, maize, and barley (Lobell and Field, 2007). Moreover, a study by the Agricultural Model Intercomparison and Improvement Project (AgMIP), using 23 maize simulation models, suggested that climate change-induced warming will lead to a yield loss of 4 to 7% per degree temperature increase (Bassu et al., 2014). The critical contribution of this research is linking these temperature trends to international migration with the results suggesting that an increase in high temperature extremes drives international migration. Low Temperature. In the low temperature group, two of five coefficients reach statistical significance (40%, medium evidence) and both indicate the expected directionality (100%, high 85

101 agreement), leading to the assignment of a confidence class C4 (high confidence). Note that the coding of the two variables leads to opposing signs of the effect, but both coefficients suggest that an increase in low temperature extremes (cooling) decreases international migration from Mexico, confirming hypothesis H4.2. Among the significant coefficients, the climate change index for the temperature of the coldest day indicates a strong positive effect on international migration. This index measures an increase in the minimum daily maximum temperature relative to the 30-year baseline period. An increase in the coldest daily temperature, meaning a general warming trend, is associated with higher probability of outmigration. Inversely, a general cooling trend in the coldest day temperature would result in a significant decline in international migration. The climate change index for the number of frost days suggests the same directionality. An increase in the number of frost days, and thereby an overall cooling, is related to a decline in the likelihood of international migration. This effect can also be explained in reference to the agricultural sector and crop production. Each crop has an optimal growing temperature (Sanchez et al., 2014) and if temperatures are above the optimum, a cooling may increase crop yields (c.f., Bassu et al., 2014). An added benefit of cooler temperatures is the drop in dew point, meaning that the amount of precipitable water increases and moisture becomes more readily available for plants (Reitan, 1963). Moreover, although the included measures don't address specific seasonality, an increase in the number of frost days in particular seasons may kill pests that would otherwise negatively impact yields (Porter, Parry, and Carter, 1991). Overall, the beneficial effects of a cooling trend on the agricultural sector helps to explain the finding that a cooling trend leads to lower levels of international outmigration, confirming hypothesis H

102 High Precipitation. Of the six measures in the high precipitation group, five (83%, robust evidence) are significant. All of these significant measures (100%, high agreement) indicate the expected directionality, resulting in confidence class C5 (very high confidence). Therefore, hypothesis H4.3 is supported and the results suggest that an increase in high precipitation extremes (more wet) decreases international migration from Mexico. Judging by the size of the coefficients, the effect of the total wet-day precipitation (total amount of rain during wet days, prcptot) is strongest associated with international migration. 17 Larger amounts of rainfall during wet days is associated with lower probabilities of international migration -- a general relationship that is in line with findings using average precipitation measures (Hunter, Murray, and Riosmena, 2013; Nawrotzki, Riosmena, and Hunter, 2013). For the agricultural sector, an increase in precipitation is largely beneficial (Lobell and Field, 2007), especially given that only a small percentage (23.15%) of arable and permanently-cropped land in Mexico is irrigated (Carr, Lopez, and Bilsborrow, 2009). Water is required for various metabolic processes such as photosynthesis (Setter, Flannigan, and Melkonian, 2001) leading to a direct relationship between evapotranspiration and crop yield (Payero et al. 2006). However, sensitivity to water stress varies by plant species and development stage (Steduto et al., 2012). Taking maize as example, water deficits significantly reduce plant growth, dry matter accumulation, and yield. When the plant experiences prolonged water stress during sensitive stages (tasselation, ear formation) grain yield losses of 66-93% can be expected (Cakir, 2004). 18 The two measures, number of days of heavy precipitation (r10mm) and number of days of very heavy precipitation (r20mm) allow gauging the impact of varying thresholds of rainfall 17 Climate change indices have been standardized and coefficients are therefore comparable. 18 These negative impacts are the cumulative result of reduced leaf area increase, delayed ear and ovule development, poor pollination, reduction in starch synthesis and accumulation, and kernel abortion (Cakir, 2004; Jama and Ottman, 1993; Setter, Flannigan, and Melkonian, 2001; Zinselmeier, Jeong, and Boyer, 1999). 87

103 increase across time. The higher threshold measure (r20mm) leads to the strongest decline in international migration, suggesting that heavier precipitation has positive livelihood impacts. The significance of the measures of maximum 5-day precipitation (rx5day) and the precipitation on extremely wet days (r99ptot) further strengthen the conclusion that precipitation extremes are beneficial. Given that the 1990s were exceptionally dry years (Stahle et al., 2009), heavy precipitation may have helped to meet the water demand of the mostly rainfed agricultural sector without surpassing thresholds that would have caused flooding. Under such unsaturated conditions, an increase in precipitation usually leads to increases in crop yield (Lobell and Field, 2007) and positively impacts rural livelihoods, reducing the need to migrate. Also urban areas will indirectly benefit from positive effects of improved water availability on the agricultural sector through reduced food prices (Revi et al., 2014). In addition, more abundant water access may increase employment opportunities and income of the urban and peri-urban agriculture (Satterthwaite et al., 2007). Finally, more abundant water access will lower energy costs from hydro power (Schaeffer et al., 2012) and will allow easier access to fresh water as production input (e.g., cement and paper production) and various other functions (e.g., engine cooling) in the manufacturing sector. As such, an increase in water supply through heavy precipitation during dry years has likely resulted in positive effects on urban livelihoods, which helps to explain the association between an increase in precipitation extremes and a reduction in international migration. However, it is likely that during a more wet period, a further increase in rainfall extremes may result in adverse impacts due to flooding. In urban areas, flooding may damage important parts of the urban infrastructure with adverse impacts on various economic activities (Hallegatte 88

104 et al., 2010). In rural areas, flooding and excess soil moisture (saturation and waterlogging) impair plant growth. Waterlogging negatively impacts plants metabolic activities through a reduction in nutrients uptake (Ashraf and Habib-ur-Rehman, 1999; Zaidi, Rafique, and Singh, 2003), increased risk of plant disease and insect infestation (c.f., Kozdroj and van Elsas, 2000), and delayed planting or harvesting due to inability to operate machinery. Growing-season precipitation and harvest yield are non-linearly related with a decline in crop-yield when cumulative precipitation exceeds 500 mm (Rosenzweig et al., 2002). However, for Mexico during the study period ( ) the cumulative amount of precipitation during the maize growing season (June, July, August) was 388 mm (range 317 to 443) and therefore did not surpass the amount of precipitation where an adverse effect could be expected. Overall, the high precipitation measures appear to capture general trends in precipitation and an increase in rainfall is generally beneficial in rural and urban areas and tends to reduce international migration, confirming hypothesis H4.3. Low Precipitation. Only one ETCCDI climate change index examines low precipitation effects: the length of the dry spell duration as indicated by the maximum number of consecutive days with precipitation below 1 mm. The coefficient is significant (100%, robust evidence) but the direction is in contrast to that hypothesized (0%, low agreement), resulting in a confidence class C3 (medium evidence). I am, therefore, unable to confirm or reject hypothesis H4.4 (An increase in low precipitation extremes (more dry) increases international migration from Mexico). Dry spells have adverse effects on crop yield when they occur during sensitive stages in the plant growth cycle (e.g., Steduto et al., 2012). However, the annual measure of the dry spell 89

105 duration does not provide information on timing, which may play a role in this measure s lack of predictive ability as related to international migration. In certain times of the year, an increase in the length of dry days might actually be beneficial. For example, dry and warm weather is the preferred climate during harvest season. Under rainy conditions, harvest machineries perform poorer, additional costs for manual drying is incurred, and the crop might be damaged by mold (c.f., Abawi, Smith, and Brady, 1995). In addition, an increase in dry periods during the winter season will not negatively impact crop yield. Moreover, it is possible that climate change leads to an increase in extreme conditions on both ends of the precipitation spectrum, such as drier winters and wetter summers (c.f., IPCC, 2013). Under such conditions the increase in precipitation during the summer may dominate the effect on migration. To test this assumption, I computed the correlation matrix between the dry spell duration index (cdd) and the measures in the high precipitation group. And indeed, most of the significant measures are positively correlated with dry spell duration (r10mm: r=0.18, r20mm: r=0.32, rx5day: r=0.44, r99ptot: r=0.23, prcptot: r=-0.12), suggesting that an increase in precipitation extremes was accompanied by a concomitant increase in dry spell duration. In addition, as for all climate change measures, threshold effects may play a role. As long as an increase in the dry spell duration occurs in a moderate range (e.g., increase from 20 days to 30 days during the winter period) with a similar increase in precipitation in a moderate range, this may not pose substantial livelihood challenges. However, if the tendency towards extremes becomes more amplified and the dry season increases substantially, livelihoods may be negatively impacted leading to increases in international migration. Climate Change Effects: Rural vs. Urban 90

106 In order to investigate the impact of climate change on international migration from rural compared to urban areas, the evidence scale (percent of significant coefficients) of the confidence matrix (Figure 4.1) was employed. However, for a more conservative test, I only consider those coefficients for which a separate test of coefficient difference (Paternoster et al., 1998) indicated a significant difference between the results of the rural and urban sample. Among the 17 climate change indices in the four relevant groups (high and low temperature, high and low precipitation), 4 (24%) significantly predict international migration from rural areas with significantly different coefficients compared to urban areas. In contrast, for urban areas none of the 17 (0%) climate change indices that are significantly different from rural areas are significant predictors of international migration. Therefore, the results confirm hypothesis H4.5, because climate change more strongly impacts international migration from rural compared to urban areas in Mexico. In line with the conceptual framework and empirical findings (Mueller, Gray, and Kosec, 2014), these findings provide evidence that climate change influences migration largely through its impact on the agricultural sector, and rural populations more strongly depend on subsistence farming and agricultural employment as compared to urban populations (c.f., Eakin and Appendini, 2008). In support of this assumption, Feng and Oppenheimer (2012) observed an effect of crop yield changes on U.S. bound migration only for more rural states in Mexico. Even so, there is also limited evidence that climate change drives migration from urban areas, in line with other authors arguments that climate change also effects the manufacturing sector (Boyd and Ibarraran, 2009) and tourism sector (Amelung, Nicholls, and Viner, 2007). The two climate change indices that are significantly associated with international outmigration (precipitation on extremely wet days, total wet-day precipitation) show a similar direction as observed for rural 91

107 areas, indicating that an increase in precipitation leads to a decline in international migration. As noted above, it is possible that an increase in precipitation is beneficial for the manufacturing sector when water is used in the production process or influences energy costs (Pereira de Lucena et al., 2009; Boyd and Ibarraran, 2009). This effect might also result from an indirect effect of the agricultural sector on urban manufacturing and production (Wackernagel et al., 2006). For example, urban coffee mill employment may be dependent upon rural productivity while urban residents may also be employed in the agricultural sector surrounding cities (Satterthwaite et al., 2007). The data used in this study confirm this assumption, showing that in urban areas about one fourth of all males (µ = 25%) are employed in agriculture, a sizeable fraction but much smaller than in rural areas (µ = 57%). However, the indirect nature of the food price effect, the conditional nature of the hydro power effect (varying degrees of dependence on hydro-power among urban centers), as well as the limited dependence on the agricultural sector for employment (urban residents will have many alternative employment opportunities) among urban residents help to explain why a smaller number of precipitation coefficients reached significance in urban areas. 92

108 93 Table 4.4: Estimates of the effect of climate change on international migration from rural Mexico, , stratified by headship status, documentation status, and location characteristics Regular Head Mig Oth Mig Undoc Mig Docu Mig Hist Reg Oth Reg Indicator Name ID b sig. b sig. b sig. b sig. b sig. b sig. b sig. Temperature (high) No. summer days su 0.73 *** 0.77 * 0.66 * 0.72 *** *** 0.93 Warm spell duration wsdi 1.13 * ** Warmest day txx ** Warmest night tnx % warm nights tn90p 1.19 *** 1.20 ** *** Temperature (low) No. frost days fd 0.65 ** 0.63 ** ** *** Cold spell duration csdi * ** Coldest day txn 1.88 *** 1.98 *** 1.89 * 2.03 *** * 1.69 ** % cool nights tn10p *** % cool days tx10p 0.89 * * ** Precipitation (high) No. days heavy precip r10mm 0.83 * ** * 0.75 *** No. days very heavy precip r20mm 0.74 ** ** 0.69 *** * 0.71 ** Wet spell duration cwd Max 5-day precip rx5day 0.80 ** 0.82 * 0.70 ** 0.74 *** *** Precip extremely wet days r99ptot 0.82 *** 0.83 *** 0.79 ** 0.78 *** *** Total wet-day precip prcptot 0.61 ** 0.66 * 0.58 * 0.56 *** *** Precipitation (low) Dry spell duration cdd 0.67 *** 0.70 *** 0.58 ** 0.64 *** *** 0.83 Temperature & Precipitation (other) Average precip aprec * ** Average temperature atemp Temperature range dtr Note: Coefficients reported in odd ratios; Regular = all cases included; Head Mig = migrant was household head; Oth Mig = migrant was other household member (e.g., son, daughter); Undoc Mig = migration was undocumented; Docu Mig = migration was documented, Hist Reg = historical regions; Oth Reg = other non-historical regions; each coefficient was estimated using the complete set of household and municipality control variables; * p<0.05; ** p<0.01; *** p<0.001

109 94 Table 4.5: Estimates of the effect of climate change on international migration from urban Mexico, , stratified by headship status, documentation status, and location characteristics Regular Head Mig Oth Mig Undoc Mig Docu Mig Hist Reg Oth Reg Indicator Name ID b sig. b sig. b sig. b sig. b sig. b sig. b sig. Temperature (high) No. summer days su Warm spell duration wsdi Warmest day txx ** 1.01 Warmest night tnx ** 0.64 ** % warm nights tn90p Temperature (low) No. frost days fd * * 1.03 Cold spell duration csdi Coldest day txn * 1.09 % cool nights tn10p % cool days tx10p * Precipitation (high) No. days heavy precip r10mm No. days very heavy precip r20mm ** Wet spell duration cwd Max 5-day precip rx5day * Precip extremely wet days r99ptot 0.83 * 0.80 * ** Total wet-day precip prcptot 0.45 ** 0.33 *** ** ** Precipitation (low) Dry spell duration cdd Temperature & Precipitation (other) Average precip aprec 0.47 ** 0.35 *** * 0.31 * ** Average temperature atemp Temperature range dtr Note: Coefficients reported in odd ratios; Regular = all cases included; Head Mig = migrant was household head; Oth Mig = migrant was other household member (e.g., son, daughter); Undoc Mig = migration was undocumented; Docu Mig = migration was documented, Hist Reg = historical regions; Oth Reg = other non-historical regions; each coefficient was estimated using the complete set of household and municipality control variables; * p<0.05; ** p<0.01; *** p<0.001

110 Headship, Documentation Status, and Historical Regions Prior studies for Mexico suggest distinct migration patterns for household heads vs. other household members (Cerrutti and Massey, 2001), documented vs. undocumented migrants (Fussell, 2004), and households residing in regions of historically high migration rates vs. other regions (Hunter, Murray, and Riosmena, 2013). To investigate whether the impact of climate change on international outmigration differs across these dimensions, I estimated models for subsets of the data. Results are presented for both rural (Table 4.4) and urban (Table 4.5) areas. 19 Of the households that reported an international move (rural: n = 1,373; urban: n = 1,048), about three fourth (rural: 76%; urban: 74%) sent the household head. Investigating the effect of climate change on the odds of migration from rural areas for the head compared to other household members (e.g., son, daughter), shows that the majority of effects are significant for both groups. However, some effects only emerge for the household head (e.g., tn90p, fd), while others impact the odds of international migration for other household members (e.g., wsdi, r20mm). This suggests that there is no systematic difference in the household specific livelihood strategy of sending the head vs. another member in response to climate change. Although less significant climate effects emerged for urban areas, the migration response of the household head is associated with precipitation indices (r99ptot, prcptot), while the migration response of other household members is impacted by a single temperature index (fd). Of those households that sent a migrant to the U.S., the vast majority in rural areas (89.0%), and to a lesser extent in urban areas (74%), reported undocumented entry to the U.S. For migration from rural areas, climate change appears to only impact undocumented moves. 19 In addition to the outlined specifications, models were estimated that also included the survey year as control to test whether recall bias may impact the results. However, inclusion of this control variable had no impact on the significance of the climate change predictors. 95

111 None of the coefficients reached significance for the documented migrants. 20 Although the number of documented migrants is small, this finding is in line with assumptions that undocumented moves are more responsive to external factors since they can be initiated faster than documented moves that require an often long visa application process (Papademetrious and Terrazas, 2009). For example, it has been shown that a general decline in economic conditions increased attempted illegal border crossings (Hanson and Spilimbergo, 1999). Climate change, through its adverse impact on various parts of the Mexican economy (c.f., Boyd and Ibarraran, 2009) may differentially drive undocumented migration in a similar way. However, this conclusion only applies to migration from rural areas. The opposite seems to be the case for urban areas. Here, undocumented migration is entirely unresponsive to climate factors, while three precipitation indices and one temperature index were significantly associated with documented migration. Perhaps urban populations have better access to embassies, which facilitates obtaining a regular work visas. In addition, research has shown that migrant networks that enable undocumented moves are predominantly available in rural areas (Fussell and Massey, 2004). Finally, climate change effects may differ by regions (Hunter, Murray, and Riosmena, 2013). Central-western states have historically contributed most of the emigrant flow (Durand and Massey, 2003) and have therefore been termed historical regions (Hunter, Murray, and Riosmena, 2013, p. 882). These historical regions comprise the states of Colima, Durango, Guanajuato, Jalisco, Michoacán, Nayarit, San Luis Potosi, and Zacatecas. The other regions, 20 About 146 documented moves were reported for rural areas. Compared to the sample of 47,262 household periods, a documented move constitutes a rare event. However, for the maximum likelihood estimation, the absolute number of events is more important than the rarity of the event (e.g., proportion of events compared to sample size), and 146 appears to be a sufficiently large number. It has been suggested that for each predictor between 5 (liberal criterion) and 10 (conservative criterion) events need to be available (Vittinghoff and McCulloch, 2007). The fitted models contain 20 substantive predictors and about 7.3 events are available for each predictor, suggesting that the maximum likelihood estimates for the models of documented migration are unbiased. 96

112 represented in the sample, comprise the states of Chihuahua, Guerrero, Hidalgo, México, Morelos, Oaxaca, Puebla, Querétaro, Sinaloa, Tabasco, Tlaxcala, Veracruz, and Yucatán. Of the sample households, about the same number are located in other regions (rural: 56%; urban: 50%) compared to historical regions. In rural areas the climate change predictors appear to have a stronger association with international migration from other regions compared to historical regions. This observation is in contrast to findings by Hunter, Murray, and Riosmena (2013) who observed stronger effects of state-level rainfall on migration in historical regions. The present study likely offers a more precise examination of this association due to the use of more complex models (multilevel models), finer scale environmental measures (municipality level), and the use of a larger sample. However, for urban areas this relationship is less clear. While in historical regions, four climate change indices are significant, three indices are significant for other regions. In historical regions, climate change related changes in temperature matter most, while in other regions the significant relationships occur for both temperature and precipitation indices. This finding suggests that for urban areas there is no clear distinction between migration-related climate sensitivity and historical vs. other sending regions. Socio-climatic Interactions In a final analytical step, I performed a full interaction screening, interacting each climate change index with each of the sociodemographic control variables. Only the interactions for which at least 6 of 17 interaction coefficients were significant (34-100%, medium to robust evidence) are shown. Interaction models for rural areas in Mexico are presented first. 97

113 Rural Areas. The following three tables show the interactions of the male labor force employed in agriculture (Table 4.6), the municipality-level wealth index (Table 4.7), and the international migrant prevalence (Table 4.8) with all 17 climate change indices. Although there is variation, some general relationships emerged. To make the explanation more tangible, I selected one temperature and one precipitation interaction and graphically visualized the relationship. I chose significant climate change variables for which a clear directionality was anticipated from published literature such as the warm spell duration (c.f., Turner et al., 2011). 98

114 Table 4.6: Interaction between climate change indices and the male labor force employed in the agricultural sector predicting the odds of international migration from rural Mexico, CC Index Male labor in Ag. Interaction b sig. b sig. b sig. Temperature (high) No. summer days su Warm spell duration wsdi * *** Warmest day txx Warmest night tnx % warm nights tn90p * *** Temperature (low) No. frost days fd ** Cold spell duration csdi *** *** Coldest day txn * % cool nights tn10p % cool days tx10p * Precipitation (high) No. days heavy precip r10mm * No. days very heavy precip r20mm *** ** Wet spell duration cwd Max 5-day precip rx5day * Precip extremely wet days r99ptot *** ** Total wet-day precip prcptot * Precipitation (low) Dry spell duration cdd * Note: Coefficients reflect log odds; a one unit increment reflects a 10% change in the male labor force employed in the agricultural sector; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; variables were not centered; * p<0.05; ** p<0.01; *** p<

115 Table 4.7: Interaction between climate change indices and the municipality-level wealth index predicting the odds of international migration from rural Mexico, CC Index Wealth index Interaction b sig. b sig. b sig. Temperature (high) No. summer days su Warm spell duration wsdi *** ** Warmest day txx Warmest night tnx % warm nights tn90p *** *** Temperature (low) No. frost days fd * Cold spell duration csdi Coldest day txn % cool nights tn10p % cool days tx10p Precipitation (high) No. days heavy precip r10mm ** * No. days very heavy precip r20mm ** * Wet spell duration cwd ** ** Max 5-day precip rx5day *** *** Precip extremely wet days r99ptot *** *** Total wet-day precip prcptot *** *** Precipitation (low) Dry spell duration cdd Note: Coefficients reflect log odds; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; variables were not centered; * p<0.05; ** p<0.01; *** p<

116 Table 4.8: Interaction between climate change indices and international migrant prevalence in the community predicting the odds of international migration from rural Mexico, CC Index Int. migrants Interaction b sig. b sig. b sig. Temperature (high) No. summer days su *** *** Warm spell duration wsdi *** *** *** Warmest day txx *** Warmest night tnx *** % warm nights tn90p *** *** *** Temperature (low) No. frost days fd *** *** *** Cold spell duration csdi *** ** Coldest day txn *** *** % cool nights tn10p *** % cool days tx10p *** *** * Precipitation (high) No. days heavy precip r10mm *** No. days very heavy precip r20mm *** Wet spell duration cwd *** Max 5-day precip rx5day *** Precip extremely wet days r99ptot *** Total wet-day precip prcptot ** *** Precipitation (low) Dry spell duration cdd *** Note: Coefficients reflect log odds; one unit reflects a 10% change in migrant prevalence; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; variables were not centered; * p<0.05; ** p<0.01; *** p<

117 Figure 4.2: Interaction between the warm spell duration (a) and number of days of very heavy precipitation (b) and the male labor force employed in the agricultural sector in predicting the odds of international migration from rural Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities 102

118 Figure 4.2 (a) shows that the warm spell duration has almost no effect on international migration when only a small number of males in the municipality are employed in the agricultural sector. However, with higher levels of male agricultural employment, warm spells make outmigration more likely. However, the difference in predicted probabilities becomes significant mostly for higher levels of the climate change variables. The observed difference in effects is in line with assumptions made by the NELM theory (c.f., Massey et al., 1993) and other research on this topic (Liverman, 1990, Adamo and de Sherbinin, 2011), suggesting that temperature sensitivity is shaped by level of employment in climate sensitive sectors. Figure 4.2 (b) illustrates that migration probabilities generally decline with increases in high precipitation. However, the relationship becomes stronger for areas with lower numbers of males employed in the agricultural sector. This may suggest that a decline in heavy rainfall also adversely impacts non-agricultural sectors. In rural areas the number of small-scale manufacturing enterprises and micro-industries are increasing (Cornelius, 1990), and rural dwellers have started to reduce their dependence on the agricultural production through employment in these rural manufacturing sectors (Cornelius and Martin, 1993). Water availability may be important for the rural manufacturing sector, if water is used as production input or when energy supply through hydropower determines production costs (Pereira de Lucena et al., 2009). Moreover, many small-scale businesses such as retail and wholesale activities, as well as manufacturing, rely on the agricultural sector. For example, retail activities in Mexico include grocery stores and butcher shops, wholesale activities consist of bulk trading of livestock and goods, while small-scale manufacturing includes tortilla mills (e.g., Massey and Parrado, 1998). As such, these enterprises may be indirectly impacted when a precipitation decline adversely affects the prices and/or supply of agricultural goods. 103

119 Figure 4.3: Interaction between the warm spell duration (a) and number of days of very heavy precipitation (b) and the municipality-level wealth index in predicting the odds of international migration from rural Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities 104

120 Figure 4.3 (a) visually displays the interaction between the warm spell duration and the municipality-level wealth index. For households located in relatively poor municipalities, warm spell duration has almost no effect on international migration probabilities. However, those households located in relatively wealthier communities are more responsive through migration to the climate signal. This suggests that the community wealth conditions, and by extension community-level access to resources, impacts households ability to employ migration as a coping strategy for climate change. Households located in wealthier municipalities may be able to easier access funding to finance an expensive international move. In addition, community wealth might have been accumulated through remittances, and may function as an indirect measure of access to migrant networks. Overall these explanations mirror at the community-level findings at the household-level suggesting that wealthier households in Mexico are initially more likely to send a migrant (McKenzie and Rapoport, 2007). For the precipitation measure, an increase in the number of days of very heavy precipitation is associated with a decline in the probability to migrate. This relationship is strongest for households located in the wealthiest municipalities and becomes weaker for increasingly poor municipalities. Similar to the temperature effect, this might suggest that access to community based resources enables household to quickly respond to changes in the climatic patterns. However, this finding seems to be unique to the Mexican case and is not supported by work in other countries such as Ecuador (Gray and Bilsborrow, 2013), Ethiopia (Gray and Mueller, 2012a) and Bangladesh (Gray and Mueller, 2012b). The overlap in confidence intervals further indicate a large amount of uncertainty in the predicted probabilities, making it difficult to distinguish climate impacts at different levels of community-level wealth. 105

121 Figure 4.4: Interaction between warm spell duration and migrant networks in predicting the odds of international migration from rural Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities Figure 4.4 explores the interaction effect of migrant networks more directly. This interaction was significant only for temperature measures. For those households located in municipalities with well-established social networks, there is no relationship between an increase in warm spell duration and international outmigration. However, for households with less access to migrant networks, the relationship becomes increasingly strong and positive. Differences in predicted migration probabilities between strong (90%ile) and weak (10%ile) access to migrant networks was most pronounced for changes in warm spell duration up to 2 standard deviation units above the long-term average. When networks are established a move is largely determined by information flowing through this networks (e.g., where to find a job, where to stay, etc.) while for households without network access, external factors such as climate change are the predominant push factors (c.f., de Janvry et al., 1997). 106

122 Urban Areas. Although only few effects of climate change were found to be significant within urban areas, the interaction screening provided some additional insight. Specifically, the effect of climate change on migration varies by municipality-level wealth (Table 4.9) and household-level occupational status (Table 4.10) as shown in the next two tables. Table 4.9: Interaction between climate change indices and the municipality-level wealth index predicting the odds of international migration from urban Mexico, CC Index Wealth index Interaction b sig. b sig. b sig. Temperature (high) No. summer days su Warm spell duration wsdi ** Warmest day txx Warmest night tnx % warm nights tn90p * Temperature (low) No. frost days fd * Cold spell duration csdi ** Coldest day txn * % cool nights tn10p ** % cool days tx10p Precipitation (high) No. days heavy precip r10mm No. days very heavy precip r20mm * Wet spell duration cwd Max 5-day precip rx5day Precip extremely wet days r99ptot Total wet-day precip prcptot ** Precipitation (low) Dry spell duration cdd * Note: Coefficients reflect log odds; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; variables were not centered; * p<0.05; ** p<0.01; *** p<

123 Table 4.10: Interaction between climate change indices and occupation predicting the odds of international migration from urban Mexico, CC Index Occu: White collar Interaction b sig. b sig. b sig. Temperature (high) No. summer days su *** Warm spell duration wsdi *** * Warmest day txx *** Warmest night tnx *** % warm nights tn90p *** Temperature (low) No. frost days fd *** Cold spell duration csdi *** Coldest day txn *** % cool nights tn10p *** * % cool days tx10p *** * Precipitation (high) No. days heavy precip r10mm *** ** No. days very heavy precip r20mm *** * Wet spell duration cwd *** Max 5-day precip rx5day *** Precip extremely wet days r99ptot * *** Total wet-day precip prcptot ** *** * Precipitation (low) Dry spell duration cdd *** ** Note: Coefficients reflect log odds; occu = occupation; blue collar used as reference category for white collar occupation; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; the interaction models included also the interaction term Occu: NLF times the respective climate change measure (not shown); variables were not centered; * p<0.05; ** p<0.01; *** p<0.001 Graphical visualizations are used to explain the interactions. Significant interactions between climate change measures and wealth emerged largely for temperature indices. 108

124 Figure 4.5: Interaction between warm spell duration and the municipality-level wealth index, predicting the odds of international migration from urban Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities Figure 4.5 shows that the interaction between warm spell duration and wealth is similar for urban regions in Mexico as compared to rural areas (Figure 4.3). There is no significant relationship between warm spell duration and the probability to migrate internationally for households located in relatively poor municipalities. However, within municipalities with higher wealth levels, households tend to have higher international migration probabilities. As discussed above, the wealth measure may capture the effect of access to community based financial and physical capital, which might facilitate a move. However, overlapping confidence intervals suggest that when comparing the effect of community-level wealth at the 10 th and 90 th percentile, no clear distinction between predicted migration probabilities at these levels is possible. In addition to the interaction with community-level wealth, occupational status was a factor that determined the effect of climate change on migration. 109

125 Figure 4.6: Interaction between warm spell duration (a) and the number of days of heavy precipitation (b) and occupational status, predicting the odds of international migration from urban Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities ;the occupation status was captured by two dummy variables, white collar, and not in labor force, using blue collar as reference. The interaction models included interaction terms for both dummy variables but only the relevant interaction of white collar with the respective climate change indices are shown. Figure 4.6 illustrates that for both temperature and precipitation measures, households with a head employed in a white-collar job responded stronger to the climate signal through 110

126 international migration than those households with a blue-collar head. The difference in predicted migration probabilities could be best distinguished for low levels of changes in the warm spell duration and larger positive changes in the number of days of heavy precipitation. In the absence of time-varying income information (not collected by MMP), the occupational status measures may capture differences in household-level income and social status. In Mexico, whitecollar employees earn about twice as much as blue-collar workers and this wage inequality increased over the study period in response to trade reforms and liberalization (Hanson and Harrison, 1999). More well-off white-collar households may more easily afford the payment for a coyote for illegal border crossing, costing between $300 and $500 during the study period (Orrenius, 2001). In addition, white-collar households may be less dependent on multiple household members for general livelihood maintenance tasks and income generation and may therefore easier afford the loss of human capital in form of a migrating household member. In conclusion, this chapter served the dual purpose of investigating the general directionality of different climate change effects on international migration as well as differences in the migration response by origin characteristics (e.g., rural vs. urban). A number of key findings expand our knowledge of the climate change migration association: (1) Overall the results suggest that warming temperatures are associated with an increase in international migration. In addition, an increase in precipitation tends to reduce the likelihood on an international move. (2) Investigating differences between rural and urban sending regions, shows that climate change more strongly drives migration from rural compared to urban areas, likely due to the effect of climate change on the agricultural sector. (3) Significant interactions demonstrate that the climate change migration association differs by subpopulations. Among 111

127 rural populations, an increase in temperature led to the strongest increase in outmigration for households located in municipalities with a high percentage of males employed in the agricultural sector, that were relatively wealthy, with limited access to international migrant networks. Among urban populations, an increase in temperature strongest increased the probability of an international move for households residing in relatively wealthy municipalities with a household head employed in a white-collar occupation. (4) Using stratified samples revealed that climate change more strongly impacts undocumented moves from rural areas while for urban areas a stronger relationship with documented migration exists. In addition, climate change is stronger associated with migration from rural communities that have traditionally not been sending areas, while this difference is not pronounced for urban municipalities. The next chapter investigates the effect of climate change on international vs. domestic moves. 112

128 CHAPTER V RESULTS: CLIMATE CHANGE AS A MIGRATION DRIVER: DIFFERENCES BETWEEN INTERNATIONAL AND DOMESTIC MIGRATION Hypotheses In this chapter, I investigate whether climate change more strongly impacts international or domestic moves from Mexico. There are a few compelling arguments to suggest that climate change may predominantly impact international moves. First, a long history of labor migration has led to the establishment of dense migrant networks connecting Mexico and the U.S. (Fussell, 2004). Such networks are known to operate as migration corridors, strongly shaping migration streams (Bardsley and Hugo, 2010; Adamo and de Sherbinin, 2011). As related to climate, researchers have indeed found a strong association between rainfall decline and international migration from rural Mexico (Feng and Oppenheimer, 2012; Hunter, Murray, and Riosmena, 2013). Second, the self-insurance function of migration, stressed by the NELM theory (Massey et al., 1993), suggests that households ideally choose a destination within which the environmental and market conditions are uncorrelated with those at the origin. A relatively longer distance move, particularly if to another country, may best assure the disconnection from local weather and market conditions. Third, within Mexico, international moves may be more likely than domestic moves given the substantially higher wages earned in the U.S. (Massey and Espinosa, 1997). Moreover, the wage differential itself may be impacted by climate change, as Mexico is generally more vulnerable towards adverse climate changes than the industrialized U.S. (c.f. Thornton et al., 113

129 2009). Hence, climate variations may increase the incentives to migrate internationally via changes in the wage ratio (Lilleor and Van den Broeck, 2011). In contrast, research mostly from the African continent shows a stronger association with internal, circular migration (e.g., Findley, 1994; Henry, Schoumaker and Beauchemin, 2004). This line of research argues that in times of environmental stress and reduced income, households may be unable to engage in costly international moves and, instead, may employ internal, circular migration for livelihood diversification. In addition, internal moves have a number of added benefits including positive returns on education investment, a familiar cultural and linguistic environment, potential for closer geographic proximity and lower travel costs, and perhaps the avoidance of costs associated with an undocumented legal status (Lindstrom and Lauster, 2001). These considerations combine to produce the hypothesis that climate change more strongly impacts international compared to domestic migration from rural Mexico (H5.1). In addition to testing this hypothesis, the chapter compares the strength of the climate change effect relative to sociodemographic predictors and investigates sociodemographic interactions. In this chapter, results for the first move are presented leaving a detailed comparison of first and last moves to Chapter VI. I focus on rural areas since prior research (Chapter IV) has demonstrated that climate change largely drives migration from rural areas in line with the observation that climate change has particularly strong impacts the agricultural sector (Boyed and Ibarraran, 2009). Model building Results and Discussion 114

130 Various factors at the household- as well as at the municipality-level impact the odds of international and domestic migration. It is necessary to control for these factors when attempting to isolate the relationship between climate change and migration. Table 5.1 shows the composition of the base models, used for the hypothesis testing. When comparing the international to the domestic migration models, a few differences became apparent. With regard to the household-level control variables, marital status is a significant predictor of domestic migration although not of international migration. A domestic move was less likely when the household head was married. At least some of the domestic moves are related to family formation or education (White and Lindstrom, 2006). Such moves are more likely for unmarried individuals, explaining the observed negative relationship. In contrast, international moves usually do not serve the purpose of marital formation or to gain education but are largely employment related (Cerrutti and Massey, 2001; Riosmena, 2009). Similarly, household heads that reported to be unemployed or not in the labor force were more likely to migrate domestically as compared to blue collar workers. This observation is supported by prior research suggesting that domestic moves are frequently employed to find jobs in urban areas and to transition from informal to formal employment (Lucas, 1997). 115

131 Table 5.1: Odds of international and domestic migration from households in rural Mexico, International Domestic b sig. b sig. Period ** 0.08 * Period * 0.09 Period * 0.08 * Period ** 0.09 Period * 0.09 * Period * 0.08 * Period * Household Level Female 0.48 *** 0.64 *** Married *** No. of children 0.89 *** 0.85 *** Education Working experience 0.72 *** 0.74 *** Occupation: NLF *** Occupation: White collar 0.31 *** 0.62 * Owns property * Owns business 0.78 * 0.91 Municipality Level International migrants 1.61 *** 1.10 Domestic migrants Road network Distance city Wealth index Land area planted Farmland irrigated Base precip Base temp 0.91 * 0.94 Male labor in Ag * Model statistics Var. Intercept (Mun) BIC N (HH-period) N (Mun) Note: Coefficients reflect odd ratios; baseline hazard is captured by the following period dummies: Period 1 ( ), Period 2 ( ), Period 3 ( ), Period 4 ( ), Period 5 ( ), Period 6 ( ), Period 7 ( ); Occupation: Blue collar used as reference; NLF = unemployed/not in labor force; all predictors were lagged by one year; low values on the variance inflation factor (VIF) suggest that multi-collinearity is of no concern; * p<0.05; ** p<0.01; *** p<

132 The contrast between white-collar and blue-collar workers was more pronounced for international moves, with household heads employed in white-collar jobs being less likely to move than blue-collar workers, a finding in line with the established literature (Massey et al., 1987; Fussell, 2004). Property ownership influences only domestic migration with ownership deterring a move, perhaps due to the added potential of property to support livelihood diversification (c.f., Massey, 1987b). In contrast, business ownership significantly reduced the risks of migration only for international moves from rural areas, likely due to the importance of international migration to overcome liquidity constraints to start a business (Massey and Parrado, 1998). As anticipated by prior research on social networks, the percentage of adults with international migration experience significantly increased the odds of an international move as a result of the benefits from having access to established migrant networks (de Janvry et al., 1997; Fussell, 2004). In addition, the temperature during the baseline period ( ) was a significant predictor with warmer rural municipalities tending to send lower numbers of international migrants. Finally, an increase in the percentage of males employed in agriculture tended to increase the odds of a domestic move, perhaps in search for better paying jobs in urban centers (Lucas, 1997). Climate Change Effects: International vs. Domestic Moves Using the above discussed modeling platform, the impact of climate change was investigated by adding one climate change index at a time to the model. The results of this exercise are reported in Table 5.2 and form the basis of the hypothesis testing. To formally test whether climate change more strongly impacts international or domestic migration, the number 117

133 of significant effects that are also significantly different between the contrasting groups (e.g., international vs. domestic) were counted and compared. Only the 17 climate change indices in the high and low temperature and high and low precipitation groups were included in the hypothesis testing. The temperature and precipitation (other) measures are shown largely to allow for a comparison with prior work. A jack-knife type procedure was employed to test the robustness of the observed effects. During each permutation, one municipality was removed and the model was estimated using the reduced sample (Nawrotzki, 2012; Ruiter and de Graaf, 2006). The percentage of times that a particular coefficient was found to be significant are reported in Appendix Table D.2. The jackknife procedure demonstrated that the findings are highly robust. 118

134 Table 5.2: Estimates of the effect of climate change indices on international and domestic migration from rural Mexico, International Domestic Indicator Name ID b sig. b sig. CoefDif Temperature (high) No. summer days su 0.73 *** 0.83 No Warm spell duration wsdi 1.13 * 1.11 No Warmest day txx No Warmest night tnx No % warm nights tn90p 1.19 *** 1.20 * No Temperature (low) No. frost days fd 0.65 ** 0.98 No Cold spell duration csdi No Coldest day txn 1.88 *** 1.07 Yes % cool nights tn10p No % cool days tx10p 0.89 * 1.02 No Precipitation (high) No. days heavy precip r10mm 0.83 * 1.08 Yes No. days very heavy precip r20mm 0.74 ** 0.99 No Wet spell duration cwd No Max 5-day precip rx5day 0.80 ** 1.05 Yes Precip extremely wet days r99ptot 0.82 *** 1.00 Yes Total wet-day precip prcptot 0.61 ** 1.21 Yes Precipitation (low) Dry spell duration cdd 0.67 *** 0.68 ** No Temperature & Precipitation (other) Average precip aprec * Yes Average temperature atemp No Temperature range dtr No Note: Coefficients reported in odd ratios; CoefDif = indicates whether coefficients of the contrasting models are significant based on Paternoster et al. (1998); each coefficient was estimated using the complete set of household and municipality control variables; * p<0.05; ** p<0.01; *** p< Table 5.2 reveals that five out of 17 (29%) climate change predictors that were significantly different for international and domestic migration based on Paternoster et al. (1998), significantly predicted international migration. In contrast, none out of 17 (0%) climate change predictors that were different between international and domestic migration were significantly 119

135 associated with domestic moves. Therefore, the results confirm hypothesis H5.1 because climate change more strongly impacts international compared to domestic migration from rural Mexico. In general, compared to an international move, domestic migration is an attractive option because of a culturally and linguistically similar environment in the destination as well as lower travel costs (Lindstrom and Lauster, 2001). It has therefore been argued that in times of environmentally-induced livelihood strains, a domestic move may be a preferred option to seek alternative income sources (Henry, Schoumaker, and Beauchemin, 2004). Although this argument has been employed for the African context to justify a stronger correlation between climatic factors and domestic relative to international moves, this argument may also be advanced to justify the limited response of domestic migration to climatic factors observed in this study for the Mexican case. The relative ease of moving domestically may make such migration an attractive strategy that is employed under circumstances unrelated to the local climatic conditions. For example, a household head more likely sends a son or daughter to attend college within Mexico compared to an international destination. In addition, most households will have access to a larger social network within Mexico than internationally through family, friends, and acquaintances, which may provide information on employment opportunities in various alternative locations within Mexico (c.f., Curran and Rivero-Fuentes, 2003). Moreover, information flows easier domestically than internationally and a rural Mexican farmworker may read in the local newspaper about a new production plant in a nearby urban center that is presently hiring. As such, education opportunities, domestic social networks, and information access are non-climatic factors that strongly influence the timing of a domestic move but are largely unrelated to international moves. Because of their low costs (monetary and socially), domestic moves may 120

136 frequently occur in response to non-climatic factors and a potential influence of climate change will be more difficult to detect. Turning to international migration, it is clear that international moves are also influenced by various non-climatic factors including cultural norms (Kandel and Massey, 2002), migrant networks (Fussell and Massey, 2004), and financial interests (Massey, 1987b; Taylor et al., 1996). While these factors may influence the timing of a move independently, there is also evidence for interactions with climatic conditions. For example, climate change may lead to a widening in the income gap, if Mexico s agriculture-dependent economy is more strongly impacted by adverse climatic conditions compared to the U.S. economy due to differences in technological advancements (e.g., irrigation systems). A larger income gap in turn may enhance incentives to migrate internationally (Lilleor and Van den Broeck, 2011). In addition, a major motivation that distinguishes international from domestic moves is the self-insurance function suggested by the NELM theory (Stark and Bloom, 1985; Massey et al., 1993). In the absence of functioning financial and insurance markets, a household might decide to send a migrant elsewhere to self-insure against climatic shocks and their impact on the local economy and related income generation options. The insurance function is most effective when the market and climatic conditions in the destination are largely uncorrelated to those in the origin (Massey et al., 1993). Such uncoupling from local conditions is often more fully accomplished for an international compared to a domestic destination and may explain the higher sensitivity of international moves to climate change. In short, because domestic moves are often more easily facilitated they may be impacted by various non-climatic conditions while there are various direct and indirect channels through which climate change may influence international moves. 121

137 However, the emergence of a few significant effects seems to indicate that climate change does have some influence on domestic moves as well. Among temperature-based climate change indices, only the percentage of warm nights (tn90p) was significantly related to domestic migration. An increase in the number of warm nights was associated with a higher probability of a domestic move, perhaps due to adverse impacts of a warming trend on crop yield and the agricultural sector (Lobell and Field, 2007; Mueller, Gray, and Kosec, 2014). For precipitationbased climate change indices, only the dry spell duration index (cdd) was a significant predictor of domestic migration. An increase in the dry spell duration decreases the odds of domestic migration. A dry climate during the harvest season may be beneficial for the agricultural sector due to the better performance of harvest machinery and no additional need for expensive crop drying (c.f., Abawi, Smith, and Brady, 1995). For the two climate change indices (tn90p, cdd) that were significant for both international and domestic moves, a remarkable similarity in direction and size of the coefficients was observed. This may suggest that the underlying causal relationship that determines these variables influence on migration responses is similar for domestic and international moves. However, it remains a matter of speculation as to why these two climate change indices are significant determinants for both international and domestic migration. These two climate change indices count among the strongest predictors and it might be that only strong effects are discernable from among the many other non-climatic factors that determine the timing of a domestic move. Overall, this analysis reveals that in rural Mexico, international migration is a more common livelihood strategy in response to climate change compared to domestic migration. 122

138 International migration may be the preferred response option to climate change due to its selfinsurance function (NELM theory) and interactions with networks and income differentials. Strength of the Observed Effects It has often been stated that environmental factors may operate as drivers of migration, but are small compared to first-order socio, political, and economic drivers (Barbieri et al., 2010; Carr, 2005; Jonsson, 2010; Suhrke, 1994). In order to investigate the magnitude of the effects revealed in Table 5.2, I computed standardized beta coefficients according to the following formula SD beta x bx * SD x y, with b x representing the unstandardized regression coefficient for a given variable x, and SD x and SD y representing the standard deviation of the given predictor variable x and the outcome variable y (Vittinghoff et al., 2012). The beta coefficients allow comparing the strength of the effect and express the change in the outcome variable for a change in one standard deviation unit of x, holding all other variables in the model constant. I computed beta coefficients for the international out-migration from rural areas since climate change effects have been shown to stronger impact rural agricultural dependent areas (Boyd and Ibarraran, 2009) and because this study has demonstrated a stronger relationship between climate change and international migration. Table 5.3 lists significant predictors, including climate change indices and sociodemographic factors, ranked by the size of the observed effect (largest effect first). Because negative odd ratios (range 0 to 1) have a different scaling than positive odd ratios (range 1 to ), I also report the inverse value for negative odd ratios (1/beta (OR)). Due to the standardization, coefficients reported in Table 5.3 reflect a change in log odds or odds of international migration from rural areas for a change in one standard deviation unit of the predictor (e.g., working experience, % warm nights, etc.). 123

139 Table 5.3: Comparison of the size of the climate change effect in relation to sociodemographic factors, predicting international migration from rural Mexico, Variables b b sig. beta beta 1/beta (Logit) (OR) (Logit) (OR) (OR) International migrants (%) *** Working experience *** Occupation: White collar *** Base period temp ( C) * Female (yes=1) *** Coldest day *** Precip extremely wet days *** % warm nights *** Warm spell duration * Dry spell duration *** Total wet-day precip ** Max 5-day precip ** No. summer days *** No. days very heavy precip ** % cool days * No. of children (< 5 yrs) *** No. frost days ** No. days heavy precip * Owns business (yes=1) * Note: beta = standardized coefficients; 1/beta (OR) = inverse of beta (OR) computed for negative effects to allow for direct strength comparison; bold font used for climate change indices; table was sorted by the absolute size of the standardized coefficients; * p<0.05; ** p<0.01; *** p<0.001 Table 5.3 shows that a number of sociodemographic variables have stronger effects on international migration compared to climate change indices, supporting a minimalist viewpoint (Suhrke, 1994) that the environment is a secondary driver of migration. The two largest effects emerge for migrant networks and working experience. For example, an increase in the percentage of adults with migration experience in a particular community by one standard deviation unit increases the odds of outmigration 27.5 times. Similarly, a decline in the working experience of the household head increases the odds of outmigration 22.0 times. Compared to these effect sizes, the effect of climate change is much smaller. For example, an increase in the 124

140 warm spell duration (wsdi) by one standard deviation unit increases the odds of outmigration 2.5 times, while a decline in the total wet day precipitation (prcptot) increases the odds of outmigration 2.2 times. Thus, the effect of climate change is roughly a tenth of the size of the strongest sociodemographic predictors. Temperature and precipitation indices appear to be similar in strength on their impact on migration. In contrast, literature on the effect of climate change on agricultural production and crop yield usually finds stronger effects of temperature than precipitation (Lobell and Field, 2007; Mendelsohn, 2007; Mueller, Gray, and Kosec, 2014). Socio-climatic Interactions For international moves from rural areas (see Chapter IV), the sociodemographic variables for which interactions were observed included the male labor force employed in the agricultural sector, the wealth index, and the migrant network. Due to the lack of significant coefficients, it is not surprising that much fewer interactions emerged for domestic migration from rural areas. Because the threshold of six interactions per group (34-100%, medium to robust evidence), used in Chapter IV, was not reached for any sociodemographic variable, I report interactions with the sociodemographic variable for which the largest number of significant interaction terms were found, this being the occupational status for domestic moves from rural areas (Table 5.4). 125

141 Table 5.4: Interaction between climate change indices and occupational status predicting the odds of domestic migration from rural Mexico, CC Index Occu: White collar Interaction b sig. b sig. b sig. Temperature (high) No. summer days su * Warm spell duration wsdi Warmest day txx * Warmest night tnx * % warm nights tn90p * Temperature (low) No. frost days fd * Cold spell duration csdi Coldest day txn ** % cool nights tn10p % cool days tx10p * Precipitation (high) No. days heavy precip r10mm * No. days very heavy precip r20mm ** Wet spell duration cwd * Max 5-day precip rx5day * * Precip extremely wet days r99ptot * * Total wet-day precip prcptot * Precipitation (low) Dry spell duration cdd * Note: Coefficients reflect log odds; Occu = occupation; blue collar used as reference; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; the interaction models included also the interaction term Occu: NLF times the respective climate change measure (not shown); variables were not centered; * p<0.05; ** p<0.01; *** p<

142 In order to facilitate the explanation of the interactions, I employ graphs and display significant interactions with both a temperature and precipitation index, when possible. The two contrasted occupation groups differ much in their composition with about 77% of the entire study population employed in blue collar jobs and only 9% working in white collar positions. Figure 5.1 shows that for blue collar workers, an increase in the temperature on the coldest day (warming trend) led to a slight increase in domestic out-migration from rural areas. Significant differences in the predicted probabilities emerged predominantly for increases in the temperature during the coldest day greater than 0.5 standard deviations above the longer term mean. Using the agricultural explanation model, this effect would be expected as it might signal a warming trend, with likely adverse impacts on crop yield and farm income (Lobell and Field, 2007; Mueller, Gray, and Kosec, 2014). In contrast, for white collar workers a warming trend (when the coldest temperature increased) led to a decline in domestic outmigration. The steeper slope for white-collar workers suggests a stronger migration response towards climate change effects. Occupational status may function as a proxy indicator for income (Hanson and Harrison, 1999) and better-off white-collar households may be more able to finance a move and to afford losing human capital when sending a household member elsewhere. Similarly, an increase in the number of days of very heavy precipitation led to a decline in the migration probability for blue-collar workers. The majority of rural blue-collar workers are employed in the agricultural sector (Massey et al., 1987), and for this sector an increase in rainfall is largely beneficial (Steduto et al., 2012), reducing the need for livelihood diversification through migration. In contrast, for white collar workers an increase in very heavy precipitation increased the probability of domestic outmigration. The difference in predicted 127

143 probabilities was distinguishable largely when the number of days of heavy precipitation declined more than 0.5 standard deviations below the long term mean. Figure 5.1: Interaction between the coldest day (a) and the number of days with very heavy precipitation (b) and occupational status predicting the odds of domestic migration from rural Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities; the occupation status was captured by two dummy variables, white collar, and not in labor force (NLF), using blue collar as reference. The interaction models included interaction terms for both dummy variables but only the relevant interaction of white collar with the respective climate change index is shown. 128

144 The observed effects for white collar workers contradict the paradigm that adverse climatic conditions should lead to higher rates of outmigration, which has been referred to as environmental-risk hypothesis (Gray, 2009, p. 458). While this paradigm works well for blue collar workers, more likely to be employed in the agricultural sector (Massey et al., 1987), domestic migration response of white collar workers may be differently motivated. These individuals may employ domestic migration more as a form of investment in education, network expansion, or as an informal insurance mechanism, under favorable conditions, a paradigm that has been referred to as environmental-capital hypothesis (Gray, 2009, p. 458). This observation is in line with findings observed for South Africa where favorable environmental conditions in terms of access to natural capital increased domestic migration (Hunter et al., 2014). In summary, this chapter offers three conclusions to our growing knowledge of the climate change - migration association for the Mexican case: (1) Confirming hypothesis H5.1, the results show that households more strongly respond to climate change with international compared to domestic migration. Households may prefer international migration in response to climate change due to its self-insurance function (NELM theory) or in response to climate s impact on income differentials. (2) Comparing the strength of the direct climate change effect on international migration from rural areas shows that the effect size amounts to one tenth of the stronger sociodemographic drivers such as social networks. (3) The relationship between climate change and domestic migration is moderated by households occupational status suggesting that better-off white-collar workers may employ migration under favorable weather conditions while blue-collar workers employ migration as livelihood strategy under adverse climatic conditions. 129

145 The next chapter investigates the effect of households internal social capital on the climate change migration association. 130

146 CHAPTER VI RESULTS: CLIMATE CHANGE AS A MIGRATION DRIVER: DIFFERENCES BETWEEN FIRST AND LAST MOVES Hypotheses In this chapter, I engage past research examining the differential impacts for individuals and households of intra-household patterns of migration. Specifically, migration has been usefully distinguished into first and last move, with the last move representing the second, third or further migration (e.g., Fussel, 2004). Research suggests these different moves have distinct motivations and impacts. As a result of the first move, for example, a household gains internal or household-specific social and human capital through the establishment of social ties that link households in sending communities to individuals and institutions in receiving areas. The migrant individual and household also gain new knowledge about the migration process itself (Massey et al., 1987; Massey, 1990). Prior research shows that social capital, built through the first move, substantially increases the likelihood of a consecutive move (Curran and Rivero- Fuentes, 2003; Fussell, 2004; Massey and Espinos, 1997; Massey and Garcia Espana, 1987). This project contributes to this literature with a focus on the determinants of first and last moves at the household level to explore the impact of household-specific social and human capital on climate change-related migration. Two theories can be identified that suggest different underlying causal mechanisms, which I term the additive theory and the multiplicative theory. Additive theory. Some researchers have suggested that the first move is more directly associated with environmental factors than later moves (Henry, Schoumaker, and Beauchemin, 2004). From this point of view, the first move represents a livelihood strategy more directly linked to the external environment since it does not have existing social capital from which to 131

147 ease migration costs. After this first migration, the obtained human and social capital (e.g., experience, social networks) becomes a stronger predictor for future moves as compared to external factors such as environmental change (Fussell, 2004). This theory is based on an additive understanding of migration drivers if the influence of one driver (e.g., social networks) increases, the effect of other drivers (e.g., environmental factors) must decline. In support of this theory, an early study found that migration from Mexican households without internal social and human capital was more strongly impacted by community and environmental characteristics compared to households that possessed internal social capital (de Janvry et al., 1997). Multiplicative theory. Another school of thought suggests that migration drivers interact in complex ways and may amplify or impede the effect of other drivers in a multiplicative fashion. For example, the concept of migration corridors (Bardsley and Hugo, 2010) proposes that migrant networks may facilitate environmentally-induced migration. Once a corridor between two regions is established, migration becomes less costly and even relatively small environmental strain may yield large scale migrations. As such this theory proposes that certain factors (e.g., migrant networks) may operate as amplifiers, increasing migratory response to other drivers. For the Mexican case, this multiplicative effect has been observed for communitybased social capital, which facilitates environmentally motivated moves (Hunter, Murray, and Riosmena, 2013; Riosmena, Nawrotzki, and Hunter, 2014). If this relationship holds for within-household social capital, then it can be expected that moves after the first migration should exhibit less association with environmental factors as a result of the accumulated social capital within the household. Hence, I hypothesize that climate change more strongly impacts first compared to last international moves from rural Mexico (H6.1). In addition to testing this hypothesis, this chapter offers an analysis of the timing of an 132

148 international move as related to climate change, and also an exploration of socio-climatic interactions. For the analysis, I focus on international moves because a last domestic move is a rare event resulting in unstable model estimates. In addition, for a more targeted analysis I focus on rural areas since these areas have been shown to be more vulnerable to climate change (Boyd and Ibarraran, 2009). Results and Discussion Model Building Numerous factors influence household decisions to send an international migrant. Only when these other factors are accounted for, can the isolated effect of climate change be studied without the risk of reporting spurious relationships. Table 6.1 shows the fully adjusted multilevel event history models that were employed as base models from which to then test the hypotheses. The majority of significant predictors were similar for first and last moves. Households headed by females and households with young children were less likely to send an international migrant. Similarly, employment in a white collar occupation and substantive working experience reduced the odds of sending a migrant internationally. Reflecting the importance of social networks, community-based prevalence of adults with migration experience is consistently among the strongest international migration predictors. 133

149 Table 6.1: Odds of a first and last international move from households in rural Mexico, First Last b sig. b sig. Period ** 0.00 ** Period * 0.01 * Period * 0.02 * Period ** 0.01 ** Period * 0.04 Period * 0.03 * Period * Household Level Female 0.48 *** 0.40 *** Married No. of children 0.89 *** 0.69 *** Education Working experience 0.72 *** 0.90 ** Occupation: NLF Occupation: White collar 0.31 *** 0.24 *** Owns property * Owns business 0.78 * 0.96 Municipality Level International migrants 1.61 *** 1.84 *** Domestic migrants Road network Distance city Wealth index * Land area planted Farmland irrigated Base precip Base temp 0.91 * 1.00 Male labor in Ag Model statistics Var. Intercept (Mun) BIC N (HH-period) N (Mun) Note: Coefficients reflect odd ratios; baseline hazard is captured by the following period dummies: Period 1 ( ), Period 2 ( ), Period 3 ( ), Period 4 ( ), Period 5 ( ), Period 6 ( ), Period 7 ( ); Occupation: Blue collar used as reference; NLF = unemployed/not in labor force; all predictors were lagged by one year; low values on the variance inflation factor (VIF) suggest that multi-collinearity is of no concern; * p<0.05; ** p<0.01; *** p<

150 A few predictors differed between the first and last migration. Property ownership was positively correlated with the odds of a last move but uncorrelated with the first move. Descriptive data suggest that prior to the first move, migrant households less frequently own property (µ=45%) compared to non-migrant households (µ=70%). However, many moves are undertaken to earn money in order to build a house and/or buy property (Massey, 1987b; Taylor et al., 1996). If this goal was successfully accomplished, households with property may be more likely to engage in a consecutive move (Massey and Garcia Espana, 1987) and will more likely possess property (purchased with migrant earnings). Similarly, a household might use initial migration (first move) to obtain the necessary funding to start a business (Massey and Parrado, 1998). Households already in possession of a business do not need to draw on migration to fund a business startup and are therefore less likely to send a migrant. In contrast, after a successful first move and business formation, the likelihood of a consecutive move is not anymore determined by physical capital but rather through internal social networks (c.f., Massey and Espinosa, 1997). The wealth index at the community level is only significant for the last international move. Through the first move, households may accumulate wealth which may allow them to settle in more affluent communities over time (Massey, 1996). In addition, positive effects of remittances on community development have been observed (Taylor et al., 1996; McKenzie and Rapoport, 2007). Both of these trends may explain the significant correlation between community wealth and a consecutive move. Finally, baseline temperature is only a significant predictor for the first move, lending initial evidence to the additive theory that environmental factors may matter more for the first than the last move (Henry, Schoumaker and Beauchemin, 2004). 135

151 In short, the observed effects of the control variables largely display the directionality anticipated by theoretical considerations and prior research on Mexican migration. Therefore, the modeling platform can be judged as a robust tool to test the effect of climate change on the first and last migration. Climate Change Effects: First vs. the Last Move To examine the association between climate change and first/last moves, I estimate fully adjusted multi-level event history models and include one climate change index at a time. Table 6.2 shows the results derived from these models, allowing for tests of the outlined hypothesis by comparing the number of significant coefficients. A jack-knife type procedure was performed to investigate the estimates robustness by iteratively removing one municipality from the sample and re-estimating the model (Nawrotzki, 2012; Ruiter and de Graaf, 2006). The results are shown in Appendix Table D.3, demonstrating a high degree of robustness. Table 6.2 shows that five out of 17 (29%) climate change indices that are significantly different between first and last moves based on Paternoster et al. (1998) are significantly associated with a first move. In contrast, only two out of 17 (12%) climate change indices that are significantly different between the contrasting models are significantly associated with a last move. Therefore, the analysis supports hypothesis H6.1, indicating that climate change more strongly impacts first compared to last international moves from rural Mexico. As such this finding lends support to the additive theory, suggesting that the first move is a livelihood strategy more directly linked to changes in the climate while for later moves, these external triggers have less importance (c.f., de Janvry et al., 1997). 136

152 Table 6.2: Estimates of the effect of climate change on the first and the last international migration from rural Mexico, First Last Indicator Name ID b sig. b sig. CoefDif Temperature (high) No. summer days su 0.73 *** 1.64 ** Yes Warm spell duration wsdi 1.13 * 1.3 * No Warmest day txx No Warmest night tnx * Yes % warm nights tn90p 1.19 *** 0.96 No Temperature (low) No. frost days fd 0.65 ** 1.1 No Cold spell duration csdi No Coldest day txn 1.88 *** 0.65 Yes % cool nights tn10p No % cool days tx10p 0.89 * 1.06 No Precipitation (high) No. days heavy precip r10mm 0.83 * 0.95 No No. days very heavy precip r20mm 0.74 ** 0.93 No Wet spell duration cwd No Max 5-day precip rx5day 0.8 ** 1.24 Yes Precip extremely wet days r99ptot 0.82 *** 1.04 Yes Total wet-day precip prcptot 0.61 ** 1.04 No Precipitation (low) Dry spell duration cdd 0.67 *** 1.09 Yes Temperature & Precipitation (other) Average precip aprec No Average temperature atemp *** Yes Temperature range dtr No Note: Coefficients reported in odd ratios; CoefDif = indicates whether coefficients of the contrasting models are significant based on Paternoster et al. (1998); each coefficient was estimated using the complete set of household and municipality control variables; * p<0.05; ** p<0.01; *** p<0.001 To illustrate, before the first move a household may have only limited information about employment opportunities in a potential destination and the decision to migrate is motivated by climate related livelihood insecurities that may have impaired a households agriculture related income options or food provision (c.f., Massey et al., 1993). In contrast, after the first move a 137

153 household may have established links to a destination and is therefore able to draw on this connection to gather information about job openings and employment opportunities. These links, connecting the household in the origin with the destination, constitute the household specific social capital that then begins to dominate the timing of a move (Fussell, 2004; Massey et al., 1987). Although, a first move is much more strongly influenced by climate factors, a few significant effects emerged for the last move as well. These significant relationships emerged largely for high temperature extremes. Perhaps this indicates the significance of temperature to livelihood endeavors (see Lobell and Field, 2007, for the importance of temperature on agricultural production), which influences migration decisions even when household internal networks are established. However, the relationship between the climate change signal and the migration response is different when comparing first and last moves. For example, a change in signs occurs for the effect of the number of summer days. While an increase in the number of summer days decreases the probability of a first international move, it increases the probability of a last move. For the agricultural sector, summer days (temperature exceeds 25 C) can be both beneficial and detrimental, depending on timing. During the sensitive growing period, the impact of summer days will be shaped by whether the optimal growing temperature is surpassed (e.g., 31 C for corn, Sanchez et al., 2014). The plant growing cycle is different for various crops and vulnerable stages occur at different times of the year (c.f., Steduto et al., 2012). During the harvest season, summer days are likely beneficial since they may help dry grain crops and allow crops to fully mature (Mendelsohn, 2007). The different signs may reflect differences across years in the 138

154 timing of migration since more precise timing of the move vis-à-vis the occurrence of summer days is not available. The effect of the warm spell duration (wsdi) indicates a similar directionality. An increase in the length of a warm spell increases the odds of a first and last international move. Extended periods of excessive heat will likely lead to a decline in crop yield especially when occurring in sensitive phonological phases and development stages (Sanchez et al., 2014). The larger effect size of the warm spell duration coefficient, comparing the last to the first move suggests that for this particular climate change effect, internal social capital facilitates the migration response. However, the observed variation in effect size may also be potentially attributed to different years during which the first and the last move occurred. Finally, an increase in the temperature during the warmest night leads to a decline in the probability of a last move but is uncorrelated with the first move. As explained above, during the harvest season an increase in temperature may be beneficial (Mendelsohn, 2007). While for the first move the significant temperature effects are largely consistent (warming increases outmigration), some inconsistencies arise for the last move. Assuming that the last move is more strongly influenced by household specific social capital, it is likely that unmeasured pull factors increasingly determine the climate-change migration relationship. For example, the increase in temperature during the warmest night may have corresponded with an unmeasured increase in job openings in the U.S. destination to which the household has established links. Unfortunately, it is not possible to include destination pull factors in the analysis. In short, confirming hypothesis H6.1, climate changes most strongly impact first moves from rural areas while only a few temperature effects also shape later moves. 139

155 Timing of Climate Change-Related Migration Migration represents only one of a range of possible adaptations to climate change. Based on multiphasic response theory (Davis and Lopez-Carr, 2010) in situ (in place) adaptation is likely the first response towards livelihood insecurities (Bardsley and Hugo, 2010). Such strategies include sales of assets, intensifying livelihood activities or adopting new ones, use of formal and informal credit, reducing nonessential expenditures, and drawing on social networks and public programs for assistance (Gray and Mueller, 2012a). If those strategies prove to be insufficient, a household may then opt to send a member elsewhere (McLeman and Smit, 2006; Warner et al., 2010). From this perspective, a delayed migration response is most likely when households employ in situ strategies during the first years of livelihood strain and only if those fail or are exhausted might they choose to engage in migration as a strategy. However, when households are unable to employ in situ livelihood strategies, due to a lack of livelihood capitals or institutional conditions that prevent their use (Eakin, 2005), then a household perhaps reverts to migration more immediately in the face of strain. The immediate use of migration may also be the result of household capabilities and preferences, especially when strong social networks exist (Fussel and Massey, 2004) or when migration is perceived as a cultural norm (Kandel and Massey, 2002). In this case, an immediate migration response to climate change might be expected, one that becomes weaker as time passes. In order to investigate the timing of migration responses, I include different time lags for the climate change predictors. Climate change in specific years 1 to 4 -- prior to the observation year -- is used to predict outmigration (Table 6.3). Since climate change shows the strongest impacts on first international moves from rural areas, I focus on this migration stream for the following analysis. 140

156 Table 6.3: Estimates of the timing of the migration response in relation to the climate signal for the first international move from rural Mexico, Lag 1 Lag 2 Lag 3 Lag 4 Indicator Name ID b sig. b sig. b sig. b sig. Temperature (high) No. summer days su 0.73 *** 0.68 *** 0.64 *** 0.65 *** Warm spell duration wsdi 1.13 * 1.14 * Warmest day txx Warmest night tnx * 0.66 * % warm nights tn90p 1.19 *** 1.13 * Temperature (low) No. frost days fd 0.65 ** Cold spell duration csdi ** Coldest day txn 1.88 *** % cool nights tn10p % cool days tx10p 0.89 * 0.86 ** 0.84 *** 0.81 *** Precipitation (high) No. days heavy precip r10mm 0.83 * 0.79 ** 0.78 ** 0.89 No. days very heavy precip r20mm 0.74 ** 0.81 * Wet spell duration cwd Max 5-day precip rx5day 0.80 ** Precip extremely wet days r99ptot 0.82 *** 0.89 ** 0.90 * 0.87 ** Total wet-day precip prcptot 0.61 ** 0.61 ** 0.67 * 0.93 Precipitation (low) Dry spell duration cdd 0.67 *** 0.73 ** Temperature & Preciptiation (other) Average precip aprec ** 0.61 ** 0.72 Average temperature atemp * 0.45 * 0.51 Temperature range dtr * 2.14 ** 1.90 * Note: Coefficients are reported in odds ratios; Lag 1-4 reflect the number of years the climate change predictors were lagged; * p<0.05; ** p<0.01; *** p<0.001 The results show that different climate change indices produce different response patterns. Three distinct types of migration response patterns to climate change can be identified, including direct response, continuous response, and lagged response. Direct Response. Climate change indices in the Direct Response group (wsdi, tn90p, fd, txn, r10mm, r20mm, rx5day, prcptot) exhibit mostly immediate effects on migration behavior. 141

157 These indices show significant effects using a 1-year lag but the effects weaken over the years and statistical significance is lost by the 4 th year. For example, an increase in total wet-day precipitation, one year (lag 1) and two years (lag 2) prior to the observation period, reduces the odds of an international move by 39%. However, the same increase in total wet-day precipitation three years (lag 3) before the observation period only reduces the odds of an international move by 33%, and the relationship becomes insignificant when using the wet-day precipitation four years (lag 4) earlier. This behavior may suggest that households have limited in situ livelihood strategies at their disposal, or that households employ migration as the preferred response mechanism due to established migration networks (Fussell and Massey, 2004) and cultural norms (Kandel and Massey, 2002). Out of 20 climate change predictors, 9 (45%) follow this pattern, making the direct response the most frequently employed response pattern. Continuous Response. Climate change indices in the Continuous Response group (su, tx10p, r99ptot) significantly impact migration behavior when lagged by 1 year but retain their significance if climate change is used two, three, or even four years prior to the observation period. Such behavior would suggest that migration is employed as a continuous adaptation strategy through multiple years. Perhaps households pursue a dual adaptation pathway that combines in situ strategies and migration. However, this behavior is much more rare and is only observed for 3 (15%) of the 20 climate change predictors. Lagged Response. For climate change indices in the Lagged Response group (tnx, csdi, aprec, atemp, dtr) a delayed migration response was observed. The coefficients are not significant when lagged by 1 year but become significant determinants of migration, two, three, or four years prior to the observation period. This pattern was common to annual average measure including average precipitation (aprec) and average temperature (atemp). A lagged 142

158 response pattern was observed for 5 (25%) of 20 climate change predictors. Similar lagged response patterns have been observed in another study that used average annual precipitation (Hunter, Murray, and Riosmena, 2013). For these variables, households appear to delay their migration response, perhaps due to the employment of in situ strategies during the first years in line with multiphasic response theory (Davis and Lopez-Carr, 2010). In summary, the direct response pattern with a strong initial migration response and a decline in strength thereafter is the most common response pattern to climate change. Lagged response as well as continuous response patterns occur but less frequently. Socio-climatic Interactions To investigate differences in the climate change-migration association among various sub populations, I conducted an interaction screening. Only the interactions for which at least six out of 17 interaction coefficients were significant (34-100%, medium to robust evidence) were considered relevant and are presented below. Interactions for the first international move from rural areas were already discussed in Chapter IV. The results suggested that the effect of climate change on the first international outmigration from rural areas differed by the percentage of the male labor force employed in the agriculture sector, municipality-level wealth, and the prevalence of migrant networks. In contrast, for the last international move from rural areas, the percentage of households with domestic migration experience and the percentage of farmland irrigated, impacted the climate change migration association. 143

159 Table 6.4: Interaction between climate change indices and municipality domestic migrant prevalence predicting odds of the last international move from rural Mexico, CC Index Domestic migrants Interaction b sig. b sig. b sig. Temperature (high) No. summer days su Warm spell duration wsdi Warmest day txx Warmest night tnx * % warm nights tn90p Temperature (low) No. frost days fd ** Cold spell duration csdi Coldest day txn ** % cool nights tn10p % cool days tx10p Precipitation (high) No. days heavy precip r10mm No. days very heavy precip r20mm ** Wet spell duration cwd Max 5-day precip rx5day *** Precip extremely wet days r99ptot ** Total wet-day precip prcptot ** Precipitation (low) Dry spell duration cdd ** Note: Coefficients reflect log odds; one unit reflect a 10% change in domestic migrant prevalence; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; variables were not centered; * p<0.05; ** p<0.01; *** p<

160 Table 6.5: Interaction between climate change indices and the proportion of irrigated farmland, predicting the odds of the last international move from rural Mexico, CC Index Farmland irrigated Interaction b sig. b sig. b sig. Temperature (high) No. summer days su * ** Warm spell duration wsdi Warmest day txx Warmest night tnx % warm nights tn90p Temperature (low) No. frost days fd Cold spell duration csdi Coldest day txn % cool nights tn10p % cool days tx10p Precipitation (high) No. days heavy precip r10mm * No. days very heavy precip r20mm Wet spell duration cwd ** Max 5-day precip rx5day * Precip extremely wet days r99ptot * ** Total wet-day precip prcptot Precipitation (low) Dry spell duration cdd * Note: Coefficients reflect log odds; each row represents a full interaction model of which only the coefficients for the terms involved in the interaction are shown; variables were not centered; * p<0.05; ** p<0.01; *** p<0.001 To facilitate the interpretation of the observed interaction effect, I graphically visualize a selection of relationships that are representative of the conditional effect. If possible, I selected a temperature and a precipitation measure for each interaction. 145

161 Figure 6.1: Interaction between the number of frost days (a) and the number of days of very heavy precipitation (b) and the percentage of domestic migrant households in predicting the odds of the last international move from rural Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities The interactions between climate change and the percentage of domestic migrants residing in a particular municipality largely emerged for precipitation indices. Figure 6.1 shows that at rural locations with only few domestic migrants, an increase in the number of days with very heavy precipitation leads to a decline in the probability of a last move, similar to the main 146

162 effect observed for the first move. However, as the percentage of domestic migrants increases, the slope of the line becomes more positive and for municipalities with a high number of domestic migrants, an increase in the number of days of very heavy precipitation increases the probability of a consecutive (last) international move. Similarly, for households located in municipalities with only few domestic migrants, an increase in the number of frost days is associated with a decline in the probability to migrate. In contrast, when many households in the particular municipality have domestic migration experience, an increase in the number of frost days increases the probability of a last international move. Although the differences in slope are statistically significant, the confidence intervals overlap and the differences in the predicted probability for the 10 th and 90 th percentile of domestic migrant prevalence cannot clearly be distinguished. Only for areas with few domestic migrants, does the relationship between climate and the last international move mirror the relationship observed for the first international move, where an increase in precipitation and a cooling favorably impacts crop yield (Lobell and Field, 2007) and thereby leads to a decline in international outmigration. However, this relationship is inversed for areas with a large proportion of domestic migrants, where favorable agricultural conditions increase the probability of a last move. The measure of the percentage of households with domestic migrant experience captures both households that have sent a migrant to a different location within Mexico as well as in-migration to the particular municipality. 21 Hence, municipalities that have attracted large numbers of domestic in-migrants over the past five years may offer alternative employment opportunities in non-agricultural sectors. These other sectors, 21 The percentage of households with domestic migration experience was constructed using census data (IPUMS). The measure was computed as the percentage of households that had at least one member residing in a different administrative unit within Mexico five years prior to the census. 147

163 such as tourism (e.g., Amelung, Nicholls, and Viner, 2007) may be differently impacted by climate change and benefit from a warmer, dryer climate. Figure 6.2: Interaction between the number of summer days (a) and the precipitation on extremely wet days (b) and the proportion of the farmland irrigated in predicting the odds of the last international move from rural Mexico, Note: blue and red ribbons around lines represent 95% confidence intervals for the predicted probabilities 148

164 A second interaction was observed for climate change as driver of the last international move from rural areas and the extent to which the farmland in a particular municipality was irrigated. Similarly to the interaction with the domestic migrant network, the interactions with the irrigated farmland largely emerge for precipitation indices. Figure 6.2 shows that an increase in the number of summer days tends to increase the probability of a last international move at areas that have only limited or average irrigation capacity. However, when the cropland is highly irrigated the relationship reverses and an increase in the number of summer days is associated with a decline in outmigration. For rainfed farming, a further warming may negatively impact the crop yield, while a warming trend may be beneficially impacting the harvest through increased plant metabolism if enough water can be provided through irrigation systems (Challinor et al., 2007). However, the predicted probabilities were distinguishable only when the number of summer days was more than 0.5 standard deviations above the long term mean. Precipitation on extremely wet days is not associated with a change in migration responses in areas that have little to average irrigation systems in place. In contrast, for locations with extensive coverage of the farmed land by irrigation systems an increase in the precipitation during extremely wet days seems to be detrimental for the agricultural sector, leading to an increase in outmigration. Perhaps, the additional precipitation is not needed in these areas and may even lead to flooding and soil erosion with damaging effects on the harvest (Eakin and Appendini, 2008). However, an overlap in the confidence intervals suggests that the observed differences in irrigation coverage at the 10 th and 90 th percentile on migration probabilities is difficult to distinguish for the impact of precipitation. 149

165 Overall, these interactions suggest that the relationship between the last move and climate change is sensitive to the technological development level (c.f., Gutmann and Field, 2010), particularly the availability of irrigation systems. In addition, a decline in precipitation and an increase in temperature appear to drive a last international move largely in areas with low levels of domestic migrants - a measure that may serve as a proxy for the presence of pull factors such as other non-agricultural employment opportunities. In conclusion, this chapter contributes three key points to our understanding of the climate change migration association: (1) As hypothesized, the first international move is more strongly impacted by climate change as compared to the last international move from rural Mexico. This finding supports views that environmental drivers matter most for the first move, while later moves are stronger influenced by other factors, as suggested by the additive theory. (2) Investigating different lag times reveals a direct response pattern. Climate change effects on migration are strongest in the following year and the response weakens thereafter, suggesting a preference for migration in the face of climate change perhaps due to a lack of in situ options or the influence of cultural factors and social network effects. (3) The socio-climatic interactions reveal that the last international move is impacted by the prevalence of domestic migrants in the municipality as well as the percentage farmland irrigate. For example, a warming trend increased the probability of a last move predominantly for municipalities with high dependence on rain-fed farming and limited number of domestic migrants. 150

166 CHAPTER VII CONCLUSIONS Theoretical Contributions: Expanding the Environmental Dimensions of the Sustainable Livelihoods Framework This study was designed to investigate nuanced measurements of climate change and their impact on various migration streams in Mexico with the goal of advancing theoretical and empirical understanding of complex environment migration dynamics. The study's central hypotheses and results were presented in three chapters. Chapter IV examined the effect of high/low temperature and high/low precipitation extremes on migration responses and tested whether climate change more strongly impacts migration from rural or urban areas in Mexico. Chapter V expanded the investigation by exploring whether climate change more strongly impacts domestic compared to international moves. Finally, Chapter VI investigated whether the first or the last move within a household was more responsive to climate change. The results of hypotheses testing are summarized in Table 7.1. In addition, Figure 7.1 displays the expanded Sustainable Livelihoods framework, updated to reflect the study's main findings. As outlined in Chapter II (Scope of Study), this research estimated the direct relationship between driver (climate change) and outcome (migration) although the framework also includes components of economic sectors and livelihood capitals as mediators. Although unmeasured, these mediators are important for conceptual understanding of why and how certain climate change effects influence migration decisions. Because the mediator function of these conceptual components was not measured directly, this part of the framework remains unchanged from the original representation (Figure 2.1). Finally, it is important to note that the refined framework depicts the 151

167 climate change migration relationship for certain regions in Mexico during the period The relationships may differ for other temporal and spatial contexts. A globally generalizable framework of the climate change migration association would only be possible through a combination of results from multiple settings and time periods. Table 7.1: Summary of the hypothesis testing results to investigate the climate change migration association in Mexico, Chapter ID Topic Hypothesis description Hypothesis confirmed IV H4.1 high temperature An increase in high temperature Yes extremes (warming) increases international migration from Mexico IV H4.2 low temperature An increase in low temperature Yes extremes (cooling) decreases international migration from Mexico IV H4.3 high precipitation An increase in high precipitation Yes extremes (more wet) decreases international migration from Mexico IV H4.4 low precipitation An increase in low precipitation Unclear extremes (more dry) increases international migration from Mexico IV H4.5 rural vs. urban Climate change more strongly impacts international migration from rural compared to urban Yes V H5.1 international vs. domestic areas in Mexico Climate change more strongly impacts international compared to domestic migration from rural Mexico VI H6.1 first vs. last Climate change more strongly impacts first compared to last international moves from rural Mexico Yes Yes 152

168 Figure 7.1: Refined Sustainable Livelihoods framework with added information on the effect of climate change on migration dynamics for Mexico, Note: Thickness of the climate change impact arrows represent anticipated strength of the climate change effects on various sectors based on Boyd and Ibarraran (2009) (not measured in this study). The interconnectedness between various livelihood capitals is illustrated by dashed lines. 153