Migration as a response to climatic stresses and shocks in ......Migration as a response to climatic stresses and shocks in Bangladesh: a quantitative study Edgar Cooke, Maxmillan

Migration as a response to climatic

stresses and shocks in Bangladesh:

a quantitative study

Edgar Cooke, Maxmillan Martin and Dominic Kniveton

Working paper 3

An output of research on climate change related migration in Bangladesh, conducted by Refugee and Migratory Movements Research Unit (RMMRU), University of Dhaka, and

Sussex Centre for Migration Research (SCMR), University of Sussex, with support from

Climate & Development Knowledge Network (CDKN)

Copyright: RMMRU and SCMR, 2013 Refugee and Migratory Movements Research Unit Sattar Bhaban (4th Floor) 3/3-E, Bijoynagar, Dhaka-1000, Bangladesh. Tel: +880-2-9360338, Fax: +880-2-8362441 E-mail: [email protected], Web: www.rmmru.org | www.samren.net Sussex Centre for Migration Research School of Global Studies University of Sussex Falmer, Brighton BN1 9SJ, UK Tel: +44 1273 873394, Fax : +44 1273 620662 Email: [email protected], Web: www.sussex.ac.uk/migration About the authors: Edgar Cooke is a PhD student at the Department of Economics at the School of Business, Management and Economics, University of Sussex, UK Maxmillan Martin is a PhD student at the Department of Geography, School of Global Studies, University of Sussex, UK Dominic Kniveton is Professor of Climate Science and Society at the Department of Geography, School of Global Studies, University of Sussex, UK, Email: [email protected]

Migration as a response to climatic stresses and shocks in Bangladesh: a quantitative study Introduction Migration often becomes an important strategy for rural people engaged in primary modes of production to cope with sudden shocks as well as climatic stresses that develop over a longer period, undermining their livelihoods and safety. The whole process becomes complex moves within countries, as well as international mobility. Migration can be viewed as a livelihood strategy generating financial and social remittances to help build resilience to existing and future climate threats for those remaining in or returning to an origin location (McLeman and Smit 2006, Barnett and Webber 2009, Tacoli 2009, Foresight 2011, ADB 2012). While adaptation can be anticipatory or reactive, autonomous or planned, structural or non-structural, and in-situ or ex-situ (Fankenhauser et al. 1999, Smit et al. 1999, McCarthy et al. 2001, Bardsley and Hugo 2010) in terms of migration it tends to be something undertaken proactively and with a high degree of agency. It has been argued that designing effective public policy for facilitating adaptation necessitates an appreciation of the adaptation decision making process in order to create conditions under which people are more likely to make the most beneficial decisions and not ones that they will later come to regret (Gowda and Fox 2002, Grothmann and Patt 2003). Qualitative research in three climatic hazard-prone districts in Bangladesh has shown that while concern is commonly expressed of climate related pressures on livelihoods the decision to migrate is rarely attributed by those undertaking them to such shocks and stresses (Martin et al. 2013). In this paper we analyse quantitative data collected from five districts in Bangladesh to develop a model of migration decision making in the context of climate-related environmental stresses and shocks. We try to probe whether migration is increased in areas affected by such threats before examining the relationship between concern over and experience of such stresses and shocks as riverbank erosion, flooding, cyclones, drought and rainfall variability and migration behaviour. The research shows that the direction and strength of the relationship between climate risk perception and migration is sensitive to gender, geographical location and type of hazard and can be at odds with the relationship between the climate variables of rainfall and migration. The context Bangladesh is a country that faces both gradual onset climate stresses and sudden shocks1, including water shortage, cyclone, floods and coastal/ delta erosion (EM-DAT, 2012a, b), making it a climate change ‘hotspot’ (Haq, 2001, Huq and Ayers, 2008). During 1991-2010, Bangladesh was among the three countries – along with Myanmar and Honduras – most

1 For the purposes of brevity from hereinafter atmospheric-related extreme events and climate stresses will be referred to as climate stresses, threats and events.

affected by extreme weather events (Harmeling 2012). Bangladesh’s low altitude, exposure to cyclones and sea level rise, high population density and poverty levels with natural-resource-based livelihoods (Agrawala et al. 2003) make such impacts especially serious. Eighty percent of Bangladesh consists of floodplains of major rivers. These floodplains are highly prone to both inland flooding and riverbank erosion. In catastrophic flood years, such as 1988, 1998, and 2004 more than 60% of the country was inundated by flood waters for the duration of nearly 3 months (CEGIS, 2002). The economic cost of these events on Bangladesh is huge with estimated losses and damages from the 1998 event alone, of over $2 billion or 4.8% of GDP. It has been estimated that between 2,000 to 3,000 kilometres of river-bank experiences major erosion annually (Islam and Islam, 1985).

Figure 1. Flood extent for 1998 and upazilas prone to river bank erosion. Data from the Bangladesh Water Development Board. Bangladesh is also affected by cyclones nearly every year, with severe cyclones striking on average every 3 years. Inundation of sea water and the associated salinization of land and water resources are the largest adverse impacts resulting from the storm surges associated with these cyclones. Salinization of land not only affects agricultural production but can have a direct impact on the human health (Khan et al 2011). In terms of economic loses it has been estimated that a single event, Cyclone Sidr in 2007 resulted in damages and losses of $1.7 billion, or 2.6% of GDP.

Bangladesh is also a country with a long history of internal and international migration and mobility (Afsar 2003; Siddiqui 2003). In terms of stresses and migration Penning Rowsell et al. (2013) found that despite major hazards – mainly cyclones, floods, riverbank erosion and drought – people did not move out or migrate in large numbers, except in the case of riverbank erosion of their place of living or farms getting salinised by salty water flooding. Similarly, using event history analysis of a longitudinal dataset from 1994 to 2010, Gray and Mueller (2012) found that flooding and non-flooding-related crop failure were the most frequent and damaging events – climate-related or otherwise. Flooding had modest effects on mobility, whereas crop failure unrelated to flooding had a much more significant effect. Gray and Mueller also found that families in severely affected areas were likely to move, even though they themselves might not be directly affected. Drivers of individual migration: evidence from a retrospective longitudinal survey In this section we outline findings based on logistic regression modelling of factors hypothesised to actually drive migration at an individual level. These findings are based on an extensive retrospective longitudinal survey conducted in three districts of Bangladesh in 2011 – Satkhira in the south-east of the country, a coastal district south of Khulna city; Munshiganj in the centre of the country, just south of the capital Dhaka, and Nawabganj, in the north-west of the country on the border with West Bengal. These districts were chosen to be broadly representative of three major sets of environmental extremes facing Bangladesh – coastal flooding and cyclones (Satkhira); river bank erosion (Munshiganj) and drought (Nawabganj). The survey included a record for each individual interviewed of where they were living every year from age 6 to the time of survey, allowing analysis of change of residence – including migration from one district or one country to another – lasting a full year or longer. The survey included a series of modules, as well as a summary grid designed to aid recall. In addition to migration, the survey recorded summary information about a range of factors that might affect migration, including time variant factors such as age, family size, education, employment and well-being, as well as time invariant factors such as gender and parental background. The selection of sample variables in the survey was informed by a series of focus group discussions conducted in the country prior to the survey phase. Focus group discussions explored how local livelihoods changed over the past 40 years, climatic and non-climatic events during the period and their impacts, migration trends, as well as current and future concerns of the local people in terms of climate, safety and livelihoods. The questions about climatic stresses and shocks included flood, drought, riverbank erosion, salinity of the soil, cyclones, based largely on recent synthesis reports that dealt with projected impacts of climate change (Walsham 2010, Foresight 2011, World Bank 2011). The survey itself was based on a model provided by the MAFE project (Schoumaker et al. 2013), adapted to include environmental variables, and to refer to internal rather than international migration. As a large and complex survey, it was necessary to make comprises on some variables in order to prevent each interview from becoming overly long and/or losing the goodwill of

participants. Amongst these compromises, an important one concerned the decision not to collect income data, which is both complex and often unreliable. Instead, respondents were asked – for each time period – to indicate merely whether they viewed their financial circumstances as ‘mostly sufficient’, ‘sometimes sufficient’ or ‘mostly insufficient’; and also whether they considered their living conditions ‘better’, ‘equivalent’ or ‘less good’ than those living around them. Such subjective and relative definitions did however provide an indication of changing levels of well-being, and coupled with more detailed information on assets (land, houses, businesses and savings in the form of gold or jewelry) provided a basis to consider the economic drivers of migration. Methods and analysis The sample was restricted to first migrations to other regions in Bangladesh and individuals aged 18 years and above. First migrations were chosen as migration is a path-dependent process, in which first migration influences the likelihood, timing and direction of subsequent migrations (Balaz and Williams, 2007) , and our intention was to isolate other factors in the migration decision. The data consists of 1,492 subjects tracked on average over a 22-year period (20 and 75 years are the median and maximum number of years respectively), leading to a person-period dataset of 34,311 records. Each line of data represents a year in the sample. To understand the factors behind first migrations in Bangladesh we adopt a discrete event history model. We model the individuals at risk using a logit model:

( )

) )

Where ) ) is the probability that individual i experienced a first internal migration conditioned on a set of covariates. The set of covariates include household characteristics, individual characteristics, experiences of weather conditions, concerns about weather conditions and rainfall data. This set of variables includes time invariant and time varying covariates. The baseline hazard function is which is specified as the log of time spent in the risk set (time prior to the migration event occurring). The logit is specified in the log-odds form; however, the odds ratio form is used in the analysis and reported at the end of the document to aid the interpretation of the results. The odds ratio form is obtained by taking the exponent of the coefficients. Thus, the odds ratio can be written as the ratio of the probabilities that the migration event occurred to the probability that the migration event did not occur:

) )

) )

In odds form, if the ratio is larger than one then the factor of interest raises the hazard of the first internal migration occurring. A ratio less than one indicates that the factor reduces the hazard of the first migration occurring.

Table 1 (see appendix) presents the descriptive statistics of the key variables used in the analysis. The mean values are shown by the different gender categories and for everyone. The mean length of time elapsed prior to first migration is 11.5 years for females, and 19.3 years for males, with the overall mean being 17.6 years. As a result, there are more male person-period observations than female person-period observations: of the 33,699 non-missing observations of gender, approximately 78% are males. The mean age for migrants is 37.6 years while the mean ages for male and female migrants are 38 and 36.8 years respectively. The mean family size for both males and females are 4-5 individuals as embodied in the mean value of 2 recorded in the table. The following variables were included in all regressions reported in the results section to control for the various household characteristics that affect first migration: age, family size, educational attainment of respondents and parents, employment status of father, presence of assets in the household, presence of migrant networks, sufficiency of income, birth of first child and death of first child. Each of these variables were expected to have an influence on the decision to migrate. For example Stark and Lucas (1988) have argued that smaller families overcome their constraints in size by broadening their relevant space. Migration of a family member also allows pooling and reduction of risks. While large family size places more demand for resources and thus migration, sometimes the family size acts against migration due to the costs of travel and the close ties involved. Migration represents an effective risk mitigation strategy (Halliday 2008, Lueck 2011). Similarly, the decision to migrate is more likely when individuals have migrant networks present (2008, de Haas 2010) and/or sufficient income to support themselves through the journey or for the first few days at destination. To address environmental factors as possible causal influences on migration, we included two distinct sets of variables: first, self-reported experiences of floods, cyclones, erosion and droughts in each place of residence, as well as self-reported concerns about these events; and second, measured rainfall data obtained from meteorological stations in each location, from which we also calculated rainfall variability as a covariate. As noted earlier, rainfall and harsh weather conditions have been linked to migration events (Henry et al., 2004). Here, rainfall is measured as the average rainfall precipitation over the whole period that data is available2. Henry et al (2004) and Henry and Dos Santos (2012) note that mean rainfall for longer periods of time captures long term weather conditions such as vulnerability to drought as well as agricultural productivity. As a result one of the explanatory variables is measured as the mean rainfall for the period rainfall data is available. Three categories of mean rainfall are included in the study. These categories are less than 100mm of rainfall per day, 100-149mm and greater than149 mm. These categories are informed by the data available on rainfall and by the classification of extreme rainfall events for India by Rajeevan et al. (2008), who consider rainfall between 5mm and 100mm per day to be

2 Recorded rainfall data is not available for all districts hence, regressions using recorded rainfall data are for a

smaller sample. The stations for which rainfall data is available include Dhaka, Khulna, Rajshahi and Satkhira (for

1970 – 2012). Additional rainfall data is available for Munshiganj and Nawabganj for the period 1980 –2012.

Divisions for which rainfall data were not available for the analysis include Chittagong, Barisal, Rangpur and Sylhet

moderate rainfall, 100-150mm to be heavy rainfall and 150mm and above per day to be very heavy rainfall. Observed rainfall patterns show that mean rainfall recorded is mostly below 50mm for the period for most areas with low to moderate rainfall patterns. Additionally, we do observe a number of regions recording heavy rainfall (100-150mm). Few areas such as the Satkhira region are observed to have mean rainfall larger than 150mm implying very heavy rainfall. In turn, rainfall variability was constructed as:

Rainfall variability captures the extent to which rainfall for period i departs from the mean rainfall over the period (Henry et al, 2004). Three categories of rainfall variability were created – less than 85%, 85–95% and above 95%. Overall, it was expected that individuals living in districts with droughts, cyclones or floods would be more likely to migrate than people living in areas where these weather conditions are absent. Observed rainfall data were collected to augment the results of the self-reported data. Rainfall events greater than 100mm or 150mm were expected to be more likely to lead to migration compared to moderate rainfall activity. Those experiencing mean rainfall greater than 150mm were found to have the highest hazard of migrating, while the lowest hazard was obtained by the category of rainfall between 100mm and 150mm. The hazard functions by rainfall variability are much closer across the categories compared to those of the mean rainfall hazards shown in figure 4. Those experiencing rainfall variability larger than 95% were found to have the highest hazard of migrating while those experiencing rainfall variability between 85% and 95% had the lowest hazard.

Figure 4: Hazard functions by Mean Rainfall

Figure

5: Hazard functions by Rainfall Variability

Regressions were carried out by gender and results were also pooled for both genders. The reason for the disaggregation by gender was to accommodate differences in causes of migration observed for example by Henry et al. (2004) and Henry and Dos Santos (2012). Additionally, estimating a single regression for both genders assumes that all the covariates included in the regression are the same for both genders. This assumption can be tested and if rejected would require regressions to be estimated separately for each gender. A t-test of the differences in the means of migration across gender returned a significant difference between males and females (t statistic=10.58). Furthermore, age, education, employed, concerns and experiences of floods, and rainfall variability reported significant t-tests indicating that the means across gender are not equal for the variables mentioned. This indicates that estimating regressions separately for each gender is the appropriate approach to adopt.3 Another issue to be wary of in estimating the regressions is unobserved heterogeneity referred to in the event history literature (for example, Mills, 2011) as frailty. Given that the data on individuals in the dataset are in panel form there might be unobservable factors determining their propensity to migrate. The event history literature recommends a number of ways of dealing with this phenomenon. One way to control for frailty is by assuming that it is normally distributed—this can be done by estimating a random effects logit. The random effects logit assumes that the unobserved heterogeneity are normally distributed and all estimates are conditional on this distribution of unobserved heterogeneity. All tables in the paper are based on a pooled logit whereby no assumptions have been made about the distribution of any unobserved heterogeneity present in the data.4 The logarithm of time is used to model the baseline hazard function. Other specifications that were tested but not reported include a quadratic of time and dummies representing each time interval. There were not many differences between the log of time and the quadratic of time as the baseline hazard. In all regressions, the baseline hazard was significantly different from zero and positive indicating positive duration dependence (the hazard increases with time).

Results of the event history analysis In table 2, all weather events are incorporated into the regression. The tables indicate that some weather phenomena do increase the hazard of migrating. These results however differ across gender. All dummies are interpreted in terms of the reference categories indicated. Also, all covariates are assumed constant while allowing the covariate being interpreted to vary as is usual of interpretation of multiple variable regressions. In terms of rainfall variability, females do not report any significant coefficients. Males on the other

3 An alternative test is to include interactions of gender with each covariate to the model and estimate an F-test for all interactions. The null hypothesis in this case would be the pooled regression and the alternate hypothesis would be to estimate separate regressions for each gender. A rejection of the null would indicate that pooling the coefficients would not be appropriate. This procedure produced the same conclusion as the t-test reported here. 4 Analysis was also conducted where unobserved heterogeneity was accounted for in the estimations. The results did not differ significantly from the pooled results.

hand, do report significant coefficients for the category 85 – 95%. The odds ratio is less than one, indicating that the odds of migrating is lower for this category compared to the less than 85% category for males and the “All” columns. Mean rainfall between 100 and 150mm have lower odds of migrating compared to mean rainfall less than 100mm. On the contrary, mean rainfall larger than 150mm report odds indicating that this category raises the hazard of migrating compared to the less than 100mm category. The result shows that very heavy rainfall tends to increase the hazard of migrating from the district of residence. Experiences of cyclones and droughts lower the odds of migrating for males while experiences of river bank erosion lowers the odds for females. Contrarily, concerns about river bank erosion among males raises the odds of migrating. Concern about cyclones rather lowers the odds for females. Large family sizes (>5) tends to raise the odds for females to migrate compared to smaller family sizes (1-3). This does not seem to be a significant predictor of male migration patterns. The next table, table 3 takes the analysis of the weather experiences further by excluding the measured rainfall variables (that is rainfall variation and the mean rainfall categories) and concerns about the weather. The various weather experience dummy variables are then introduced separately in the various regressions in the table. Family size is now a predictor of male migration. Both medium and large families now show increased odds of migrating compared to smaller families. Furthermore, the employment status of the respondent’s father is also significant in the case where the father is “self-employed/employer” and the “other” category compared to being an “employee”. This is the case for females but does not seem to matter for males. Again, river bank erosion raises the odds of migrating for males. The remaining significant odds ratios indicate a reduction in the odds of migrating for the other experiences. Table 4 focuses on the concerns about the weather. Concerns about erosion again raise the odds of migrating for males but is insignificant in the case of females. Concerns about cyclones just as those reporting experiences about cyclones lower the odds of migrating. The last but one table, table 5 reports the results for the rainfall data obtained from the meteorological stations. Two definitions of rainfall variability and mean rainfall for the period are used. The first definition of rainfall variability used is a year’s lag while the second definition measures variability in terms of the second lag of mean annual rainfall. The second definition incorporates the log of the mean rainfall for the period instead of the three defined categories. Both measures of rainfall precipitation are significant in most cases. The category of rainfall above 150mm raises the odds of migration compared to the category less than 100mm for males. The second category 100mm – 149mm lowers the odds of migrating for both males and females compared to the base category. The logged mean rainfall tends to follow the same pattern. Increases in long-term rainfall lower the odds of migrating from the district of residence for both genders. This might be attributable to the fact that small increments in long-term rainfall might increase agricultural yield for respondents who own or work on farms (see for example, Henry et al., 2004). Additionally, the community also benefits from any increased agricultural yield. Rainfall variability for

the second category (85-95%) lowers the odds of migrating for men. In the table, having sufficient income is now significant for males. However, it lowers the odds of migrating for men with no effect for women in terms of migrating behaviour. Environmental drivers of migration in different districts The final table, table 6 in the appendix provides a regional breakdown of the results to allow for the unpacking of any spatial differences in the response to climatic conditions. This is done for the three main districts—Munshiganj (flood and erosion), Satkhira (cyclones) and Nawabganj (drought). The impact of climatic conditions on migration in these districts might, therefore vary by the main conditions prevalent in the districts. Looking first at Munshiganj, a flood and erosion prone district, as for the three districts overall, concern about floods is more important for males than for females, and raises the odds of migrating. This is understandable given that the area is prone to floods and erosion. However, experiences of floods and other weather conditions are not significant, whilst for females, concerns about cyclones do not raise the odds of migrating but rather lower them. Turning to rainfall data, rainfall variability (85-95%) in the Munshiganj area is seen to reduce the odds of males migrating compared to rainfall variability less than 85%5. Turning to Satkhira, a cyclone prone district, unlike Munshiganj, rainfall variability is not a significant predictor of migration. In contrast, both concerns about cyclones, and the experience of them, are significant predictors of first migration, but only for men, and somewhat surprisingly, both reduce rather than increase the odds of men migrating. Mean rainfall also report some significant coefficients. In this region, mean rainfall in the 100-149mm range lowers the odds of migrating for males compared to mean precipitation larger than 149mm. Again, the measured rainfall variables have no effect on women in the sample. In addition, and somewhat surprisingly, although Satkhira is a region not heavily affected by riverbank erosion or drought, experience of either lowers the odds of migrating, whilst concerns about the former also reduce the odds of migrating. Finally, in Nawabganj a drought prone district, rainfall variability is unsurprisingly significant, although again the effect is gender-specific, and variability over 95% reduces rather than increases the odds of migrating compared to variability lower than 85% for men. For women, the rainfall coefficients are not significant. Meanwhile, mean rainfall lower than 100mm does raise the odds of migrating for men compared to mean rainfall between 100mm and 150mm, all things equal, but again it has no impact on women. However, women concerned about floods have higher odds of migrating relative to those who have no concerns about floods. The remaining concerns and experiences about the weather are not significant. Thus in terms of the rainfall variables, rainfall variability has an impact in the flood and drought prone districts (that is, Munshiganj and Nawabganj respectively), but it does not play a significant role in migration in the cyclone prone region of Satkhira. Measured

5 Mean rainfall is less than 100mm in this region, so due to the lack of variation in the rainfall precipitation measure across the sub-districts, we are unable to estimate a coefficient for rainfall categories.

rainfall on the contrary has an impact in both Satkhira and Nawabganj. In Munshiganj we are unable to estimate a coefficient for the mean rainfall variable since there is no variation in the recorded mean rainfall figures. For the estimated coefficients, moderate rainfall in Satkhira lowers the odds of men migrating compared to heavy rainfall. In Nawabganj, low rainfall raises the odds of men migrating compared to moderate rainfall. One of the key findings which is corroborated in the literature is the differential impact of rainfall on migration by men and women. Rainfall and weather conditions tend to influence migration among men to a much greater extent than it does among women. Figure 6 below corroborates the evidence presented in the tables. Munshiganj and Nawabganj both have higher hazards for migrating compared to Satkhira. The evidence seems to point to floods and droughts being the main climatic conditions increasing the likelihood of migratory events. Satkhira a cyclone prone area shows a much lower hazard for migrating. This implies migration from the Satkhira region is not driven strongly by climatic conditions and that inhabitants might have adjusted to living with this weather condition.

Figure 6 Baseline hazard by areas Discussion This paper has adopted a multi-method and time-sensitive approach to environmental influences on migration in Bangladesh, demonstrating that whilst environmental factors are correlated with internal migration over time, their effect is gender and region specific, and that different environmental hazards have different effects. The paper illustrates the value of an integrated approach, in which environmental factors are considered alongside

the perception of climate risks and non-environmental influences on migration. Whilst some findings vis-a-vis the relationship between climate related threats and migration are intuitive, others are surprising, notably that fears of, and the experience of cyclones actually reduce rather than increasing the propensity to migrate. This may reflect a sense of fatality in the face of extreme hazard risk, or alternatively may be due to other factors not tested in the model – notably concerns to assert or maintain title to land Furthermore when one compares how people appear to respond to concerns of climate related threats this appears at times to be contradictory to the relationship between their experience of the climate related stresses and shocks and migration. For example while women are statistically more likely to migrate if they have concerns about riverbank erosion they are less likely to migrate if they have experienced riverbank erosion. . While in Munshiganj concern over floods seems to have the opposite effect to a low variability in rainfall in terms of male migration. A highly probable explanation of this latter finding is that the damage that the likelihood of riverbank erosion is poorly related to local rainfall conditions, instead representing hydrological conditions higher up the catchment. One explanation of the variation between the relationship between migration and concern of a riverbank erosion and that of migration and experience of the phenomena is that people are acting adaptively before an event by migrating but are unable to do so once the event has struck. However this explanation ignores first why the relationships are not the same for all climate related threats. For example while migration is likely to increase with a fear of floods and riverbank erosion it decreases for a concern of cyclones. Second, evidence from focus group discussions described in Martin et al (2013) of the same areas reveals that while climate threats are acknowledged by individuals the decision to migrate is seldom attributed to such events. Third, given that the migration discussed in this study includes all migration types there remains the question of whether the ability to migrate is really restricted by a lack of assets.

References Adger, W. N., Jon Barnett, F. S. Chapin III, and Heidi Ellemor (2011) This Must Be the Place: Un-derrepresentation of Identity and Meaning in Climate Change Decision-Making, Global Environ-mental Politics 11(2), May 2011. Afsar R, (2003) Internal migration and the development nexus: the case of Bangladesh', paper presented at the Regional Conference on Migration, Development and Pro-Poor Policy Choices in Asia, 22 -24 June, Dhaka; Refugee and Migratory Movements Research Unit, Dhaka, and UK Department for International Development, London Asian Development Bank (2011) Climate Change and Migration in Asia and the Pacific Manila: Asian Development Bank, Balá ž, V. and Williams, A.M. (2007) Path-dependency and Path-creation Perspectives on Migration Trajectories: The Economic Experiences of Vietnamese Migrants in Slovakia, International Migration Vol. 45 (2): 37 – 66 Barnett, J. and Webber, M. (2010). "Accommodating Migration to Promote Adaptation to Climate Change." World Bank Policy Research Working Paper: no. WPS 5270, Apr 2010. Washington DC: World Bank de Haas, H (2010) The Internal Dynamics of Migration Processes: A Theoretical Inquiry, Journal of Ethnic and Migration Studies 36(10): 1587 - 1617 EM-DAT (2012a) Top 10 Natural Disasters in Bangladesh for the period 1900 to 2012, sorted by numbers of total affected people, The OFDA/CRED International Disaster Database, Université Catholique de Louvain, Brussels, available at http://www.emdat.be/result-country-profile#sumtable accessed on November 8, 2012. EM-DAT, (2012b) Disaster data of Bangladesh, The OFDA/CRED International Disaster Database, Université Catholique de Louvain, Brussels. http://www.emdat.be/database accessed on November 8, 2012 Garip (2008) Social Capital and Migration: How do Similar Resources Lead to Divergent Outcomes? Demography 45(3): 591–617. Gray, C, and Mueller. V (2012) Natural disasters and population mobility in Bangladesh. Proceedings of the National Academy of Sciences. Halliday, T.J. (2008) : Migration, risk and the intra-household allocation of labor in El Salvador, IZA Discussion Papers, No. 3322

Henry, S., Schoumaker, B., & Beauchemin, C. (2004). The impact of rainfall on the first out-migration: A multi-level event-history analysis in Burkina Faso. Population and Environment, 25, 423–460. Henry, S., & Dos Santos, S. (2012). Rainfall Variations and Child Mortality in the Sahel: Results from a Comparative Event History Analysis in Burkina Faso and Mali. Population and Environment 34 (4) (May 25): 431–459. doi:10.1007/s11111-012-0174-4. Lueck, M.A.M. (2011) United States Environmental Migration: Vulnerability, Resilience, and Policy Options for Internally Displaced Persons, in Leighton, M. Shen, X. and Warner, K (eds) Climate Change and Migration: Rethinking Policies for Adaptation and Disaster Risk Reduction, Bonn: UNU Martin, M., Billah, M., Siddiqui, T., Kniveton, D., Black, R (Forthcoming) ‘Climate-related migration in rural Bangladesh: a behavioural model’, Population and Environment Mills, M. (2011). Introducing Survival and event history analysis, SAGE publications. Rajeevan, M, Jyoti Bhate & A.K. Jaswal (2008). Analysis of variability and trends of extreme rainfall events over India using 104 years of gridded daily rainfall data, Geophysical Research Letters, Vol. 35, L18707, doi:10.1029/2008GL035143. Rammel, C, Sigrid Stagl and Harald Wilfing (2007) Managing complex adaptive systems — A co-evolutionary perspective on natural resource management, Ecological Economics 63 (1): 9 - 21 Stark, O and Lucas, R E B (1988) Migration, Remittances, and the Family, Economic Development and Cultural Change, 36 (3): 465-481 Stern, N. (2006). "Stern Review on The Economics of Climate Change (pre-publication edition). Executive Summary". HM Treasury, London Walsham, M. (2010) Assessing the Evidence: Environment, Climate Change and Migration in Bangladesh. Geneva: International Organization for Migration. World Bank (2011) The Cost of Adapting to Extreme Weather Events in a Changing Climate, Bangladesh Development Series. Paper No. 28, Washington DC: The World Bank Appendix Table 1: Descriptive Statistics of Key Variables Variable Description Female Male Total

(mean) (mean) (mean)

Dependent Variable

Migration First Internal Migration (Yes/No) 0.06 0.03 0.04

Covariates

_t Time 11.50 19.31 17.56

Q1 Q1: Respondent Gender (Male/Female) 0 1 0.78

newfamsize Family Size (1: 1-3 2: 4-5 3: >5) 2.10 2.12 2.12

feduc Father’s Education (1: Yes 0: No) (Base: No) 0.24 0.26 0.26

meduc Mother’s Education (Yes/No) 0.10 0.13 0.12

education Education (Yes/No) 0.40 0.55 0.52

newQ8 Father’s Employment Status (1: Employee; 2: Self; 3: Other) (Base: Employee)

2.04 2.04 2.04

age Age of Respondent 36.8 40.8 39.9

ln_age Age (in logs) 3.49 3.58 3.56

ln_agesq Age squared (in logs) 6.98 7.16 7.12

birth_child Has first child? (Yes/No) 0.96 0.87 0.89

death_child Has first child died? (Yes/No) 0.03 0.02 0.02

employed Employed? (Yes/No) 0.45 0.63 0.59

netwrk Is Migrant Network Present? (Yes/No) 0.65 0.63 0.63

newQ405 Sufficient Income (Yes/No) 0.65 0.57 0.59

asset Asset Present? (Yes/No) 0.19 0.51 0.43

exp_floods Q308: Experienced Floods (Yes/No) 0.74 0.65 0.67

exp_erosion Q308: Experienced river bank erosion (Yes/No) 0.54 0.53 0.53

exp_cyclones Q308: Experienced cyclones (Yes/No) 0.50 0.47 0.47

exp_droughts Q308: Experienced droughts/lack of rain (Yes/No) 0.19 0.36 0.32

con_floods Q310: Concerned about floods (Yes/No) 0.53 0.44 0.46

con_erosion Q310: Concerned about river bank erosion (Yes/No) 0.42 0.44 0.44

con_cyclones Q310: Concerned about cyclones (Yes/No) 0.33 0.32 0.32

raincat_1 Mean Rainfall: < 100mm (Base category) 0.41 0.61 0.57

raincat_2 Mean Rainfall: 100 – 149mm 0.58 0.39 0.43

raincat_3 Mean Rainfall: 150mm and above 0.01 0.00 0.01

variabilitya_1 Rainfall Variability (1 year lag): < 85% (Base Category) 0.25 0.24 0.24

variabilitya_2 Rainfall Variability (1 year lag): 85 – 95% 0.21 0.17 0.18

variabilitya_3 Rainfall Variability (1 year lag): Over 95% 0.54 0.59 0.58

Table 2: Logit Estimates of First Migration: Rainfall Precipitation, Weather Experience & Concerns (1) (2) (3) (4) (5) (6) (7) (8) (9) Males Males Males Female

s Female

s Female

s All All All

Baseline Hazard:

Time (in logs)

6.473*** 6.647*** 6.642*** 7.723**

* 7.169**

* 7.769**

* 5.953*

** 5.942*** 5.983***

(1.114) (1.004) (1.177) (1.275)

(1.230)

(1.301)

(0.661)

(0.599) (0.667)

Rainfall Variability: 1 year lag (Base: <85%)

85 - 95% 0.697* 0.738* 0.726* 0.942 0.979 0.963 0.764* 0.808 0.784* (0.102) (0.108) (0.107) (0.177

) (0.182

) (0.181

) (0.088

) (0.093) (0.090)

Over 95% 0.853 0.841 0.862 1.001 1.023 1.019 0.875 0.875 0.878 (0.100) (0.097) (0.102) (0.156

) (0.161

) (0.160

) (0.082

) (0.081) (0.082)

Mean Rainfall (Base:

<100mm) 100-149mm 0.669* 0.534**

* 0.722 0.760 0.783* 0.728*

(0.108) (0.098) (0.124)

(0.125)

(0.090) (0.095)

150mm and above

11.17** 10.16** 1.753* 1.460 4.565***

3.497***

(8.385) (9.069) (0.499)

(0.393)

(1.515) (1.255)

Weather Experience

Q308: Experienced Floods (Base: No)

0.878 0.919 0.993 1.103 0.862 0.958

(0.133) (0.109) (0.206)

(0.175)

(0.105)

(0.092)

Q308: Experienced river bank erosion (Base: No)

0.891 1.150 0.493** 0.519**

* 0.730* 0.897

(0.169) (0.162) (0.119)

(0.077)

(0.101)

(0.086)

Q308: Experienced cyclones (Base: No)

0.607** 0.619** 0.628* 0.680* 0.669*

* 0.599***

(0.095) (0.105) (0.134)

(0.116)

(0.085)

(0.077)

Q308: Experienced droughts/lack of rain (Base: No)

0.525*** 0.470*** 0.729 0.717 0.559*

** 0.531***

(0.100) (0.087) (0.136)

(0.137)

(0.069)

(0.067)

Concerns about the Weather

Q310: Concerned about Floods (Base: No)

1.089 1.180 1.103 1.205 1.127 1.177*

(0.153) (0.114) (0.218)

(0.178)

(0.123)

(0.094)

Q310: Concerned about river bank erosion (Base: No)

1.462* 1.830*** 1.138 0.729 1.352* 1.322***

(0.238) (0.178) (0.309)

(0.129)

(0.179)

(0.104)

Q310: Concerned about cyclones (Base: No)

0.554*** 0.506*** 0.898 0.694 0.647*

* 0.560***

(0.087) (0.081) (0.240 (0.152 (0.089 (0.072)

) ) ) Q1 - Respondent Gender

1 1 1 1 1 1 0.316*

** 0.306*** 0.311***

(.) (.) (.) (.) (.) (.) (0.032)

(0.031) (0.033)

Family Size (Base: 1-3)

Family Size: 4-5

1.188 1.182 1.176 1.174 1.215 1.155 1.167 1.201 1.173

(0.130) (0.129) (0.128) (0.266)

(0.276)

(0.266)

(0.114)

(0.119) (0.115)

Family Size: >5

1.021 1.038 0.976 1.911** 1.910** 1.881** 1.203 1.244* 1.165

(0.124) (0.125) (0.118) (0.377)

(0.379)

(0.374)

(0.115)

(0.120) (0.112)

Father's Education (Base: No)

0.937 0.964 0.989 0.856 0.813 0.886 0.894 0.907 0.921

(0.107) (0.117) (0.111) (0.153)

(0.159)

(0.163)

(0.088)

(0.092) (0.092)

Mother's Education (Base: No)

0.940 0.813 0.795 1.242 1.209 1.273 1.018 0.954 0.966

(0.149) (0.135) (0.132) (0.288)

(0.285)

(0.294)

(0.132)

(0.125) (0.124)

Father’s Employment Status (Base: Employee)

Fathers employment status: Employer/Self-Employed

1.342 1.198 1.315 0.563** 0.526** 0.571** 0.973 0.904 0.964

(0.263) (0.225) (0.253) (0.114)

(0.112)

(0.120)

(0.147)

(0.133) (0.145)

Fathers employment status: Other

1.310 1.089 1.260 0.532* 0.551* 0.598 0.856 0.795 0.931

(0.300) (0.252) (0.295) (0.141)

(0.156)

(0.166)

(0.158)

(0.148) (0.170)

Age (log) 0.0501*

** 0.0451*

** 0.0501*

** 0.694 0.673 0.743 0.103*

** 0.0945*

** 0.103***

(0.022) (0.017) (0.022) (0.270)

(0.279)

(0.297)

(0.029)

(0.024) (0.029)

Has first child? (Base: No)

1.046 1.097 1.020 1.061 0.984 1.062 0.875 0.890 0.857

(0.143) (0.148) (0.140) (0.335)

(0.313)

(0.323)

(0.102)

(0.102) (0.100)

Has first child died? (Base: No)

0.980 1.127 0.944 1.379 1.465 1.560 1.181 1.263 1.260

(0.369) (0.427) (0.350) (0.667)

(0.755)

(0.759)

(0.332)

(0.373) (0.335)

Migrant Network

1.037 1.064 1.032 0.793 0.841 0.818 0.945 0.999 0.971

Present? (Base: No) (0.092) (0.094) (0.092) (0.130

) (0.148

) (0.143

) (0.073

) (0.077) (0.076)

Sufficient Income (Base: No)

0.980 0.943 0.947 1.089 1.131 1.063 0.964 0.932 0.944

(0.089) (0.086) (0.087) (0.185)

(0.184)

(0.185)

(0.072)

(0.070) (0.071)

Asset Present? (Base: No)

1.011 0.935 1.041 1.139 0.974 1.105 0.960 0.867 0.968

(0.092) (0.086) (0.095) (0.212)

(0.183)

(0.217)

(0.078)

(0.069) (0.078)

Observations 16137 16137 16137 4733 4733 4733 20870 20870 20870 Clusters 776 776 776 344 344 344 1120 1120 1120

Exponentiated coefficients (Odds ratios reported); Clustered standard errors in parentheses. * p < 0.05, ** p < 0.01, *** p < 0.001 Table 3: Logit Estimates of First Migration: Harsh Weather Experience (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Male

s Male

s Male

s Male

s Fema

les Fema

les Fema

les Fema

les All All All All

Baseline Hazard

Time (in logs)

4.071***

4.124***

4.311***

4.111***

7.321***

7.991***

7.857***

7.267***

4.568***

4.526***

4.820***

4.544***

(0.377)

(0.385)

(0.406)

(0.394)

(1.227)

(1.322)

(1.294)

(1.220)

(0.339)

(0.333)

(0.361)

(0.338)

Weather Experience

Q308: Experienced Floods

0.829*

0.806 0.812**

(0.067)

(0.110)

(0.057)

Q308: Experienced river bank erosion

1.320***

0.582***

1.025

(0.109)

(0.077)

(0.070)

Q308: Experienced

0.514***

0.639**

0.576***

cyclones (0.04

7) (0.09

8) (0.04

5)

Q308: Experienced droughts/lack of rain

0.722***

1.176 0.836*

(0.068)

(0.178)

(0.065)

Observations

21437

21437

21437

21437

6233 6233 6233 6233 27670

27670

27670

27670

Clusters 841 841 841 841 380 380 380 380 1221 1221 1221 1221 Odds ratios reported; Clustered standard errors in parentheses; The coefficients for family size, father’s education, mother's education, father's employment status, age (in logs), has first child, has first child died, migrant network present, sufficient income, asset present as suppressed here. These are the same variables as defined in Table 1. * p < 0.05, ** p < 0.01, *** p < 0.001

Table 4: Logit Estimates of First Migration: Concerns about the Weather (1) (2) (3) (4) (5) (6) (7) (8) (9) Males Males Males Female

s Female

s Female

s All All All

Baseline Hazard

Time (in logs)

4.048**

* 4.196**

* 4.296**

* 7.285*** 7.459*** 7.866*** 4.536**

* 4.552**

* 4.810**

* (0.371) (0.399) (0.404) (1.228) (1.304) (1.269) (0.335) (0.336) (0.359) Concerns about Weather

Q310: Concerned about Floods

0.955 0.860 0.928

(0.078) (0.122) (0.065) Q310: Concerned about river bank erosion

1.498**

* 0.724 1.181*

(0.126) (0.121) (0.085) Q310: Concerned about cyclones

0.481**

* 0.611* 0.550**

*

(0.049) (0.119) (0.050) Observations

21437 21437 21437 6233 6233 6233 27670 27670 27670

Clusters 841 841 841 380 380 380 1221 1221 1221

Odd ratios reported; Clustered standard errors in parentheses; The coefficients for family size, father’s education, mother's education, father's employment status, age (in logs), has first child, has first child died, migrant network present, sufficient income, asset present as suppressed here. These are the same variables as defined in Table 2. * p < 0.05, ** p < 0.01, *** p < 0.001

Table 5: Logit Estimates of First Migration: Rainfall Precipitation (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Male

s Male

s Male

s Male

s Fema

les Fema

les Fema

les Fema

les All All All All

Baseline Hazard

Time (in logs)

5.957***

6.149***

5.734***

5.600***

6.688***

6.879***

6.546***

6.826***

5.672***

5.873***

5.516***

5.444***

(0.864)

(0.976)

(0.832)

(0.722)

(1.124)

(1.187)

(1.103)

(1.140)

(0.566)

(0.638)

(0.551)

(0.499)

Rainfall Variability 1 year lag (Base: <85%)

85 - 95% 0.798

0.784

0.968 0.972 0.844

0.834

(0.118)

(0.116)

(0.178)

(0.178)

(0.097)

(0.095)

Over 95%

0.818

0.814

1.030 1.040 0.873

0.873

(0.096)

(0.096)

(0.160)

(0.162)

(0.081)

(0.081)

Mean Rainfall (Base: <100mm)

100-149mm

0.507***

0.493***

0.499***

0.646**

0.654**

0.655**

0.589***

0.583***

0.582***

(0.054)

(0.052)

(0.053)

(0.101)

(0.105)

(0.101)

(0.051)

(0.051)

(0.050)

150mm 9.63 10.4 9.12 1.424 1.518 1.337 3.17 3.33 2.99

and above

7** 6** 0** 6*** 2** 7***

(7.439)

(8.248)

(6.644)

(0.363)

(0.424)

(0.330)

(1.114)

(1.229)

(0.988)

Q1 - Respondent Gender

1 1 1 1 1 1 1 1 0.283***

0.283***

0.280***

0.281***

(.) (.) (.) (.) (.) (.) (.) (.) (0.028)

(0.028)

(0.028)

(0.027)

Rainfall Variability 2 year lag (Base: <85%)

85 - 95% 1.092

0.633*

0.883

(0.157)

(0.114)

(0.099)

Over 95%

0.819

0.770 0.789*

(0.107)

(0.117)

(0.077)

ln_mnrain

0.739***

0.827*

0.796***

(0.038)

(0.061)

(0.033)

Observations

16137

15632

16137

16612

4733 4551 4733 4906 20870

20183

20870

21518

Clusters 776 771 776 784 344 340 344 352 1120 1111 1120 1136 Odds ratios reported; Clustered standard errors in parentheses; The coefficients for family size, father’s education, mother's education, father's employment status, age (in logs), has first child, has first child died, migrant network present, sufficient income, asset present as suppressed here. These are the same variables as defined in Table 2. * p < 0.05, ** p < 0.01, *** p < 0.001

Table 6: Logit estimates of first migration: Munshiganj, Satkhira and Nawabganj

Munshiganj Satkhira Nawabganj

1 2 3 4 5 6 7 8 9 10 11 12

Male Male

Female

Female

Male Male Female

Female

Male Male Female

Female

Baseline Hazard

Time (log)

9.530***

8.739***

91.12***

82.24***

5.542***

7.677***

5.172***

5.825***

9.812***

9.340***

26.64***

26.38***

-1.91

7

-1.59

4

-55.184

-51.601

-1.58

5

-2.308

-0.96

6

-1.02

2 -3.2

-3.02

2

-18.752

-18.199

85 - 95% (Base: <85%)

0.657*

0.653*

1.019

1.039

0.689

0.719 0.92

1 0.95

7 0.75

4 0.76

5 1.66

9 1.62

8

-0.13

3

-0.13

2

-0.32

4

-0.33

4

-0.25

9

-0.275

-0.27

6 -0.29

-0.23

5

-0.23

5

-1.00

8

-0.98

9 Over 95% 0.83 0.83 0.82 0.85 1.49 1.501 1.14 1.18 0.63 0.64 1.36 1.31

8 6 5 3 9 8 4 9* 6* 8 4

-0.14

1 -0.14

-0.23

6

-0.24

2

-0.44

8

-0.456

-0.28

9

-0.29

7

-0.13

6

-0.13

6

-0.41

2

-0.41

5 < 100mm (Base: 100-149mm)

1.38

6* 0.68

-0.22

4

-0.39

2 100-149mm (Base: >149mm)

0.072

3*** 0.56

8

-

0.048

-0.24

7

Q310: Concerned about Floods

1.576**

0.82

1.042

0.98

1.206

2.763**

-0.24

1

-0.21

3

-0.22

1

-0.25

3

-0.18

5

-0.98

9

Q310: Concerned about river bank erosion

1.075

0.776

0.631*

0.471*

1.369

0.561

-0.27

1

-0.19

4

-0.11

9

-0.14

8

-0.47

-0.29

6

Q310: Concerned about cyclones

0.74

0.253***

0.684*

0.877

0.543

2.144

-0.38

-0.07

2

-0.12

8

-0.18

2

-0.21

2

-0.94

2

Q308: Experienced Floods

0.98

6 1.13

3 0.605

0.55

5 1.05

4 1.78

5

-0.17

3

-0.40

1

-0.209

-0.19

4

-0.22

-0.68

9 Q308: Experienced river bank erosion

0.81

7 0.68

1 0.457

*** 0.425***

1.183

0.586

-0.37

3

-0.17

5

-0.075

-0.1

-0.3

-0.26

1 Q308: Experienced cyclones

0.63

5 0.84

2 0.177

*** 0.79

5 0.9

1.71

8

-0.17

4

-0.67

-0.073

-0.15

6

-0.19

8

-0.53

4 Q308: Experienced droughts/lack of rain

0.13

8 0.40

5 0.342

* 0.75

6 0.91

8 0.75

1

-0.2

-0.20

1

-0.164

-0.32

-0.26

9

-0.27

3 Observations

3615 3615 1213 1213 5972 5972 2523 2523 6485 6485 929 929

Clusters 251 251 108 108 215 215 159 159 306 306 70 70

Clustered standard errors in brackets, Odd ratios reported; The coefficients for family size, father’s education, mother's education, father's employment status, age (in logs), has first child, has first child died, migrant network present, sufficient income, asset present as suppressed here. These are the same variables as defined in Table 2. * p < 0.05, ** p < 0.01, *** p < 0.001