2. Background and contribution

There is a large body of literature that examines migration as a coping strategy to natural disasters. This literature is implicitly motivated by the theoretical model of migration suggested by Harris and Todaro (Reference Harris and Todaro1970) and further developed by Stark and Bloom (Reference Stark and Bloom1985). A main assumption is that households exposed to risk consider migration as a mechanism to diversify income. Since inter-household family transfers are an important source of insurance in low-income countries, a household, as long as it can afford the cost of migration, may decide to relocate one or more members to reduce the gap of income shocks between origin and destination [Rosenzweig and Stark (Reference Rosenzweig and Stark1989)].

Given this theoretical background, a growing number of studies have provided empirical evidence of migration induced by these types of shocks.Footnote ¹ Following the seminal work by Rosenzweig and Stark (Reference Rosenzweig and Stark1989) in India, the impact of weather shocks on migration has been found in both low-income [Findley (Reference Findley1994), Meze-Hausken (Reference Meze-Hausken2000), Ezra and Kiros (Reference Ezra and Kiros2001), Henry et al. (Reference Henry, Schoumaker and Beauchemin2004)] and high-income countries [McLeman and Smit (Reference McLeman and Smit2006), Feng et al. (Reference Feng, Krueger and Oppenheimer2010)].Footnote ² However, the size of this impact is not clear with migration being a highly contextual phenomenon, depending on political, economic, and social factors. For example, Barrios et al. (Reference Barrios, Bertinelli and Strobl2006) show that rainfall shocks increase rural–urban migration in sub-Saharan Africa, whereas Marchiori et al. (Reference Marchiori, Maystadt and Schumacher2012) show that migration in sub-Saharan Africa can be constrained by access to credit, available technologies, and macroeconomic conditions. This suggests that further empirical evidence of the nexus between climatic events and migration is warranted, particularly in societies that are vulnerable to climatic events and which are heavily reliant on agriculture for their livelihood.

Although most studies focus on internal migration [e.g., Bohra-Mishra et al. (Reference Bohra-Mishra, Oppenheimer and Hsiang2014), Bohra-Mishra et al. (Reference Bohra-Mishra, Oppenheimer, Cai, Feng and Licker2017), Dallmann and Millock (Reference Dallmann and Millock2017), Peri and Sasahara (Reference Peri and Sasahara2019)], a number of recent studies have examined international migration induced by shocks. For example, Baez et al. (Reference Baez, Caruso, Mueller and Niu2017a, Reference Baez, Caruso, Mueller and Niu2017b) investigated youth migration caused by droughts and heat exposure in Latin America and the Caribbean. The authors find a significant impact of shocks, although this impact is mediated by social protection and government policies. Studying both natural disasters and long-term climatic factors, Beine and Parsons (Reference Beine and Parsons2015) find little evidence of an association between the two, although a later study shows that natural disasters tend to increase emigration to neighboring countries [Beine and Parsons (Reference Beine and Parsons2017)]. Other studies using cross-country data reach similar conclusions [Drabo and Mbaye (Reference Drabo and Mbaye2015), Cattaneo and Peri (Reference Cattaneo and Peri2016), Gröschl and Steinwachs (Reference Gröschl and Steinwachs2017)].

The empirical approach adopted in most studies testing the relationship between natural disasters and migration also builds on the implicit assumption that the occurrence of an extreme event will reduce household income. Most studies employ two-stage least squares techniques, whereby the first stage examines the relationship between shocks (as an instrument) and household income, whereas the second stage estimates the impact of household income on the probability of migration. In doing so, two important issues need addressing. First, an extreme event can directly impact migration decisions by damaging or destroying household infrastructure. Natural disaster variables, therefore, should also be included in the second stage of these empirical models. However, this approach is problematic because of the expected correlation between natural disasters and household income. Multicollinearity can lead to the following symptoms: (i) coefficients may have very high standard errors; and (ii) coefficients may have the “wrong” sign or be of implausible magnitudes [see Greene (Reference Greene2012), p. 129]. In other words, parameter estimates can be unreliable and the standard errors attached to the parameters will be inflated. This presents an empirical challenge since including only one of these variables in the model will result in omitted variable bias.

The second issue relates to the measurement of extreme events or weather shocks.Footnote ³ Most studies to date employ long historical data on temperature and precipitation to measure weather shocks [Maccini and Yang (Reference Maccini and Yang2009), Datar et al. (Reference Datar, Liu, Linnemayr and Stecher2013), Baez et al. (Reference Baez, Lucchetti, Genoni and Salazar2017c)]. Specifically, the occurrence of an extreme event is captured by a binary variable, taking the value of one when temperature or precipitation deviates from an arbitrary distance from the long-term average. One limitation of this approach arises when a disaster occurs with trivial loss. Such a case is unlikely to lead to migration. Conversely, households might face little choice but to migrate in the advent of a shock of large magnitude. Consequently, there is an increasing number of studies using more detailed information on natural disasters to capture their intensity [e.g., Beine and Parsons (Reference Beine and Parsons2017), Bohra-Mishra et al. (Reference Bohra-Mishra, Oppenheimer, Cai, Feng and Licker2017), Dallmann and Millock (Reference Dallmann and Millock2017)].

This study contributes to the literature by addressing these two issues. Regarding the empirical model, we adopt a novel two-stage econometric technique that accounts for the direct impact of natural disasters on migration as well as its potential indirect impact via reduced household income. The technique is known as the residual generated regressor approach.Footnote ⁴ First, agricultural output is regressed on the disaster variables with other controls following the empirical model of Mendelsohn et al. (Reference Mendelsohn, Nordhaus and Shaw1994). We use agricultural output as a proxy for income since we are interested in farmers living in rural areas where agriculture is the main source of income. Findings from the estimation of this model provide an indication of the extent that natural disasters have impacted on agricultural output. The residual from this model is then employed as an explanatory variable in a second model that estimates the probability of migration. Note that the residual captures the variation in agricultural output that is not explained by shocks and will not therefore be correlated with the shock variables. This approach: (i) allows for the reliable identification of the direct impact of shocks on migration; and (ii) controls for changes in agricultural output and income (not due to shocks but other factors) which can also impact on migration decisions.

We follow the recent literature by examining the severity of natural disasters by using the number of deaths, injuries, houses damaged and houses destroyed, rather than examining the impact of the occurrence of natural disasters. By employing the magnitude of disasters rather than their occurrence, we account for the possibility that shocks with higher intensity are more likely to trigger migration, whereas shocks causing less damage are less likely to lead to farm-household members relocating. Since migration is an economic decision, what matters is the way that disasters translate into tangible outcomes. This suggests that to properly quantify weather shocks it is useful to focus on outcome variables rather than on deviations in temperature or precipitation alone. That said, as a robustness check, we also employ binary shock variables defined as extreme deviations of temperature and precipitation.

We choose Vietnam as a case study to investigate the relationship between natural disasters and migration. The country is one of five countries deemed most-affected by climate change and suffers from a high frequency of natural disasters [ISPONRE (2009)]. These extreme events have posed a significant threat to large portions of the population living in rural areas with agricultural production being their main source of income. Given their vulnerability to disasters, affected farm-households are likely to consider migration as a possible coping strategy. The data used in this study are sourced from the Disaster Inventory System (DesInventar) provided by the United Nations Office for Disaster Risk Reduction (UNISDR). This dataset provides unique information on the damage from disasters at the province level. We then match this disaster data with household data in Vietnam by using the Vietnam Household Living Standard Survey (VHLSS) for the period 2006–2008. We test the hypothesis that a higher severity of natural disasters, measured by the four indicators identified above, is associated with a higher probability of migration.

We further explore the relationship between natural disasters and migration by looking at different household characteristics. Specifically, we assess separately poor and non-poor households using the internationally accepted poverty line of PPP $1.25-a-day (in 2008). This is motivated by contradictory findings from the recent literature. For example, some studies demonstrate that low-income people often settle in areas vulnerable to extreme climate events which force them to subsequently migrate [Morrow-Jones and Morrow-Jones (Reference Morrow-Jones and Morrow-Jones1991), Koerber (Reference Koerber2006)]. Conversely, Myers et al. (Reference Myers, Slack and Singelmann2008) find that wealthier households are more likely to migrate from flood prone areas, since poorer families are unable to meet the costs of migration.

Finally, we contribute to the literature by examining the relationship between migration and natural disasters in the context of public support. Vietnam provides an interesting case study with different post-disaster support programs from the government and international organizations. For example, the government together with the World Bank has established a “Community-based disaster risk management program” in the Mekong Delta to build flood and storm mitigation infrastructures, aimed at strengthening the capacity of people in this region to cope with extreme events. This study seeks to evaluate the effectiveness of the program by testing whether households receiving benefits from the government program in Mekong Delta have had better coping abilities following a disaster and therefore a lower probability of migration.

3. Natural hazards in Vietnam

This paper provides a case study of Vietnam. Located in a tropical monsoon region combined with a diverse and complex topography, Vietnam has suffered different types of natural disasters: both hydro-meteorological (floods, storms, and droughts) and geophysical (landslides and earthquakes). Figure 1 summarizes the most common types of natural hazards that occurred during the period 1990–2010. According to Figure 1a, floods are the most frequent events accounting for 48% of disasters, followed by hailstorms (20%), storms (13%), and flash floods (7%). Cyclones (or typhoons), landslides, and other disasters account for 12% of the reported events. In terms of disaster damage, Figure 1b shows that floods account for 67% of deaths and floods, hailstorms, storms, and flash floods combined are responsible for nearly 90% of the loss of life from disasters [IMHEN and UNDP (2015)]. By region, Figure 2 shows that most affected provinces by natural disasters are located in the North and Central Vietnam. Provinces in the Mekong River Delta have a smaller number of reported events than the average. However, they suffer from a higher number of people affected and households damaged [IMHEN and UNDP (2015)].

Figure 1. (a) Proportion of disasters (%) in terms of reported events in 2006–2008. (b) Proportion of disasters (%) in terms of damage in 2006–2008.

Notes: Flash flood is defined as an event that occurs within 6 hours following the end of the causative event, whereas flood is defined as an event that occurs after 6 hours following the end of the causative event.

Source: IMHEN and UNDP (2015).

Figure 2. Number of disasters across provinces.

Source: Author's calculation using UNISDR database.

Despite its vulnerability to shocks, there is little quantitative research examining disaster induced migration in Vietnam. One important exception is Gröger and Zylberberg (Reference Gröger and Zylberberg2016) who use panel data for 2,200 households in three provinces across 3 years (2007, 2008, and 2010). The authors find that households exposed to Typhoon Ketsana in 2009 experienced an average reduction in income of 10%, and around 17% of households responded by sending members away to diversify income. Another exception is Koubi et al. (Reference Koubi, Spilker, Schaffer and Bernauer2016) who examine different types of environmental stressors, namely, short-term and long-term events. The authors show that sudden-onset environmental events, such as floods and typhoons, are more likely to trigger migration, whereas long-term environmental problems, such as salinity, reduce the likelihood of migration. This suggests that individuals respond to longer-term events with other forms of adaptation, rather than migration.

This paper builds upon these previous studies by using national household surveys collected across Vietnam during the period 2006–2008. This study differs from Gröger and Zylberberg (Reference Gröger and Zylberberg2016) by examining the effects of all disasters occurring over this period instead of concentrating on specific events. Although Koubi et al. (Reference Koubi, Spilker, Schaffer and Bernauer2016) use the occurrence of disasters to examine its impact on migration, we account for the possibility that shocks with higher intensity will trigger migration, whereas shocks leading to less damage are less likely to lead to migration. Finally, this study also proposes a novel econometric framework to assess both the direct and indirect impacts of natural disasters on migration.

4. Model specification and data

4.1 Model specification

We use a cross-sectional sample of households located in the agricultural sector to investigate the impact of natural disasters on the migration decisions of farm households in Vietnam. Following previous studies, we start with a basic framework by employing a traditional IV model to examine the relationship between weather shocks and migration decisions [Cai et al. (Reference Cai, Feng, Oppenheimer and Pytlikova2016), Kubik and Maurel (Reference Kubik and Maurel2016)]. The first stage of this model examines the impact of natural disasters on agricultural output as presented in equation (1). The second stage then tests if a reduction in agricultural output caused by shocks leads to an increase in the probability of migration:

(1)$${\rm Crop}\_{\rm revenu}{\rm e}_{i, j, t-1} = \alpha _0 + {\alpha }^{\prime}_1{\rm Shoc}{\rm k}_{\,j, t-1} + {\alpha }^{\prime}_2X_{i, j, t} + \alpha _3C_{i, j} + \alpha _4E_{i, j, t} + \upsilon _{i, j, t-1}$$

(2)$$ {\rm Migratio}{\rm n}_{i,j,t}\; = \beta _0 + \beta _1{\rm Crop}\_{\rm revenu}{\rm e}_{i,j,t-1} + {\beta }^{\prime}_2X_{i,j,t} + \beta _3C_{i,j} + \beta _4E_{i,j,t} + \sigma _{i,j,t}$$

In equation (1), Crop_revenue_{i ,j,t−1} of household i in province j in period t – 1 is calculated by taking the total gross crop revenue (or total sales for each crop) less all costs, divided by the area of agricultural land.Footnote ⁵ The values of total sales and total expenditures are converted into constant 2010 prices using the CPI. With respect to household demographics, X_{i ,j,t} is a set of variables representing household size as well as the age and education level of the household head. Household characteristics are chosen based on those of previous studies examining migration and climate change in developing countries [Kubik and Maurel (Reference Kubik and Maurel2016), Koubi et al. (Reference Koubi, Spilker, Schaffer and Bernauer2016)]. Following Kubik and Maurel (Reference Kubik and Maurel2016), this paper also includes C_{i ,j} as the cost of migration proxied by the distance from the district capital of province where household i lives to the nearest province. E_{i ,j,t} is a dummy variable capturing previous migration experience. It is equal to one if household i had previous members moving out during the previous period t – 1. Finally, we include province characteristics. These include the secondary school completion rate, the number of hospitals, and the provincial poverty rate to control for differences in the level of development.

In equation (2), Migration_{i ,j,t} is a binary variable which equals one if household i in province j has at least one migrant in period t. Shock_{j ,t−1} is a vector of variables representing the severity of natural disasters in period t – 1 in province j. In this study, it is posited that the migration decision in period t will depend on reductions of agricultural production caused by natural disasters in the previous period t – 1. The reason is that migration is a costly decision which may take time to save for and enact. In other words, a disaster during period t – 1 reduces agricultural production in period t – 1, which leads to a decision to migrate in period t.

Four measures of the severity of shocks are included: (i) number of deaths; (ii) injuries; (iii) houses damaged; and (iv) houses destroyed. As discussed in section 3, this represents a departure from most previous studies which examine the occurrence of shocks rather than their intensity [Koubi et al. (Reference Koubi, Spilker, Schaffer and Bernauer2016)]. One may argue that densely populated areas are likely to have higher exposure to shocks. To account for this, the numbers of deaths and injuries are divided by the population of province j, whereas the number of houses damaged and houses destroyed are divided by the total number of households in the province.

As discussed above, one potential limitation of using the IV framework is that an extreme event can directly impact on migration decisions by damaging or destroying household infrastructure as well as reducing their income from a loss of agricultural output. Natural disaster variables, therefore, should also be considered in the second stage of the IV framework. This, in turn, leads to a problem of collinearity as natural disasters are strongly correlated with agricultural output. If both variables are included in the model, their coefficients might have the wrong sign or implausible magnitudes [Greene (Reference Greene2012)]. However, if one of these variables is omitted from the model, it becomes mis-specified and parameter estimates will be subject to omitted variable bias. To account for this, we adopt a residual generated regressor approach. The first stage of our approach models agricultural output (crop revenue) in period t – 1 as a function of weather-shock outcome variables and other controls, as indicated in equation (3). One advantage of using the residual generated regressor approach is that it allows us to control for idiosyncratic factors that affect agricultural output in the first stage, whereas the traditional IV approach requires the same set of control variables in both stages:

(3)$${\rm Crop}\_{\rm revenu}{\rm e}_{i, j, t-1} = \gamma _0 + {\gamma }^{\prime}_1{\rm Shoc}{\rm k}_{\,j, t-1} + {\gamma }^{\prime}_2X_{i, j, t-1} + \gamma _3Z_{i, j} + \beta _4I_{i, t-1} + \upsilon _{i, j, t-1}$$

where Z_{i ,j} is a set of variables representing the quality of arable land. We assume that arable land is a function of soil type and the irrigation technology available to the household. Type of soil is measured using a dummy variable equal to one if the farm-household has rich ferralitic soil, implying low silica and high aluminum and iron contents. Following Schlenker et al. (Reference Schlenker, Hanemann and Fisher2005) and Kurukulasuriya and Mendelsohn (Reference Kurukulasuriya and Mendelsohn2008), we include an irrigation binary variable (I_{i ,t−1}) which equals one if the household has access to an irrigation system and equals zero if they use rain water. In all model specifications, standard errors are clustered at the province level to control for serial correlation in the error terms across households in the same area.Footnote ⁶

The residual υ_{i ,j,t−1} in equation (3) captures the variation in agricultural output that is not explained by weather shocks (and other control variables). This residual is then employed in the second stage model explaining the decision to migrate. If the “transmission” variable (in this case crop revenue) has a negative impact on migration and shocks have a negative impact on crop revenue, this method will provide larger coefficients on the shock variables. Moreover, the shock variables and the residual capturing agricultural output variables will not be highly correlated allowing for the accurate identification of their parameters:

(4)$${\rm Migratio}{\rm n}_{i, j, t} = \delta _0 + {\delta }^{\prime}_1{\kern 1pt} {\rm Shoc}{\rm k}_{\,j, t-1} + \delta _2\upsilon _{i, j, t-1} + {\delta }^{\prime}_3X_{i, j, t} + \delta _4C_{i, j} + \delta _5E_{i, j, t} + \lambda _{i, j, t}$$

4.2 Data

This study uses migration data from the VHLSS in 2006 and 2008.Footnote ⁷ There are 9,189 households interviewed in each survey, with 4,138 households interviewed in both rounds. In each survey, households were asked if there is any member in their families that has moved to another province since the last interview.Footnote ⁸ The survey also includes the reason why they left. This paper identifies migration as those households with a member who left for work due to economic factors. Other reasons for migrating, such as education or marriage, are not considered because they are unlikely consequences of weather shocks.Footnote ⁹ It is possible that the entire household might migrate after a disaster. This study does not examine this type of migration because entire households moving to a different location in 2006 would not be captured in the 2008 survey.

Since we focus on farm-households, it is likely that migration of household members may be affected by the seasonality of agricultural production. To address this issue, we control for the month of interview in all model estimations. Figure 3 provides a pie chart for the migration types in our sample. Approximately 8% of households have migrants for economic reasons, whereas the proportion of migration for education and marriage are 8.7% and 5.5%, respectively.Footnote ¹⁰ The overall sample in this study includes 1,003 farm households from 34 provinces in Vietnam.Footnote ¹¹

Figure 3. Types of migration.

Information on other household characteristics is also provided in the VHLSS. Province level variables, to capture differences in economic development and infrastructure, are collected from the Statistical Yearbook of Vietnam. They include the average secondary school completion rate, the number of hospitals, and the incidence of poverty.

Natural disasters data are sourced from the DesInventar, provided by the United Nations Office for Disaster Risk Reduction (UNISDR).Footnote ¹² This database contains a historical inventory of disasters and their impacts on goods, infrastructure, lives, and social relations at province and national levels. One advantage of the DesInventar database is its focus on disasters at the province level, whereas other widely used databases, such as EM-DAT, are collected at a country level [Soto (Reference Soto2015)].Footnote ¹³ Weather shocks are defined as events or forces of nature that have catastrophic consequences on people. This study uses four measures of disaster intensity discussed above. In this analysis, we focus on disasters occurring during the period 2006–2008.Footnote ¹⁴ By employing the magnitude of disasters rather than their occurrence, we account for the possibility that shocks with a higher intensity will trigger migration, whereas shocks causing less damage are less likely to lead to migration. Tables 1 and 2 provide the descriptive statistics of the demographic and disaster variables in our sample, respectively. In the empirical model, the natural disaster variables are measured relative to the population of the province in the year prior to the investigation period (to control for the fact that disasters will have a greater impact in highly populated areas).

Table 1. Descriptive statistics of demographic variables

Table 2. Descriptive statistics of natural disasters by provinces

Notes: This table shows the actual number of deaths, injuries, house destroyed, and house damaged. In the empirical model, all variables are normalized by population and number of households.

To examine whether migrants were different from non-migrants in terms of observed characteristics, we conduct a simple t-test using the sample in 2006 (before migration occurred) and 2008 (after migration) separately. The results are presented in Table A1 (Appendix) which show that in most cases, the two groups are comparable. Still, we acknowledge that it is not possible to completely rule out differences between two groups in terms of unobserved characteristics.