Temporary international migration, shocks and informal finance: analysis using panel data

There is a large literature that examines the causes and consequences of international migration.

See Kerr and Kerr (2011) and Gaston and Nelson (2013) for reviews of the literature. A large number of studies examining the determinants of migrants have used data for migrants from Mexico to the U.S. These studies include Chiquiar and Hanson (2005), Ibarraran and Lubotsky (2007), Mishra (2007), Mckenzie and Rapoport (2010), Vincenzo (2011), and Kaestner and Malamud (2014) among others.

We examine the migration response of households to different types of natural shocks and provide evidence on potential mechanisms driving the effects. We use a unique household panel data from the Life in Kyrgyzstan (LiK) surveys, which allows us to observe the dynamics of temporary migration (LIK, 2010/2013) at the household level. The LiK surveys provide a nationally representative panel data comprising about 3,000 households in Kyrgyzstan. We use data from four waves of the survey (2010–2013). To our knowledge, the LiK survey data is the only panel data available for a low income source country that allows for tracking and analyzing temporary international migration decisions of households.

For details on the LiK survey data see Brück and Esenaliev (2014).

We use the information on whether a household was affected by one of five natural shocks – drought, rain and landslide, winter and frost, earthquake, and pest. We restrict our analysis to natural shocks as they are most likely to be exogenous to a household's migration decision. Further, given that weather shocks are more likely to affect the livelihoods of rural households, we restrict our analysis to households residing in villages.

We begin our analysis by estimating the effect of each shock on a household's decision to have a migrant. We find that while receiving a drought shock in the last period increases the likelihood of migration, winter shocks reduce the likelihood of migration.

While most previous studies in the literature find an increase in migration in response to shocks, Halliday (2006) finds that adverse agricultural conditions increase net migration while earthquakes reduce net migration from El Salvador to the United States. Our findings of differences in effects of shocks for Kyrgyzstan are similar to these results.

When examining temporary migration, an important distinction is between the decision to send a migrant and to recall a migrant. The longitudinal nature of the LiK data allows for examining these dynamics of temporary migration and evaluate whether the effect of shocks on the two decisions depends on the nature of the shock. For analyzing the effect of shocks on recalling a migrant, we compare two households that both have a migrant in the current period and study the change in their migrant status in the next period. Analogously, for analyzing the effect of shocks on a household's decision to send a migrant, we compare two households that do not have a migrant in the current period and one of them switches to be a migrant in the next period. We find that while a drought shock affects the decision to send a migrant, a winter shock does not affect this decision. Even though the effect of winter shocks on the recall decision is large and positive, it is imprecisely estimated.

To understand why the migration response to these shocks differ, we explore the underlying mechanisms. We argue that depending on the nature of income generating activities that households are engaged in, local labor markets, and liquidity constraints faced by a household, a natural shock might reduce or increase migration. For instance, a drought shock is likely to reduce income for rural households that are largely engaged in agricultural activities. Migration, in such a situation, helps affected households to mitigate the adverse income effect of the shock (Morten, 2016). On the other hand, a potential explanation for the migration-lowering effects of winter shocks could be that colder months lead to an adverse labor situation by increasing the need for household labor in rural Kyrgyzstan. For instance, we find that time spent on various labor-intensive household activities is much higher in colder months than in milder months.

We examine the mechanisms driving different responses to different shocks by studying a household's access to informal finance. It is well established that in the absence of financial markets, rural households use informal networks to insure against income shocks (Townsend, 1994).

A large literature examines the importance of informal networks in providing insurance in rural areas, see Morten (2019) for a brief review of the literature.

We contribute to the existing literature in two ways. First, with the exception of Halliday (2006), previous studies have focused on only one shock as a determinant of migration [see, e.g., Angelucci (2015), Dillon et al. (2011), and Gröger and Zylberberg (2016)]. Second, we discuss the underlying mechanisms that explain differences in the migration response of households to different shocks. We then empirically investigate these mechanisms and find evidence in support of the proposed mechanisms.

The rest of the paper is organized as follows. Section 2 discusses the data, and provides an overview of migration and shock experiences of households. Section 3 outlines the empirical specification that we use for our analysis. Section 4 reports our findings for the effects of shocks on the decision to migrate. Section 5 provides a discussion of the mechanisms driving the findings and empirically investigates the mechanisms. Finally, Section 6 provides a brief conclusion.

We use four waves of the LiK survey, a panel data collected annually between 2010 and 2013. The survey was conducted in roughly 120 communities across the country and covers all provinces of Kyrgyzstan. We study the effects of weather and other natural shocks on a household's decision to have a migrant member. These shocks are more likely to affect the livelihoods of rural households predominantly employed in agriculture and related activities. Hence, we restrict our analysis to households that reside in villages. To do so, we follow Chakraborty et al. (2015) and use a variable in the data that provides information on whether a household resides in a city or village. The data consists of about 3,000 households, of which about 59% reside in villages.

We consider the decision to send a migrant to another country as a joint household decision. Accordingly, our outcome variable is the international migrant status of a household in a specific year. We construct this variable from the survey question that asks each household whether any of the household members lived in another country for more than 1 month (excluding business trips, vacations, and visits) during the last 12 months.

Based on the information available in the LiK data, more than 90% of the migrants from Kyrgyzstan go to Russia. While there were some requirements for workers to register in Russia, there was free mobility of workers between Central Asian countries and Russia over the period of our analysis.

The data also provide information on the month in which a migrant leaves home to go to another country. In Figure 1, we use this information to plot the percentage of migration in each month in 2010. The figure indicates that while migrants leave all year round, the highest percentage of migration is around the month of March (Spring) and September (Fall). In comparison, migration is very low in winter between November and February. This pattern of migration is similar for other years in the data.

Figure 1

While the information on migration status, in any year, is the primary outcome of interest, we also examine the dynamic nature of temporary migration decisions. In Table 2 we summarize the four types of year-to-year migration status changes for households. First, if a household was a non-migrant (migrant status = 0) in the previous year and continues to be a non-migrant in the current year, it is indicated as (0 to 0) in Column 1. Thus in 2011, 2012, and 2013, roughly 76%, 74%, and 70%, respectively, of the households did not have a migrant and also did not have one in 2010, 2011, and 2012, respectively. Second, if a household was a non-migrant in the previous year but decides to send a migrant (migrant status = 1) member in current year it is indicated by (0 to 1) in Column 2. In 2011, roughly 9% of the households had a migrant member in 2011 but did not have one in 2010. The third change is (1 to 0), that is, the household recalls a migrant. In 2011, roughly 6% of the households recalled the migrant. In other words, there were 15% of households that had a migrant in 2010, of which 6% recalled a migrant in 2011; or, conditional of having a migrant, about 40% of households recalled the migrant between 2010 and 2011. Finally, if a household had a migrant in the previous year and continues to have a migrant in the current year it is indicated as (1 to 1) in Column 4. Between 2010 and 2011, roughly 9% of the households continued to have a migrant member.

The pattern of migration indicates that there are households of all types in the data and there are substantial changes in the migration status of households over the period of our analysis. We also find that almost all of the international migration from Kyrgyzstan is temporary in nature: there are only 3% of households that have a migrant status of 1 in all 4 years of the data.

Following the literature on the socio-economic determinants of migration, in our analysis, we control for household demographics, education, and wealth. The demographic control variables we include are: total number of members in a household (Household size); gender composition of the household using the ratio of the number of male members in the household and the household size (Male fraction); age composition of the household using two variables: the fraction of members in the household older than 18 years of age (Adult fraction) and the fraction of household members older than 65 years of age (Elderly fraction). To account for education as a determinant of migration, we construct a measure of education at the household level. We use the highest years of schooling achieved by any member within the household (Education years). For wealth of a household, we follow Chakraborty et al. (2015) and combine various asset indicators to create a wealth index using a principal component method that serves as a proxy for household income.

Specifically, we use a weighted average of whether the household owns a house, car, refrigerator, gas-stove, microwave, washing machine, vacuum cleaner, television, computer, mobile-phone, and livestock.

Table 3 provides a summary of the data for migrant and non-migrant households for the year 2010.

We only present summary statistics for 2010 as household migration status changes across years. Summary statistics for migrant and non-migrant households for other years are similar to that for 2010.

We examine the relation between shocks and the decision to migrate by estimating variations of the following baseline empirical model: (1) hhmigi,t=β0∑j=15γjshocki,j,t−1+Xi,tβ+ϕt+ϕt*region+ηi+εi,t, hhmi{g_{i,t}} = {\beta _0}\sum\limits_{j = 1}^5 {{\gamma _j}\,shoc{k_{i,j,t - 1}} + {X_{i,t}}\beta + {\phi _t} + {\phi _t}*region + {\eta _i} + {\varepsilon _{i,t}},} where the dependent variable hhmig represents migration status for household i in year t. It takes a value 1 for a migrant household; and a value 0 for a non-migrant household. We include each shock j separately as well as all five shocks together in the regression equations that we estimate. The control variables, X, included in the regression are household size, fraction of males, fraction of adults, fraction of elderly, years of education, and wealth quintile of households.

As illustrated in Figure 1, while migrants leave all year round, the highest percentage of migration is around the month of March (Spring) and September (Fall), and the survey rounds have typically been fielded between October and December. Thus, when a household is surveyed in 2013, specifically between October 2013 and January 2014, the migrant would have most likely left the household during March–June 2013. This suggests that within a year shocks and migration could occur at the same time. To ensure that for a household the shock precedes the migration decision, we use one-period lagged values (t − 1) of shocks as explanatory variables. Even though households might want to migrate in the months immediately after the shock, it is likely that there will be some lag in international migration as it takes time to make arrangements to find a job and travel.

Another significant identification challenge arises from the possibility of unobserved heterogeneity. We cannot be sure that the observed correlates of the migration decision are not picking up the effects of other unobserved household characteristics. The longitudinal nature of our data enables us to address this issue by introducing household FE (η_i), which allows us to control for any time-invariant differences between households. In addition, we include time FE (ϕ_t) and the interaction of time and region FE (ϕ_t^*region) to control for time-varying unobserved heterogeneity at a regional level. We use the oblast of residence for the household as the region. There are nine oblasts in our data.

While we also have community identification in the data, weather shocks are unlikely to vary across households within these small geographic areas with very few households.

Including household FE in non-linear models, like probit or logit, can lead to severely biased estimates because of an incidental parameter problem, particularly given that we have limited time variation for each household (Greene et al., 2002; Greene, 2004). Hence, the regressions are estimated using a linear probability model. In addition, to address spatial correlation, we cluster the standard errors (SE) for all regressions at the community-year level since migration outcomes of households are likely to be correlated across households from the same geographic region in a specific year.

We did not cluster at the oblast level since there are very few oblasts and the estimated standard errors are known to be biased with few clusters (Colin and Miller, 2015).

We begin by estimating a baseline specification in which we estimate the independent effect of each shock on the decision to migrate, that is, we estimate regression Eq. (1) separately for each shock without including any control variables. As discussed in Section 2, we consider the effect of natural shocks that are available in the data and restrict our attention to households residing in villages.

We estimated the effects of shocks on migration for households residing in cities. The estimated coefficients for the shocks were not statistically significant, suggesting that these shocks had no effect on the decision to migrate for households residing in cities (Table A1 in Appendix).

Table 6 presents the estimates for this baseline specification. In Columns 1–5 an indicator for whether a household experienced each of the shocks in the previous year is introduced separately. Column 1 indicates that a drought in the previous year increases migration for households residing in villages. On the other hand, negative coefficients in Columns 3 indicate that a household is less likely to have a migrant in the current year, if it experienced a severe winter in the previous year. Excessive rain or floods (in Column 2), earthquake (in Column 4), and pest infestation (Column 5) do not have a statistically significant effect on migration. In Column 6, we include all the shocks simultaneously and find that the estimated coefficients are similar to those in Columns 1–5. This indicates that the occurrence of each of these shocks is independent of those of the others.

The estimates presented in Table 7 control for household characteristics that are established determinants of migration. We find that larger households and households that have a higher fraction of individuals in the employable age groups, 18–65 years, are more likely to have migrant members. The signs of the estimated coefficients for the control variables are similar to those reported in previous studies [see, e.g., Kaestner and Malamud (2014)].

The estimated coefficients for the shocks remain unaffected by the inclusion of controls, both when shocks enter the specification individually and when shocks are introduced in the regression simultaneously. This suggests that the shocks are very likely exogenous with respect to household characteristics. Nevertheless, we estimate the full specification that includes all the household level control variables for all the analysis that follows.

Our findings indicate that there are significant differences in the effect of shocks on a household's migration decision. While a drought shock increases the probability of a household to have a migrant, a winter shock reduces the probability of a household to have a migrant. In this section, we discuss potential mechanisms that might underlie these differences.

Depending on the nature of income generating activities that households are engaged in, local labor markets, and liquidity constraints faced by the household, a natural shock might reduce or increase migration. For instance, Gröger and Zylberberg (2016) show that flooding, caused by typhoons, lead to a drop in household income, which increases migration for non-migrant households in rural Vietnam.

Excessive rainfall is also likely to damage crops and affect agricultural income in a way similar to that of a drought shock. However, the estimate for the shock in Table 7, though positive, is not statistically significant.

In rural Kyrgyzstan, most households raise livestock or engage in growing crops as their main source of income. For instance, in the LiK data, 80% of the households in the rural region report being engaged in agricultural activity or owning livestock or both. Hence, in rural regions, where the labor markets are predominantly agricultural, a drought shock is likely to reduce income and release labor. Moreover, since these shocks are aggregate in nature, they are likely to reduce household income for both migrant and non-migrant households that are engaged in agricultural activities. Members of the non-migrant households affected by these adverse shocks could choose to temporarily migrate in order to insure themselves against the income reductions, in the absence of alternate job opportunities in the local market. On the other hand, for households with existing migrant members, the loss of household income at home is likely to make the household more dependent on outside income, reducing the possibility of recalling the migrant member. In other words, migration could be used by households as a means to buffer against shocks that reduce income from their primary activities.

One possibility is for individuals to migrate to nearby urban regions. However, the migration destination is likely to depend on a combination of factors like the returns to and availability of unskilled work in the destination market, migration network that determines both access to information regarding the destination labor market as well as the migration process itself, and the monetary cost of migration. In the case of Kyrgyzstan, given the history of migration to Russia and Kazakhstan (Dzhooshbekova et al., 2021; Gerber and Zavisca, 2020), and relatively lower returns to unskilled work domestically, a drought shock is likely to trigger international migration to established destinations.

While this does not rule out some domestic migration, we do not combine the two forms as the cost implications of domestic migration are quite different from those of international migration.

While drought increases migration, we find that cold winters reduce migration. In Kyrgyzstan, a bulk of agricultural work and livestock rearing activities take place in the warm months (Atamanov and Berg, 2012). Hence, winter shocks are unlikely to affect agricultural income. A potential explanation for the reduction in likelihood of migration due to a winter shock could be that such shocks require additional labor input at home. An adult member of the household is likely to contribute to the household not only in terms of earnings but also by providing labor services. The extent of benefit from such labor services, and equivalently the opportunity cost of losing them due to migration, is likely to depend on the need for such services at home. Colder winters could increase the need for additional labor at home. For instance, households in rural Kyrgyzstan have to make heating arrangements at home to keep themselves warm in winters since electric heating is rarely available in rural regions. Hence, the migration decision of the adult member will depend on the availability or cost of these services in the local market. If labor markets are imperfect, such services cannot be purchased from the local market, constraining the migration decision of adult members.

Using a time-use data, collected as part of the 2013 LiK survey, we provide an estimate of the approximate time spent on activities that are labor intensive compared to year-round usual activities in Table 9. The survey was conducted across various households over the period November 2013–January 2014. Hence, while it does not allow us to compare winter and summer, it does allow us to estimate the time-use separately for the coldest periods and the relatively milder weeks preceding these. We use mid-December as a transition from mild to cold winter. We find that the number of hours spent by households in repair and maintenance of the house is much higher in the colder months compared to the milder months. Specifically, households spend an extra hour per day on maintenance and repair of the house in colder periods compared to warmer periods. On the other hand, the average time spent on regular activities like eating, resting, and laundry is identical across these months. This evidence provides support for our argument that a household might anticipate greater labor needs when winters are colder, making it less likely for adult members to migrate.

Thus, depending on the nature of the shocks, households could be faced with an adverse income situation that is likely to increase migration or they could experience an adverse labor situation leading to a decrease in migration. The net effect on the migration decision depends on the relative strengths of these two effects.

Using panel data for households residing in rural Kyrgyzstan, we investigate the role of temporary international migration as a risk mitigation strategy. We ask whether the decision to migrate depends on the shock experiences of a household. We use a household FE model to account for unobserved correlates of a household's shock experiences and its decision to have a member migrate. In line with the previous literature, we find that shocks with adverse income effects increase migration, indicating that households might use temporary migration as a coping strategy. However, in addition, we find that winter shocks reduce migration.

We argue that the effect of a shock on the decision to migrate depends on the shock effect on households’ income and need for labor services. While a drought shock is likely to reduce agricultural production and income, a winter shock is less likely to affect agricultural income as agriculture is primarily done during the warmer months in Kyrgyzstan. Instead, a winter shock increases labor requirement in rural Kyrgyzstan. To substantiate our claim that migration is used as a mitigation strategy in the face of negative income shocks, we explore whether the decision to migrate responds to shocks differently in the presence of alternate coping mechanism. We find that a household's migration response to negative income shocks decreases when access to informal finance is easier, that is, when households find it easier to borrow. However, access to informal finance has no impact on a household's migration response to a winter shock. These findings provide evidence in favor of our proposed mechanisms through which shocks affect temporary migration.