Modeling destination choices during temporary migration.

For the first research objective, we utilize a three-step conditional probit selection model with subsamples in different equations. In the first migration step, equation (2), we utilize the entire individual dataset of 2,793 observations to model the participation of individual i in migration versus non-migration (M_i). In the second step, equation (3), we utilize the subsample of 605 individual migrants to model their participation in temporary versus longer-term migration (TM_i), incorporating their self-selection into migration (imr1_i) that we calculate from equation (2) based on equation (1). In the third step, equation (4), we use only the subsample of 385 temporary migrants to model their choice of rural versus urban destinations (R_i), incorporating their self-selection into temporary migration (imr2_i), as calculated from equation (3). The three-step conditional probit equations are specified as below:

(2)

(3)

(4)

In equation (2), we account for the characteristics of individual i and household j () relevant for individual i’s participation in migration, as conceptualized in the literature [9,10,38]. This vector includes migrant i’s individual characteristics such as age, education, gender, and primary occupation type, and household j’s characteristics such as its experience of seasonal employment fluctuations, farm labor or family obligations, and the size of migrant networks at the origin. In this equation, we also account for relevant other controls for household j and village k (). They include household size, wealth, access to alternative livelihoods, and proximity to nearby migration hubs. Additionally, they include some relevant village-level controls such as, whether the village is in a flood-prone area, and village-level fixed effects. For consistent estimates, this vector of is controlled for in all subsequent equations.

Equation (3) models the individual migrant i’s selection into temporary versus longer-term migration () by accounting for similar vectors of and , and the selection effects in migration- , as calculated from equation (2). Since we use similar sets of explanatory variables in both equations (2) and (3), for a robust estimation of the selection effect- , we utilize an exclusionary variable () in equation (2), as suggested by [37]. We use households’ experience of random economic shocks in their crops, livestock, and assets in the past 12 months as the exclusionary variable. These idiosyncratic economic shocks can sometimes restrict their capability of sending migrants [9,10]. However, these shocks are unlikely to affect households’ choice between the physical labor-based temporary and longer-term migration, if they have already decided about migration [9]. Another type of idiosyncratic shock includes the sudden death or severe accident of a working household member—family demographic shocks that may affect individual migrant’s choice between temporary and longer-term migration due to increased family obligations at the origin [29]. We separate these shocks and use only economic shocks as the exclusionary variable here. Table A2 in the Online Appendix confirms that the experience of random economic shocks differs significantly between migrant and non-migrants (), but not between temporary and longer-term migrants (), confirming our hypothesis.

Finally, equation (4) models the temporary migrants’ choice of rural versus urban destinations in their latest migration episode by correcting their self-selection into temporary migration (). In this equation, we incorporate v_i as the vector of relevant explanatory variables for individual migrant i’s choice of destination, informed by the relevant literature [9,29]. This vector includes migrant i’s individual characteristics- I_i, negative perception of urban areas- U_i, experience of the latest destination- D_i, migrant networks- N_i, and distance travelled in the latest migration episode- Dist_i. These variables are described in Table 1. The parameters to be estimated in the respective equations are represented by , β, and ẟ, and the error term by u_i, , and . Since equations (3) and (4) include distinct sets of explanatory variables, we do not introduce any additional exclusionary variable in equation (3), apart from the self-selection into migration ().

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Variables for analyzing destination choices during temporary migration.

https://doi.org/10.1371/journal.pone.0352665.t001

Regarding individual characteristics (I_i), studies find that individuals with higher education are more prone to longer-term migration, or at least to urban destinations during temporary migration [9,29], as education rarely brings extra benefits in rural destinations. Educated individuals often possess increased life-skills making them confident about better opportunities in cities. Conversely, individuals with lower or no education often lack life-skills, leading to a preference for simpler settings like in rural destinations. For better estimates about the association of individuals’ education with their destination choices, we also account for their skills (discussed in detail below) in this model.

Similarly, individuals engaged in agriculture labor at the origin may prefer on-farm jobs in other rural destinations due to socio-cultural familiarity [25]. We use individuals’ engagement in farm labor sales at the origin as a proxy for this preference (Table 1). However, any physical sensitivity to agricultural jobs—such as an inability to bend the waist for rice harvesting—may discourage sensitive individuals from choosing rural destinations, which we control for in the model. Moreover, we account for other relevant factors both at the individual (e.g., age) and household levels (e.g., household size, agricultural landholdings, family demographic shocks, and engagement in crop farming, livestock farming, business, safety-nets, and microcredit).

Regarding urban negativity (U_i), individuals with lower education or lacking skills beyond agriculture often view urban destinations as a difficult place for earning and living [39]. This negative perception of cities may discourage aspiring migrants with lower education or skills from choosing urban destinations. Therefore, apart from education, we also account for migrants’ lack of life-skills beyond agriculture.

For individual migrants’ experiences at their latest migration destination (D_i), we collect their subjective experience ratings on a 1–10 Likert scale, where 1 denotes ‘worst’ and 10 denotes ‘best’. For example, a migrant rating 10 for income-to-cost ratio characteristic means they could save most of their daily earnings at the latest migration destination. This often occurs in rural destinations in Bangladesh, where employers frequently offer free accommodation and meals for migrant laborers. Conversely, in urban destinations, around half of daily wages typically go toward accommodation and meal expenses [29], shifting their income-to-cost experience closer to ‘worst.’ Therefore, although wage opportunities are higher in urban areas, rural destinations may offer greater psychological satisfaction from saving ‘hard-earned’ income, influencing poor migrant laborers’ destination choices—a concept similar to ‘loss-aversion,’ which suggests that ‘losses’ have a greater influence on preferences than ‘gains’ [40].

Similarly, while jobs in both types of destinations can be physically demanding, agricultural tasks in rural destinations may offer comparatively better physical comforts to the migrants from rural origins. In contrast, urban destinations often provide longer-duration wage opportunities than rural ones [4]. This is particularly encouraging for temporary migrants from flood-prone villages, where lean periods are often prolonged due to weather extreme [5,9]. Therefore, we also account for migrants’ subjective experiences concerning wage opportunity duration at their latest migration destination, the geographic location of the village in flood-prone areas, and village fixed effects.

Regarding migrant networks (N_i), in addition to the influence of migrant kin, the size of the migrant group from the origin may play a key role in choosing destinations [29]. A larger group size reduces rural poor’s risk-aversion toward migration and makes their migration pleasurable. Rural-bound migrants often travel in large groups, which is frequently required for employment in rural destinations. Employers in these areas tend to prefer hiring larger groups of migrant laborers to keep up with their crop calendar. Conversely, migration in larger groups may raise competition for jobs at rickshaw garages or construction sites in urban destinations. Group migration, therefore, could be more closely associated with rural-bound temporary migration, potentially encouraging risk-averse rural poor to prefer rural destinations. For this analysis, we use data on group size from migrants’ latest migration episode.

Regarding the implications of migration distance (Dist_i) for destination choices, earlier studies have found that longer-distance permanent migrations move towards urban centers [11,26]. However, the relationship between distance and destination choices during temporary migration remains unclear, which we address here. We collect data on the destination sub-district for each migrant’s latest migration episode. The physical distance in kilometer (km) between the origin and destination sub-districts is then calculated using a geo-referencing system. Here, we mainly use the bus-road distance, as buses are the common transport mode across the country. Migration distance is treated as a proxy for migration costs. Moreover, we control for the proximity of individual migrants’ household to the nearest migration hub, often the closest sub-district.

Due to the relatively large number of explanatory variables included in the model, we test for multicollinearity by calculating the variance inflation factors for each equation. The results are shown in Table A3 in the Online Appendix. They do not indicate a high correlation among the explanatory variables and selection effects.

To check the robustness of our findings, we employ a system of simultaneous mixed-process equations using limited information maximum likelihood (LIML), following [41]. When multiple equations are mutually interdependent and deal with subsamples in different equations, as in our case, this analytical approach proves useful. For this analysis, we use the cmp command in Stata, incorporating equation (2)–(4) while excluding their respective imrs. We skip the likelihood-ratio test, use five random draws for the Geweke-Hajivassiliou-Keane (GHK) simulator, and apply the Newton-Raphson method for optimization.

Furthermore, migrant network (N_i) and urban negativity (U_i) variables may remain potentially endogenous in both models discussed above. To assess their sensitivity, we re-estimate the main model (Eq. 4) excluding these variables sequentially.

Modeling the income effects of different destination choices.

For our second research objective (fourth step, Fig 1), we utilize the same individual-level dataset to measure the income effects. We use the lean period income of household j (Inc_j) as the outcome variable in this analysis, as temporary migration takes place mainly during the lean period [5,7,42]. The income effects of the individual migrant i’s choice of rural versus urban destination during temporary migration () is captured in equation (5) below:

(5)

Studies have shown that earning a lot of remittance is often not a priority for constrained poor temporary migrants [39]. Therefore, to better understand the income effects (Inc_j) of destination choices during temporary migration, we use three indicators of income: i) total lean-period income of household j from all sources (tot_inc_j), ii) income earned exclusively from migration remittances (remit_inc_j), and iii) income from the origin’s labor market (loc_inc_j = tot_inc_j- remit_inc_j). Income amounts, categorized by seasons and sources, were collected in Bangladeshi Taka (BDT) currency. For this analysis, we apply a logarithmic transformation to the income data.

To obtain a more consistent estimate of the income effects, we control for certain relevant household and village characteristics for income () in equation (5). These characteristics include the household head’s age, education, and gender, household size, and experience of seasonal employment fluctuation. They also include some village-level factors such as the location of the village in flood-prone areas and village fixed effects. The parameters are represented by θ, and the error term by . From equation (5), we report the coefficient (θ) for choosing rural versus urban destinations () as the estimated effects on the households’ lean period income (Inc_j).

In equation (5), we also account for the migrant’s self-selection into destinations- imr3_i, which is calculated from equation (4) based on equation (1). However, imr3_i appears insignificant in equation (5) for all indicators of income, as presented in Table A7 in the Online Appendix. This suggests that self-selection into destinations may not be a challenging issue when estimating the income effects of different destination choices. Nevertheless, we cannot entirely rule out the possibility of endogeneity in destination choice, particularly arising from unobserved heterogeneity. To address this challenge, we apply a control function approach, which is effective in correcting this type of endogeneity [43].

In this approach, we calculate the control function or residuals (res) from equation (2)–(4) and incorporate them into subsequent equations in place of imrs. While imr is useful to correct self-selection bias, residuals account for the endogeneity arising from unobserved factors, mentioned above, by capturing the part of participation that is not explained by the controlled variables in the respective equation. To calculate residuals, after estimating a regression, we predict the probability of participation for individual i (). Next, we calculate the residual for individual i’s participation (res_i) as the difference between the observed value of participation () and the predicted probability of participation (), as outlined in equation (6) below:

(6)

In our multi-step control function analysis, we calculate res1_i from migration equation (2) based on equation (6) and incorporate it into temporary migration equation (3), replacing imr1_i. Similarly, res2_i is calculated from equation (3) and used in destination choice equation (4). Finally, we calculate res3_i from this equation (4) and incorporate it into income effects equation (5). This analysis uses a similar exclusionary variable (ev_j) design, as discussed earlier in the multi-step conditional probit selection model.

Using control function approach, res3_i in equation (5) appears significant for some income indicators, as shown in row (1) of Table 5 in the results section, indicating the presence of endogeneity. Additionally, res1_i and res2_i appear significant in the respective equations, as presented in Table A8 in the Online Appendix.

To check robustness of our results from the multi-step control function analysis, we employ a similar approach of simultaneous mixed-process equations with LIML, as discussed earlier. Additionally, we use a two-stage least square (2sls) analysis with an instrumental variable (IV) design. The general two-stage equations for this analysis are outlined in equation (7) and (8) below:

(7)

(8)

where, represents migrant i’s instrumented choice of rural versus urban destination, and IV_i denotes the instruments. We use the presence of rural-bound temporary migrant kin or relatives (see Table 1 for details) as an instrument here. This instrument is expected to influence migrant i’s choice of rural over urban destinations (R_i) through network effects, but not to directly affect household income through channels other than networks. The first-stage regression results, presented in Table A10 in the Online Appendix, confirm the relevance of the instrument, with an F-statistic of 76.47 indicating its strength.

Destination choices during temporary migration.

As mentioned earlier, our three-step conditional probit selection model addresses three key questions: (i) why the rural people choose to migrate (M_i, eq. 2), (ii) why they opt for temporary as opposed to longer-term migration (TM_i, eq. 3), and (iii) why they prefer rural over urban destinations during temporary migration (R_i, eq. 4). Results from equation (2) and (3) are presented in Table A6 in the Online Appendix. The choices of migration (eq. 2) and temporary versus longer-term migration (eq. 3) have already been studied in the literature, employing both qualitative and quantitative methodologies [9,13,16,29,38,44,45]. In brief, existing studies find that poor earnings at the origin and the presence of functional migration networks are important factors influencing the decision to migrate. Conversely, the presence of farm labor and family obligations discourages constrained households from migrating. However, when migration is necessary to diversify risks in a less-diversified economy, temporary migration is preferred, as it maximizes utility without exacerbating constraints. Our regression results align with existing literature despite drawing on different contexts and datasets.

The results from equation (4), which addresses our first research objective, are presented in column (1) of Table 4 below. Overall, most of our hypotheses on destination decision-making hold.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Factors for destination choices during temporary migration.

https://doi.org/10.1371/journal.pone.0352665.t004

Individual characteristics, such as low education and engagement in agricultural labor at the origin, are significantly associated with a preference for rural destinations during migration, as expected. Relevant household-level factors also play remarkable roles: livestock farming is associated with rural-bound migration, while crop farming has the opposite effect. This contrast likely reflects differences in labor requirements—livestock farming demands year-round family labor, discouraging longer duration migration such as to urban destinations, whereas crop-farming households can rely on hired local labor and still send migrants to urban areas for longer-duration wage opportunities [9,46]. We discuss this association further in the income effects section.

Negative perceptions of cities before the first migration, along with a lack of life-skills beyond agriculture, are significantly associated with choosing rural destinations, even in the latest migration episode. Moreover, a higher daily income-to-cost ratio (i.e., higher daily wages relative to expenses) appears to favor rural over urban destinations, although reverse causality is plausible. However, urban destinations are more strongly associated with wage opportunities, particularly for longer-duration employment. This suggests that rural destinations may be preferred for shorter and less costly migration episodes, whereas urban destinations may attract migrants seeking more sustained income-earning opportunities despite higher living costs. This also helps explain why many crop-farming households with relatively lower labor constraints are more inclined toward urban destinations during temporary migration.

Regarding migrant networks, the presence of kin or relatives migrating to rural destinations affects aspiring migrants’ destination preferences through network effects. Similarly, larger migrant group sizes are strongly associated with choosing rural over urban destinations, as expected, although endogeneity remains plausible in this context as well.

Also, distance is significantly associated with the destination choice. Our data reveal that rural-bound temporary migration is significantly more common over shorter distances, likely to minimize migration costs, which aligns with loss-aversion theory [40] and classical migration theories [11]. Shorter-distance migration also reduces the duration of migrants’ separation from their left-behind families. This is particularly encouraging for households that often face greater labor constraints for livestock farming and labor sales, along with increased family obligations due to a less flexible member structure (Table 4), consistent with the existing literature [9].

Selection effects (imr) are significant at every stage of the model, as shown in Table 4 above and Table A6 in the Online Appendix, justifying their correction. Results from the simultaneous mixed process equations, presented in column (2) of Table 4, are consistent with the main model, confirming its robustness. Moreover, the main results remain insensitive to potentially endogenous networks and perception variables, as demonstrated in Table A11 in the Online Appendix.

Income effects of the destination choice during temporary migration.

Results from the multi-step control function analysis, showing the income effects of destination choices during temporary migration (equation 5), are summarized in row (1) of Table 5. Results from equations (2)–(4) using this approach are presented in Table A8, and the full regression results for equation (5) are in Table A9 in the Online Appendix.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Income effects of the destination choice during temporary migration.

https://doi.org/10.1371/journal.pone.0352665.t005

While temporary migration generally generates positive income gains for poorer households [9,47], we find that choosing rural over urban destinations is not significantly associated with higher total household income (tot_inc_j). While rural destinations are often associated with better income-to-cost ratio than urban ones (Table 4), urban-bound migration generates higher remittances (remit_inc_j). At the same time, we find that households choosing rural destinations earn significantly more from local labor market activities, particularly livestock farming and labor sales at the origin (Table A5). Similarly, equation (4) results show that engagement in these activities is positively associated with choosing rural over urban destinations (Table 4).These findings suggest that rural-bound temporary migration functions less as a remittance-maximization strategy and more as a complementary livelihood strategy that allows households to maintain stronger participation in origin-based economic activities while still benefiting from migration income, and more importantly, managing their labor constraints better through shorter-duration migration. By contrast, urban-bound migration appears to rely more heavily on remittances as a primary income source. In this sense, destination choice during temporary migration reflects different forms of household risk diversification and labor allocation, rather than a simple trade-off between higher and lower income.

One possible explanation is that some households face stronger farm labor or family constraints, limiting their ability to disengage from origin-based livelihoods for extended periods. For these households, rural destinations may offer a more compatible balance between migration and local economic participation, particularly given their more favorable income-to-cost conditions. Urban destinations, although potentially generating higher remittances, may require prolonged separation from origin-based labor activities and family obligations, with stronger dependence on migrant earnings alone.

These findings make a two-fold contribution to the literature: first, they clarify [29]’s ambiguous conclusion about the better income-to-cost ratio characteristic of rural destinations; and second, they shed light on the complex destination decision-making processes during temporary migration, which are often misunderstood from a monetary return’s perspective.

Results from the simultaneous mixed process equations, presented in row (2) of Table 5, are consistent with our main model. However, results from 2sls, presented in row (3) of Table 5, show significantly negative association between rural-bound temporary migration and total income, possibly because this group of households often belongs to the poorest segment of the community [4,5]. Similarly, although 2sls does not find a significant association between rural-bound migration and household income from the local labor market, it suggests a generally positive relationship, consistent with our main model.