2. Conceptual framework and hypotheses

A renewed interest in the role of social protection to facilitate mobility stems from the nascent findings that liquidity constraints pose implicit barriers for the rural poor to take advantage of employment opportunities in distant towns, cities or even abroad (Bryan et al., Reference Bryan, Chowdhury and Mobarak2014; Angelucci, Reference Angelucci2015; Bazzi, Reference Bazzi2017; Morten, Reference Morten2017). Stecklov et al. (Reference Stecklov, Winters, Stampini and Davis2005) and Angelucci (Reference Angelucci2015) exploit the occurrence of a conditional cash transfer program in Mexico, Oportunidades, which randomized eligibility status, to identify the causal effect of a positive income shock on migration. Stecklov et al. (Reference Stecklov, Winters, Stampini and Davis2005) first show that the receipt of transfers mediates international migration. Restricting the focus to unskilled workers, Angelucci (Reference Angelucci2015) finds that having access to additional income encourages international migration, with virtually no impact on domestic migration. Angelucci (Reference Angelucci2015) argues that the divergent effects on migration patterns lend credence to the role of financial constraints affecting migration, presumably since the costs of moving a long distance are likely more pronounced than those of moving a short distance. Debt patterns confirm her claims. Eligibility into the program does not affect loans at the extensive margin, yet those that already incur debt and are eligible to receive the transfer are more likely to increase the amounts that they borrow. Effects are most pronounced for the poor.

The previous literature focuses on the role of social protection during periods of climate normalcy. Two recent working papers explore how social protection programs affect migration during climate anomalous conditions. Chort and de la Rupelle (Reference Chort and de la Rupelle2017) evaluate the independent consequences of receiving an agricultural unconditional cash transfer (PROCAMPO) and a disaster fund stipend (FONDEN) in Mexico on documented and undocumented Mexico-US migration flows. Both instruments reduce the flow of undocumented migrants induced by lower than average rainfall levels. Hoddinott and Mekasha (Reference Hoddinott and Mekasha2017) show that the Productive Safety Net Program (PSNP) in Ethiopia also reduces migration during anomalous years but only for girls aged 12 to 18 years. It should be noted that the latter of the two studies has a few distinct features from our CGP evaluation in Zambia. First, the majority of the beneficiaries (85 per cent) are required to work off-season in public works programs under the PSNP (Hoddinott and Mekasha, Reference Hoddinott and Mekasha2017). Thus, the condition of work in exchange for food and cash payments constrains the household by removing a potentially productive worker. Girls may be retained to substitute the absent household labor. Second, the PSNP issues a combination of food and cash payments. The payment modality can bear differential impacts, at least, on consumption patterns (Hidrobo et al., Reference Hidrobo, Hoddinott, Peterman, Margolies and Moreira2014). Since food payments are less fungible, programs offering cash may provide a stronger push if indeed financial barriers inhibit movement.

These preliminary studies suggest the following first hypothesis: unconditional cash transfers discourage migratory responses to climate variability in the short term. We test this hypothesis in the context of the CGP in Zambia. Previous evaluations of the CGP suggest that cash transfers ameliorate push factors (Handa et al., Reference Handa, Seidenfeld, Davis and Tembo2016, Reference Handa, Natali, Seidenfeld, Tembo and Davis2018; Palermo et al., Reference Palermo, Handa, Peterman, Prencipe and Seidenfeld2016), such as food insecurity, known to influence migration (Pankhurst et al., Reference Pankhurst, Dessalegn, Mueller, Hailemariam, Pankhurst, Rahmato and Van Uffelen2013). Asfaw et al. (Reference Asfaw, Carraro, Davis, Handa and Seidenfeld2017) measure how the effect of the cash transfers on food insecurity varies with climate variability. In particular, the authors find that households facing a short-term rainfall shock were more protected from the negative consequences on consumption, caloric intake and dietary diversity. We provide additional evidence that consumption patterns at baseline are negatively associated with short-term temperature shocks (see table A1 in the online appendix). If households in these districts use migration as a risk management strategy to smooth consumption, these findings might lead one to expect a reduction in migration behavior among households receiving the cash transfer and exposed to climate variability.

An additional feature of the CGP that is likely to affect those at risk of migration and apt to move in response to the cash transfer is that women in the household were allocated the cash transfer, given assumptions of their status as primary caregiver (Seidenfeld and Handa, Reference Seidenfeld and Handa2011). We therefore expect women to be less likely to migrate in households receiving the cash transfer, if concerns over risk of losing eligibility status exist. Thus, we differentiate the effects of the cash transfer by gender, expecting women's mobility to be differentially affected by receipt of the cash transfer under conditions of extreme climate due to expected income losses.

We next examine the validity of a second hypothesis: the effect of unconditional cash transfers on the use of migration as an adaptation strategy will be positive for the poor. Bazzi (Reference Bazzi2017) provides a theoretical justification for heterogeneous migratory responses to cash transfers. Intuitively, while a positive shock to income can reduce the financial burden of moving, it can also increase the relative returns of staying. Given the observed behavioral changes among the poor in Mexico (Angelucci, Reference Angelucci2015) and Indonesia (Bazzi, Reference Bazzi2017), the perceived gains in income from financing the migration of a family member outweigh the potential returns from using the migrant's labor at home, especially during years of climate normalcy in Mexico. The intent of the second hypothesis is to understand whether the poor deem investments in migration as beneficial, when simultaneously exposed to a climate shock and a positive income shock.

To examine the longevity of the previous claim, we propose a third hypothesis: the effects of unconditional cash transfers on the use of migration as an adaptation strategy among the poor will be short-lived. Short-term financial incentives to migrate have been shown to generate greater migration rates over time, even after the payments are terminated (Bryan et al., Reference Bryan, Chowdhury and Mobarak2014). As migrant networks develop, the cost of moving declines and the probability of securing a job at the destination increases, encouraging migration (Carrington et al., Reference Carrington, Detragiache and Vishwanath1996; Munshi, Reference Munshi2003). Furthermore, individual patterns of migration are reinforced through experiential learning (Bryan et al., Reference Bryan, Chowdhury and Mobarak2014). The above would suggest that the provision of cash transfers to the poor may reinforce the migratory responses to climate variability through the creation of migrant networks. Alternatively, the migrant networks created from individuals historically taking advantage of job opportunities abroad to help their families adapt offer an additional mechanism for communal resilience through their provision of remittances (see Mbaye and Zimmermann (Reference Mbaye and Zimmerman2016) for a review). Based on insights in rural Mexico (Nawrotzki et al., Reference Nawrotzki, Riosmena, Hunter and Runfola2015), we suspect that households residing in communities vulnerable to climate variability will suppress (rather than amplify) migration responses in the long term, if the CGP provokes network formation and communities benefit from remittance receipts.

3. Data

3.2 Migration data

We define the migration behavior of individual household members at baseline through their absence in subsequent rounds. Given the focus of migration as an adaptation strategy, we only consider the moves of those who reportedly left the village. Thus, the migration outcome is a binary variable, which receives a value of one if the person interviewed at baseline is absent from the household and reported to be at a destination outside of the village at the time of the interview. If the person remained in the village (living in the baseline household or with another household) in a given round, then s/he receives a value of zero for the migration outcome. We further distinguish moves by short-distance (to a nearby village) and long-distance (to another town in Zambia or abroad) in alternative specifications.Footnote ¹ It is possible that some individuals might have moved and returned to the household in between survey rounds. Thus, our measures of migration would tend to underreport the true number of moves inherent to the sample.

The conceptual framework described in the paper is most appropriate for describing individual migration behavior that is in result of decisions made by the household to manage risk. We focus on the migration responses of 5,128 individuals, who were ages 15 through 65 at baseline. We observe moves 24, 36, and 48 months following the baseline. Of the sample of working aged individuals, 1.3 per cent was missing migration information (65 individuals)Footnote ² and 5.1 per cent were not interviewed in subsequent rounds due to the relocation of a household (263). Thus, the final sample used for the analysis consists of 4,802 individuals.

Due to the absence of individual migration information and household attrition, we may be concerned that the observed impacts from the intervention reflect those of more resilient households, or households able to withstand climate variability. The low fraction of individuals who were missing migration information does not differentially impact the composition of the treatment and control groups: 98.9 per cent of the control group and 98.6 per cent of the treatment group have complete migration information. However, attrition appears to be slightly more pronounced in the treatment group, where 93.9 per cent of the original sample is included in our analysis compared to 95.9 per cent of the control group. A higher proportion of resilient households in the control group could potentially attenuate the effects of the program on migration.

To determine the extent to which our results may be influenced by attrition bias, we estimate a linear probability model which correlates the probability of an individual leaving the panel due to household relocation, with the treatment dummy; individual baseline characteristics (female, ages 19–35, 36–55, or >55 years old); household baseline characteristics (number of household members ages 6–12, 13–18, 19–35, 36–55, 56–69, and whether the household's per capita consumption expenditures were above the median value); contemporaneous rainfall and temperature (levels or anomalies expressed in z-scores); and district (Kaputa, Shangombo) and time (36, 48 months) fixed effects. We present the results from stepwise regressions in table 1, where the simplest model presented in column (1) includes the treatment indicator, district and time fixed effects. The results in columns (1)–(4) show that there is no robust relationship between attrition and the treatment indicator. There is a small, weakly significant positive effect of the treatment on attrition in model (1), which disappears after conditioning on the remaining variables. Thus, the attrition regressions validate that our migration effect estimates are less susceptible to bias from household attrition (table 1).

Table 1.

Attrition analysis

Notes: Unit of analysis is person-year. The dependent variable reflects whether the person sampled at baseline was removed from the final sample due to the movement of a household. Village-clustered standard errors in parentheses.

* p < 0.1, ** p < 0.05, *** p < 0.01.

3.3 Climate data

Using GPS points collected at each household's baseline location, we extract monthly climate data from 1981–2014 from two sources. Temperature values were extracted from the Climatic Research Unit's (CRU) time-series at ~50 km resolution, created via spatial interpolation from over 4,000 global weather stations including a large number in Sub-Saharan Africa (UEACRU et al., Reference Harris and Jones2017), where CRU data are considered to provide reliable climate information (e.g., Zhang et al., Reference Zhang, Kornich and Holmgren2013). Precipitation values were extracted from the Climate Hazards Group Infrared Precipitation with Stations' (CHIRPS) dataset, which integrates high-resolution satellite imagery to produce a ~5 km resolution product specifically developed for drought monitoring in Africa (Funk et al., Reference Funk, Peterson, Landsfeld, Pedreros, Verdin, Shukla, Husak, Rowland, Harrison, Hoell and Michaelsen2015).Footnote ³ These monthly data were transformed into running means at 12-month time scales, and these means were subsequently standardized into z-scores capturing local deviations from the pre-intervention climate (1981–2009). Z-scores represent exogenous climate shocks that are not confounded by baseline climate and have previously been shown to better predict migration in Africa compared to raw climate values (Gray and Wise, Reference Gray and Wise2016).

We link the raw climate values and their corresponding z-scores to the person-year data via baseline location and month of follow-up interview. Additional specifications link individuals to climate observed 12 months before the month of interview in order to observe lagged effects. Figures 1 and 2 present the distributions of monthly average rainfall (in millimeters per month) and temperature (in degrees Celsius), respectively, one year prior to the observed migration episodes. Over the period of study, there was considerable variation in rainfall, ranging from a monthly average of 51 to 121 millimeters (figure 1). The wide range in rainfall values is largely driven by regional differences in climate, where Kaputa district typically experiences an abundance of rainfall relative to the other two western districts (Asfaw et al., Reference Asfaw, Carraro, Davis, Handa and Seidenfeld2017). Sufficient overlap occurred in the rainfall exposure of treatment and control groups as expected, given that the randomization was implemented at the household level. In contrast, temperature values remain concentrated in the range of 22 and 24 degrees Celsius, with the occasional incidence of a ‘colder’ event at 20 degrees Celsius (figure 2).

Figure 1.

Distribution of contemporaneous rainfall.

Figure 2.

Distribution of contemporaneous temperature.

Since our primary motivation is to understand migration related to environmental risks, it is fruitful to examine the extent to which the climate values presented in figures 1 and 2 relate to the conditions normally experienced by the sample. We therefore also present the distributions of contemporaneous rainfall and temperature z-scores, respectively, in figures 3 and 4. The observed rainfall anomalies appear within the realm of normal, since the majority of the z-score values lie between one standard deviation below and above the mean (figure 3). From the perspective of risk, the sample was much more subjected to colder spells (figure 4). Although our preferred specifications focus on this component of risk, our empirical specifications of migration consider the influence of climate levels, lagged climate factors, as well as differences between positive and negative anomalies.

Figure 3.

Distribution of contemporaneous rainfall z-score.

Figure 4.

Distribution of contemporaneous temperature z-score.

3.4 Descriptive statistics

Table 2 displays the mean and standard deviations of the outcomes and explanatory variables utilized in the analysis. The majority of working age individuals in the CGP sites are between 19 and 35 years old, living in households with an average of 2 children (ages 6 through 18). From these households, 8 per cent of males and 5 per cent of females migrated on average at each round. Both men and women slightly favored destinations near their local village relative to another town in Zambia. International migration was rare. Only 0.1 per cent of men and women reported moving abroad.

Table 2.

Summary statistics

Notes: Standard deviations in parentheses. The absolute value of the anomaly is reported for both the negative and positive rainfall and temperature anomalies.

4. Methodology

Our primary interest is to quantify whether receiving an unconditional cash transfer T affects migration adaptation strategies, using the following linear probability model:Footnote ⁴

(1)\begin{equation} M_{ivt} = \alpha + \delta _t + \delta _d + \beta _TT_v + \beta _{TC}T_vC_{vt} + \beta _CC_{vt} + \beta _XX_i + \varepsilon _{ivt}.\end{equation}

The outcome M documents the movement of individual i out of village v in survey year t. We further differentiate between short-distance (destination is in a nearby village) and long-distance (destination is in a town in Zambia or abroad) moves. A vector of climate anomalies C is added to the model to detect the extent to which migration is used by the household to manage risk. The vector T × C allows us to capture how the migration coping changes upon receipt of the cash transfer.

Additional explanatory variables are included in (1) to improve the precision of our primary estimates of interest. Vector X _i includes individual and household demographic control variables collected at baseline: categorical variables for age (19–35, 36–55, >55 years old), number of household members by age category (6–12, 13–18, 19–35, 36–55, 56–69 years old), and a dummy variable for above median per capita household consumption expenditure. Fixed effects at the district δ _d and year δ _t account for all time-invariant characteristics at the district level and time-varying macroeconomic variables that may be correlated with climate shocks. Standard errors are clustered at the village level to account for within-village correlation between unobserved factors that affect migration.Footnote ⁵

We restrict the sample to those who were ages 15 through 65 at baseline to focus on the working population. Separate regressions are estimated for male and female adult samples. Additional specifications replace the contemporaneous climate z-scores C, with lagged climate z-scores, with current and lagged raw climate values, and differentiate positive from negative anomalies.

To accept our first null hypothesis, the estimate of β _TC in (1) would be negative. The fact that the assignment of who receives the cash transfers is randomly stratified equally across treatment and control communities ensures our ability to interpret β _TC as a causal estimate. To internally validate the experimental design, we demonstrate that the baseline outcome and covariates are balanced across treatment and control households in table 3, and reject the joint statistical significance of the baseline variables on treatment using an F test (p-value = 0.29 for the male sample and p-value = 0.18 for the female sample).

Table 3.

Average individual and household baseline characteristics by treatment

Notes: Standard errors reported in parentheses. P values in brackets for t tests of difference in means. F statistic testing joint significance of all variables for male sample is 1.19 (p-value = 0.29). F statistic testing joint significance of all variables for female sample is 1.37 (p-value = 0.18).

The estimate β _TC can be interpreted in two ways. For the purpose of testing our first hypothesis, we assume it reflects how risk management strategies change with the receipt of a cash transfer. However, it is important to note that the parameter may also be indicative of how migratory responses to the cash transfer vary with climate conditions. To confirm that the financial instruments are affecting migration adaptation strategies, we also provide F statistics testing the hypothesis β _TCT _v + β _C = 0.

To examine the validity of our second hypothesis, we estimate (1) for different wealth subgroups, using baseline consumption as a proxy.Footnote ⁶ Wealth is defined by whether the household had below or above median per capita consumption expenditure at baseline. We then evaluate whether β _TC is positive for the asset poor subgroup to determine if the intervention rendered countervailing effects on the migration of individuals given heterogeneity in wealth.

Finally, we add three more terms to (1) to test our third hypothesis: $\beta _T^H T_iH_i$, $\beta _{TC}^H T_iC_{vt}H_i,\; {\rm \;} \beta _{TH}^H C_{vt}H_i$. Here, H _i is an indicator for the 2013 survey year. The linear term of H _i in this alternative specification is already present in (1) in our survey fixed effect δ _t. Our third hypothesis explicitly tests whether the triple interacted variable is positive, or $\beta _{TC}^H \gt 0$. In other words, we are assessing whether tendencies towards climate migration dampen at the end of the CGP.