3.1 Data collection, sample size, and selection procedures
We collected data using the following procedure: In 2009, we conducted participatory rural appraisals (PRA) in ten villages using a checklist of 113 questions addressing the causes and trends of migration, different types of migrants, impacts of migration, remittances and their uses for livelihoods and investments, on-farm labor, off-farm rural activities, agricultural technologies, community activities, and natural resource management.
The Directorate of Extension in Aleppo provided us with lists of farm households, regardless of their asset ownership. Based on that, we selected the households’ sample for the formal survey using a multi-stage sampling procedure. Due to a lack of prior data on variances of income and number of migrants within and across the villages, we included 25 % of the total 120 villages in the study area in our research sample.
The online “sample size calculator” formula determined the size of our sample at 577 households as the minimum sample size to ensure 95 % confidence and 4 % precision levels. The second step was to distribute the total sample size proportionally among selected villages using, as suggested by the theory, a 50-50 weighting between the population and the number of households in each village. We considered five as a minimum sample size from each village. Accordingly, in cases where the minimum sample size determined is below five, it was increased to five for which the total sample size became 608. Finally, we used a simple random sampling procedure to select respective numbers of households from each village (Abdelali-Martini and Hamza 2014).
We targeted heads of households (almost all men) in our survey, and we defined migrants as individuals who spent any period of time away from home during the past 12 months at the time of the survey. Migration included off-farm activities performed through daily commuting to Aleppo City or neighboring towns. Additionally, ten focus group discussions were held with women and men in the communities to complete our understanding from women migrants and non-migrants. We gathered detailed information about household capitals (human, social, physical, natural, and financial capitals), in addition to migrants’ profiles (individual characteristics, destination of migration, and types of work performed by each member) as well as information about remittances.
The paper focuses on analyzing migrants as individual observations for further understanding migrants’ behavior regarding the choice of migration’s destination and sector of employment as well as the use of their remittances. These observations constitute a sample of 349 migrants of which 49 are women (14 %). We replicated the community and household characteristics (target area, household type, household members and total land, number of sheep and goats, number of income sources, land reclamation, development project beneficiaries, as well as the number of migrants from the household) for each migrant. Additional variables such as socio-demographic (age, sex, education level (number of years)), institutional (relationship to the head of the household, destination of migrants), and economic (type of work performed and the amount of remittances) were included.
We used three categories of variables in our empirical analysis:
-
1.
Individual characteristics of migrants including migrant’s age, educational level in years, sex, marital status, and whether the migrant interviewed was the head of the household or not
-
2.
Socio-demographic and economic household characteristics such as household size, land per capita, number of sheep, and the economic status of households (poor or less poor) and the number of other migrants in the household
-
3.
Community characteristics (e.g., the agricultural stability zone)
3.2 Analytical framework
In this study, we simulated an empirical multinomial logit model (Mora and Taylor 2006). In their paper, Mora and Taylor (2006) intended to estimate “the differential net effects of individual, family and community variables on migration outcome” using the probability that individual j is combined with destination-and-sector regime d as follows:
$$ \mathrm{prob}\left({U}_d^i\forall j\ne d\right)=\frac{e^{\beta_d{Z}^i}}{{\displaystyle {\sum}_{j=0}^J}{e}^{\beta_j{Z}^i}} $$
(1)
where Z
i is a vector of i’s personal, household, and community characteristics.
In our paper, we incorporated personal, household, and community characteristics to estimate the factors determining migrants’ destination and sector of employment. We used two logit models in our analysis: The first one includes the destination and the other the sector of employment. Consequently, in order to estimate the factors affecting the amount of remittances, we used the Heckman (two-stage) model. When modeling the main drivers of migrant remittances, it is important to indicate that only part of the households that reported members’ migration received a transfer (of any amount). Given this special case on which the dependent variable is censored, the application of the ordinary least squares (OLS) method would not be satisfied. This problem has been treated for a long time in the econometric literature, and two alternative approaches are used to solve it. The first approach consists of modeling the remitting decision and interprets the factors affecting the probability that a household would ever receive a transfer and then use the corrected OLS to model the amount transferred. In fact, this is an application of the standard Heckman (1979) two-step procedure and has been used by Banerjee (1984), Cox (1987), Hoddinott (1994), and Zhu (2002) in modeling remittances. This approach has the advantage of treating the remitting as a two-stage process. In the first stage, the decision whether to remit or not takes place and in the second stage follows the decision on the amount of transfer.
However, as noted by Hoddinott (1994), in none of the theoretical literature on migration and remittances, there is a distinction made between factors influencing the decision to remit and the level of remittances. It is possible to avoid such a challenge by adopting a second approach, which assumes that the decision to remit and level of remittances are made simultaneously. We used a censored Tobit model that uses data from both remitters and non-remitters, where the independent variable has two effects: it affects the probability of migrants falling in the remitting sub-sample and the amounts they remit. The maximum likelihood estimation of this model yields parameter estimates that are consistent in the context of modeling remittance behavior and has the disadvantage that a given determinant is restricted to having the same sign effect on the decision to remit as on the size of the remitted amount (Hoddinott 1994). It is therefore possible to explore both econometric procedures in this case.
As using the first approach, we define a two-stage sequential remitting process to correct the selectivity bias:
$$ {r}_i=\gamma {X}_i+{u}_i $$
(2)
and
$$ {R}_i=\beta \hbox{'}{X}_i+{\varepsilon}_i $$
(3)
where i indexes households, r
i
is the binary variable denoting the decision to send remittances: r
i
= 1 if a migrant sends remittances and r
i
= 0 if the migrant remits zero, R
i
is the size (amount) of remittances received by the household i, γ and β are vectors of parameter estimates, X
i
is a vector of remittance determining variables and characteristics for household i, and u
i
, ε
i
denote the error terms. Following Hoddinott (1994), the estimation of two separate equations, Eqs. (2) and (3), implicitly assumes that emigrants take the decision whether or not to remit and how much to remit sequentially. Thus, in order to obtain consistent and efficient estimates, we used the Heckman (1979) two-step procedure.
The second model in our empirical analysis of remittance functions is the so-called censored Tobit, specified as follows:
$$ \begin{array}{cc}\hfill {R}_i^{*}=\beta \hbox{'}{X}_i+{u}_i\hfill & \hfill {u}_i\approx N\left(0,{\sigma}^2\right)\hfill \end{array} $$
(4)
where
$$ {R}_i=\left\{\begin{array}{cc}\hfill {R}_i^{*},\ \mathrm{if}\ \beta \hbox{'}{X}_i+{u}_i\succ 0\hfill & \hfill \left(\mathrm{the}\ \mathrm{observed}\ \mathrm{values}\right)\hfill \\ {}\hfill 0,\ \mathrm{otherwise}\hfill & \hfill \left(\mathrm{the}\ \mathrm{unobserved}\ \mathrm{values}\right)\hfill \end{array}\right. $$
(5)
Let \( {R}_i^{*} \) the partial latent dependent variable that captures the ith individual’s propensity to remit and R
i
the observed value of amount remitted by the ith individual. Equation (5) indicates to us that R
i
is either positive or zero. It is important to indicate here also that X
i
denotes the vector of remittance determining variables for the ith individual (characteristics) and u
i
the error term. We estimated β and σ using the maximum likelihood method.
3.3 Selection of variables and specifications
We selected a set of variables for analysis using the logit models. They include community, households, and individual characteristic variables.