Does access to credit alter migration intentions?

Abstract

This study investigates whether access to credit affects migration intentions. Using Gallup survey data for the years 2009 and 2010 for 17 African countries, we document a negative link between the ability to borrow and the desire to migrate. Being able to borrow reduces the likelihood of reporting wanting to migrate, especially for those with some education, those with lower income, for individuals with a bank account, and for those who feel their assets are safe. To deal with endogeneity, we assume that migration desire is driven by borrowing which in turn is determined by access to financial services. Therefore, we estimate a two-equation system for migration and borrowing, using variables describing access to financial services as instruments. We verify our findings using identification through heteroscedasticity and using a geographical instrument. Our results indicate that efforts to increase credit access in developing economies can cement residents’ attachment to their home country.

Appendices

A: Wealth Index construction

Using joint correspondence analysis (JCA) on the following Yes or No questions, we construct a wealth index such that higher values imply higher wealth.

• Have there been times in the past 12 months when you did not have enough money to provide adequate shelter or housing for you and your family?

• Have there been times in the past 12 months when you did not have enough money to buy food that you or your family needed?

• Does your home have a cellular phone?

• Does your home have a television?

The proportion of variance explained by the first component is 68.37. It is worth noting that in Dustmann and Okatenko (2014) the authors use additional variables capturing internet access and electricity. However, our data do not include such information. Additionally, while Dustmann and Okatenko (2014) use polychoric principal component analysis to construct their index, we use JCA.

Table 9 Weak instruments tests (Appendix B)

B: Two tests for weak instruments

We run two more tests to check whether the IVs are weak and report them in Table 9. We report the test results for specifications in columns 1, 2, and 3 of Table 4. As suggested by Stock et al. (2002), the first test checks for relative bias in our 2SLS estimator compared to the standard OLS. We find that our 2SLS estimator suffers at most 5% of the bias delivered by the OLS estimator. This is evident by the fact that our test statistic reported in Table 9 is larger than the critical value for 5% relative bias. Therefore, our 2SLS estimator is superior to OLS and we can thus conclude that our instruments are not weak. The second test checks for the presence of large size distortions in our 2SLS estimates that can arise if our instruments are weak. As can be seen in Table 9, our test statistic is larger than the critical value, indicating that a Wald test at the 5% level can have an actual rejection rate of no more than 10%. Therefore, we supply additional evidence rejecting the null hypothesis of weak instruments.

Table 10 Bivariate probit estimation (Appendix C)

C: Biprobit model

We rewrite the two equations given by 1 and 2 as follows:

$$\begin{aligned} {\textrm{prob}}[m_{ij}=1]&=\Phi [x_{ij}' \beta + b_{ij} \alpha + c_j' \delta _1] \end{aligned}$$

(3)

$$\begin{aligned} {\textrm{prob}}[b_{ij}=1]&=\Phi [z_{ij}' \gamma + c_j' \delta _2], \end{aligned}$$

(4)

where $\Phi$ is the evaluation of the standard normal cumulative distribution function (cdf). Then, we estimate Eqs. 3 and 4 jointly by Maximum Likelihood (ML) as a biprobit. To do so, we assume that the random errors $u_{ij}$ and $v_{ij}$ follow a bivariate normal distribution and allow that ${\textrm{cov}}(v_{ij},u_{ij})=\rho$. Given our discussion above on omitted variables bias and reverse causality, we suspect that $b_{ij}$ is correlated with the error term $u_{ij}$ and therefore is not exogenous, we expect the parameter determining the correlation between the aforementioned random errors ($\rho$) to not be zero. If $b_{ij}$ were exogenous, then the correlation would be zero. Although we can use a two-step procedure to estimate a linear probability model using 2SLS, we cannot replicate this procedure here since the endogenous variable in our model is binary (Wooldridge 2010). In any case, Angrist (1991) shows that 2SLS and biprobit can deliver similar results.

The bivariate probit model is econometrically identified without an exclusion restriction. This provides another avenue to evaluate the identification arising from the exclusion restriction by comparing the estimated model that includes the exclusion restriction, to the estimated model that is identified purely from functional form. Therefore, we calculate the Akaike’s information criterion (AIC) and Bayesian information criterion (BIC) for two different models. The first model is the one with the exclusion restriction where we use bank account ownership and cost of banking variables as instruments. The second model is one with the migration and borrowing equations determined by the same exogenous variables and no exclusion restrictions are imposed (variables are listed in Table 10, column 1). We find that the model with the exclusion restrictions delivers a smaller AIC and BIC (AIC = 26753.39, BIC = 27153.29), compared to the model without an exclusion restriction (AIC = 27275.9, BIC = 27660.43). Therefore, the model with the exclusion restriction fits the data best according to these criteria.¹⁹

D: Descriptive statistics

In Table 11, we report the average, standard deviation, minimum, and maximum of all variables used in the analysis.

Table 11 Descriptive statistics (Appendix D)