Labor Migration from North Africa Development Impact, Challenges, and Policy Options

Transcription

1 Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Public Disclosure Authorized Middle East and North Africa Region Labor Migration from North Africa Development Impact, Challenges, and Policy Options A project implemented by the World Bank Volume 2 Statistical Appendix A project funded by the European Union This report as well as the background research underlying the analysis and conclusions of this report constitute part of an EC- Funded World Bank Program of International Migration from Middle East and North Africa and Poverty Reduction Strategies, a program of migration-related research and activities to identify and support the implementation of projects, policies, regional arrangements, and institutional reforms that will maximize the benefits of international migration flows and reduce their costs. The views herein are those of the authors and should not be attributed to the World Bank, the European Commission, or the institutions and countries they represent. Keller MNA vol2.indd 1

2 Labor Migration from North Africa Development Impact, Challenges, and Policy Options Keller MNA vol2.indd 2

3 Table of Contents Appendix 1: Appendix 2: Appendix 3: Appendix 4: Appendix 5: Appendix 6: Appendix 7: Appendix 8: Appendix 9: Measuring Growth, Accumulation, and TFP Growth...1 Description of Migration/Remittances and Poverty Analysis in Egypt: Methodology and Results...5 Description of Remittances and Poverty Analysis in Morocco: Methodology and Results...19 Description of Migration/Remittances and Labor Market/ Employment Analysis in Egypt: Methodology and Results...21 Description of Migration/Remittances and Labor Market/ Employment Analysis in Morocco: Methodology and Results...69 Description of Migration/Remittances and Decisions Affecting Children in Egypt: Methodology and Results...77 Description of Migration/Remittances and Decisions Affecting Children in Morocco: Methodology and Results...89 Impact of Remittances on Growth: Methodology and Results...93 Return Migration and Occupational Mobility...97 Appendix 10: Return Migration and Entrepreneurship Analysis Appendix 11: Review of Institutional And Legal Framework for Migration in Spain and the Netherlands Appendix 12: Computable General Equilibrium Analysis of Impact of Increasing MENA to Europe Migration iii Keller MNA vol2.indd 3

4 Labor Migration from North Africa Development Impact, Challenges, and Policy Options Keller MNA vol2.indd 4

5 Appendix 1: Measuring Growth, Accumulation, and TFP Growth (From Keller and Nabli, 2007, with estimates updated by Keller 2009) To examine how the MENA 1 region s growth has changed since it began its comprehensive structural reform process, we made simple calculations of the change in both the region s rate of accumulation, as well as the region s total factor productivity (TFP) growth. TFP growth is the residual of what cannot be explained by investments if we assume those investments (both physical and human) earn a reasonable rate of return. TFP growth is often thought of as technical progress, but in fact, as the residual of a growth accounting estimation, it not only embodies the differences across countries in their progress in the adoption of better technology, but also reflects a host of nontechnological differences, including changes in the utilization of both capital and labor, changes in schooling quality, and changes in the overall efficiency with which factors are allocated in the production process. Because of the many other factors that can potentially affect the growth residual, much empirical work has focused on reducing those elements of the residual (TFP) that do not reflect actual shifts in technological opportunities in the economy. For example, adjustments for the business cycle have been introduced to account for the short-term fluctuations in capacity utilization (Griliches, 1979; Lefort and Solimano; 1994; Fajnzylber and Lederman, 2000). An alternative procedure employed by Griliches and Lichtenberg (1984) has been to estimate growth over five-year periods, and to only allow the TFP series to increase or stay constant (resetting any values to the previously observed peak level) to maintain the assumption that true productivity can only improve and that measured reductions in TFP can only reflect short-term fluctuations. For our purposes, we have adopted a more casual approach about our measurements. Our interest is to explore how MENA and North Africa s overall growth has improved or deteriorated since it began the structural reform process. In the end, growth will be determined by both accumulation of physical and human capital, as well as the overall manner in which those factors are put to production. For the MENA region, things such as improved capacity utilization of capital and human capital by the region are precisely the elements we believe may be heavily affected by structural reform, and thus we would like to have this effect reflected in our estimates. At the same time, as we discuss in the subsequent section, we have controlled for global shocks. Under many circumstances, the environment created to encourage investment would also correspond to an environment in which those investments could be productive. But in the MENA region, accumulation and productivity have often gone in opposite directions, such as during the period of massive public sector investments that yielded rates of return well below international 1 The Middle East and North Africa region (MENA) comprises Algeria, Bahrain, Djibouti, Egypt, Iran, Iraq, Jordan, Kuwait, Lebanon, Libya, Morocco, Oman, Qatar, Saudi Arabia, Syria, Tunisia, United Arab Emirates, West Bank and Gaza, and Yemen. 1 Keller MNA vol2.indd 1

6 Labor Migration from North Africa Development Impact, Challenges, and Policy Options norms. Examining growth alone will mask these very different effects, and the somewhat anemic growth that has characterized the region since reform may be more a reflection of significantly lower public investments than of continuing poor productivity performance. And from the standpoint of evaluating the impact of the region s structural reform, it is precisely TFP growth that we would expect to be most influenced by changes in national policies that enhance the efficiency of capital and labor. Data and Methodology TFP growth estimates were made utilizing panel data of capital stock accumulation, human capital stock accumulation, and GDP growth from Estimates of the physical capital stock for a sample of 83 economies from 1960 to 1990 come from Nehru and Dhareshwar (1993), which was created by a perpetual inventory method from investment rates from 1950 forward, with initial assumptions about the capital/output ratio, and assuming a common fixed annual geometric depreciation rate of These capital stock data were extended to 2005 using the growth rates of constant price local currency investment from the World Bank s World Development Indicators database, 2 and applying similar assumptions on the depreciation rate. Capital stock estimates for another 12 economies, including four economies in the MENA region of particular interest to us, 3 were created according to a similar methodology, using investment rates from 1960 forward. Real GDP in constant local currency also comes from World Bank data. The human capital-augmented labor stock was estimated, using both labor force estimates from the World Bank s World Development Indicators, and estimates of the educational attainment of the adult population from Barro and Lee. 4 The functional form of human capital augmented labor has been assumed as: H = L e (r * S) where L is the labor force and S is the average years of schooling of the adult population, and r is the rate of return to schooling. According to international evidence, a reasonable approximation of that rate of return is 10 percent, which we have assumed for the purposes of our analysis. TFP growth was calculated over ten-year periods from , and then from , rather than on an annual basis, to minimize the error that is inherent in current capital stock measurements. National accounts would attribute any investment expenditures made over the year, even the last day of the year, to that year s capital stock. However, it is unlikely that that investment expenditure would contribute to economic growth immediately, but rather would only create the potential to contribute to growth into the future. To reduce this lag-effect that physical capital exhibits, we calculated TFP growth based on ten-year averages (except for the final period, which is a five-year average). Production was assumed to follow a Cobb- Douglas specification with constant returns to scale between physical and human-capitalaugmented labor: Y t = A (t) * K ta *H t (1 a) where Y is output, A is an index of total factor productivity, and K and H are the stocks of physical and human-augmented labor, respectively. Dividing both sides by the work force, taking logs, and first-differencing, growth of output per laborer can be related as follows: ln (y i / y i 1, ) = a ln (k t / k t 1 ) + (1 a) ln (h t / h t 1 ) + ln (A t / A t 1 ) To determine the coefficients on capital and human-capital augmented z, a and (1-a), the average annual rate of GDP per capita growth over the decade was regressed on average growth of physical capital per worker and human-capital 2 In the case of MENA economies, where there were inconsistencies, the World Bank MENA regional database investment series was preferred. 3 The four focus countries of this study include Algeria, Morocco, Tunisia and Egypt. 4 Barro and Lee, Educational attainment data (available until 1999) were extended to 2000 assuming constant growth between Keller MNA vol2.indd 2

7 Appendix 1: Measuring growth, accumulation, and TFP growth per worker with a least squares trend over the entire period of availability ( ). From our estimation, the elasticity of output of physical capital was estimated to be 0.49, somewhat higher than the average estimated coefficient from previous research, but within the range of accepted parameters. This may be due to the inclusion of several more developing countries than in the original Nehru-Dhareshwar physical capital stock dataset, made possible using World Bank data. At the same time, our purpose here is not to break new ground in measuring TFP, but to evaluate the region s performance in factor allocation and efficiency. Thus, we have calculated the TFP using three distinct calculation of factor shares ak=0.3, ak=0.4, and ak=0.5, to check the sensitivity of the region s growth performance to the assumptions made on the output elasticities. The results, in terms of orders of magnitude, are robust to changes in these elasticities. Final TFP estimates utilized an assumed output elasticity with respect to capital of Keller MNA vol2.indd 3

8 Labor Migration from North Africa Development Impact, Challenges, and Policy Options Keller MNA vol2.indd 4

9 Appendix 2: Description of Migration/ Remittances and Poverty Analysis in Egypt: Methodology and Results (The following description is taken from Roushdy, Assaad, and Rashed, 2009) The analysis mainly relies on data from the Egypt Labor Market Panel survey of 2006 (ELMPS 06), which is one of the first true nationwide longitudinal surveys to be carried out in Egypt. It attempted to track households and individuals first interviewed in 1998 as part of the Egypt Labor Market Survey of 1998 (ELMS 98) and re-interview them in Both the ELMPS 06 and ELMS 98 were conducted by the Economic Research Forum (ERF) in cooperation with CAPMAS. The ELMS 98 was carried out on a nationally representative sample of 4,816 households. The ELMPS 06 tracks the labor market and demographic characteristics of the households and individuals interviewed in 1998, and any new households that might have formed as a result of splits from the original households. The ELMPS 06 sample consists of a total of 8,349 households distributed as follows: (i) 3,684 households from the original ELMS 98 survey, (ii) 2,167 new households that emerged from these households as a result of splits, and (iii) a refresher sample of 2,498 households. Of the 23,997 individuals interviewed in 1998, 17,357 (72 percent) were successfully re-interviewed in 2006, forming a panel that can be used for longitudinal analysis. The 2006 sample contains an additional 19,743 new individuals. Of these, 2,663 individuals joined the original 1998 households, 4,880 joined the split households, and 12,200 were part of the refresher sample of households. 5 Our analysis of the attrition process that occurred in the panel tracked from 1998 to 2006 revealed that there are two distinct attrition processes at play. The first is if the entire household could not be located in 2006 and the second is when an individual who split from one of the households that were successfully tracked could not be found. The rate of the first type of attrition was about 23.6 percent (1,138 households) at the household level. More than 54 percent (615 households) of this first stage attrition cases resulted from the loss of identifying records of the households between 1998 and 2006, but, luckily, the process by which they were lost was almost entirely random (see Barsoum, 2008). The remaining attrition cases were due to the total relocation of the household, the death of all household members, or, in a few cases, refusal to participate in the survey. On the other hand, the second attrition process results from the inability to locate individuals who split from their original households, conditional on finding the original households in the first stage. The rate of this second type of attrition was about 15.4 percent. Of the 18,856 members of the 1998 households found in 2006, 14,661 were still in their original households, 790 had died, 220 had left the country, and 3,185 had split off to form separate households within Egypt. Of those splits, we successfully located 2,694 individuals, implying that the remaining 491 of the splits could not be located. 5 The data description and attrition analysis presented here is based on Assaad (2007) and Assaad & Roushdy (2008). 5 Keller MNA vol2.indd 5

10 Labor Migration from North Africa Development Impact, Challenges, and Policy Options An examination of the household and individual correlates in 1998 of those two attrition processes revealed that some household characteristics in 1998 were in fact systematically associated with the first type of attrition, but no individual characteristics in 1998 were associated with the second type of attrition (see Assaad and Roushdy 2008 for a detailed comparison of individuals/households who left versus those who stayed in the sample). We expect that very few cases of those missing households were due to migration of all of the household members, since migration in Egypt is often of a short-term nature by a single member in the household. However, we expect that households that migrate in their entirety would tend to be richer, since the whole family can afford to relocate. Hence, not correcting for this household-level attrition, when using the panel data, might lead to a downward biased estimate of the effect of migration on poverty. Accordingly, weights based on the probability of non-response were constructed to adjust the cross-sectional and panel samples from the ELMPS 06 for these attrition processes. Only the variables that were found to impact the probability of the first type of attrition in a significant way were used to predict the weights that correct for attrition. Those weights are applied whenever panel data is used in the analysis of this paper. The ELMPS 06 and ELMS 98 provide detailed information on household housing conditions, ownership of durables, access to basic services and the neighborhood infrastructure. It also contains a great deal of information on the household members education, employment status, time allocation, job mobility, earnings, migration, and household enterprises. With regard to migration questions, each round of the Egypt Labor Market Surveys (ELMSs) contains information on internal and international migration history (e.g., place of birth, year leaving place of birth, and the place and date of the previous two moves if different from the current place of residence). ELMS 98 includes only one (yes/no) question on whether the household receives remittances from relative(s) living abroad. However, in ELMPS 06, a new module on current migrants and remittances was added and it includes questions on whether the household receives remittances from household members living abroad, the amount and type of these remittances, and which household member receives the remittances. ELMPS 06 also includes information on the place and reason of migration for individuals who were in the household in 1998 but were not found in 2006 because they migrated between the 1998 and Although the two ELMSs are rich sources of information on labor market dynamics and individual and household characteristics, the ELMSs samples were not designed to measure migration. Accordingly, the number of migrants appearing in each of the ELMSs is fairly small. The ELMPS 06 sample contains about 603 return migrants (who migrated and returned before the 2006 survey interview) and 396 current migrants (who were still living abroad during the 2006 interview). While in the ELMS 98 there are only about 471 return migrants and no information was collected about current migrants. Hence, we do not expect to obtain accurate trends of migration and remittances flows from the ELMSs data that would coincide with official estimates. However, to the best of our knowledge, the ELMPS 06 is the only recent national household survey that collects information on incidence of international migration and remittances. The focus of the anlaysis uses the ELMPS 06 sample in the cross-sectional analysis, since it provides richer information, relative to ELMS 98, on international migration and remitances. Determinants of Migration and Remittances Before investigating the effect of migration and remittances on household poverty status, we are interested in exploring the household characteristics that might motivate the decision to migrate and remit. In this section, a Probit specification is used to model the likelihood of migration (receiving remittances) at the household level. The dependent variable takes the value 1 if the household, h, is a migrant household (remittances-recipient household) and zero otherwise. The explanatory variables consist of a set of the household and household head characteristics. It is worth mentioning here 6 Keller MNA vol2.indd 6

11 Appendix 2: Description of Migration/Remittances and Poverty Analysis in Egypt: Methodology and Results that we need to restrict the analysis to variables that are less likely to be caused by the migration decision per se. For instance, one should try to avoid variables such as: the number of children in the household below age 5, household wealth, residence, and current household head characteristics. Such variables are arguably endogenous to migration decision. The number of newly born/ young children is obviously affected by the spouse absence from the household. Household wealth and residence often change after migration. Also, the household head and his/her characteristics change if the original head is the migrant member. In the regression analysis of this section we try to avoid such variables. Instead of using current household head s characteristics in the regression analysis, we introduce a migration-neutral head as a substitute. If the current head is a male, we use the household head s spouse characteristics regardless of whether the household has a migrant or not. If the head is not married we use the characteristics of the oldest female (above age 15) living in the household. Only when the head is a male living alone, we use his own characteristics. We are aware that the characteristics of the migration-neutral head would have less explanatory power in comparison to that of the current household head, since under this definition the substitute head might have a marginal role in household decisions. 6 However, contrary to the current household head, we believe that the characteristics of this migration-neutral head are arguably exogenous to migration decision, since our sample shows that women generally do not migrate alone. Also, in Egypt generally, there exists a correlation between the characteristics of the household members; and hence we expect the characteristics of the migration-neutral head to be similar to that of the current household head. In the regression analysis, the household composition is captured by five variables: number of children age , number of unmarried males age 16 30, number of unmarried females age 16 30, number of elderly aged 64+, average years of schooling of males above age 18, and average years of schooling of females above age 18 in the household. The substitute head s characteristics include: age, marital status and education. The substitute head education is measured by the three dummies variables (illiterate or no degree is the omitted category): primary or preparatory degree, secondary degree, and above secondary degree. Marital status is captured by the two dummies (not married is the omitted category): married, and divorced or widowed. Moreover, since migration is a chain phenomenon, it is often expected that households belonging to traditionally migrant sender communities are more likely to have better social networks abroad which can potentially help in the migration process of other household members. Accordingly, in this analysis we include the following two variables to proxy for migration networks: the percent of households with at least one current migrant in the village/shiakha of the household and its interaction with the average years of schooling of adult members of the household. The percent of households with at least one current migrant in the village/shiakha of the household is obtained from the 2006 Census. 8 Such proxies have been frequently suggested in the literature. We believe that, in Egypt, these variables are good proxies of the size of the household s migration network abroad. We also expect that the adult members of the households, specifically those who are more educated, would make better use of the information available through their networks. Results Table A1 shows the regression results of the migration and remittances decisions. In this paper we report marginal effects as well as Huber- White adjusted standard errors to account for heteroskedasticity in all tables. 9 In both tables, 6 A better alternative for the migration-neutral head is to use the characteristics of the household head before migration. Unfortunately, this information is not available in the data. 7 As discussed in the previous section, migration in Egypt is often of a short term nature; hence, the number of children above age 6 (relative to the number of children less than age 5) are less likely to be affected by the spouse absence from the household. 8 As has been suggested in the literature (Section 4.2), it would have been better to use the lagged/historical migration levels instead of the same year of the household survey, but, unfortunately, migration information was not collected in censuses prior to that conducted in Marginal effects are based on marginal change for continuous variables and change from 0 to 1 for dummy variables. Coefficients are available upon request. 7 Keller MNA vol2.indd 7

12 Labor Migration from North Africa Development Impact, Challenges, and Policy Options Table A1: Determinants of Egyptian migration and receiving remittances Household level, 2006 Migration Remittances Variables (1) (2) (3) (4) No Children (0.002) No. Males *** (0.003) No. Females *** (0.003) No. Elderly (0.006) Avg. Male 18+ Years of schooling 0.006*** (0.000) Avg. Female 18+ Years of schooling (0.001) Household substitute head characteristics Age (0.001) Age square (0.000) Married(d) (0.016) Divorced or Widowed(d) *** (0.009) Primary or preparatory degree(d) (0.010) Secondary degree(d) * (0.013) Above secondary degree(d) ** (0.017) (0.002) 0.008*** (0.002) 0.006** (0.003) (0.005) 0.005*** (0.000) (0.001) (0.001) (0.000) (0.014) 0.029*** (0.008) (0.009) (0.012) 0.028* (0.015) % of HHs with Migrants in Shiakha/village from Census *** (0.132) % of HHs with Migrants in Shiakha/village x Avg. Yrs of schooling of *** (0.015) (0.001) 0.003* (0.002) 0.003* (0.002) (0.004) 0.004*** (0.000) 0.002*** (0.001) (0.001) (0.000) (0.011) 0.025*** (0.004) (0.005) (0.007) (0.008) (0.001) 0.003* (0.002) (0.002) (0.003) 0.004*** (0.000) 0.002*** (0.000) (0.001) (0.000) (0.010) 0.021*** (0.003) (0.004) (0.006) (0.006) 0.242*** (0.080) 0.041*** (0.009) Observations Pseudo R-squared Notes: Marginal effects are reported and robust standard errors in parentheses (d) for discrete change of dummy variable from 0 to 1 *** p<0.01, ** p<0.05, * p<0.1 1 reference category: never married 2 reference category: no educational certificate column 1 and 3 control for the household composition and the substitute head s characteristics, while column 2 and 4 investigate the effect of the network variables. 10 The table shows that, in both specifications, households with larger numbers of males and 10 Unfortunately, the ELMPS data does not provide the remittances senders characteristics or the type of relationship of the sender to his/her home family. Thus, it is important to note here that in the absence of such variables, it is difficult to interpret these results as different motives for sending remittances (see Acosta 2006 for a discussion). 8 Keller MNA vol2.indd 8

13 Appendix 2: Description of Migration/Remittances and Poverty Analysis in Egypt: Methodology and Results females (age 15 29) are more likely to have a migrant member and receive remittances. Also, in all models, adult males average years of schooling decreases the likelihood of migration and receiving remittances, while the females average years of schooling only increases the likelihood of receiving remittances. These results should be taken with caution, since these might be the results of migration per se. As mentioned earlier, if migration selects on education and gender, adult males with higher education levels would be the ones who are more likely to migrate which in turn would lead to poorer endowment of human capital among males who stay in the household. Incidents of migration and remittances are less common among households with widowed or divorced substitute heads. The household substitute head education is only significant in the migration regression. A household whose substitute head has above secondary education, relative to illiterate heads, has a higher likelihood of having a migrant member by about four percentage points in specification 1 and by about three percentage points in specification 2. Controlling for the network variables improves the fit of the migration and remittances models. In both models, the migration network variables increase the likelihood of being a migrant and a remittances recipient household. Belonging to a village/shiakha that is traditionally migrant-sending increases the likelihood of migration and receiving remittances. More specifically, a one percent increase in the fraction of migrants in the village/shiakha increases the probability of migration by 58 percentage points (column 2) and the probability of receiving remittances by 24 percentage points (column 4). While, the interaction term of the percent of migrants and average years of schooling further increase the likelihood of migration by 7.6 percent and the likelihood of receiving remittances by 4 percent. This fits with our expectation that the more educated members of the household are those who are more likely to make use of the migration information available through their network. It is not surprising that the results of the regressions explaining the likelihood of receiving remittances are remarkably similar to those of the regressions explaining the likelihood of being a migrant household, since both dependent variables are highly correlated (R 2 = 0.686). In fact, as mentioned above, 66 percent of the households with at least one current migrant member receive remittances; although the data does not show whether remittances are actually received from those migrant family members. On the other hand, 86 percent of households receiving remittances have at least one current member abroad. Impact of migration and remittances on household poverty This section investigates the effect of migration and receiving remittances on household poverty status. The variable used to investigate the effect of remittances in the regression analysis is whether the household receives transfers from abroad, instead of the amount of remittances in order to avoid possibilities of recall bias. 11 The outcome variable of interest in this analysis is whether the household is poor or not. The following Probit regression is estimated to explain the poverty status of the household: Pr( Poor = 1 X, I ) = Φ( X β + I γ + e ) h h h h h h The outcome is a binary variable which takes the value 1 if the household h belongs to the lowest quintile of the wealth distribution and zero otherwise. X h is a vector of the household and the household head characteristics. The set of household and household head characteristics included in this poverty equation consists of: the household region of residence, number of children age 0 5, number of children age 6 15, number of unmarried male age 16 30, number of unmarried females age 16 30, number of elderly age 64+, average years of schooling of males age 18+, average years of schooling of females age 18+ in the household, and the substitute head 11 Since international transfers are generally considered another source of income, they traditionally tend to be underreported in household surveys in comparison to macroeconomics balance of payment figures. For a detailed discussion of this issue, see Freund and Spatafora (2005) and Acosta et al. (2006). 9 Keller MNA vol2.indd 9

14 Labor Migration from North Africa Development Impact, Challenges, and Policy Options age, age square, marital status, and education. Four interaction terms are also included: the interaction of migration (remittances) with a rural dummy of the household residence, and with the household head education dummies. Those interaction terms would allow us to investigate whether poverty alleviation impact of migration and remittances are higher for migrants from urban household versus those from rural households and whether this impact differs depending on the education status of the household. I h is an indicator of whether the household has a migrant member (receive remittances, respectively) and e h is the error term. Migration and remittances may be endogenous to household poverty. Also households may not be randomly selected into being migrant households or remittances recipient households. The literature has often depended on instrumental variables (IV) techniques to overcome such endogeneity and selection bias problems. However, since both poverty and migration (receiving remittances, respectively) are binary variables, the model estimation strategy is not a trivial choice. Newey (1987) argues that using a two-stage least square (2SLS) in case of a binary dependent outcome and a binary endogenous variable might lead to inconsistent estimates, and instead suggests the use of Amemiya s generalized least square (GLS) estimator (provided under the IVprobit command in STATA packages) in such occasions. Nevertheless, later on, Angrist (1991) provided certain conditions under which a two-stage linear model (2SLS) can perform well with binary endogenous variables models (Acosta, 2006). In this analysis, as a robustness check, we estimate a simple one equation Probit, a 2SLS and a GLS models. We also estimate a bivariate Probit (two equation Probit) model using the biprobit command in STATA but implement it as an IV estimation. This specification allows us to account for the binary nature of poverty and migration and, at the same time, deal with selfselection and endogeneity of migration (remittances) by allowing the error terms in both the poverty and migration (remittances) equations to be correlated. In the first-stage of each of the two-equation model estimations, we estimate the full model specification of the migration (remittances) equation presented in column 2 (4) of Table We use the two migration social network variables discussed above (the percent of households with at least one current migrant in the village/shiakha of the household and its interaction with the average years of schooling of adult members of the household) to instrument for migration and remittances. We believe that these instruments are good proxies of the local migration network, since households belonging to traditionally migrant sender communities are more likely to have better social networks abroad, which can potentially help in the migration process of other members. However, it is not easy to defend that the number of migrants at the community level impacts household living standard only through affecting migration; since, for instance, among the most important determinants of migration are labor market opportunities which affect both migration and poverty. One possible improvement, to reduce the effect of this potential problem, is to include others controls at the household community-level in the poverty equation. Accordingly, we include the following five variables to control for labor market structure at the cluster-level: the percent of unemployed adult males age 18 64, percent of males age working in agriculture, percent of males age working in the public sector, percent of males age working in private wage work, and the percent of males age with secondary or higher education. Moreover, for each specification of the bivariate and the ivprobit (corrected) models, we test the exogeneity of migration (remittances) to household poverty. The null hypothesis here is that the correlation between the error terms of the poverty and migration (remittances) equations, rho, is zero. If we cannot reject this null hypothesis, than we cannot reject that migration is exogenous to household poverty, that is, migration (remittances) is uncorrelated to the error 12 We also investigated other specifications and found that similar results are obtained for the poverty equation, when using any of those specifications. 10 Keller MNA vol2.indd 10

15 Appendix 2: Description of Migration/Remittances and Poverty Analysis in Egypt: Methodology and Results term of the poverty equation. In such case, the results of the single equation Probit model would be more efficient than those of the bivariate Probit model. On the other hand, if the error terms are strongly correlated (i.e., we cannot reject that the unobservables that affect the poverty status also influence the decision to migrate), we expect the size of coefficient of the migration (remittances) variable to be substantially larger in the corrected models than in the uncorrected single equation model. Additionally, the 2SLS estimation allows us to perform both an over-identification test and a weak instruments test. The Sargen s test for over-identification of the instrumental variables tests the null hypothesis that both instruments are valid; i.e. could be excluded from the poverty equation. A statistically significant Sargen s test statistic indicates that the instruments may not be valid. On the other hand, a test based on the Cragg Donald minimum eigenvalue statistic created by Cragg and Donald (1993) can be used to test for the weakness of instruments. The value of this statistic is compared to critical values provided by Stock and Yogo (2005). It provides measures of goodness of fit of the first-stage equation (migration and remittances). It also uses an F-statistic to test the null hypothesis that the coefficients on the instruments are equal zero in the first-stage equation. The F-statistic is often compared, in the literature, to the threshold of 10, which is suggested by Staiger and Stock (1997). An F-statistic below the threshold of 10 suggests the existence of a weak-instrument problem. Results Table A2 and A3 present the regression results of the effect of migration and receiving remittances on household poverty. Once again both tables report the marginal effects and the Huber- White standard errors. Column 2, 4, and 6 add the rural and education interaction terms to investigate whether those interactions have additional significant effects on household poverty. Column 1 and 2 present the uncorrected single equation Probit results, column 3 and 4 present the biprobit results, while column 5 and 6 present the 2SLS results. The GLS estimates are not reported in the tables as they yield similar results to that of the biprobit model. At the bottom of the tables, the goodness of fit measures, the p- value of Sargen s test for over-identification of the instruments and the statistics of the weak instrument test are reported. The first stage results of the migration (remittances) equation closely resemble the results presented in column 2 (4) of Table 10. Both instrumental variables are individually strongly significant (at 1 percent level of significant). Also, based on all the 2SLS models specifications of both the migration and remittances effects on poverty, Sargen s test for over-identification of the instruments does not reject the null hypothesis that both instruments are valid (p-values are substantially higher than the 10 percent level). Additionally, the weak identification test provides an F-statistic that is substantially higher than the threshold rule of thumb of 10. All the R 2 statistics of the first-stage regression are also relatively high, so they do not imply a weak-instrument problem. Hence, we can reject the null hypothesis that our two instruments are weak. On the other hand, for each of the biprobit specifications in Table A2 and A3, the value of the correlation between the error terms of the poverty and migration (remittances) equations, rho, and its significance level are reported. In the remittances and migration analysis, both the biprobit and ivprobit (GLS) model specifications lead to a p-value larger than 0.1 for the Wald-test of significance of rho (except for the biprobit specifications in the migration table the p-value is 0.07). 13 Hence, we cannot reject the null-hypotheses that rho=0 at 5 percent significance level. In other words, we cannot reject that the error term of the migration (remittances) equation is uncorrelated to the error term of the poverty equation. Accordingly, in this case we expect the coefficient results of both the corrected and uncorrected models to be considerably close. 13 The results of the ivprobit (GLS) estimation lead to p-values over 0.8 in all models. 11 Keller MNA vol2.indd 11

16 Labor Migration from North Africa Development Impact, Challenges, and Policy Options Table A2: The impact of Egyptian migration on poverty status of the household (1) (2) (3) (4) (5) (6) Variables Probit Probit Biprobit Biprobit 2SLS 2SLS International migrant in HH 0.075*** (0.011) Community controls % unemployed males age * (0.075) % males age working in agriculture 0.195*** (0.024) % males age working in public sector (0.030) % males age working in private wage work 0.121*** (0.026) % males age with secondary + education (0.021) Household characteristics No Children (0.004) No Children 6_ (0.003) No. Males (0.003) No. Females ** (0.004) No. Elderly (0.007) Avg. Male 18+ Years of schooling 0.007*** (0.001) Avg. Female 18+ Years of schooling 0.009*** (0.001) 0.077*** (0.011) 0.138* (0.075) 0.195*** (0.024) (0.030) 0.121*** (0.026) (0.021) (0.004) (0.003) (0.003) 0.010** (0.004) (0.007) 0.007*** (0.001) 0.009*** (0.001) 0.091*** (0.010) 0.135* (0.075) 0.200*** (0.025) (0.030) 0.121*** (0.026) (0.022) (0.004) (0.003) (0.003) 0.009** (0.004) (0.007) 0.007*** (0.001) 0.009*** (0.001) 0.087*** (0.009) 0.130* (0.071) 0.191*** (0.024) (0.029) 0.115*** (0.025) (0.021) (0.004) (0.003) (0.003) 0.009** (0.004) (0.007) 0.007*** (0.001) 0.008*** (0.001) 1.769*** (0.568) (0.132) 0.402*** (0.044) (0.042) 0.114*** (0.037) 0.097** (0.038) (0.006) 0.014*** (0.005) (0.007) 0.016** (0.007) (0.014) 0.017*** (0.002) 0.012*** (0.002) 1.554** (0.637) (0.080) 0.342*** (0.041) (0.035) 0.133*** (0.031) (0.028) (0.006) (0.005) 0.014*** (0.006) 0.011* (0.006) (0.012) 0.010*** (0.001) 0.015*** (0.002) Alexandria and Suez(d) (0.017) (0.017) (0.017) (0.016) (0.015) (0.012) Urban Lower Egypt(d) *** 0.070*** 0.072*** 0.069*** 0.071*** (0.020) (0.020) (0.020) (0.019) (0.023) (0.015) Urban Upper Egypt(d) *** 0.178*** 0.179*** 0.172*** 0.118*** 0.076*** (0.024) (0.024) (0.024) (0.023) (0.018) (0.016) Rural Lower Egypt(d) *** 0.105*** 0.106*** 0.101*** *** (0.019) (0.019) (0.019) (0.019) (0.022) (0.021) Rural Lower Egypt(d) *** 0.247*** 0.250*** 0.241*** 0.164*** 0.252*** (0.027) (0.027) (0.027) (0.027) (0.025) (0.024) Rural x Migrant HH *** 1.362*** (0.049) (0.050) (0.050) (0.049) (0.564) (0.489) Household substitute head characteristics Age 0.003** 0.003** 0.003** 0.003** *** (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) (continued on next page) 12 Keller MNA vol2.indd 12

17 Appendix 2: Description of Migration/Remittances and Poverty Analysis in Egypt: Methodology and Results Table A2: The impact of Egyptian migration on poverty status of the household (continued) (1) (2) (3) (4) (5) (6) Variables Probit Probit Biprobit Biprobit 2SLS 2SLS Age square 0.000** 0.000** 0.000** 0.000** *** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Married(d) *** 0.108*** 0.112*** 0.106*** 0.080** 0.071** (0.032) (0.032) (0.032) (0.031) (0.040) (0.034) Divorced or Widowed(d) *** 0.049*** 0.053*** 0.050*** 0.083** (0.015) (0.015) (0.015) (0.014) (0.042) (0.038) Primary or preparatory degree(d) (0.012) (0.012) (0.012) (0.011) (0.020) (0.027) Secondary degree(d) * 0.023* 0.022* 0.021* * (0.013) (0.013) (0.013) (0.013) (0.023) (0.030) Above secondary degree(d) *** 0.097*** 0.096*** 0.091*** 0.050* 0.066** (0.008) (0.008) (0.009) (0.008) (0.030) (0.033) Primary or Prep degree x Migrant HH ** (0.077) (0.073) (0.434) Secondary degree x Migrant HH ** (0.041) (0.045) (0.378) Above secondary degree x Migrant HH 0.085*** 1.404** (0.006) (0.559) Observations Pseudo R-squared rho 0.236* 0.245* Wald-test 0f rho=0 (p-value) Sargen s test of over-identification (p-value) Test of weak Instruments min eigenvalue statistic R 2 Adjusted R 2 F-test *** *** *** *** Notes: Marginal effects are reported and robust standard errors in parentheses (d) for discrete change of dummy variable from 0 to 1 *** p<0.01, ** p<0.05, * p<0.1 1 reference category: never married 2 reference category: Greater Cairo Region 3 reference category: no educational certificate 4 The interaction term Above secondary degree x Migrant HH predicted failure perfectly in 64 observations in the probit estimation. Hence, to avoid STATA dropping those cases, the interaction of secondary degree and above secondary degree has been combined in the probit specification of column (2). As shown in Table A2, the coefficient of interest, the effects of migration on household poverty in both the corrected and uncorrected Probit models are very close. Migration significantly decreases the likelihood of household poverty by about 8 percentage points in the uncorrected models (column 1 and 2) and by 9 percent points in the corrected models (column 3 and 4). Similarly, both the corrected and uncorrected models specification (Table 12) show that receiving remittances has the same effect on reducing poverty (around eight percentage points) as that of migration. Hence, this analysis suggests that, in Egypt, migration and receiving remittances have 13 Keller MNA vol2.indd 13

18 Labor Migration from North Africa Development Impact, Challenges, and Policy Options Table A3: The impact of remittances on poverty status of the Egyptian household (1) (2) (3) (4) (5) (6) Variables Probit Probit Biprobit Biprobit 2SLS 2SLS Household receive remittances 0.083*** 0.086*** 0.088*** 0.088*** 2.071*** 2.318** (0.008) (0.008) (0.008) (0.007) (0.672) (0.914) Community controls % unemployed males age * 0.137* 0.134* 0.132* 0.209** (0.075) (0.076) (0.075) (0.073) (0.085) (0.084) % males age working in agriculture % males age working in public sector % males age working in private wage work % males age with secondary + education Household characteristics 0.194*** (0.024) 0.197*** (0.025) 0.196*** (0.025) 0.192*** (0.024) 0.411*** (0.045) 0.329*** (0.043) (0.030) (0.030) (0.030) (0.029) (0.039) (0.035) 0.126*** (0.026) (0.021) 0.128*** (0.027) (0.022) 0.126*** (0.026) (0.021) 0.123*** (0.026) (0.021) 0.160*** (0.036) 0.093** (0.038) 0.121*** (0.033) (0.031) No Children (0.004) (0.004) (0.004) (0.004) (0.006) (0.006) No Children 6_ ** (0.003) (0.003) (0.003) (0.003) (0.005) (0.006) No. Males ** (0.003) (0.003) (0.003) (0.003) (0.006) (0.005) No. Females ** 0.010** 0.010** 0.010** 0.016** 0.013** (0.004) (0.004) (0.004) (0.004) (0.006) (0.006) No. Elderly * (0.007) (0.007) (0.007) (0.007) (0.012) (0.012) Avg Male 18+ Years of schooling 0.007*** 0.007*** 0.007*** 0.007*** 0.018*** 0.009*** (0.001) (0.001) (0.001) (0.001) (0.003) (0.001) Avg Female 18+ Years of schooling 0.009*** 0.009*** 0.009*** 0.008*** 0.010*** 0.015*** (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) Alexandria and Suez(d) (0.017) (0.017) (0.017) (0.017) (0.014) (0.014) Urban Lower Egypt(d) *** 0.071*** 0.071*** 0.069*** 0.073*** (0.020) (0.020) (0.020) (0.019) (0.022) (0.017) Urban Upper Egypt(d) *** 0.179*** 0.177*** 0.174*** 0.107*** 0.082*** (0.024) (0.024) (0.024) (0.024) (0.016) (0.015) Rural Lower Egypt(d) *** 0.105*** 0.104*** 0.102*** *** (0.019) (0.019) (0.019) (0.019) (0.019) (0.019) Rural Lower Egypt(d) *** 0.247*** 0.245*** 0.241*** 0.173*** 0.242*** (0.027) (0.027) (0.027) (0.027) (0.022) (0.023) Rural x HH receive remittances *** 1.956*** (0.096) (0.100) (0.097) (0.098) (0.664) (0.713) (continued on next page) 14 Keller MNA vol2.indd 14

19 Appendix 2: Description of Migration/Remittances and Poverty Analysis in Egypt: Methodology and Results Table A3: The impact of remittances on poverty status of the Egyptian household (continued) (1) (2) (3) (4) (5) (6) Variables Probit Probit Biprobit Biprobit 2SLS 2SLS Household substitute head characteristics Age 0.003** 0.003** 0.003** 0.003** *** (0.001) (0.001) (0.001) (0.001) (0.002) (0.002) Age square 0.000** 0.000** 0.000** 0.000** 0.000* 0.000*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Married(d) *** 0.113*** 0.113*** 0.111*** 0.085** (0.032) (0.032) (0.032) (0.032) (0.039) (0.038) Divorced or Widowed(d) *** 0.051*** 0.052*** 0.051*** 0.102** (0.015) (0.015) (0.015) (0.014) (0.043) (0.042) Primary or preparatory degree(d) (0.011) (0.012) (0.011) (0.011) (0.019) (0.026) Secondary degree(d) * 0.026** 0.024* 0.025** (0.013) (0.013) (0.013) (0.012) (0.021) (0.030) Above secondary degree(d) *** 0.098*** 0.097*** 0.095*** * (0.008) (0.008) (0.008) (0.008) (0.024) (0.033) Primary/Prep degree x HH receive remittances (0.090) Secondary degree x HH receive remittances 4 (0.064) Above secondary degree x HH receive remittances (0.089) (0.074) 0.084*** (0.005) 1.757** (0.709) 1.441*** (0.559) 2.144*** (0.820) Observations Pseudo R-squared rho Wald-test 0f rho=0 (p-value) Sargen s test of over-identification (p-value) Test of weak instruments min eigenvalue statistic R 2 Adjusted R 2 F-test *** *** *** *** Notes: Marginal effects are reported and robust standard errors in parentheses (d) for discrete change of dummy variable from 0 to 1 *** p<0.01, ** p<0.05, * p<0.1 1 reference category: Greater Cairo Region 2 reference category: never married 3 reference category: no educational certificate 4 The interaction term Above secondary degree x Migrant HH predicted failure perfectly in 64 observations in the probit estimation. Hence, to avoid STATA dropping those cases, the interaction of secondary degree and above secondary degree has been combined in the probit specification of column (2). 15 Keller MNA vol2.indd 15