Less is more?
Migrant rights and the impact of
remittances on poverty in the home
countries.
Ma lgorzata Majewska
A Master’s thesis under the supervision of
prof. dr. Joop Hartog
Abstract
International remittances have been consistently found to reduce poverty in developing countries. This thesis hypothesises that the effect might be influenced by the immigration policies in the migrant-receiving countries. The empirical strategy exploits the variation in the remittances received from the countries that grant the fewest and the most rights to the immigrants among low- and middle-income countries and bases on the methodology of Adams and Page (2005). Replication of Adams and Page (2005) with an extended and updated dataset yields results consistent with the original study, with the exception of country fixed effects models where the impact of remittances on poverty becomes small and insignificant. The empirical analysis does not provide a definitive answer to the research question but the results do suggest that the differences between the high- and low-rights countries affect the effectiveness of remittances with regard to poverty reduction.
Faculty of Economics and Business
University of Amsterdam
Statement of Originality
This document is written by Student Ma lgorzata Majewska who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in
creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Contents
1 Introduction 5
2 Literature Review 8
2.1 Impact of remittances on poverty . . . 8
2.1.1 Country-level studies . . . 8
2.1.2 Cross-country studies . . . 9
2.2 Determinants of remittances . . . 9
2.3 Temporary and permanent migration . . . 10
3 Methodology 11 3.1 Adams and Page (2005) methodology . . . 11
3.2 Changes to AP05 . . . 12
3.2.1 Variable of interest . . . 12
3.2.2 High- and low-rights countries . . . 13
3.2.3 Measures of poverty . . . 14
3.2.4 The model . . . 15
3.3 Data . . . 17
3.3.1 Sample . . . 17
3.3.2 Bilateral remittance matrices . . . 18
3.3.3 Other variables and their data sources . . . 18
3.3.4 Robustness checks . . . 19
4 AP05 Replication 20 4.1 Reconstructed AP05 dataset . . . 20
4.2 The new dataset . . . 22
5.1 OLS . . . 27 5.2 IV . . . 28 5.3 Panel regressions . . . 29 5.4 Robustness checks . . . 32 6 Conclusion 33 Bibliography 35 A Appendix 38 A.1 Data . . . 38
A.1.1 AP05 replication: description of the datasets . . . 38
A.1.2 Sample countries and their main remitting areas . . . 39
A.1.3 BRM generation . . . 40
A.2 AP05 replication additional results . . . 41
A.3 Robustness checks . . . 42
A.3.1 Weak instrument IV bias checks . . . 42
A.3.2 First extension of the high- and low-rights groups . . . 42
A.3.3 Second extension of the low-rights group . . . 48
1.
Introduction
In 2013 migrants around the world sent home as much as 560 billion dollars, and this num-ber only reflects the official data from the World Bank. This is over three times as much as the total Official Development Assistance flows in the same year (167 billion, OECD). Undoubtedly, such a large cash stream must play an important role in the development of the poorest countries.
The literature on the impact of remittances is not unequivocal in this regard, however. The effects on most dimensions of development are disputed, and evidence from different countries points in different directions. There is one notable exception – international remittances have been found to have a small, but negative effect on poverty, consistent across the literature. The magnitude of this effect might be dependent on other factors, such as the financial development of the home country (Majeed, 2014). On the other hand, it is possible that the characteristics of migrant’s destination affect the efficiency of remittances as well.
Donor countries have been found to have a large impact on how aid contributes to the development of the recipient countries. Moreover, the immigration policy of the destina-tion country could play an important role for the efficiency of internadestina-tional remittances in terms of poverty reduction. The Arab states of the Persian Gulf – one of the most popular migration destinations – serve as a good example of how such an effect could possibly come about.
The discovery of oil brought immense wealth to countries surrounding the Persian Gulf. However, particularly for the members of the Gulf Cooperation Council (GCC), the ex-traction of the benefits from the newly found resource required a substantial number of foreign workers. The demographic results of the process are striking, with the population in the six member states of the GCC increasing tenfold over little more than 50 years, mainly due to the influx of workers from abroad. Table 1.1 illustrates the unique features of the Gulf demographics: with the exception of Oman and Saudi Arabia, nationals are a minority, and in all countries they constitute less than a half of the employed popula-tion. This is even more pronounced in the private sector, which is largely dominated by foreigners.
However, this has not led to integration efforts. Conversely, dual societies have emerged, where the distinction between citizens and immigrants is very clear. The numerous limi-tations for foreigners (e.g. prohibition of intermarriage, lack of access to public education for migrants’ children) and the kafala (sponsorship) system are all supposed to ensure their stay in the host country is only temporary (Fargues, 2011). This feature of the labor market in the GCC and its implications resonate in the highly publicized instances of exploitation and human rights abuse in the region.
In contrast, in the policies followed by many other developed countries, guest-worker
programs are virtually nonexistent1 and migrants’ families are given priority in issuing
1In the postwar period, Western European countries ran guest-worker programs aimed at attracting
Table 1.1: The composition of GCC labor force.
% in total population % in employed population % non-nationals by sector
Country population nationals non-nationals nationals non-nationals private public
Bahrain 1,314,562 48.0 52.0 25 75 83 30 Kuwait 4,161,404 30.8 69.2 17 83 95 29 Oman 4,149,917 56.0 44.0 23 77 86 14 Qatar 1,699,435 14.3 85.7 6 94 99 56 Saudi Arabia 30,770,375 67.3 32.7 44 56 78 5 UAE 8,264,070 11.5 88.5 7 93 n.a. n.a. Total 50,359,763 51.9 48.1 - - -
-Source: Gulf Labour Markets and Migration (GLMM) programme. Most recent national data.
visas, allowing family reunion. Once admitted to the country, migrants enjoy similar rights to those of the citizens and temporary stay in some countries can lead to a permanent residency.
In my thesis, I hypothesise that the vast differences between the immigration policies in the major labour-importing regions can influence the remitting behaviour of the migrants and the decision on how to invest the transfers by remittance-receiving households in the home countries, thus translating into a differential effect on poverty reduction. In the GCC, the perspective of only limited time to exploit the wage differentials might induce the migrant and his family to invest the remittances in a more productive way, while remittances from European countries could be treated more as a gift from a distant relative.
To the best of my knowledge this question has not yet been addressed in a quantitative study. The attempt to provide an answer is made with a full consciousness of the binding data limitations, in particular regarding remittances. Nevertheless, bilateral remittance matrices, based on the stocks of migrants and the income differentials between their origin and destination countries, can provide an estimate of flows from countries with relatively few migrant rights and from those that grant them the most. By exploiting the variation in these flows across a group of nearly 100 developing countries over the period from 1990 to 2013, it is possible to determine whether they have a different effect on poverty reduction. Should substantial differences emerge, it could suggest that the money is invested in various manners and has a different impact on lifting the countries from poverty. The methodology follows Adams and Page (2005) who also estimate the impact of remittances on poverty in a large set of developing countries and is adjusted to provide the answer to the research question of this thesis.
Had a differential impact of remittances on poverty depending on host country’s im-migration policies been identified, an important policy implication would follow, as the balance in the “numbers vs. rights” trade-off would be affected. This term stands for the phenomenon of either admitting a large number of (low-skilled) migrants or granting them many rights. The trade-off has been noted by the academic literature both across countries and in the history of immigration policies in the US (Martin, 2004; Ruhs and
Martin, 2008)
Its emergence is explained by the downward sloping demand curve for low-skilled foreign labour (with regard to rights). On one hand, there is excess supply of low-skilled workers and, taking the situation in the home country as their reference point, they are willing to accept even very harsh conditions. On the other hand, granting them more rights represents additional cost for the employer thus lowering the demand. Moreover, the state might be interested in limiting the number of unskilled immigrants admitted, as their contribution to the tax system is likely to be lower than the benefits they would receive having been granted a status equal to the nationals (Ruhs and Martin, 2008). The structure of this thesis is as follows: Section 2 discusses the relevant literature on the impact of remittances on poverty, their determinants and the characteristics of immigra-tion policies and what it implies for the potential poverty-reducing impact of remittances flowing from different regions. Section 3 presents the methodology used and the dataset. Section 4 replicates Adams and Page (2005) whose methodology is the backbone of the empirical approach used in this thesis. The new hypothesis is tested in Section 5 and Section 6 concludes.
2.
Literature Review
2.1
Impact of remittances on poverty
Poverty reduction seems to be the one effect of remittances over which there is a consen-sus in the scholarly literature. A steady inflow of money allows households to cope with the everyday hardships, manage risk and accumulate capital assets (Maimbo and Ratha, 2005). A 2011 review of 50 recent empirical studies of the effect of international remit-tances on economic development (Adams, 2011) concludes that all the research suggests a small, but statistically significant poverty-reducing impact of remittances. As presented below, this is true on both country- and cross-country level.
2.1.1
Country-level studies
The evidence reviewed in this section has been collected on five continents, both through large, nationally representative surveys in major labour-exporting countries and smaller studies conducted in regions where migration is particularly important.
In Asia, one fifth of the poverty reduction in Nepal between 1995 and 2004 is suggested to have occurred due to the high levels of remittances sent home by migrant workers (Lokshin et al., 2010). Yang and Martinez (2007) are the only authors in this strand of literature to take advantage of a natural experiment – the 1997 Asian financial crisis. Sudden exchange rate changes in its aftermath varied across the locations of Filipino migrants. Exploiting this variation, they relate a 10 percent improvement in the exchange rate to an 0.6 percentage point decline in the poverty rate among migrants’ households and smaller in magnitude spillovers to households without migrant members.
With regard to Africa, Ghana has been the focus of Adams (2006) and Adams, Cuecuecha, and Page (2008). The first study, controlling for selection bias, concludes that internal and international remittances contribute to the decline of the level, depth and severity of poverty (see definitions in section 3.1.), international remittances being more effective at reducing the depth and severity of poverty. The latter study confirms these results and suggests that the receipt of remittances is related to greater spending at the margin on investment goods and smaller on food. Similarly, an early study on rural Egypt shows that the receipt of remittances allows households to invest in durables, especially improved housing (Adams, 1991).
In Latin America, the example of Mexico shows that again, international remittances decrease poverty more than the internal ones, and that a 10 percent increase in interna-tional remittances can lead to an 0.77 percent decrease in the poverty headcount ratio (Taylor et al., 2005). This is on par with the results from Guatemala, which, similarly to Ghana, indicate that remittances have a greater negative impact on the severity of poverty (Adams and Cuecuecha, 2010). Moreover, Bertoli and Marchetta (2014) point out, based on their research in Ecuador, that the poverty reducing effect is larger for
households with a migrant member as opposed to households that are only remittance recipients but do not have a member abroad.
Two recent papers focus on regions much less often studied in terms of the impact of remittances. Brown, Connell, and Jimenez-Soto (2014) find that in Fiji and Tonga, re-mittances are a substitute for the absent formal system of social protection and have a
poverty reducing effect. M¨ollers and Meyer’s (2014) study on Kosovo employs Propensity
Score Matching to construct counterfactual incomes and the analysis strongly suggests that remittances are an important tool in rural households’ fight against poverty.
2.1.2
Cross-country studies
The evidence from cross-country studies analyzing the impact of international remittances
and poverty supports these findings. On a regional level, Acosta, Calder´on, Fajnzylber,
and Lopez (2008), using household survey data from 10 Latin American and Caribbean countries show that remittances have small inequality- and poverty-reducing effects. Sim-ilarly, Imai, Gaiha, Ali, and Kaicker’s (2014) annual panel data for 24 countries in Asia and Pacific reveals that remittances alleviate poverty. Gupta, Pattillo, and Wagh’s (2009) study focusing on a sample of 24 Sub-Saharan countries confirms these results in the African setting.
Expanding the sample further does not change much. Adams and Page (2005), having collected survey data from 71 developing countries, find that a 10 percent increase in per capita international remittances leads to an average decline of the poverty headcount ratio of 3.5 percent. A similar magnitude (2 percent reduction in poverty headcount related to 10 percent increase in remittances to GDP ratio) is found in the 2005 World Economic Outlook (Spatafora, 2005). A more recent study, Majeed (2014), uses panel data covering 65 countries and the period from 1970 to 2008 to show that the impact of remittances on poverty depends on the level of financial development of the remittance-receiving country, with no positive effect of remittances on economies that are characterized by low financial development.
2.2
Determinants of remittances
Migration can be expected to alleviate poverty when it leads to transfers of money from the migrant to the country of origin. The decision to do so has been extensively studied and while it cannot be explained by a single motive, some distinctive patterns emerge from the research done all over the world.
Rapoport and Docquier (2006) consolidate previous theoretical discussions and present several general reasons to remit: altruism, exchange, strategic motives, insurance and moral hazard, repayment of loans for investments in education or migration and securing inheritance. Regardless of the dominating motive, they predict that a larger income of the migrant and adverse short-run shocks in the home country will increase the amounts remitted. Time since arrival and distance from the family is expected to lower the amounts
remitted, rather than increase them. The effect of migrant’s education will be negative if the exchange motive dominates and positive if the investment motive dominates.
These are largely on par with, among others, the results of a pioneering study of remittance behaviour in Botswana (Lucas and Stark, 1985), which finds remittances to be mainly intended as repayment of education loans and securing inheritance and to cease in the long term. Likewise, the investment motive and gradual decrease over time is stressed in a later study focused on Mexican migrants to the USA (Durand et al., 1996). The insurance role of remittances has been identified by research conducted in Mali (Gubert,
2002), the Philippines (Yang and Choi, 2007) and the Dominican Republic (De La Bri`ere
et al., 2002).
2.3
Temporary and permanent migration
Glytsos (1997) assumes that Greek migration to Germany in the period of 1960-1993 can be characterised as temporary while that to Australia as permanent. He then proceeds to hypothesise that the temporary migration to Germany had only been a way to accumulate capital and therefore the migrants would send more money home, while in the case of permanent migrants to the more distant Australia the remittances would only resemble gifts and in consequence be of smaller amounts. This claim is supported by his empirical analysis; a similar approach will be taken in this thesis.
In general, the immigration policies of host countries such as the GCC member states and a few East Asian countries, e.g. South Korea, Taiwan, Singapore (low-rights countries)
resemble guest-worker programmes (Djaji´c and Michael, 2012) and they explicitly aim at
keeping migration low-skilled and temporary. In this case, next to Glytsos’ predictions, one could predict relatively higher amounts remitted (due to the shorter time since arrival and distance, as in GCC majority of migrants are Asian).
In the member countries of OECD (high-rights countries), on the other hand, most mi-grants do have the right to family reunion and can obtain long-term residence (Huddleston et al., 2015). This makes it easier to settle down in the host country and increases the probability that the migration will be permanent. Once again, both Glytsos’ hypothesis and the rest of the literature suggest smaller amounts remitted. Moreover, contrary to a temporary migrant, a permanent migrant might not be considered a member of any household in the home country and, as suggested in Bertoli and Marchetta (2014), the effectiveness of remittances in terms of poverty reduction could be lower.
3.
Methodology
Ideally, the research question would be answered by empirical analysis using survey data that would allow to identify sources of remittance income, precisely estimate the amounts remitted and control for household and migrant characteristics. Unfortunately, none of the available datasets contains all the necessary information or, as in the case of the Migration and Remittances Household Surveys conducted for the Africa Migration Project, includes enough migrants to the low-rights countries to enable inference. Therefore, this thesis exploits the cross-country variation in shares of migrants to the two destinations of interest and uses only the official remittance flows.
The official remittance flows, defined by the IMF Balance of Payments as personal trans-fers and workers’ compensation (seasonal or other short term, where the worker is not a resident in the country of employment), are a gross underestimation of the real remittance flows. Many of the transactions occur in the informal financial sector that exists paral-lel to the official banking sector (e.g. hawala) or through family members and friends. Therefore, it has to be noted that if the official flows do not represent the real flows well enough, the estimates could be significantly biased. Still, this data is widely used in empirical studies as there are no better alternatives.
3.1
Adams and Page (2005) methodology
Due to the closeness of research questions, the empirical strategy in this study is based on that of Adams and Page (2005), henceforth referred to as AP05. The validity of their approach has been asserted by many studies since their publication, as it is one of the most often cited articles within the remittance impact literature. Moreover, it has also been the basis of the estimation in Gupta et al. (2009) which has also gained prominence itself.
The paper adapts Ravallion and Chen’s (1997) growth-poverty model and measures the impact of remittances by estimating the regression:
log Pit= αi+ β1log µit+ β2log ρit+ β3log xit+ υit (3.1)
Where P is the measure of poverty, µ - mean per capita income, ρ - inequality (Gini coefficient), x - variable of interest (migration or remittances) in country i in year t. The migration variable is defined as the share of migrants in the country population while remittances are introduced to the model in per capita terms (in constant 1995 US dollars), thus controlling for the size of the migrant-sending and remittance-receiving country. The inclusion of the inequality measure in the model is supported by two main argu-ments in Ravallion (1997). First, inequality matters through its impact on growth – more unequal distribution of income may hamper growth of the economy, therefore hindering poverty reduction. Second, higher inequality might be reflected in a lower share of the
poor in that growth in absolute terms, even if all income levels grow at the same rate, and thus lower poverty reduction rate.
As the data in AP05 comes straight from household surveys, the authors estimate sev-eral variations of the model using two different measures of income – one from national accounts and one calculated from the survey data. Additionally, they extend the model by including five regional dummies (Middle East and North Africa, Latin America and the Caribbean, East Asia, South Asia, Europe and Central Asia). These estimations are carried out for three measures of poverty. Poverty headcount ratio is the basic (α = 0) measure from the FGT class (Foster et al., 1984) described by equation (3.2) below. It simply reflects the share of population that is poor, i.e. H – the number of households
whose income (yi) falls below the poverty line z ($1.00 per day in AP05) divided by the
total number of households N , thus is a measure of the level of poverty. Poverty gap is also an FGT measure, with α = 1. By taking into account how far from the poverty line the poor households are (z − yi), it represents the depth of poverty. Finally, with α = 2,
squared poverty gap is a measure of the severity of poverty, as it is more sensitive to the situation of the poorest households than poverty gap.
F GTα= 1 N H X (i=1) z − yi z α (3.2)
In their basic specification, AP05 use Ordinary Least Squares but later, to overcome the possible endogeneity problems, they turn to Two-Stage Least Squares. First, they test three instruments for migration and remittances (separately and all together): distance from the remittance-sending area (main region for each country), percent of population over 25 years old that has secondary education and government stability (as measured by ratings published by the PRS Group in the International Country Risk Guide). Basing on the results of the first stage regressions, they decide to continue their analysis using all three instruments together, as these regressions perform best at explaining the variation in migration. The first stage IV estimates include also the exogenous variables: a measure of income, the Gini coefficient and the regional dummies.
3.2
Changes to AP05
The nature of the research question itself requires that some vital changes to the AP05 methodology be made. First, regarding the independent variables and in consequence, the instrumentation strategy. Moreover, owing to the improved availability of data, the sample size and composition will be changed.
3.2.1
Variable of interest
In order to allow for a differential impact of the remittances from the high- and low-rights countries, the total remittances are disaggregated using bilateral remittance matrices, fur-ther discussed in section 3.3.2. In the baseline specification, instead of official remittances per capita as in AP05, there are two variables of interest: the sum of official remittances
per capita flowing from high-rights countries and the sum of official remittances per capita flowing from low-rights countries.
3.2.2
High- and low-rights countries
Facing the lack of one comprehensive migrant rights index that would cover all the main migration destinations, the composition of the high- and low-rights groups of countries is determined based on the available measures, relevant literature and legal obligations of the migrant-receiving countries.
The low-rights group of countries, as suggested by Ruhs and Martin (2008), consists of GCC member states: Bahrain, Kuwait, Oman, Kingdom of Saudi Arabia, Qatar and United Arab Emirates. The sponsorship visa system in these countries ties migrants to specific employers who are then responsible legally for the workers. Importantly, it also implies that the termination of an employment contract equates the cancellation of the visa. This system has been primarily put in place to assure the temporality and rotation of migration and protection of the interests of the native population (Baldwin-Edwards, 2011).
Turning to the high-rights countries, Ruhs and Martin (2008) suggest Sweden and Canada as granting relatively many rights to the migrants. The exact composition of the high-rights group of countries is supported by their MIPEX 2015 score. MIPEX is a multi-dimensional index measuring migrants’ opportunities to participate in society across a wide range of policies: labour market mobility, education, political participation, access to naturalization, family reunion, health, permanent residence and anti-discrimination, developed by the Barcelona Centre for International Affairs and the Migration Policy Group. It covers 38 countries in Europe and North America as well as Australia, New Zealand and South Korea. In the basic specification, only the countries with the highest scores, categorized as Slightly Favourable (Sweden, Portugal, New Zealand, Canada, Fin-land, the Netherlands, Belgium, Australia, Norway, Spain, United States of America and Germany), are included in the high-rights group.
This division is supported by the countries’ own declarations towards migrants. Among the countries under consideration only those that scored high in the MIPEX 2015 ranking ratified ILO’s two conventions that specifically address migrant issues: 1949 Migration of
Employment Convention1 and 1975 Migrant Worker Convention2. The 1966 International
Covenant on Civil and Political Rights, which introduces a ban on slavery and forced labour, has been ratified by all the high-rights countries. At the same time, among the low-rights countries, the major migration destinations such as Saudi Arabia and United Arab Emirates, but also Oman and Qatar, did not sign the Covenant.
1Portugal, the Netherlands, Belgium, Norway, Spain, New Zealand 2Sweden, Portugal, Norway
3.2.3
Measures of poverty
The FGT poverty measures are calculated using household-level data and household sur-veys indeed are the main source of data for AP05. However, the datasets constructed for the replication of AP05 and to test the hypothesis of this study are compiled using readily available data on the country level. The source of data on poverty – World Bank’s World Development Indicators – does contain the measurements of the poverty headcount and poverty gap indices, but there is no information on the squared poverty gap as defined in section 3.1.
A simple way to approximate the measure of the severity of poverty would be to square the FGT poverty gap index, as it would put more emphasis on the poorest countries by inflating large values of the measure more than the smaller ones. In order to check how the FGT squared poverty gap and the square of FGT poverty gap are related, the latter measure is calculated using the AP05 data and compared to the first one, already included in the dataset.
The modified measure will surely operate on a different scale. While, by construction, all the FGT measures are expressed as shares and in the AP05 dataset are transformed to take values between 0 and 100, the maximal possible value of the square of the FGT poverty gap would be 10000. Moreover, some of the diversity of the FGT squared poverty gap will be lost – the modified measure cannot take on different values for the same value of poverty gap. Yet, the two measures of the severity of poverty are closely related, with the coefficient of their correlation amounting to 0.96.
The models in AP05, however, do not use the poverty measures themselves, but their natural logarithms. Graphs A and B of Figure 3.1 plot the quintiles of the logs of the two measures against quintiles of the normal distribution. Taking logs appears to be successful at normalizing the distributions, with the deviations from the normal distribution very similar for both the measures and concentrated on the tails. It can be therefore inferred from Graphs A and B that the distributions of the logs of the two measures are very similar in shape.
Graph C of Figure 3.1 plots the values of the two measures corresponding to the given values of the log FGT poverty gap, thus keeping the order of the observations the same for both measures. The observations of the two measures create lines of slightly differ-ent slopes that cross around the cdiffer-entre of the graph. The differences between the two variables are the biggest in the tails, with the square of the FGT poverty gap being an underestimation of the FGT squared poverty gap for low values of the poverty gap and overestimation for its large values – a consequence of the different scale of the modified measure.
Overall, the square of the FGT poverty gap seems to be a good approximation of the FGT measure of the severity of poverty. The close association between them allows to conclude that the modified measure can be used as a substitute of the FGT squared poverty gap.
Figure 3.1: FGT and modified measures of the severity of poverty
3.2.4
The model
In the baseline specification, a simple adjustment to equation (3.1) is made, where the
AP05 remittance variable is substituted by the two new variables, x(h,it) (remittances
flowing from the high-rights countries) and x(l,it) (remittances flowing from the low-rights
countries). The basic predictions about β1 and β2 remain the same as in AP05, higher
mean income is supposed to lower poverty and higher inequality increase it. Both β3
and β4 are the main coefficients of interest and the relationship between them, whether
they are significantly different from each other, is particularly important for the research
question. According to the hypothesis outlined above, β4 should be more negative than
β3.
log Pit = αi+ β1log µit+ β2log ρit+ β3log x(h,it)+ β4log x(l,it)+ υit (3.3)
This baseline specification does not, however take into consideration the remittance flows from other countries than those included in the high- and low-rights groups. Ignoring them would likely result in omitted variables bias. Two solutions are suggested to account for them.
In the first one, regression (3.3) is extended by an additional control – residual remit-tances i.e. the remitremit-tances from high- and low-rights countries subtracted from the total remittance inflows. This solution, formalized in equation (3.4) is straightforward in OLS, yet it poses a big challenge to the instrumentation strategy, as the additional variable
encompasses transfers from many different countries that are likely determined by many factors. Moreover, it might be the case that the remittance flows from one of the consid-ered sources are systematically higher and the effect picked up by the coefficient is the consequence of larger sums at the disposal of remittance recipients and not the way the immigration policies influence the investment of this money.
log Pit = αi+ β1log µit+ β2log ρit+ β3log x(h,it)+ β4log x(l,it)+ β5log x(r,it)+ υit (3.4)
The second solution is an extension of (3.1) by interaction terms of total remittance flows
with the shares of remittances coming from the high and low rights countries that, if β4
or β5 turn out statistically different from zero, would capture the differential impact of
high and low rights on poverty:
log Pit= αi+ β1log µit+ β2log ρit+ β3(1 + β4sh(h,it)+ β5sh(l,it)) log xit+ υit (3.5)
The hypothesis would be supported if β3 is smaller than zero, indicating the
poverty-reducing impact of total remittance flows, and β5 is larger than both zero and β4,
indi-cating that the higher share of remittances is coming from the low rights countries, the stronger the poverty-reducing impact of remittances.
As the level of total remittances per capita is explicitly included in the model, the issue of transfers from one destination being systematically higher is no longer a problem. To strengthen this argument, the correlation coefficients between the two share variables and the total per capita remittances were calculated. With the values of 0.071 for the share of remittances coming from low-rights countries and 0.153 for the share of remittances coming from high-rights countries, they suggest only a weak dependence between these
two pairs of variables.3 It cannot be ruled out, however, that there are other systematic
differences between the countries that receive a large portion of their remittance flows from either of the sources that might also influence the poverty rates.
Endogeneity
The various new variables introduced in the AP05 model are hardly exogenous. There-fore, AP05’s instrumentation strategy has to be expanded to account for the different migration destinations. First, as in the original paper, the total remittance flows remain instrumented for by the distance to the main source of remittances, percentage of adults with secondary education and government stability in the migrant-sending countries, al-though the data sources need to be changed due to data availability issues. Next, for the two new variables (regardless of whether they are used as levels or shares), new in-struments are developed. The main instrument for each will be the country’s distance to, accordingly, the nearest high-rights or low-rights country. Additionally, following the McKenzie and Sasin’s (2007) overview of instruments used in the migration literature, historical and cultural factors are included. For high-rights countries this would be a past colonial relationship, for low-rights – the share of Muslim population (Islam is the state
3Interestingly, in the dataset, higher total remittances are associated with lower share of the flows
from countries that are classified as neither high- nor low-rights (correlation equals -0.185). This could be due to the fact that the high- and low-rights groups consist of many of the biggest remitters.
religion in all the GCC countries; in Saudi Arabia where the most migrants are, it is the only religion that can be publicly worshipped). For both groups of countries a common language spoken by at least a fraction of the populations could as well facilitate migration. With multiple endogenous variables and more than one instrument suggested for each of them, weak instruments become a serious problem that requires special attention. Weak instruments can lead to a bias of IV results towards OLS. As presented in Stock and Yogo (2005), in the case of multiple endogenous regressors, the rule of thumb of the F statistic on excluded instruments larger than 10 is not sufficient to assert their strength. Therefore, the choice of the instrumentation strategy for each of the models is guided by the value of Cragg and Donald (1993) F statistic and Stock and Yogo’s (2005) critical values for it, in order to minimize the bias of the IV estimator. Additionally, the process of selecting and testing the instruments is carried out according to Angrist and Pischke’s (2008) advice. First stage regressions are reported to reflect on the sign and magnitudes of the coefficients on the instruments and how they correspond to the theory motivating their use. Furthermore, the appendix contains LIML and just-identified estimates that are less prone to the weak-instruments bias.
Moreover, as the new dataset is composed of a larger number of observations than AP05’s, the panel dimension of the dataset can be exploited. Following the results of the Hausman specification test, either random or fixed effects regressions are estimated, allowing to control for time-invariant country characteristics such as cultural and historical ties and thus giving a better image of how the remittances that do change over time affect poverty.
3.3
Data
AP05 collected data directly from available household surveys and this effort lead to a sample of 71 countries with 184 observations over the period from 1980 to 1999. The dataset of this thesis bases on widely available databases of institutions such as the World Bank and UN Population Division and is substantially larger than that of AP05. Increased data availability, especially in the period since the publication of AP05, allows to study an extended sample and results in more observations over a similar period of time.
3.3.1
Sample
From the initial sample of 135 countries currently classified by the World Bank as low-and middle-income, 37 had to be dropped due to the lack of data on the dependent variable (24) or only one observation of it over the studied period (13). Nevertheless, despite constituting over one fourth of the initial group of countries, they are the origin of only around 10 percent of migrants. Despite this procedure, data scarcity remains a major problem, in particular with regard to the poverty and inequality measures that are reported irregularly, especially in the first decade of the studied period (1990-2000).
3.3.2
Bilateral remittance matrices
The empirical strategy of this study relies on disaggregation of remittance inflows by the remittance-sending countries. Typically, the sources of remittance inflows are not accu-rately recorded in the national accounts and could well be confusing, as money coming through an international bank from different regions could be attributed to the country where the bank has its headquarters (Maimbo and Ratha, 2005). To provide at least an idea of the bilateral flows, the World Bank has been publishing the Bilateral Remittance Matrices (BRMs) since 2010. These are estimates, generated according to the method-ology developed by Ratha and Shaw (2007), by weighing the total remittance inflows by the stock of immigrants from the remittance-receiving country in the remittance-sending country and the incomes of the two countries. As the BRMs made available by the World Bank cover only recent years, Ratha and Shaw’s (2007) methodology has been applied to construct new BRMs for a longer period of time. The details of the procedure can be found in the section A.1.3. of the Appendix.
The data on migrant stocks by the country of origin and destination is based on the
matrices published by UN Population Division for 1990, 2000, 2010 and 20134. Given
the irregular measurements of other variables, four data points would be insufficient to carry out analysis and additional data had to be generated. Assuming that the stocks of migrants are not subject to violent fluctuations, the data was interpolated to obtain annual observations from 1990 to 2012. Next, using well available annual data on income and aggregate remittance flows from the World Bank’s World Development Indicators database, the remittance flows have been generated on a yearly basis.
3.3.3
Other variables and their data sources
The data on the dependent and explanatory variables: poverty headcount ratio, poverty gap, GDP per capita and Gini coefficient all come from the World Development Indicators. These are all exact equivalents of AP05’s variables with the slight modification of the poverty variables that are currently calculated by the World Bank at the $1.25 per day poverty line. Instead of the FGT squared poverty gap, the square of FGT poverty gap is used. For the IV regressions, data on distance, colonial relationships and common language were obtained from the CEPII database, percentage of Muslims in the society from the CIA World Factbook and data on education attainment by age group from the Baro-Lee dataset available through the World Bank’s Education Statistics. AP05’s PRS Group’s Country Risk Guide government stability index is substituted by World Bank’s Political Stability and Absence of Violence index from the World Governance Indicators.
4Estimates. Other data on migrant stocks by the country of origin and destination has been published
by the OECD and the World Bank and while it is certainly of better quality (OECD’s calculations are based on primary data from national censuses) it covers fewer countries and is only available for the last few years (OECD:2010, WB: 2010 and 2013)
3.3.4
Robustness checks
To test the robustness of the results, the effect of several to the dataset will be estimated. First, the composition of high- and low-rights groups of countries will be changed. The high-rights countries will include also these that fall into the Halfway Favourable category (Italy, Denmark, Luxembourg, United Kingdom, France, the Republic of Korea, Austria, Ireland, Switzerland, Estonia, Hungary, Iceland, Czech Republic, Romania, Slovenia, Bul-garia, Poland, Greece, Japan, Croatia) in the MIPEX ranking, while low rights countries group will be extended to cover Jordan and Lebanon that also use the sponsorship system and the destinations in East Asia that, similarly to GCC, admit mostly unskilled, tem-porary migrants: Hong Kong, the Republic of Korea, Singapore, Thailand and Malaysia (Djaji´c, 2014)5.
The Republic of Korea appears to be on the border of high and low rights, with some policies drawing it to the bottom end but the overall quality of immigration policies placing it among relatively high-rights countries. As the expansion of the high- and low-rights groups, if the hypothesis is true, should lead to the conversion of the differential impact to the mean (or towards zero in the case of the interaction terms in the second model), Korea will be assumed to be a low-rights country.
In order to check how the data generation influences the results, the estimation will be limited to the four data points for which there are available migrant stocks. With such a reduced sample, the significance might disappear completely, yet the magnitude of the coefficients remaining roughly the same would lend support to the validity of the results.
4.
AP05 Replication
The analysis in AP05 is based on an original dataset, compiled mostly from household survey data, that covers 71 low-income and middle-income countries and years 1980 to 1999. Other cited sources include US Census and OECD for migration, IMF Balance of Payments for remittances and PRS Group’s Country Risk Guide for the government stability instrument.
A part of the dataset is included in the appendix of the paper, but it does not cover any of the income measures or the instruments. Therefore, the replication of the study is done in two steps. First, the AP05 dataset is reconstructed as closely as possible using available measures for the missing variables. Next, the same methodology is applied to an extended dataset covering a longer time period and more countries.
The exact definitions of variables and data sources in the three datasets, as well as the time periods and numbers of countries covered are presented in Table A.1 in section A.1.3 of the Appendix. The primary focus of this replication will be on the effect of international remittances on poverty.
4.1
Reconstructed AP05 dataset
The estimation in this section is performed using the dataset included in the appendix of AP05, with the lacking data substituted with variables best matching the description in the paper.
For the OLS estimates, only the income variable used by AP05 is not available. In the paper, two versions of this variable are used. The first, survey mean income, could not be reconstructed due to data availability issues. The second, the national accounts data on GDP per capita has been taken from the World Bank’s World Development Indicators
and performs best when expressed in constant 1995 dollar terms1. Table 4.1 presents the
OLS results estimated using the reconstructed dataset and the new dataset juxtaposed to their equivalents published in AP05.
As shown in columns (2), (5) and (8) of Table 4.1, the reconstructed dataset has a few more observations and some of the estimates are more precise. In general, however, the coefficients are very close to each other and there are no significant differences between them. This is the case for all the models, with the exception of the regression of poverty headcount ratio on per capita official international remittances where the coefficient on re-mittances is marginally significant in AP05 but falls just below the 10 percent significance level in the reconstructed dataset.
1In the text, AP05 write that the per capita GDP income variable is in purchasing power parity
terms (PPP), while the tables describe the income variable as “per capita GDP (constant 1995 dollars)”. Indeed, constant 1995 dollars provide estimates much closer to those in the paper than does GDP per capita PPP in constant 1995 dollars.
The replication of the IV estimates poses a greater challenge. The first instrument is de-scribed in AP05 as “distance (miles) between the remittance-sending area (United States,
OECD (Europe)2, or the Persian Gulf) and the remittance-receiving country”. For Latin
American countries the paper assumes the remittance-sending area to be the US, for North Africa – Europe and for South Asia – the Persian Gulf. As neither the exact source of the distance data nor the method of its calculation is specified, the data used to reconstruct this variable is taken from the CEPII database. The measure used is the distance between the most populous city of the remittance-receiving country and the nearest most populous city of a country in the relevant remittance-sending area. Which area is relevant for each country in the sample is determined using the above mentioned rule, and for countries that do not fall into any of these categories, it is individually decided basing on World Bank’s Bilateral Remittance Matrix for 2012. Section A.1.2 of the Appendix contains information on what area of reference has been used for each country in the sample. The second instrument, “the percent of the population over age 25 that have completed secondary education” is reconstructed using the Barro-Lee’s data defined exactly this way. The measurements are available from 1980 to 2010 in 5 year intervals and have been linearly interpolated to cover all the years.
Finally, even though AP05 exactly describe their measure of government stability, the PRS Group data is not available for the purposes of this thesis and an alternative measure has to be employed. The World Governance Indicators, produced by researchers from the World Bank and the Brookings Institute provide a measure of political stability and absence of violence or terrorism that is closely related to PRS Group’s government stability. However, this measure is only available from 1996 on, which poses a serious problem, as it severely affects the number of complete observations, driving it down to only 22. Using the 1996 values for all earlier observations would be indefensible as they reach as far as 1980. Nevertheless, it could be argued that the 1996 values are a good proxy for the situation in the first half of the 1990’s. This is the approach taken here – the 1996 values of the index are copied to the observations from years 1990-1995 and the estimations using the stability instrument are run only on the data from 1990 on.
Table 4.2 presents the first stage IV regressions for international remittances. The results of the replication show that the performance of the instrumentation strategy is dependent on the exact definitions and data sources of the instruments. The coefficient on the reconstructed distance is much smaller and less precise that in the original study, possibly due to the lack of clarity in AP05 about the reference area for a number of countries in the sample, and the F statistic is over six times smaller. The coefficient on education, where the measure used has exactly the same definition but the data source could be different (as it is not specified in AP05), is much smaller and the F statistic is near zero. The last instrument, government stability, severely reduces the number of observations. Here, the point estimate is larger than AP05’s and significant but of the opposite sign.
Inclusion of all three instruments, the strategy followed by AP05 in further estimations, improves slightly the fit but not the precision of the first stage estimates. In fact, the F statistic in the last replicated model, as those in all other first stage regressions, is well
2Defined as a group of 21 countries: Austria, Belgium, Czech Republic, Denmark, Finland, France,
Germany, Greece, Hungary, Ireland, Italy, Luxemburg, Netherlands, Norway, Poland, Portugal, Slovak Republic, Spain, Sweden, Switzerland, and United Kingdom
below 10, suggesting that the instruments are very weak, what is confirmed by the value of Cragg-Donald Wald F statistic that indicates that the maximal IV bias is larger than 10 percent.
The results of the second stage IV estimation in columns (2), (6) and (10) of Table 4.3 sharply contrast with the results of AP05. None of the coefficients on the instrumented variables is significant and the point estimates in the poverty headcount ratio and squared poverty gap regressions are closer to zero than in OLS. Unlike the OLS results, in the IV estimation AP05 do not present or comment their results for the models that use the national accounts measure of income, undermining the basis for comparison and making it difficult to determine how much of the differences should be attributed to the discrepancies in the instrumentation strategies and to the weakness of instruments. However, comparing the results to the just-identified (using distance instrument only) and LIML estimations (Table A.4 in section A.2 of the Appendix), it appears that the latter is less of a culprit. Indeed, the coefficients estimated using these two methods are larger than in Table 4.3 but still far from those presented in AP05.
4.2
The new dataset
The dataset compiled for the purposes of this thesis benefits from improved data availabil-ity and covers a longer period of time – from 1980 to 2013, as well as more countries. As shown in Table A.1 in the Appendix, this is also true for the IV estimation, which, by the construction of the government stability instrument is limited to observations from 1990 on (22 years and 104 countries in the new dataset compared with at most 19 years and 71 countries in AP05). The composition of the sample has been determined according to the current classification by the World Bank; out of 135 low- and middle-income economies, the dataset includes complete observations of 106.
Similarly to AP05, the remittance data comes from the IMF Balance of Payments and the data on poverty and inequality available through the World Development Indicators is based on primary data coming from household surveys. Instead of the FGT squared poverty gap, the square of the FGT poverty gap is used. Instead of constant 1995 dollars, the year 2005 is used as the point of reference, as it fits better the studied period. The instruments are constructed as described in the previous section.
Performing the same analysis as in AP05 but with this larger dataset provides some support to their argument. The larger number of observations (up to 699 compared with maximum 111 in AP05) results in improved fit and precision of the estimates. In the regressions of poverty headcount and squared poverty gap the coefficients on the variable of interest are almost equal to those presented by AP05, while the impact of remittances on poverty gap is much smaller in the new dataset – columns (3), (6) and (9) of Table 4.1.
Table 4.1: OLS estimates of the effect of international remittances on poverty: AP05, the reconstructed dataset and the new dataset
(1) (2) (3) (4) (5) (6) (7) (8) (9) AP05 Reconstructed New AP05 Reconstructed New AP05 Reconstructed New
poverty headcount ratio poverty gap squared poverty gap Per capita official international
remittances -0.077* -0.081 -0.078*** -0.208*** -0.143*** -0.076*** -0.164** -0.158** -0.163*** (0.045) (0.053) (0.017) (0.081) (0.051) (0.015) (0.081) (0.076) (0.033) Per capita GDP -0.852*** -0.830*** -0.840*** -0.961*** -0.976*** -0.665*** -0.929*** -0.968*** -1.347*** (0.138) (0.133) (0.035) (0.182) (0.174) (0.030) (0.220) (0.199) (0.066) Gini coefficient 1.882*** 1.813*** 2.139*** 3.184*** 2.939*** 1.666*** 3.271*** 3.304*** 3.236*** (0.481) (0.500) (0.166) (0.637) (0.658) (0.148) (0.737) (0.698) (0.318) East Asia dummy 0.065 -0.054 -0.154 -0.306 -0.432 -0.523*** -0.991* -1.073*** -1.210***
(0.342) (0.193) (0.118) (0.450) (0.277) (0.105) (0.508) (0.333) (0.224) Europe and Central Asia dummy -1.928*** -1.982*** -0.755*** -2.198*** -2.238*** -0.740*** -1.826*** -1.816*** -1.582***
(0.364) (0.403) (0.114) (0.483) (0.567) (0.097) (0.553) (0.658) (0.209) Latin America and the Caribbean
dummy
-0.147 -0.156 -0.415*** -0.128 -0.203 -0.409*** -0.314 -0.339 -1.025*** (0.313) (0.240) (0.095) (0.413) (0.332) (0.092) (0.483) (0.411) (0.197) Middle East and North Africa dummy -2.099*** -2.207*** -1.045*** -1.748*** -2.233*** -1.186*** -2.101*** -2.199*** -2.649**
(0.337) (0.363) (0.130) (0.446) (0.380) (0.096) (0.569) (0.536) (0.198) South Asia dummy 0.077 -0.014 0.280** 0.165 -0.093 -0.201 -0.384 -0.496 -0.563* (0.296) (0.222) (0.125) (0.393) (0.279) (0.133) (0.492) (0.388) (0.295) Constant 10.575*** 10.474*** 1.210* 11.437*** 11.377*** 0.954 10.567*** 10.925*** 2.256*
(1.002) (1.032) (0.661) (1.328) (1.387) (0.597) (1.623) (1.534) (1.271)
Observations 99 102 699 99 102 699 89 91 699
Adjusted R-squared 0.744 0.741 0.793 0.679 0.682 0.762 0.606 0.608 0.738
Standard errors presented in parentheses, robust for the replication regressions, recalculated from t-ratios for the original AP05 estimates. All variables in logs. In column (9) the dependent variable is log square of the FGT poverty gap. * p < 0.1, ** p < 0.05, *** p < 0.01
Table 4.2: First stage IV results of the effect of remittances on poverty: AP05, the reconstructed dataset and the new dataset
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) AP05 Recons. New AP05 Recons. New AP05 Recons. New AP05 Recons. New Distance -1.565*** -1.145** -0.187** -1.929*** -0.952** -0.430*** (0.251) (0.461) (0.095) (0.309) (0.434) (0.126) Education -0.232 -0.0737 0.407*** -0.694* -0.185 0.538*** (0.400) (0.310) (0.107) (0.356) (0.397) (0.126) Government stability 0.611 -1.227* 0.130 0.328 -1.120* 0.237 (0.546) (0.617) (0.277) (0.475) (0.620) (0.314) Gini coefficient 1.272 0.932 1.644*** -0.593 -0.371 1.160** -0.552 0.170 1.230*** 0.264 0.955 1.510*** (0.949) (0.982) (0.394) (1.289) (1.129) (0.514) (1.284) (1.159) (0.383) (1.148) (1.067) (0.537) Per capita GDP 0.505 0.0781 0.139* 0.773 0.492 0.125 0.141 0.451 0.219*** 1.358*** 0.316 -0.0996 (0.262) (0.324) (0.077) (0.444) (0.341) (0.080) (0.371) (0.277) (0.075) (0.417) (0.299) (0.104) East Asia dummy 2.149*** 1.085 0.627*** 0.660 0.575 0.352 0.587 -0.196 0.603*** 3.084*** 0.695 0.596** (0.724) (0.783) (0.206) (1.015) (0.600) (0.235) (0.889) (0.595) (0.207) (0.961) (0.618) (0.254) Europe and Central Asia
dummy
-2.000*** -1.607* 1.862*** -0.392 -0.242 1.227*** 0.205 -1.343** 2.036*** -3.164* -1.893** 0.653 (0.725) (0.848) (0.240) (1.188) (0.933) (0.371) (1.206) (0.633) (0.234) (1.212) (0.886) (0.418) Latin America and the
Caribbean dummy
0.084 1.579** 1.182*** 1.528 1.638** 0.987*** 1.687* 0.960 1.342*** -0.303 1.311* 1.215*** (0.646) (0.752) (0.190) (0.849) (0.650) (0.227) (0.823) (0.676) (0.189) (0.842) (0.668) (0.226) Middle East and North
Africa dummy
-0.864 1.277 2.012*** 2.504* 2.715*** 2.115*** 2.813*** 1.901* 2.294*** -1.742 1.208 1.700*** (0.823) (1.240) (0.289) (0.998) (0.868) (0.301) (0.839) (1.088) (0.260) (1.131) (1.372) (0.336) South Asia dummy 1.378* 0.947 1.300*** 1.957* 1.720** 1.031*** 1.852* 0.899 1.336*** 2.054* 0.753 0.864***
(0.560) (0.807) (0.217) (0.851) (0.654) (0.272) (0.802) (0.607) (0.249) (0.790) (0.746) (0.329) Constant 10.914*** 10.62** -3.314** -4.645 -2.814 -3.812** -0.791 -0.260 -4.057*** 7.598* 9.285* -0.583 (2.231) (4.860) (1.579) (3.271) (2.084) (1.852) (2.825) (1.884) (1.427) (3.501) (4.672) (2.050) Observations 101 98 667 91 98 552 91 72 653 84 66 496 Adjusted R-squared 0.463 0.2608 0.2443 0.193 0.2211 0.2229 0.177 0.2829 0.2703 0.436 0.3027 0.2441 F statistic 39.42 6.16 3.87 4.92 0.06 14.51 4.23 3.96 0.22 8.90 2.09 7.04
Dependent variable: per capita official international remittances. Standard errors presented in parentheses,robust for the replication regressions, recalculated from t-ratios for the original AP05 estimates. All variables in logs. * p < 0.1, ** p < 0.05, *** p < 0.01
Table 4.3: Second stage IV and FE results of the effect of remittances on poverty: AP05, the reconstructed dataset and the new dataset
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) AP05 Recons. New FE New AP05 Recons. New FE New AP05 Recons. New FE New
poverty headcount ratio poverty gap squared poverty gap Per capita official
international remittances
-0.351*** -0.105 -0.306*** -0.045 -0.396*** -0.219 -0.102 -0.025 -0.283* 0.060 -0.113 -0.051 (0.099) (0.194) (0.099) (0.033) (0.136) (0.259) (0.077) (0.036) (0.126) (0.319) (0.175) (0.087) Per capita GDP -0.904*** -0.754*** -1.405*** -1.000*** -0.611*** -1.210*** -1.120*** -1.251*** -2.529***
(0.145) (0.045) (0.146) (0.173) (0.034) (0.140) (0.253) (0.074) (0.334) Survey mean income -1.590*** -1.986*** -2.072
(0.174) (0.240) (0.249)
Gini coefficient 2.950*** 1.759*** 2.095*** 2.026*** 4.407*** 3.026*** 1.502*** 1.769*** 4.700*** 3.340*** 2.849*** 3.436*** (0.447) (0.602) (0.273) (0.315) (0.616) (0.779) (0.216) (0.308) (0.648) (0.946) (0.467) (0.717) East Asia dummy -0.230 -0.180 -0.179 -0.688 -0.542 -0.623*** -1.373*** -1.284*** -1.460***
(0.303) (0.240) (0.157) (0.417) (0.356) (0.126) (0.422) (0.466) (0.271) Europe and Central Asia
dummy
-1.608*** -2.225*** -0.500** -1.790*** -2.683*** -0.890*** -0.929 -1.805** -2.107*** (0.401) (0.549) (0.229) (0.554) (0.803) (0.164) (0.615) (0.719) (0.363) Latin America and the
Caribbean dummy
-0.021 -0.131 -0.160 -0.038 -0.140 -0.433*** -0.303 -0.599 -1.226*** (0.300) (0.394) (0.174) (0.422) (0.564) (0.144) (0.446) (0.535) (0.317) Middle East and North
Africa dummy
-0.614 -2.316*** -0.644** -0.533 -2.241*** -1.259*** -1.026 -3.082*** -3.063*** (0.495) (0.597) (0.282) (0.683) (0.726) (0.210) (0.666) (1.087) (0.459) South Asia dummy 0.443 -0.162 0.512*** 0.363 -0.119 -0.239 -0.145 -0.943 -0.777* (0.346) (0.363) (0.194) (0.478) (0.480) (0.182) (0.518) (0.817) (0.416)
Constant 17.27*** 11.04*** 1.38 5.33*** 20.21*** 11.77*** 1.32 3.99*** 20.26*** 11.71*** 3.07* 8.87*** (1.394) (1.189) (1.022) (1.506) (1.923) (1.534) (0.825) (1.407) (2.012) (1.766) (1.765) (3.201) Observations 81 66 496 699 81 66 496 699 75 59 496 699 Adjusted R-squared 0.811 0.657 0.722 0.756 0.602 0.749 0.744 0.513 0.720
Standard errors presented in parentheses, robust for the replication regressions, recalculated from t-ratios for the original AP05 estimates. All variables in logs. In columns (11) and (12) the dependent variable is log square of the FGT poverty gap. * p < 0.1, ** p < 0.05, *** p < 0.01
Although the results of the first stage IV estimation in columns (6) and (12) of Table 4.2 show some improvement over the replication attempt with a reconstructed AP05 dataset, both the distance and stability instruments perform much worse in the new dataset. The effect of distance on the amounts remitted is much smaller both in the just-identified and overidentified models and the F statistic on the excluded stability is close to zero. Still, the education instrument is much stronger in the extended dataset and more consistent with the general theoretical prediction in AP05 of a positive relationship between education of the population and migration. The F statistic on the excluded instruments in the overidentified model (column (12)) is also larger than the corresponding value in column (11).
Columns (3), (7) and (11) of Table 4.3 present the results of the IV estimation of the effect of international remittances on poverty using AP05’s instrumentation strategy (all three instruments), however it still remains that AP05’s second stage IV results are not available for the appropriate measure of income.
As in the reconstructed dataset, in most cases instrumentation produces smaller coef-ficients on the endogenous variable than in AP05, yet the difference, especially in the poverty headcount and squared poverty gap regressions is smaller. Moreover, only one of the coefficients on remittances is as precisely estimated as in AP05. These discrep-ancies with the results of AP05 seem to stem in a large part from the weakness of the instruments as the coefficients of the just-identified (using education instrument only) are of much larger magnitudes and the results of LIML estimations also differ from TSLS (results in Table A.5 in section A.2. of the Appendix).
AP05 include in their models five regional dummies in an attempt to account for fixed effects, given their small, relative to the number of countries analyzed, number of observa-tions. Using the same approach in the extended dataset yields results partially consistent with AP05, often with smaller and less significant coefficients on the variables of inter-est. However, the number of the observations in the extended dataset allows also for estimating a panel regression. Following the results of the Hausman specification test, fixed (country) effects results are presented in columns (4), (8) and (12) of Table 4.3. Interestingly, even though the point estimates remain negative, they are much closer to zero and insignificant.
5.
Testing The Hypothesis
5.1
OLS
The basic OLS results provide little new information about the impact of immigration policies on poverty in migrant-sending countries. Yet, for both models estimated in Table 5.1 the coefficients on the income and inequality variables are of expected signs and magnitudes, consistent with AP05, and there is some indication of a negative effect of remittances on poverty.
Table 5.1: OLS estimates of the impact of remittances from high- and low-rights countries on poverty.
PANEL A: LEVEL MODEL
(1) (2) (3) (4) (5) (6) poverty headcount ratio poverty gap squared poverty gap Remittances from high rights
(per capita)
-0.046** -0.008 -0.068*** -0.038* -0.167*** -0.084* (0.020) (0.021) (0.018) (0.020) (0.040) (0.044) Remittances from low rights
(per capita)
0.009 0.011 -0.071*** 0.018 -0.165*** 0.061 (0.034) (0.040) (0.022) (0.029) (0.048) (0.060) Residual remittances (per
capita) -0.152*** -0.115*** -0.107*** -0.079*** -0.211*** -0.167*** (0.024) (0.024) (0.022) (0.021) (0.047) (0.044) GDP per capita -0.961*** -0.840*** -0.751*** -0.643*** -1.532*** -1.291*** (0.027) (0.036) (0.025) (0.031) (0.053) (0.067) Gini coefficient 2.580*** 2.349*** 2.224*** 1.780*** 4.294*** 3.442*** (0.121) (0.178) (0.108) (0.155) (0.238) (0.333) Constant 0.241 0.443 -0.794* 0.401 -1.018 1.182 (0.519) (0.700) (0.456) (0.629) (0.993) (1.340) Regional dummies No Yes No Yes No Yes Observations 642 642 642 642 642 642 Adjusted R-squared 0.773 0.807 0.745 0.785 0.715 0.763
PANEL B: SHARE MODEL
poverty headcount ratio poverty gap squared poverty gap Share high rights ∗ Total
remittances
0.086*** 0.071*** 0.052** 0.040* 0.085* 0.082 (0.025) (0.026) (0.023) (0.023) (0.050) (0.051) Share low rights ∗ Total
remittances 0.094 0.113* -0.038 0.089** -0.117 0.210*** (0.058) (0.060) (0.037) (0.038) (0.082) (0.078) Total remittances -0.196*** -0.124*** -0.166*** -0.107*** -0.345*** -0.227*** (0.026) (0.026) (0.023) (0.021) (0.051) (0.046) GDP per capita -0.979*** -0.845*** -0.768*** -0.650*** -1.570*** -1.307*** (0.026) (0.036) (0.024) (0.031) (0.052) (0.067) Gini coefficient 2.576*** 2.358*** 2.215*** 1.796*** 4.278*** 3.478*** (0.125) (0.179) (0.114) (0.157) (0.252) (0.337) Constant 0.402 0.447 -0.613 0.409 -0.629 1.192 (0.534) (0.704) (0.474) (0.639) (1.036) (1.365) Regional dummies No Yes No Yes No Yes Observations 642 642 642 642 642 642 Adjusted R-squared 0.769 0.804 0.739 0.783 0.708 0.760
Robust standard errors in parentheses. All variables in logs. In columns (5) and (6) the dependent variable is log square of the FGT poverty gap. * p < 0.1, ** p < 0.05, *** p < 0.01
coeffi-cients on the two main variables of interest usually only in the specifications without the regional dummies, as presented in Panel A of Table 5.1. Nevertheless, higher remittance flows (especially from countries other than the low-rights countries) remain associated with lower poverty rates and the effect persists for all three measures of poverty, the re-lationship being the strongest and most precisely estimated for the residual remittances. The point estimates on remittances from low-rights countries are in most cases positive and insignificant.
However, the estimation of the share model (equation (3.5)) casts doubts on the validity
of these claims. With negative coefficients on remittances per capita, the larger the
(positive) coefficient on the interaction terms, the more poverty-reducing is the effect of higher shares of remittances from high- or low-rights countries. When regional dummies are included in the model, the theoretical predictions expressed in section 3.2.4 are all supported. The coefficients on interaction terms associated with remittances from low-rights countries in panel B of Table 5.1 are significantly larger than on these associated with remittances from high-rights countries. Importantly, the magnitude of the coefficient on total remittances is consistent with AP05.
5.2
IV
Causal inference from the OLS results would rely on the very strong assumption that all the remittance variables are orthogonal to poverty. Moreover, possibly due to endogeneity issues, the two models yield opposite results. In an attempt to overcome this problem, instrumental variables are employed, and the two models are estimated using Two-Stage Least Squares.
In the case of the level model, where only per capita official remittances from high- and low-rights countries are instrumented, weak instruments are not an issue. F statistics on the excluded instruments in both first stage regressions are well above 10 (columns (1)-(2) in Table 5.2) and the value of Cragg Donald Wald F statistic suggests that the maximal IV bias is below 5 percent.
The share model, however, requires more adjustments. As presented in columns (3)
and (5) of Table 5.2, the F statistic on the excluded instruments is rather low for the interaction term of share of remittances from low-rights countries and the total official remittances per capita. Moreover, for both the interaction terms the sign of the coefficient on one of the relevant instruments is opposite to the expected (common language with a low-rights country and colonial relationship with a high-rights country). Adjusting the instrumentation strategy by restricting the number of the instruments to the five best performing (Strategy 2, columns (6)-(8)) brings all the F statistics above 10, driving the maximal IV bias down below 10 percent.
For both considered models, even a brief look at panels A and B of Table 5.3 makes it clear that instrumentation significantly affects the results. Again, the coefficients on both the income and inequality variables are of the expected signs and magnitudes.
disag-gregated remittance flows remain mostly insignificant but all but one of the coefficients become positive. Higher remittance flows from the low-rights countries are associated with significantly higher levels of poverty, while those from the high-rights countries turn out to increase the severity of poverty. Yet, inference from the TSLS results of this model is limited by the lack of a good instrument for the residual remittances variable, whose endogeneity is likely to bias the estimates.
In the share model, presented in Panel B of Table 5.3, the signs and the magnitudes of the coefficients on the variables of interest are also affected by instrumentation and the theoretical predictions are no longer supported. While the coefficients on the interaction terms of total remittances with their share coming from the high-rights countries remain positive and are much larger than in OLS (thus pointing to a poverty reducing impact of these remittances across different measures of poverty), those on the low-rights interaction terms become negative in the models that include the regional dummies. On the other hand, in columns (2) and (4) of Panel B, even though the point estimate on the coefficient on the level of total remittances is negative, it is no longer significant, making it more difficult to infer about the effect of the shares themselves.
As weak instruments bias is a particular worry in the case of the share model, the es-timation is repeated using a just-identified model and Limited Information Maximum Likelihood method that, as suggested in Angrist and Pischke (2008), are less prone to that bias. The estimates (Tables A.6 and A.7 in section A.3.1 of the Appendix) are very similar to those presented in Panel B of Table 5.3, lending further support to the TSLS results.
5.3
Panel regressions
In the IV results, in particular for the share model, inclusion of the regional dummies substantially changes the estimates, suggesting that some fixed characteristic of the ge-ographical regions significantly affect the poverty reduction effectiveness of remittances. Prompted by this, the panel dimension of the dataset is exploited. The Hausman specifi-cation test rejects fixed effects in favor of random effects and these results are presented in Panel C of Table 5.3. The odd-numbered columns correspond to the level model, and the even-numbered columns to the share model.
In the case of the level model, coefficients on all the remittance variables are again negative and most of the poverty-reducing effect is picked up by the flows from the low-rights countries, while the coefficients on remaining two are smaller in magnitude and much less precise.
In the share model, the level of remittance flows appears to be the more important for the reduction of the level of poverty than their source composition. For the depth and severity of poverty, however, the opposite is true. In columns (4) and (6) remittance flows are significant only when interacted with the share from low-rights.
Table 5.2: First stage IV results of the effect of remittances from high- and low-rights countries on poverty.
I. LEVEL MODEL II. SHARE MODEL
Strategy 1 Strategy 2
(1) (2) (3) (4) (5) (6) (7) (8)
Remittances Remittances share low rights share high rights total share low rights share high rights total low rights high rights *total remittances *total remittances remittances *total remittances *total remittances remittances
Distance low rights -0.376** -0.729*** -0.436** -0.597** 0.000 -0.447** -0.695*** -0.116 (0.166) (0.160) (0.213) (0.285) (0.357) (0.187) (0.196) (0.318) Percentage Muslim 0.076*** 0.020 0.034 0.028 -0.073 0.033** -0.033 -0.102**
(0.016) (0.030) (0.022) (0.045) (0.061) (0.013) (0.034) (0.051) Common language low 0.097 0.396*** -0.142 0.401* -0.487*
(0.108) (0.143) (0.173) (0.221) (0.295)
Distance high right 0.077 -1.094*** 0.422*** -1.709*** -1.559*** 0.226*** -1.589*** -1.094*** (0.065) (0.121) (0.124) (0.229) (0.269) (0.077) (0.154) (0.182) Colonial relationship high
rights
0.413*** 0.060 0.149* -0.134 -0.334* (0.089) (0.134) (0.086) (0.187) (0.199) Common language high
rights 0.878*** 0.423** 0.907*** 0.340 0.729** 0.621*** 0.439** 0.170 (0.202) (0.179) (0.273) (0.275) (0.344) (0.143) (0.176) (0.252) Distance main remittance-sending area -0.189*** 0.140 0.289** (0.052) (0.112) (0.145) Government stability 0.084 0.658*** 0.383 (0.142) (0.165) (0.234) Education 0.027 -0.190 0.412*** 0.022 0.015 0.432*** (0.094) (0.124) (0.151) (0.052) (0.094) (0.109) Constant 2.338 11.323*** 3.182 12.453*** 7.089 3.212* 13.194*** 5.259 (1.667) (2.131) (2.222) (3.605) (4.386) (1.833) (2.597) (4.043) Observations 615 615 337 337 337 484 484 484 F stat 13.27 21.27 7.55 17.34 8.80 14.39 35.10 12.94
Robust standard errors in parentheses. All regressions include GDP per capita, Gini coefficient and regional dummy variables, regressions in columns (1) and (2) include also residual remittances per capita. All variables in logs.
* p < 0.1, ** p < 0.05, *** p < 0.01
