Remittances, Poverty and Institutions
Is the effect, of remittances on poverty, conditional on
institutional quality?
Thesis MSc Economics
International Economics and Globalization
15th of August 2016
Ilja Rensje Noordam
Faculty Economics and Business Student number: 10108416 Ilja.noordam@student.uva.nl
Thesis supervisor: drs. N.J. Leefmans Second reader: dr. D.J.M. Veestraeten
Statement of Originality
This document is written by Student Ilja Rensje Noordam 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.
Abstract
This study investigates the relationship between remittances and poverty and in particular,
this study examines whether the effect of remittances on poverty is conditional on
institutional quality in the remittance-receiving country. It is acknowledged that remittances
help to reduce poverty in the remittance-receiving countries, but the question whether the
effectiveness of remittances depends on institutional quality has never been studied. To
investigate the effect of remittances on poverty and the conditionality of institutional quality,
a pooled ordinary least squares regression is performed. The problem of endogeneity will be
discussed and instruments for remittances are constructed. However, after performing a
Durbin and Wu-Hausman test, it can be concluded that an ordinary least squares regression is
preferred over the two-stage least squares estimation. Empirical evidence for the poverty
reducing effect of remittances is found. Whether the effectiveness of remittances is
conditional on institutional quality highly depends on how institutional quality is defined and
measured and what measures for poverty are used. Results also show that the conditionality of
institutions in the effect of remittances on poverty is different for Sub-Saharan African
countries compared to low- and middle-income countries from other geographic regions.
Overall, evidence is found that higher institutional quality in Sub-Saharan African countries
make remittances more effective in reducing poverty. Again, the existence of this effect
depends on how institutional quality is defined and measured.
Table of Contents
1. Introduction
1
2. Literature review
4
2.1
Remittances and poverty
4
2.2
The role of institutions
7
2.2.1 Defining institutions and institutional quality 8
2.2.2 Institutions, remittances and poverty 8
3. Empirical Model and Methodology
11
3.1 Model specification
11
3.2 Methodology
13
3.3 Testing for Heteroskedasticity
13
3.4 Testing for Multicollinearity
14
4. Data
16
4.1 Sample selection
16
4.2 Data on poverty
16
4.3 Data on remittances
17
4.4 Data on institutional quality
18
5. Endogeneity and Instrumental Variables
20
5.1 Problem of Endogeneity
20
5.2 Instruments used in other studies
21
5.3 Instruments in this study
22
6. Regression Results
24
7. Conclusion
32
References
34
Appendices
Appendix A Sample Countries 37
Appendix B Definitions and Sources 38
Appendix C Descriptive Statistics 40
Appendix D Tests for model specifications 42
Appendix E Testing for relevant instruments 44
1. Introduction
The Millennium Development Goals, established in 2000 as an initiative of the United Nations, set the target that the proportion of people whose income is less than 1 dollar a day must be halved in 2015 compared to 1990. Fifteen years later we observe that, at the global level, the number of people living in extreme poverty has declined by more than half, falling from 1.9 billion in 1990 to 836 million in 2015 (United Nations, 2015). However, Sub-Saharan Africa (SSA) has fallen behind and did not reach the goal of halving the poverty rate in 2015. Using the new measure of extreme poverty, for which the World Bank sets the poverty line at 1,90 dollar a day, a decline in the poverty headcount rate as percentage of the population in SSA from 58 percent in 1999 to 42.7 percent in 2012 can be observed. However, as a result of the rapid population growth in SSA, the number of people living in extreme poverty has increased by more than 100 million in 2012 compared to 1990 (Beegle, Christiaensen, Dabalen, & Gaddis, 2016). Figure 1 shows the development of poverty in Sub-Saharan Africa, both relative and absolute, over time.
Source: Graph constructed by the author using the World Bank PovcalNet database.
Recently, migration and more specific the financial flows sent by migrants to their left-behind families became a topic of interest for the World Bank. These financial flows called remittances are, according to the World Bank, lifelines to the poor especially in African countries (World Bank, 2015). Figure 2 shows the evolvement of different financial inflows to SSA countries over time. It can be observed that financial flows received from migrants increased over time. The growing interest in remittance flows can partly be explained by the fact that remittances are perceived as stable currency
200 250 300 350 400 450 20 25 30 35 40 45 50 55 60 65 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 N um be r o f po o r pe o pl e ( m il io ns )
P
o
v
ert
y
h
ea
d
co
u
n
t ra
te
(%
)
Number of poor Poverty headcount rate
flows to developing countries. ‘Remittances to SSA are not only consistently less volatile than official aid, they are also less volatile than FDI’ according to Gupta, Pattillo and Wagh (2009, p. 105). Another reason for the growing interest in the topic of remittances is that they are expected to be countercyclical. In times of crises migrants living abroad are even more likely to send money to their home country to help their needy family members (Ratha, 2013; Gupta et al., 2009).
Source: Graph constructed by the author using the World Bank PovcalNet database.
This thesis will focus on the relationship between remittances and poverty. The academic literature points towards a negative effect of remittances on poverty (Gupta et al., 2009; Adams, 1991; Adams & Page, 2005). The World Bank also believes in the poverty-mitigating effect of remittances and is facilitating a reduction in the cost of making remittances by helping to improve the infrastructure for domestic and cross-border payments (World Bank, 2015b). If it is indeed true that remittances have a positive impact on poverty reduction, stimulating remittances by reducing the cost of remittances can help the poor people living in developing countries and more specific, it can help SSA-countries to reach the target of halving the poverty rate in the forthcoming years.
To infer, however, that stimulating remittance flows is a useful strategy for fighting poverty can be a misapprehension, since the effect of remittances on poverty may also depend on countries’ institutional quality. Caterinescu, Leon-Ledsma and Piracha (2009) found results that support the view that in countries with higher institutional quality remittances are more likely to positively affect economic growth. While they did not study the impact of remittances on poverty, the positive effect of remittances on growth has the potential to stimulate a decline in the poverty level. In addition, it may be that in the presence of high quality institutions remittances can be channeled more efficiently,
Figure 2. Financial inflows to Sub-Saharan Africa
0 5 10 15 20 25 30 35 40 45 50 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 20 10 20 11 20 12 20 13 20 14 U S $ in b illio n s
Personal remittances, received (current US$)
Foreign direct investment, net inflows (BoP, current US$)
which will lead to poverty reduction directly. For these reasons, one might expect that remittances as an instrument for reducing poverty are likely to be more successful in countries with good institutions.
To test for the effectiveness of remittance inflows as a strategy to reduce poverty in developing and, specifically, SSA-countries, this thesis will focus on the question whether the effect of remittances on poverty is conditional on institutional quality of the recipient country. The thesis tries to answer this question by performing an empirical analysis using a panel dataset of 105 developing countries of which 39 are Sub-Saharan African countries over the period 1996 to 2013. This study will look at the impact of remittances on poverty for all low- and middle-income countries as well as SSA-countries only. Based on the basic growth-poverty model suggested by Ravallion (1997) and the more extended model used by Adams et al. (2005) the effect of remittances on poverty will be studied. In this study the model of Adams et al. (2005) will be extended by including institutional quality variables to examine whether remittances’ impact on poverty is determined, at least in part, by the quality of the receiving country’s government institutions. The possibility of an endogeneity problem will be discussed and instruments for remittances are constructed. These instruments for remittances are used in a two-stage least squares (2SLS) estimation. The Durbin and Wu-Hausman test results, however, indicate that an ordinary least square estimation is preferred over the two-stage least squares estimation. Therefore, this thesis only reports the results of the ordinary least squares regression. The results of the 2SLS regression can be found in appendix F.
The impact of remittances on poverty has been studied several times before. However, this study will contribute to the existing literature by investigating the impact of remittances on poverty conditional on institutional quality. To my knowledge, there are no other studies that focus on the conditionality of institutional quality when investigating the poverty impact of remittances.
The remainder of this thesis is structured as follows. Chapter 2 provides an overview of the related academic literature. Chapter 3 describes the model and the empirical methodology used. Information about the data and the variables used in this study can be found in section 4. Chapter 5 addresses the possibility of an endogeneity problem caused by the two-way relation between remittances and poverty. Chapter 6 presents the main results of the empirical analysis. Finally, chapter 7 provides a conclusion.
2. Literature review
This chapter gives an overview of the existing literature on the relation between remittances, poverty and institutional quality. In section 2.1, the focus will be on the effect of remittances on poverty. In section 2.2 the role and importance of institutions will be discussed.
2.1 Remittances and poverty
Some empirical studies find no significant effect of remittances on poverty. Stahl (1982) argues that ‘international migration is an expensive venture’ and therefore it is more likely that international migrants come from better-off households. These better-off families will receive a larger part of the remittance inflows, while the poorest households are not able to benefit from such remittance inflows. As a result, inequality within the country increases and poverty alleviation among the poorest households in the society will lag behind. In a panel study of 149 developed and developing countries Cattaneo (2005) found no significant effect of remittances on poverty. The focus of the research was on the effect of migration on poverty and not on the remittances poverty nexus, however remittances as independent variable, was included in the model. The paper points out that in the analysis no significant effect of remittances on poverty can be found.
However, overall it seems that the empirical evidence points towards a negative effect of remittances on poverty. From the beginning, researchers studied the poverty effects of remittances at the micro level. These localized studies found evidence for a negative effect of remittances on poverty. For example, in contrast to the reasoning of Stahl (1982), Stark et al. (1989) found that not the members of the better-off households but the members of the ‘relatively deprived’ households are more likely to migrate. In their research, based on data of 61 Mexican households, they used a multivariate probit model and instrumental variable estimation and find that ‘migration is an effective mechanism for achieving income gains in households that send migrants to the US’.
Additional evidence of a negative impact of remittances on poverty can be found in a study of Adams (1991). The study is based on a household survey conducted in 1986/1987 including data of 1000 households living in rural Egypt. To examine the first order effect of international remittances on poverty, predicted income equations are used to estimate the changes in income that occur between the situation in which remittances are excluded in household income and the situation in which remittances are included in household income. The results show that when international remittances are included in household income, the number of poor households declines by 9.8 percent.
A more recent country-specific study, using data from the 2005-2006 Ghana Living Standard Survey, also demonstrates that it is beneficial for the poor to receive remittances from family abroad (Adams & Cuecuecha, 2013). By using a multinomial probit model, the paper estimates the probabilities of being poor and not being poor conditional on whether households receive remittances.
They identify the difference between these two probabilities as the effect of remittances on the probability of being poor. They find that the probability of being poor is reduced by 96.6% when a household receives international remittances.
A shortcoming of the aforementioned studies is that they make use of a small sample size; they only include country specific data at a single point in time. Therefore, their conclusions are of limited usefulness. As a reaction to these limitations, researchers extended the scope and conducted studies for larger sample sizes by putting together data of multiple developing countries.
Adams et al. (2005) constructed a new dataset of 71 developing countries to analyze the effect of migration and international remittances on poverty in the developing world. The dataset includes data on international migration, remittances, inequality and poverty for low- and middle-income countries. The analysis of Adams et al. is based on the basic growth-poverty model suggested by Ravallion (1997) and Ravallion and Chen (1997), where poverty as dependent variable depends on income and inequality. Including remittances as a percentage of GDP extends the basic growth-poverty model and makes it possible to investigate the effect of remittances on poverty. In addition, regional dummies are included in the model to control for geographic regional fixed effects.
First, Adams et al. (2005) perform an OLS regression and the estimate shows that international remittances have a strong, statistically significant negative effect on poverty in the developing world. However, analyzing the impact of remittances on poverty by performing a simple OLS regression may give biased results due to the problem of endogeneity. It may be that international remittances have a negative impact on poverty. On the other hand, it may also be the case that the level of poverty in the developing world has an effect on the amount of remittances sent to the home country. For this reason an instrumental variable analysis is performed. Education in the remittance-receiving country and the distance between the remittance-sending area and the remittance-remittance-receiving country are used as instruments. The regression outcome shows that ‘a 10% increase in per capita official international remittances will lead, on average, to a 3.5% decline in the share of people living in poverty’ (Adams et al., 2005, p. 1660).
Instead of merely focusing on the average effect of remittances on poverty in all developing countries, Gupta et al. (2009) studied the poverty effect of remittances for SSA-countries. The study is performed based on the same basic growth-poverty model used by Adams et al. (2005) with the exception that Gupta et al. (2009) include an interaction term between remittances and a dummy for SSA-countries. After performing a three-stage least squares estimation on data of 76 countries, including 24 SSA-countries, the paper finds a significant negative effect of the inflows of remittances on the poverty headcount ratio.
Acosta et al. (2008) conducted a cross-country study, based on data of 59 industrial and developing countries spanning the years 1970-2000. Poverty can be expressed as a function of two factors; average income level and income inequality. Hence, the effect of remittances on poverty can be determined by estimating the effects of remittances on these two factors. Using GMM estimation,
their results show that remittance inflows have a positive effect on per capita GDP growth and a small negative effect on income inequality. Combining these outcomes, the paper finds that a reduction in poverty in the range of 0.03% to 1.5% will result from a 10% increase in the remittances to GDP ratio.
A recent study performed by Azam, Haseeb and Samsudin (2016) found foreign remittances to have a negative effect on poverty, but only for the upper middle-income countries. They used an OLS model with panel data to study a dataset covering the period 1990 to 2014 for 39 lower middle, upper middle and high income countries. They constructed a model for each income group to investigate the impact of foreign remittances on poverty. They do not find a significant effect of remittances on poverty for the lower middle-income countries and the high-income countries, but they do find that a one percent increase in foreign remittances decreases the poverty headcount ratio by 0.201 percent for the upper middle-income countries.
Research is not only conducted on the question whether the impact of remittances is different depending on the income level of a country, there are also studies that have examined whether the effect of remittances on poverty is different depending on the size of remittances received as a percentage of GDP. Using a three-stage least squares method Banga and Sahu (2010) first estimate the impact of remittances on poverty for a data set of 77 developing countries. In the second part of their paper they looked at the effect of remittances on poverty for a subsample of 29 developing countries that have 5 percent or more share of remittances in GDP. In the first part of their research they found the familiar result that remittances help to reduce poverty in the remittance-receiving country. A ten percent increase of remittances reduces the poverty headcount ratio by 0.9 percent. They show that these results are even more reliable for countries with remittances greater than 5 percent of GDP. In these countries, a reduction of the poverty headcount ratio by 3.1 percent can be realized when remittances increase by ten percent.
More recently, researchers also investigated whether the effectiveness of remittances in reducing poverty depends on specific conditions in the remittance-receiving country. Akobeng (2016) for example employed a macro-level study to investigate whether the effectiveness of remittances in reducing poverty and income inequality in Sub-Saharan Africa depends on the level of financial development. The study is based on the same model used by Adams et al. (2005). Innovative in this study is that the effectiveness of remittances is studied conditional on the level of financial development in the remittance-receiving country by including an interaction term between remittances and a measure for financial development in the model. In addition, variables for inflation, finance, democracy, age-dependency ratio, trade openness and investment in physical assets are included to control for these effects. A 5-year non-overlapping average panel dataset for 41 Sub-Saharan African countries is used in the period 1981 to 2010. To overcome the problem of endogeneity caused by the possibility of reverse causality, omitted variables and measurement errors, a two-stage least squares estimation as well as a two-step dynamic system GMM are performed. The instruments used to perform the analysis are variables of the remittance-sending countries: GDP per capita and the
unemployment rate, weighted by the inverse of distance. Distance is measured by the geographic distance between the remittance-receiving country and the five main OECD countries having the greatest migrant population from the remittance-receiving country. The outcomes of both regression analyses show that remittances have a significant negative effect and these effects are augmented by financial development in the remittance-receiving country.
Besides the level of financial development, instability in the country of origin also plays a role in terms of the effectiveness of remittances in reducing poverty. In the introduction of this thesis it is mentioned that migrants living abroad are more likely to send remittances to their left-behind families when their families are suffering from a crisis (Ratha, 2013; Gupta et al., 2009). Considering remittances as a sort of ‘insurance’, results in the hypothesis that the amount of remittances received would be more important for households facing higher risk.
In line with this, Goff (2010) investigates the impact of remittances on poverty conditional on economic instability in the remittance-receiving country using a panel sample of 65 developing countries over the period 1980 to 2005. The empirical research is based on the same model as Adams et al. (2005) and to test for the conditionality of instability an interaction term between remittances and a variable measuring for instability is included. Goff tests for three types of instability: (1) macro-economic stability, (2) trade instability measured and (3) climatic instability. The OLS estimation results, in case instability is measured by climate instability and trade instability, indicate that the effect of remittances on poverty reduction will be larger in countries facing higher instability.
Remarkable in this study is that the problem of endogeneity, caused by the possible two-way relation between remittances and poverty, is refuted and in contrast to most other studies, only an ordinary least squares estimation is used. After proving that the instruments are not weak, Goff (2010) performs a Hausman test that confronts the ordinary least squares model against the instrumental variable model. The instrumental variables are first regressed on the endogenous variable, namely remittances. Subsequently, the residuals from the first stage are included in the original model and the coefficient appears not to be significantly different from zero. Therefore it can be concluded that there is no endogeneity bias and that the OLS method is preferred over the instrumental variables estimation technique. A similar test will be performed in this thesis.
2.2 The role of institutions
As the above-mentioned discussion shows, most of the empirical results present evidence for a negative effect of remittances on poverty. In addition, some studies investigated whether the effectiveness of remittances in reducing poverty depended on specific conditions, such as financial development or economic stability in the remittance-receiving country. However, none of the aforementioned studies incorporate the role of institutional quality in determining the poverty effects of remittances. For this thesis, the main question is whether good institutions, which are believed to have an impact on poverty, make remittances more effective. Section 2.2.1 will provide information
on how to define institutional quality. Subsequently, section 2.2.2 will focus on the relation between institutional quality, poverty and remittances.
2.2.1 Defining institutions and institutional quality
North (1991) defines institutions as the ‘rules of the game’ and states that ‘institutions are the humanly devised constraints that structure political, economic and social interaction’ (North, 1991, p.97). These constraints can be divided in both formal and informal constraints. Constitutions, laws, regulation and property rights primarily define formal constraints. On the other hand, informal constraints refer to unwritten laws such as for example sanctions, taboos, traditions and codes of conduct.
Instead of distinguishing between formal and informal institutions, institutions can also be divided into political, economic and social institutions. Political institutions define by whom and how authoritative decisions are made and in which way power in a country is obtained. Institutions that create networks that can either promote or frustrate economic activity are called economic institutions. Whether economic institutions have a positive or negative effect on economic activity usually also depends on the interaction with political and social institutions. Social institutions define how private and communal behaviors and interactions between individuals and amongst social groups are shaped (Leftwich & Sen, 2010).
North (1990) explains that human beings and organizations are the creators of institutions and institutions establish a predictable, but not necessarily efficient, structure for human interaction. Institutions shape solid behavior patterns over time, however institutions can change over time through people’s actions (Roland, 2004). These changes in institutions determine the way society evolves, which can have both positive or negative development outcomes (North, 1990).
In line with this, the quality of institutions is determined by the extent to which these institutions are able to positively affect a country’s overall performance. High quality institutions are able to create an incentive structure that reduces uncertainty and promotes efficiency. Therefore, ‘good’ institutions are viewed as those that contribute to stronger economic performance (North, 1990).
2.2.2 Institutions, remittances and poverty
Many studies focus on the relation between poverty and remittances, indicating that remittances and poverty in the remittance-receiving country are negatively related. The relation between institutions and poverty is also studied extensively. Nevertheless, there is no literature that investigates the interplay between institutions, remittances and poverty.
There are some studies however focusing on economic growth, institutions and remittances, trying to answer the question whether the effect of remittances on economic growth is conditional on institutional quality (Catrinescu et al., 2009; Singh, Haacker, & Lee, 2011). In these studies, it is shown that the effectiveness of remittances in stimulating economic growth depends on the quality of
institutions in the receiving countries. A sound institutional environment in the remittance-receiving countries positively affects the volume and efficiency of investment. Hence, in the presence of good institutions remittances can be channeled more efficiently, which positively affects countries economic performance. This larger positive impact of remittances on economic growth in countries with sound institutions, might positively affect the reduction of poverty.
Goff (2010) mentioned that the theoretical framework on which these studies are based is not that strong and developed. He argues that it is not that realistic to think that the quality of institutions in the remittance-receiving country can have a significant impact on the effectiveness of remittances in stimulating economic growth, since remittance flows are directly transferred to households and do not pass through the State. However, one can think of the important role institutions play in providing efficient channels for households to send and receive remittances. It is commonly known that the transaction costs of remitting can be very high relative to the usually low incomes of the migrant workers and the remittance recipients. High remittances cost are partly the result of an underdeveloped financial infrastructure, limited competition, regulatory obstacles and lack of access to the banking sector by remittance senders and receivers (World Bank, 2015a). According to Massimo Cirasino (2013), public authorities should improve efficiency by making payment system infrastructure easily and widely available to consumers. In addition, excess to information on remittances and the costs of remitting should be improved in order to enable customers to make informed decisions. More competition between companies providing services to remit can help to a direct reduction in costs and an improvement of the quality of services. Better institutions in the remittance-receiving country can help to achieve these changes needed to reduce remittance cost. Reducing the transaction costs results in efficiency gains that could have a direct impact on poverty reduction and a positive effect on economic growth in the remittance-receiving countries (World Bank, 2016c).
In the absence of an uniform theoretical framework on the poverty, remittances and institutional quality nexus and since there exist no studies empirically examining the effect of remittance on poverty conditional on institutional quality, it is hard to come up with one straightforward hypothesis. Different possible explanations can be given for why on the one hand high quality institutions or, on the other hand, low quality institutions will make remittances more effective in reducing poverty.
One possible explanation for why high quality institutions make remittances more effective in reducing poverty is that in countries with a sound institutional environment, institutions are better able to provide efficient channels through which household can send and receive remittances. As an example, sending remittances to a SSA-country is most expensive compared to other regions in the world (World Bank, 2016c). Conceivably, by reducing the cost of remitting, international transfers can be a more efficient method to alleviate poverty. The reason is that the incentive to remit increases, which positively affects both the volume and value of remittances, because it is more beneficial for remittance-receiving households to receive money from abroad. It can be expected that a reduction of
the cost of sending remittances is easier to realize in countries with higher institutional quality. Therefore, the effectiveness of remittances in reducing poverty may depend on the quality of institutions and this argumentation favors the hypothesis that when remittance-receiving countries have higher quality institutions, remittances will be more effective in reducing poverty.
However, one can also think of situations in which the opposite is true, namely, that lower institutional quality results in higher effectiveness of remittances in reducing poverty. The reasoning behind this is that it is likely that households living in countries with low quality institutions have fewer options to obtain sufficient financial resources to meet their consumption needs. For example, in countries with high levels of corruption, the powerful people in society arrogate unfair earnings and the poorest people in society, as a result, have less income. Or, in countries with low quality of public services and a lack of institutions that provide social security services in case people do not have a job, households need to find other ways to pay for their basic needs. As a result, these poor households are more dependent on the financial flows sent by their families living abroad and remittances will contribute more to the reduction in poverty. This leads to the hypothesis that remittances become more important when institutions are of low quality. In other words, remittances are more effective in reducing poverty when the quality of institutions is low.
Empirical research is needed to find out whether or not the effect of remittances on poverty is conditional on institutional quality and, if this conditionality exists, empirical research helps to find out in which way institutional quality affects the effectiveness of remittances. Therefore, an empirical analysis will be performed in the next chapters of this thesis.
3. Empirical model and Methodology
To test for the effectiveness of remittances as a strategy to reduce poverty in low- and middle-income countries this thesis will focus on the question whether the effect of remittances on poverty is conditional on institutional quality of the recipient country. To answer this question, an empirical analysis is performed based on the research of Adams et al. (2005). Section 3.1 will specify the model used by Adams et al. (2005). In addition, section 3.1 shows in which way this thesis is innovative; it shows how the model of Adams et al. (2005) will be extended in order to control for institutional quality. In section 3.2 the methodology of this empirical research will be presented. In the sections 3.3 and 3.4 several tests will be described in order to determine the correct specification of the model.
3.1 Model specification
From the literature it is known that a country's change in poverty is determined by changes in economic growth and inequality. Given a fixed income distribution, a reduction in poverty can be realized when a country experiences growth, or in other words, an increase in the level of income. The size of this growth effect of income on poverty also depends on the level of inequality. Ravallion (2005) argues that the rate of poverty reduction at a given level of economic growth is influenced by the initial income distribution and changes in the income distribution over time. In line with this, this research is built on the basic growth-poverty model (Ravallion, 1997; Ravallion & Chen, 1997), which can be specified as:
log 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖= 𝛼𝛼𝑖𝑖+ 𝛽𝛽1log(𝜇𝜇𝑖𝑖𝑖𝑖) + 𝛽𝛽2log (𝑔𝑔𝑖𝑖𝑖𝑖) + 𝜀𝜀𝑖𝑖𝑖𝑖 (1) Here, 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖 indicates the poverty level in country 𝑖𝑖 at time 𝑃𝑃. 𝜇𝜇𝑖𝑖𝑖𝑖 is per capita income and 𝑔𝑔𝑖𝑖𝑖𝑖 indicates income inequality measured by the Gini coefficient. Ravallion and Chen (1997) state that the level of poverty reduction is positively affected by an increase in per capita income and negatively affected by higher levels of income inequality. As a result, 𝛽𝛽1 is expected to be negative and 𝛽𝛽2 is expected to be positive.
Just like Adams et al. (2005) and most other researchers studying the effect of remittances on poverty, this study incorporates remittances into the growth-poverty model. To analyze the effect of remittances on poverty, model (1) is expanded by including remittances as a percentage of GDP, received by country 𝑖𝑖, at time 𝑃𝑃:
log 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖= 𝛼𝛼𝑖𝑖+ 𝛽𝛽1log(𝜇𝜇𝑖𝑖𝑖𝑖) + 𝛽𝛽2log(𝑔𝑔𝑖𝑖𝑖𝑖) + 𝛽𝛽3log(𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖) + 𝜀𝜀𝑖𝑖𝑖𝑖 (2) All variables in this basic model are included as logarithms, alike the model used by Adams et al. (2005). This makes interpreting the effects easier; the coefficients 𝛽𝛽1, 𝛽𝛽2 and 𝛽𝛽3 can be interpreted as elasticities, whereby the outcomes can be seen in proportional changes.
In what follows, it is presented in which way the empirical analysis of this thesis contributes to the existing literature. To test whether the impact of remittances on poverty is conditioned by the quality of institutions, equation (3) includes an additional term in which institutional quality is interacted with remittances.
log 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖= 𝛼𝛼𝑖𝑖+ 𝛽𝛽1log(𝜇𝜇𝑖𝑖𝑖𝑖) + 𝛽𝛽2log(𝑔𝑔𝑖𝑖𝑖𝑖) + 𝛽𝛽3log(𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖) + 𝛽𝛽4(log 𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖 × 𝐼𝐼𝐼𝐼𝐼𝐼𝑃𝑃𝑖𝑖𝑖𝑖) + 𝜀𝜀𝑖𝑖𝑖𝑖 (3)
The marginal effect of remittances on poverty can be found by taking the first derivative with respect to remittances of equation (3):
𝜕𝜕log (𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖)
𝜕𝜕log (𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑅𝑅𝐼𝐼𝑅𝑅𝑃𝑃𝐼𝐼𝑖𝑖𝑖𝑖) = 𝛽𝛽3+ 𝛽𝛽4𝐼𝐼𝐼𝐼𝐼𝐼𝑃𝑃𝑖𝑖𝑖𝑖
When the coefficient 𝛽𝛽4 is significant, the effect of remittances on poverty is conditional on the quality of institutions. When higher values of the institutional quality variable indicate higher institutional quality in the remittance-receiving country, a significant negative sign indicates that the existence of better institutions, given a fixed level of remittances, results in a larger poverty reducing effect of remittances. A positive sign however, indicates that remittances are more effective in reducing poverty when institutions are of low quality. When including the interaction term, one should be aware that the total effect of remittance on poverty is now given by 𝛽𝛽4 multiplied by the value of the institutional quality measure, plus 𝛽𝛽3. Remittances will contribute to a reduction in poverty only when the total of this sum is less than zero.
As will be discussed in more detail in chapter 4, the sample used in the empirical analysis includes data of low- and middle-income countries across the world. Since the main question of this research also concentrates on Sub-Saharan Africa, equation (2) will be adjusted by including an interaction term between the logarithm of remittances and a dummy for Sub-Saharan Africa to see whether the effect of remittances on poverty for Sub-Saharan African countries may be different:
log 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖= 𝛼𝛼𝑖𝑖+ 𝛽𝛽1log(𝜇𝜇𝑖𝑖𝑖𝑖) + 𝛽𝛽2log(𝑔𝑔𝑖𝑖𝑖𝑖) + 𝛽𝛽3log(𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖) + 𝛽𝛽4(log 𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖 × 𝑆𝑆𝑆𝑆𝑆𝑆) + 𝜀𝜀𝑖𝑖𝑖𝑖 (4) In addition, to test whether the conditionality of institutions in the effect of remittances on poverty is different in Sub-Saharan Africa, a three-way interaction term between remittances, institutional quality and a dummy for SSA is included in a separate model:
log 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑖𝑖𝑖𝑖= 𝛼𝛼𝑖𝑖+ 𝛽𝛽1log(𝜇𝜇𝑖𝑖𝑖𝑖) + 𝛽𝛽2log(𝑔𝑔𝑖𝑖𝑖𝑖) + 𝛽𝛽3log(𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖) + 𝛽𝛽4 (log 𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖× 𝐼𝐼𝐼𝐼𝐼𝐼𝑃𝑃𝑖𝑖𝑖𝑖) +
𝛽𝛽5(log 𝑅𝑅𝑃𝑃𝑅𝑅𝑖𝑖𝑃𝑃𝑖𝑖𝑖𝑖× 𝐼𝐼𝐼𝐼𝐼𝐼𝑃𝑃𝑖𝑖𝑖𝑖× 𝑆𝑆𝑆𝑆𝑆𝑆) + 𝜀𝜀𝑖𝑖𝑖𝑖 (5) A significant coefficient of the three-way interaction term 𝛽𝛽5, indicates that the two-way interaction varies across different levels of the third variable. In this case it means that the interaction term
between remittances and institutional quality is different for Sub-Saharan Africa compared to the other geographic regions in the world. Whether the total effect of remittances, conditional on the quality of institutions, has a positive or negative effect on the poverty level in SSA eventually depends on the sign of the sum of 𝛽𝛽3, 𝛽𝛽4 and 𝛽𝛽5 in equation (5), taking into account the value of institutional quality and remittances. Regression models (2) till (5) will be used in the empirical analysis. The estimation results of the different regression models are presented in chapter 6.
3.2 Methodology
This section explains the estimation procedure used to test for the effect of remittances on poverty and the conditionality of institutional quality in the effect of remittances on poverty. Just as Adams et al. (2005), the first step of the analysis consists of performing a pooled ordinary least squares (OLS) regression. Pooled OLS ignores the time dimension of the panel data and treats all the observations for all of the periods as a single sample. In order to control for fixed effects by geographic region, regional dummies are included in the model. Dummy variables for Sub-Saharan Africa, East Asia, Europe and Central Asia, Middle East and North Africa and South Asia are included. Latin America and the Caribbean is the omitted regional dummy.
The OLS estimation technique will only yield unbiased and consistent estimations when certain assumptions are met. The first assumption states that the error term in the regression should have zero conditional mean. In other words, the conditional distribution of the residuals, 𝜀𝜀𝑖𝑖𝑖𝑖, given the right-hand side variables, on average should be zero. A mean of the error terms equal to zero indicates that the right-hand side variables are uncorrelated with the error term. When this is the case, the independent variables are called exogenous. This assumption will be discussed in more detail in section 6.
The second least squares assumption states that the observations are independent and identically distributed random variables. This assumption is likely to be met, since observations are drawn from different countries across the world. The third assumption states that large outliers are unlikely and the fourth least squares assumption states that there is no perfect multicollinearity (Stock & Watson, 2012). The econometric package used for this study, Stata, automatically solves the problem of perfect multicollinearity by omitting one of the regressors that is a perfect linear function of the other regressor. However, Stata does not solve the problem when two variables have an almost perfect, linear correlation with one another, therefore section 3.4 tests whether there are problems caused by imperfect multicollinearity.
3.3 Testing for Heteroskedasticity
One important assumption for performing an ordinary least squares regression is that the variance of the residuals is homogeneous, meaning that the variance of the residuals is constant. If the variance of the error term is non-constant, the residuals are called heteroskedastic. Heteroskedasticity does not
result in biased estimates, however the estimated standard errors will be biased which in addition leads to biased t-statistics and confidence intervals. Using OLS, when heteroskedastic residuals are presented, may not be the most efficient method and biased standard errors can lead to false inferences (Stock & Watson, 2012).
There are graphical and non-graphical methods for detecting heteroskedasticity. Figure 4 included in appendix D shows a graph in which the residuals of the basic model are plotted against the fitted predicted values. From this graph, one can see a clear pattern indicating that the variance of the conditional distribution of the error term, given 𝑋𝑋𝑖𝑖, is not constant. This already gives a clear indication that the residuals are heteroskedastic. In addition, the Breusch-Pagan/Cook-Weisberg-test is applied to formally test whether the residuals are heteroskedastic. The null hypothesis that the variance of the residuals is homogeneous is tested against the alternative hypothesis that the error variances are a multiplicative function of one or more variables (Breusch & Pagan, 1979). Table 11 shows the test results and based on the outcome, one should reject the null hypothesis. This confirms the findings of the graphical method, namely that the residuals are heteroskedastic. As mentioned above, heteroskedasticity causes standard errors to be biased. Therefore, this study uses heteroskedasticisity-robust standard errors when running the regression to correct for the bias in the estimated standard errors. Robust standard errors relax both the OLS assumption that the errors are independent and identically distributed and therefore show estimated standard errors that are more reliable.
3.4 Testing for multicollinearity
In this section, a test for multicollinearity among the independent variables in the regression model is employed. Multicollinearity arises when two or more variables have a perfect, or almost perfect, linear correlation with one another. When the degree of multicollinearity increases, the estimated coefficients become unstable, meaning that the estimated coefficients may vary a lot with small changes to the input data (Stock & Watson, 2012). To test for multicollinearity among the independent variables, the variance inflation factor (VIF) is calculated. The VIF measures the increase in the variance of the estimated coefficients as a result of correlation between the right-hand side variables. There is no well-defined VIF threshold value that states whether multicollinearity arises. However, most often a value of ten is used as critical value (Stine, 1995). Table 13 of appendix D reports the calculated VIFs for equations (2). Since none of the VIF’s is larger than ten, it can be assumed that there is no problem of multicollinearity in this regression model.
To test for the conditionality of institutional quality, a total of six variables indicating the quality of institutions in a country will be included in regression model (3) and regression model (5). This will be discussed in more detail in section 4.4. By including six institutional quality variables in one model, multicollinearity is more likely to occur. Tables 14 and 15 in appendix D show the calculated VIF’s for the models (3) and (5). From table 14 it can be seen that the calculated VIFs for two of the interaction terms of remittances and institutional quality included in model (3) is larger than
10. Because the calculated VIFs are only slightly larger than ten, this will not be considered as a problem when doing the regression based on equation (3).
However, multicollinearity becomes a problem in equation (5) in which an interaction term between remittances and institutional quality and a three-way interaction term between remittances, institutional quality and a dummy for SSA are included. The calculated VIFs for most of the three-way interaction terms are much larger than ten. Therefore, to test whether the conditionality of institutions in the effect of remittances on poverty is different in Sub-Saharan Africa, only one of the six measures of the institutional quality indicators is included in each regression model. Thus, regressions models based on equation (5), only include one three-way interaction term and one two-way interaction term with one of the six institutional quality indicator. The results of this regression can be found in table 4 of chapter 6.
4. Data
Chapter 4 of this thesis discusses the underlying data used to perform the empirical analysis and provides more details on the variables included in the model. Section 4.1 gives more insights on the dataset constructed for this study. Sections 4.2, 4.3 and 4.4 provide detailed explanations of how the three most important variables of this study, poverty, remittances and institutional quality respectively, are constructed.
4.1 Sample selection
The sample includes data of low- and middle-income countries from all over the world. According to the World Bank, low-income countries are defined as countries with a gross national income (GNI) per capita of $1025 or less in 2015. Middle-income countries are defined as countries with a GNI per capita of more than $1026 but less than $12475 in 2015 (World Bank, 2016b). Just like the paper of Adams at al. (2005), this study defines an observation as any point in time for which data on income, poverty, inequality, remittances and institutional quality exists. The included sample countries and the number of observations are mainly based on the availability of poverty data and remittances data. For most countries, data on these two variables are not reported on a yearly basis and are even completely missing for some countries. This results in an unbalanced panel dataset including a total of 581 observations from 105 different countries over the period 1996 to 2013. The dataset includes a sub-sample of 95 observations of 39 Sub-Saharan African countries. More information on the sub-sample countries can be found in appendix A. Definitions of the variables included in the model as well as their sources can be found in table 7 and 8 of appendix B.
4.2 Data on poverty
The dependent variable in the empirical analysis is poverty and three different measures for poverty are used: the poverty headcount ratio, the poverty gap and the squared poverty gap. The poverty headcount index measures the share of the population whose income is below the poverty line of 1.90 dollar a day. The headcount ratio is widely used because it is easy to measure and interpret. However, this ratio has some shortcomings (Haughton & Khandker, 2009). One shortcoming of the poverty headcount ratio as measure for poverty is that not every increase in income will affect the poverty headcount ratio, since an income increase for the poorest does not necessarily bring them above the poverty line. Therefore, even though poverty in a country is reduced, it can be that the poverty headcount ratio will be unchanged. Yet, another limitation is that the poverty headcount ratio does not say anything about the ‘depth’ of poverty; the ratio does not indicate how far the poor are from the poverty line (Adams et al, 2005).
To overcome these shortcomings an additional poverty measurement is employed in this study, namely the poverty gap index. The poverty gap index measures the extent to which the average
incomes of the poor fall below the poverty line as a proportion of the poverty line (Haughton & Khadker, 2009). A poverty gap index of 25 percent, for example, indicates that the average income of the poor is 75 percent of the poverty line (in this study this means 75 percent of 1.90 dollar).
Although the aforementioned poverty gap index is a more comprehensive poverty measure compared to the headcount ratio by measuring the ‘depth’ of poverty, it does not capture differences in the severity of poverty amongst the poor. To overcome this limitation, the squared poverty gap index, as additional poverty measure, is employed in this study. The squared poverty gap index is the mean of the squared distance below the poverty line as a percentage of the poverty line. By squaring the poverty gaps, different weights are given depending on the size of the poverty gap; the lowest weighting is given to the smallest poverty gap and the highest weighting is given to the largest poverty gap. In this way, the squared poverty gap index gives more insight about the distribution of poverty below the poverty line (Haughton & Khadker, 2009).
Data on these three poverty measures is obtained from the World Bank’s PovcalNet database. PovcalNet is an interactive computational tool that makes it possible to replicate poverty calculation made by the World Bank. Poverty measures of the Povcalnet database are based on more than 800 nationally representative household surveys and data is available for 126 developing countries
4.3 Data on remittances
In this research, remittances are expressed as personal remittances received as a percentage of GDP. Personal remittances, as shown in figure 3, are the sum of three components; (a) personal transfers, (b) compensation of employees and (c) capital transfers between households. Personal transfers (a) consist of all current transfers in cash or kind between resident households and nonresident households. All transfers between resident and nonresident households are taken into account, disregarding the underlying income source, the purpose of the transfer and the relationship between the households. Compensation of employees refers to income of border, seasonal and other short-term workers who are employed a country other than their country of origin. In addition, compensation of employees consists of income of resident workers who are employed by a non-resident entity (International transactions in remittances, 2009).
Source: International Transactions in Remittances: Guide for Compilers and Users, 2009.
Most countries do not report data on the third component (c), since data is difficult to obtain. As a result, personal remittances mainly consist of data on personal transfers and compensation of employees (International transactions in remittances, 2009). Remittances data in this research is obtained from the World Bank who publishes data on personal remittances based on the balance of payments data of the International Monetary Fund (IMF) and GDP estimates of the World Bank and the Organization for Economic Co-operation and Development (OECD).
As mentioned in the Migration and Remittances Factbook 2016 (World Bank, 2016a), users of remittances data should be aware of the limitations of the currently available data. It is commonly known that large amounts of remittances are channeled through informal routes of which no or limited data is available. As a result, official remittances data is expected to understate the actual remittance flows. Freud and Spatafora (2005) for example found that informal remittances amount to about 35-75 percent of official remittances. They estimated that the people living in Sub-Saharan Africa, Eastern Europe and Central Asia use these informal channels more frequently than people from other geographic regions do.
In addition, when performing a panel data analysis using remittances data, one should keep in mind that it is hard to determine whether a higher recorded amount of remittances is indeed the result of remittances growth or the result of a lower amount of the unrecorded portion of remittances. The decrease in the unrecorded portion of remittances is a consequence of better technology, decreased transfer transaction costs, and better policies to prevent money laundering (Catrinescu, 2009).
4.4 Data on institutional quality
The discussion in section 2.2.1 shows that there exists no uniform definition of the term ‘institution’ or ‘institutional quality’. As a result, it is impossible to construct just one single indicator measuring institutional quality. Looking at empirical work on institutional quality we indeed observe that researchers use different indicators measuring the quality of institutions.
This research will use the World Governance Indicators (WGIs) constructed by the World Bank in which governance is defined as ‘the traditions and institutions by which authority in a country is exercised’ (Kaufman, Kraay and Masturzzi, 2010, p. 4). The WGIs consist of six measures of governance: (i) voice and accountability, (ii) political stability and absence of violence, (iii) government effectiveness, (iv) regulatory quality, (v) rule of law, and (vi) control of corruption. The indicators range from a scale of −2.5 to 2.5, with higher scores indicating better institutional arrangements. Alike the empirical analysis of Kwon and Kim (2014), in which the effect of governance on poverty reduction is studied, this thesis includes all six WGIs in the regression model based on equation (3). In addition, a regression will be performed in which the average value of the six WGIs, as measure for institutional quality, is included in the model. However, when all six WGIs would be included in equation (5), in which an interaction term between remittances and institutional quality and a three-way interaction term between remittances, institutional quality and a dummy for
SSA are included, multicollinearity becomes a problem. Therefore, as previously discussed in section 3.4, regressions models based on equation (5), only include one three-way interaction term and one two-way interaction term for one of the six WGIs.
The WGI dataset covers the period 1996 to 2014 in which observations for the years 1997, 1999 and 2001 are missing. Since changes in governance over short year-to-year periods are usually small, linear interpolation is used to fill the gaps (Kaufman et al., 2010). By averaging the values of the indicators of the years 1996 and 1998, a value for the year 1997 is included. In the same manner, values for the years 1999 and 2001 are included. More details on the six WGIs can be found in appendix B, table 7.
The WGIs are constructed based on 30 perceptions-based governance data sources. The underlying data sources are surveys employed by households and firms, commercial business information providers, non-governmental organizations and public sector organizations (Kaufman et al., 2010). Although perception-based governance indicators are widely used, they are subject to various criticisms. An overview of the main critiques of others researchers on the WGIs is given in the paper of Kaufman, Draay and Mastruzzi (2007). The critiques mainly focus on the fact that comparison of the indicators over time and across countries is impossible, mainly because of different underlying sources for each country. Kaufman et al. (2007) directly refute these critiques in the same article by arguing that ‘they primarily think that these critiques are based on a misunderstanding of the aggregate indicators and their interpretation’ (Kaufman et al., 2007, p. 30). Other critiques are refuted by the argument that ‘the critiques either are entirely lacking in empirical support, or even if the empirically supported to some extent, the effects are so small as to be practically irrelevant’ (Kaufman et al., 2007, p. 31).
Based on the fact that the WGIs are widely used in the academic literature and the reasonable counter arguments as a reaction on the limitations mentioned above, this research will also use these indicators as measure for countries’ institutional quality. The WGIs are widely used since they provide broader country coverage than most of the other sources. They also have the advantage that they are constructed based on many different data sources and by averaging they are able ‘to smooth out some of the inevitable idiosyncrasies of individual measures’ (Kaufman et al., 2007, p. 1). In this way, the WGIs are able to provide a broader view on governance and institutional quality.
However, by using the WGIs or, in general, any sort of institutional quality indicator in empirical research, it is important to keep in mind that measuring governance and institutional quality is difficult, and that measures can be imprecise, subject to margins of error, and therefore require interpretative caution (Kaufman et al., 2010).
5. Endogeneity and instrumental variables
This chapter addresses the possibility of an endogeneity problem. Section 5.1 explains what causes this problem and introduces the two-stage least squares estimation technique as a solution for the endogeneity problem. In order to perform 2SLS estimations, one needs good instruments. Section 5.2 gives a brief overview of instruments for remittances used in empirical research performed by other researchers. Section 5.3 shows what instruments for remittances are constructed for the empirical research in this thesis. In addition, in section 5.3 a test for the relevance of these instruments is described as well as a test for variable endogeneity.
5.1 Problem of Endogeneity
As mentioned in section 3.3, an important underlying assumption of performing an OLS regression is that the right-hand side variables of equations 2, 3 and 4 are exogenous to poverty. However, it may be the case that remittances are endogenous to poverty and that there exists a problem of endogeneity. The endogeneity problem results from the possibility of reverse causality between poverty and remittances. In this thesis we expect remittances to have a causal effect on poverty. However, one can also expect that the poverty level in a country affect the number of migrants and the level of remittances that the country receives. Therefore it may be possible that there exists a causal effect of poverty on remittances.
Simultaneous causality bias occurs in case both causal effects appear to be present. This will be a problem when one wants to do a pooled OLS regression, because the independent variable will be correlated with the error term. As a result, the outcomes of the regression can be biased. One solution of this problem is to use Instrumental Variable (IV) estimation (Stock & Watson, 2012, p. 368). When more than one relevant instrument for each endogenous variable is available one can use the two-stage least squares (2SLS) estimation model. The 2SLS model consists of two stages. In the first stage, the available instruments and the exogenous variables are regressed on the endogenous regressor to see if there is a significant effect of the instruments on the endogenous variable and to get predicted values for the endogenous variable. The predicted values obtained in the first stage are used in the second stage to estimate the effect on the dependent variable.
Any instrument used in the first stage should meet two requirements (Stock & Watson, 2012, p. 481). First, the instrument chosen in the first stage should be relevant, meaning that the instrument must have significant influence on the endogenous variable even after controlling for the exogenous regressors. This requirement can be tested in the first stage regression using a simple test of weakness of the instrument. The second requirement is that the instrument is exogenous, meaning that the instruments are valid in case they are uncorrelated with the error term. In contrast to the first requirement, the exogeneity of the instrument can in general not be tested.
5.2 Instruments used in other studies
In order to test for the effect of remittances on poverty using a 2SLS regression, valid instruments have to be found first. As in any study using instrumental variables, it is challenging to find instruments that are both valid and strong. In their empirical research Adams et al. (2005) use different instruments for remittances. One of the instruments used in their study is the level of education in the remittance-receiving country. Education is measured as the percentage of the population over twenty-five that have completed secondary education. The level of education in the remittance-receiving country is positively related to remittances; migrants with higher education typically enjoy greater job opportunities in labor-receiving countries and therefore higher educated migrants are able to remit more. The second instrument used by Adams et al. (2005) is the distance between remittance-sending countries and the remittance-receiving country. Distance between the remittance sending and receiving countries is usually negatively related to the level of international migration, since one can argue that it is more expensive to migrate when countries are further apart. Fewer international migration results in fewer remittance flows. In addition, one can argue that distance does not directly affect the poverty level in the remittance-receiving country, indicating that this instrument meets the second requirement. Akobeng (2016) uses two instruments that are based on the economic conditions in the remittance-sending countries. The level of poverty in the remittance-receiving countries is not directly affected by the economic conditions in remittance-sending countries, whereas remittance flows are affected by the economic conditions in the remittance-sending countries. Akobeng (2016) argues that migrants are able to remit more when the economic conditions in the remittance-sending country are better. The measures he uses to indicate the economic conditions are GDP per capita and the unemployment rate in the remittance-sending country. Both measures are weighted by the inverse of distance between the remittance-receiving country and the remittance-sending country, because of the inverse relation between distance and remittances.
A good suggestion of an instrument for remittances would be the transaction costs of sending remittances since the cost of remitting is believed to be negatively correlated with remittance flows (Freud & Spartafora, 2008; Avcinena, Martinez, & Yang, 2010). At the same time, it can be expected that remittance costs are uncorrelated with poverty in the remittance-receiving country. Unfortunately, data on these transaction costs are only available for a limited number of countries and only for a few years. Therefore remittances cost are not suitable to use as an instrument for remittance flows in this study. Since 2008 the ‘Remittances Prices Worldwide’ database is establish as an initiative of the World Bank to improve data availability on the cost of remitting and to monitor the remittances costs. I hope that this will lead to sufficient observations in the future, which makes it possible to use remittance costs to instrument for remittances.
5.3 Instruments in this study
For the purpose of this research four instrumental variables are constructed based on the instrument used in the studies mentioned before: distance, education and two instruments indicating the economic situation in the remittance-sending country. Distance is measured as the average distance between a remittance-receiving country and the five OECD countries that are the top receivers of migrants for that remittance–receiving country. These average distances are calculated for each remittance-receiving country included in the analysis. Education is measured just as in the study of Adams et al. (2005): the percentage of the population over age 25 that completed secondary education. Data on education is obtained from the Barro-Lee educational database. Data is available from 1980 to 2010 in 5-year intervals. Linear interpolation is used to cover the missing years. This study constructed two variables to indicate the economic conditions in the remittance-sending country: real GDP per capita and the unemployment rate of the top 5 OECD countries that are the top receivers of migrant for each remittance-receiving country, weighted by the inverse of distance. More information on these instruments can be found in appendix B.
A first simple check to test for instrument relevance is done by looking at the correlation between the possible instruments and the endogenous variable. As can be seen from the cross-correlation table included in appendix C, significant cross-correlation between remittances and the instruments distance and education exists. However, the constructed instruments indicating the economic conditions in the remittance-sending country both appear to be irrelevant since they are uncorrelated with remittances. Therefore, these variables are dropped and only distance and education are used as instruments.
To test for the strength of the distance and the education variables as instruments for remittances one can look whether the effect of these two variables on remittances is significant. Table 16 in appendix E reports the first stage results in which the instruments and the exogenous variables are regressed on remittances. Again, all variables are included in the model as logarithms. As expected, the coefficient of distance appears with a significant negative sign, indicating that larger geographic distance between the remittance-sending and remittance-receiving countries results in less remittance flows. Education also appears with the expected sign. The positive significant coefficient indicates that remittance-receiving countries with a higher share of educated people, receive larger amounts of remittance flows. By controlling for regional fixed effects, the size of the effects become a little smaller and the effect for education becomes slightly less significant.
When the two instruments are included in the model, both the coefficients of education and distance show the expected sign and are significant. If the joint F-statistic in the first stage regression is larger than 10, instruments are not weak (Stock & Watson, 2012). Based on this rule of thumb and a F-statistic of 19.02, when both education and distance are included in the model, it can be concluded that these two instruments can be considered as relevant.
An additional test is performed to compare the OLS and the 2SLS model coefficients and to test whether remittances should indeed be treated as endogenous. If remittances are in fact exogenous, the OLS estimator is more efficient compared to the 2SLS model and simple OLS is preferred over 2SLS. After performing the first stage regression including both instruments, the residuals from this regression are obtained. Thereafter, the residuals are included as explanatory variable in the second stage model. The coefficient of the residual appears not to be significantly different from zero. Therefore, it can be concluded that the remittances variable is not endogenous in the effect on the poverty headcount ratio. This analysis is also performed for the two other measures of poverty and the results are the same as for the poverty headcount ratio. The outcomes are presented in table 17 of appendix E.
In addition, a Durbin and Wu-Hausman test is performed as a robustness check. The Durbin and Wu– Hausman tests assume that the error term is independent and identically distributed. This research, however, uses robust standard errors and therefore Stata will report Wooldridge’s score test and a regression-based test of exogeneity when performing the test for endogeneity. Heteroskedastic and auto-correlated errors are tolerated by both tests. The null hypothesis states that the regressor, in this case remittance, is exogenous. Based on the test results reported in tables 18-20 of appendix E, the null hypothesis cannot be rejected, indicating that OLS regression is preferred over the 2SLS method. These test results are in line with the test results of Goff (2010) and just like his research, the conclusions of this study are based on the results of the pooled OLS regressioni. The results of the ordinary least squares regressions are presented in chapter 6.
iThe results of the 2SLS regression are presented in table 22 and table 23 of appendix F. For half of the
institutional quality indicators, the results of the 2SLS corroborate the results of the OLS regression presented in chapter 6. In addition, tables 22 and 23 report the results of the Durbin and Wu-Hausman test for each model. These test results indicate that an OLS regression is preferred over a 2SLS regression for almost all models (exception: model 3 in table 23).
