Determinants of remittance prices : evidence from 201 country corridors

(1)

University of Amsterdam

F

ACULTY OF BUSINESS AND ECONOMICS

AMSTERDAM SCHOOL OF ECONOMICS

MASTER THESIS

Determinants of Remittance Prices:

Evidence from 201 Country Corridors

Amsterdam 2015

Author: Martin erný UvA id: 10826637

Supervisor: Prof. J. Hartog Academic Year: 2014/2015

(2)

Statement of Originality

Statement of Originality This document is written by Student Martin erný 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. Amsterdam, August 27, 2015

(3)

Bibliography Reference

erný, M. (2015). Determinants of remittance prices: Evidence from 201 country corridors., 54 p. Master Thesis (Msc.) University of Amsterdam. Faculty of Economics and Business, Amsterdam.

(4)

Abstract

Despite the clear importance of remittances, migrant workers are often charged very high prices for sending money home. This decreases the total volume of remittances received by the developing countries and stimulates the informal sector. What explains these prices on a global scale? Why is it cheaper to send money to Ghana from Netherlands than from Germany? Fixed eect and Mundlak estimation is used to analyze the determinants of remittance prices in 201 country corridors over the period 2011-2014. In contrast to previous ndings, this thesis nds a non-linear eect of competition among remittance service providers and GDP per capita in sending and receiving countries on the remittance price. The negative eect of nancial development on price is conrmed. The analysis also reveals that the absolute number of migrants seems not play a role. Instead, the relative share of corridor migrants to the total population has a signicant and negative eect. The results are robust to product and provider heterogeneity as well as to dierent price components or remittance value.

JEL classication: F24, F30, O11

(5)

Acknowledgments

I would like to thank to my supervisor Prof. J. Hartog for valuable suggestions and guidance throughout the project delivery. Further, I would like to extend extra appreciation to Prof. Menno Pradhan for everything he has taught me this year. I am also grateful to my parents for a support during my entire studies. Last but not least, this thesis would not be written without the help and life support of my girlfriend Ymy Vu Thi.

(6)

List of Figures

1 Typical Remittance transfer process . . . 4

2 Global Average and Global Weighted Average price . . . 9

3 Average prices across 201 corridors, year 2014 . . . 11

4 Average Price of Remittances from the USA to 27 Receiving Countries . . . 12

5 Average Price of Remittances from 13 sending countries to Philippines . . . . 12

6 Prices by individual RSPs in the Australia-Philippines corridor, year 2014 . . 13

7 Price variation across products . . . 13

8 Comparison of average remittance prices by time, transaction value and price component . . . 14

9 Comparison of average remittance prices by RSP type and price components 15 10 Comparison of average remittance prices by product type and region . . . . 15

11 Development of average price over time by RSP type . . . 17

12 Development of price over time, corridor level . . . 18

13 The eect of a positive bilateral migrant-stock shock on remittance price in a corridor A . . . 20

14 RSP competition by corridor and country . . . 21

15 MTOs by global presence in corridors . . . 21

16 Nonlinear relationship of competition, GDP per capita and price . . . 22

17 Partial plots . . . 32

18 Components Required for Compiling Remittance Items and Their Source . . 42

19 Remittances sent from or received by the Netherlands, 2014 . . . 43

List of Tables

1 Summary statistics of explanatory variables . . . 25

2 Correlogram . . . 26

3 Baseline estimates . . . 29

4 Additional Variables . . . 31

5 MTOs, Banks and all RSPs . . . 33

6 200vs500, fee vs FX margin, MoneyGram only . . . 34

(9)

List of Abbreviations

AGRP Average Global Remittance Price BP11 Beck and Peria (2011)

BSC Banking Sector Competition

CEPII Centre d'Etudes Prospectives et d'Informations Internationales CZK Czech Koruna

EUR Euro FE Fixed Eect

FDI Foreign Direct Investment FS08 Freund and Spatafora (2008) FX Foreign Exchange

GDP Gross Domestic Product

GMM Generalized Method of Moments

IFAD International Fund for Agricultural Development IMF International Monetary Fund

IV Instrumental Variable LCU Local Currency Unit

MIDAS Mixed Data Sampling Models MG MoneyGram

MR Migrant Worker

MTO Money Transfer Operator ODA Ocial Development Assistance

OECD Organization for Economic Co-operation and Development PCA Principle component analysis

RE Random Eect

RPW Remittance Price Worldwide RSP Remittance Service Provider

(10)

TC Total Costs UK United Kingdom USD United States Dollar WB World Bank

WDI World Development Indicators WU Western Union

(11)

1 Introduction

Ocial remittances ows the money sent home by migrant workers are constantly growing and are projected to reach over USD450 billion in 2015 and thus represent an essential source of external funds for households in developing countries. The volume is three times larger than Ocial Development Assistance (ODA) and, excluding China, signicantly exceeds foreign direct investment ows (FDI) to developing countries (World Bank 2014). For countries such as Tajikistan, ocial remittances represent close to 50% of the country's GDP.

Remittances are less volatile than FDI and other private capital ows and are reliably bringing in foreign currency that helps sustain the balance of payments. Further, unlike ODA and FDI, remittances go directly to recipient households, augmenting the resources at their disposal and generating strong multiplier eects across local markets. Moreover, empirical evidence suggests that remittances support social and economic development; leading to increased investments in health, education, and small businesses in receiving countries.

Despite the clear importance of remittances many countries face incredibly high charges for sending migrants money home over the border. The global average price is around 7.9% of the value sent but for example migrant workers from Mozambique sending money home from South Africa, or Ghanaians remitting money from Nigeria can face charges well in excess of 20%. Some remittance service providers (RSP) however charge over 50% of the value sent (World Bank, 2014). In an age of mobile banking, internet transfers and rapid technological innovation, such high charges are unreasonable.

The high charges associated with remittance transfers have long been recognized as a con-straint on development. Yet international eorts to reduce those charges have achieved limited results. In 2009 at the L'Aquila summit, G8 adopted the 5x5 objective which refers to a reduction of the global average total cost of migrant remittances from the 2009 value of 10% down to 5% in 2015 (RPW, 2015). Due to this joint eort, the global average cost of sending remittances has been declining from 2009 but is still well above the 5% goal.

Policy makers stress that remittance costs can be reduced by encouraging banks, micro-nance institutions, credit unions, and post oce saving banks to oer remittance services; by sharing existing payment platforms and networks; and by avoiding exclusive partnership arrangements between major money transfer operators such as Western Union and post of-ce networks (World Bank, 2005). Cirasino (2014) emphasizes (i) underdeveloped nancial infrastructure, (ii) scarce transparency, (iii) regulatory obstacles and (iv) lack of access to the principles for international remittance services as the crucial elements holding remittance prices high.

If remittance charges were reduced, there would be a double benet: the overall ow of trans-fers would increase and a greater share of the transfer would reach the intended beneciaries. Watkins and Quattri (2014) estimate that reducing the charges down to the 5% goal would bring Africa an extra 40% of the UK aid budget or in other words enough money to pay for the primary school education of 14 million children in the region. Further, due to the high

(12)

charges, it has been estimated that the ows through formal channels may account for only up to 50% of the total transfers (Amjad et al., 2013). Reducing the prices and thus increasing the use of formal channels would also help to track nancing of illegal or terrorist activities. What explains these high prices on a global scale? Why is it cheaper to send money to Ghana from Netherlands than from Germany? How come that the price varies substantially even within a country, with for example a Singapore - Bangladesh transfer being almost 10 times cheaper than Singapore - Pakistan (RPW, 2014)? What explains these price dierences? In this thesis I will take the approach of Freund and Spatafora (2008) and Beck and Peria (2011), later on abbreviated as FS08 and BP11 respectively, who analyzed the determinants of remittances prices based on cross-sectional data.

I will enrich the existing literature in a number of ways;

Firstly, I will present a novel diagram of a remittance sending process and a clear analytical framework for the calculation of the remittance prices which helps the reader to fully understand the complex nature of remittances.

Secondly, I will take the advantage of the newly available panel data on remittance prices collected by the World Bank. This will enable me to explore the causal relation-ships and thus conrm or reject previous literature ndings. Moreover, I will analyze over 200 country corridors which is a 40% increase from BP11. This will shed more light on the global picture of the remittance market.

Thirdly, I will present a simple but a novel macroeconomic model of a downward sloping remittance supply curve and tests its predictions.

Fourthly, I will add to the existing literature by distinguishing between the xed fee and the foreign exchange margin component of the total remittance costs.

Lastly, I will control for product and provider heterogeneity and thus present the most accurate estimates insofar.

The rest of the thesis is organized as follows. The next section explains the complex nature of the remittance sending process. Section three summarizes the existing literature on re-mittances with a special focus on remittance prices. The following section gives a detailed graphical analysis of the remittance prices. Section ve describes the model, explanatory variables and the methodology that will be used for the estimation. Section ve presents the empirical ndings. Section six then concludes.

2 Remittances - How It Works

2.1 Denition of Remittances

Generally, remittances are dened as ...household income being generated by economic activ-ity in another than the home economy, which subsequently is transferred to or earned on the

(13)

account of the household in the home economy. As a consequence remittances can be either based upon long-term residence in another country (usually combined with economic activity as migrant worker), or upon short-term work engagements as seasonal worker... (Eurostat 2015)

According to the International Monetary Funds (IMF 2008), total remittances are divided into several categories: (i) Personal transfers, (ii) Compensation of employees less expenses related to border, seasonal, and other short-term workers, (iii) Capital transfers between households, (iv) Social benets and (v) Current and capital transfers to non-prot institutions serving households. These items are recorded individually by each country in its balance of payment statistics. See gure 18 in the appendix for full details.

In the rest of the thesis, if not stated otherwise, remittances and the corresponding transfer prices will refer solely to a category (i) Personal transfers or to a transfer of household funds to a non-resident household, usually situated in the migrant's home economy (Eurostat 2015). Bilateral remittances, or remittances within country corridors are then for the purposes of this thesis dened as personal transfers from a country A to a country B.

2.2 Remittance Transfer Process

It is beyond the scope of this thesis to provide a detailed structure of the complex remittance-market. A brief summary follows. A typical remittance sending process takes place in 6 steps. After a migrant worker (MR) earns money in a foreign country, the MR (1) decides to choose a RSP based on several criteria, (2) the MR pays the principle amount to a chosen RSP, (3) the MR contacts relatives and the RSP contacts its agent in a destination country, (4) a destination country agent delivers remittances to the beneciary, (5) the RSP settles funds with its destinations country agent and (6) conrmation.

The rst step involves a dicult decision making for the migrant worker. She can choose between formal, regulated, or informal1 _{RSPs. Formal ones include commercial banks,}

money transfer operators (MTOs), or another nancial institutions, a post oce, a money changer, or a merchant (e.g., a gas station, grocery store). Informal channels include cash carried in person or via mail over the border or a so called Hawala transaction. 2

1_{Freund and Spatafora (2005) estimate that informal remittances amount to 35 to 75 percent of formal}

remittances to developing countries. Main reasons for the informal sector utilization include higher speed, trust and lower price of the transfer. Another advantage of the informal channels is the anonymity as nor the sender or the receiver does have to present her identication card or her bank account number. It is important to note that the word informal does not implicitly mean illegal. However informal channels have increasingly been debated due to concerns about potential misuse for criminal ends, including money laundering, the nancing of terrorism and smuggling (Siegel and Lucke, 2013)

2_{Hawala (or Hundi) systems rely on trust and an extensive use of connections such as family relationships}

or regional aliations. Hawala can be dened as a Money transfer without money movement as throughout the whole procedure, often no money crosses the border, and no ocial records exist for this transaction. A simple example is netting a debt of two parties via a migrant worker. Assume that person A in country X wishes to send money to person B in Country Y. Further assume that person C in country Y owes money to a person D in country X. Person A can pay person D and person C can pay person B. No money crosses the border yet the remittance transfer (and the debt repayment) take place. For illustrative examples on Hawala system see Jost and Sandhu (2003) or Passas (2006).

(14)

Further not all RSPs transfer money to all countries. Some are global, such as Western Union (WU) or MoneyGram (MG), some work only within a single corridor. A decision has to be also made on how the beneciary receives the money. Will the money be delivered directly to the beneciary`s house in cash, will it be transferred to an account or will she have to travel to the capital city to withdraw cash from a bank? Will she receive the amount in local currency unit (LCU) or will it be in USD or in EUR? How long will the transfer itself take? The speed of the transfer varies from less than one hour to more than six days. Lastly, the price that the RSP charges for their service plays an important role. To attract customers, RSPs also oer special products that link remittances to other nancial products such as insurance, pensions or bill payments.

In step 2, the MR walks into the RSP oce and lls out a form regarding her personal information, receiver personal information, amount that she wants to send, in which currency and form etc. The RSP takes the remitted amount from the MR and charges fees for the service. MR gets a receipt.

In step 3 the RSP in the sending country informs the agent in the destination country to deliver the remittance to the beneciary. The MR also noties the beneciary about the transaction and its details. In step 4, the RSP agent in destination country makes the payment to the beneciary. In step 5, the sender receives a conrmation that the payment was received (from recipient) and delivered (from RSP). In most cases, there is no real-time fund transfer; instead, the balance owed by the sending RSP to the destination RSP is settled periodically according to a mutually agreed schedule (Ratha and Riedberg, 2005). Hence the step 6 is the settlement. See a gure 1 below for a detailed diagram of the process.

Figure 1: Typical Remittance transfer process

Source: Author`s own compilation based on Ratha and Riedberg (2005) and IMFEC (2009) *Note: Not all steps are always included

(15)

2.3 Recent Developments

With the expansion of the internet, many formal online platforms, such as WorldRemit, Xoom or TransferWise, have emerged that provide an alternative to the traditional RSPs. With the emergence of mobile money such as M-Pesa3 _{in Kenya and Tanzania, it is possible}

to send the transfer directly to the recipient's phone using local mobile money platforms. The sender can send money through a mobile app or online. She immediately gets a conrmation that the transaction has been processed (implying that it has been taken from their account and has reached the recipient). The recipient can either immediately retrieve the money in cash from an M-Pesa agent or just uses it in a mobile money form. The advantage here is that the beneciary does not have to have any formal bank account nor has to travel to the nearest RSP agent location.

Some remittance systems do not pay in money, but pay directly in some kind of a prede-termined good. For example via MamaMikes company, you can send anyone with a mobile phone in East Africa Airtime credit for phone or even prepay certain amount of hours of electricity (Airpesa, 2015). It is also possible for example to pay for a goat to be remitted in Nepal (Brinkerho, 2008). The drawback is that it gives the seller a very strong position in terms of setting excessive prices, as the beneciary must choose amongst the goods at predetermined prices. If the beneciary was paid in cash, she could shop around for better goods or prices.

With HealthWallet application developed by Dutch Non-governmental Organization together with Safaricom, you can send money or vouchers that can be used only for specic health services such as an anti-malaria treatment within pre-determined health facilities. With these in-kind transfers, the family or donor is sure that the transfer was used for the right purpose (PharmAccess, 2015).

3 Literature Review

This thesis restricts the literature review to a quite limited and not yet much studied work considering the costs of remittances.4

3_{M-Pesa (M for mobile, pesa is Swahili for money) is a mobile-phone based money transfer, launched in}

2007 by Vodafone for Safaricom, the largest mobile network operators in Kenya. It has since expanded to 10 countries including India and Romania. M-Pesa allows users to deposit, withdraw, transfer money and pay for goods and services (Lipa na M-Pesa) easily with a mobile device (Vodafone, 2015).

4_{A central question for development economists has been the eect of remittances on poverty, development}

outcomes and economic growth. A comprehensive overview of the evidence available has been provided by Ratha (2013). Generally, remittances are considered to have a positive eect on alleviating poverty (Adams and Page 2005, Gupta et al. 2009) and on economic growth (Giuliano and Ruiz-Arranz, 2009). However, some studies show that remittances may also have negative eects. For example, Acosta et al. (2009) show that remittances can lead to a decline in labor supply and a shift in consumption demand toward non-tradables, which can induce the economic phenomenon known as the Dutch disease.

(16)

3.1 What Role Do Costs Play in Sending Remittances?

Gibson et al. (2006) estimate on the case of New Zealand-Tonga that cost-elasticity of sending remittances is thought to be lower than -1 with respect to the xed fee component of money transfer costs. This implies that a decrease in cost of sending by 5% will increase the volume send by more than 5%. Freund and Spatafora (2008) nd that high transaction costs signicantly reduce recorded remittances: a one percentage point reduction in transaction costs raises recorded remittances by 1423%. Further they nd that the number of migrants is the primary determinant of ocial remittances: an increase in the stock of migrants residing in OECD countries leads to a roughly proportionate increase in recorded remittances. Both FS08 and Kosse and Vermeulen (2013) show that when formal remittance costs decrease there is a large shift in consumers from informal to formal methods of remittances.

Using bilateral data on remittance ows to Pakistan for 23 major host countries during the period 2001-2013, Ahmed and Martínez-Zarzoso (2015) found that the eect of transaction costs on remittance ows is negative and signicant. Aycinena, Martinez, and Yang (2010) run a randomized eld experiment among migrants from El Salvador in the Washington DC area. They estimate, that the remittance fee reductions led to higher transaction frequency by remitters: each $1 fee reduction led to an additional 0.11 transactions per month. There was no change in the dollar amount remitted per transaction, resulting in an increase in total remittances sent.

Contrary to these ndings, Mookerjee and Roberts (2011) nd that transaction costs have no impact on remittances in 64 countries studied and that a crucial determinant is the nancial sector development measured by bank branches per 1000 km2. This nding goes against the need to follow the above mentioned 5x5 objective to decrease remittance costs.

3.2 Determinants of Remittance Prices

Only a limited number of studies have looked at the cost of the remittances itself. Orozco (2002) was the rst to analyze the remittance market and the costs of sending remittances. This report analyzed (not empirically) the market of remittances from the United States to four Central American and Caribbean countries. The main observation is that remittances appear less costly when competition is greater. Ratha and Riedberg (2005) interviewed major RSPs across the world. Their major contribution is the analysis of not only the costs to clients (migrant workers and recipients) but also the costs to RSPs. Sta, marketing, technology, rent and security expenditures seem to be the most important for RSPs. They also point out that the prices to consumers are high due to lack of competition and restricted access of some consumer categories to existing remittance services. They further advocate that smaller nancial institutions could play useful role in channeling remittances, but they do not typically have access to national clearing and settlement systems.

FS08 was the rst to analyze empirically the cost of remittances. They collected data from Western Union U.S. and UK and MoneyGram U.S. on the service fees associated with send-ing remittances (200USD) to 60 countries in 200506. Ussend-ing ordinary least squares (OLS) estimation on the cross-sectional data, they regressed transaction costs on several variables.

(17)

The nd that higher nancial development and dollarization have negative and signicant eect on the transaction costs. Remittance ows and the stock of migrants have also negative eect but remain insignicant. On the other hand GDP per capita, higher nancial risk and higher bank concentration increase the price although they remain insignicant throughout their estimations.

Closest to this thesis, BP11, analyze 119 country corridors to explore the factors associated with the price of remittances. In contrast to FS08, they nd that corridors with larger numbers of migrants, lower barriers to access banking services and less restrictive regulation of RSPs exhibit lower prices for remittances. Remittance prices are higher in richer corridors and in corridors with greater bank participation in the remittance market. In contrast to FS08, no robust association is found between remittance prices and measures of exchange rate stability or the presence of capital controls on remittances. Similarly, the distance between sending and receiving countries, the extent of bilateral trade, and whether countries share a common language or the share of educated migrants are not correlated with remittance prices. BP11 do not nd any signicant dierence between a USD200 and USD500 remittance. The BP11 analysis is however based only on cross-sectional data on prices from 2009. Further, often out-dated explanatory variables such as migrants' education from year 2000 or access to nancial services variables from 2003 are used. This thesis will thus shed more light on the price determinants by analyzing panel data, including the most up-to date measures and looking at the eect of new variables. BP11 and FS08 also do not take into the account the heterogeneity of various remittance products and analyze only aggregated prices. Another caveat of their approach is not exploring the non-linear relationships of some key variables. It is important to note that both PB11 and FS08 likely suered from a small sample bias as FS08 had only around 60 observations and BP11 119 but for some variables only 50.

4 Data On Remittance Prices

4.1 Denition

The total that is charged by the RSP r at time t for transferring an amount a via a product p from country i to country j T Cr(t, a, p, i, j) is a simple sum of an administration fee (F ee)

and an exchange rate margin (F X margin) expressed in percentages of the value sent: T Cr(t, a, p, i, j)(%) = F eer,t,a,p,i,j(%) + F X marginr,t,a,p,i,j(%) (1)

The F ee is the most visible cost component, and can dier signicantly among market players. This fee usually represents the charge the sender pays at the initiation point, and usually varies with the amount sent, within set bands. This transfer fee charged can range anywhere between 0 to more than 50% of the value sent and is calculated as follows:

F eex(%) =

LCU f eex

LCU amountx

(18)

where LCU fee refers to a fee denominated in local currency unit, e.g. CZK 100 and LCU amountto a value sent in local currency unit e.g. CZK 2000. x is a vector of r, t, a, p, i, j as explained above.

An important portion of the remittance cost is the exchange rate spread, which is not quoted in the transfer fee. Even though the remittances can be paid in US dollars in some countries, the majority of remittance transactions are paid in local currencies, and, thus, an exchange operation is required. The exchange rate (FX) fee or the F X margin of a RSP i at time t is calculated as a percentage dierence between the foreign currency exchange rate applied to the transaction by the RSP (F X) and the interbank exchange rate (F X bank) at time t.5 The F X margin varies between 0 (if the home and receiving currencies do not dier) and well over 10%. Some RSPs oer even a better rate than the interbank rate making the margin negative. The formula for computing the F X margin is as follows:

F X marginx(%) =

F X bankt− F Xx

F Xx

∗ 100 (3)

Naturally one RSP can be cheap regarding the administration fee but expensive regarding the FX margin. In the rest of the thesis, if not stated otherwise, price or transaction costs of remittances is understood as the sum of the administration fee and the FX margin.

4.2 Average Global Remittance Price

Average global remittance prices of sending USD 200 at time t (AGRP 200t) is calculated

by the RPW(2015) as a simple average of total costs across all products of all RSPs and corridors at time t according to the formula:

AGRP 200t= 1 k · n · z k X c=1 n X r=1 z X p=1 T C200c,r,p,t (4)

where p = 1...z is the number of products a RSP oers, r = 1...n is the number of RSPs per corridor, c = 1...k is the number of corridors and t is time and T C200 is the total costs of sending USD200 as calculated in eq. (1)

This serves as a baseline value for the 5x5 objective described above. The value has been declining, but is still above the 5% target. 6 _{Figure 4 below also depicts a weighted average}

price (red line), which accounts for a relative size of the bilateral remittance ows in each 5_{It is important to note that the FX mark-ups uctuate daily.}

6_{Evidence suggests that informal remittance channels are cheaper than formal channels, especially banks}

and MTOs like WU and MG. For informal remittance channels as a whole, Sander (2003) reports the average cost of remitting at 35% globally, although it can be higher in specic cases. Likewise, Swanson and Kubas (2005) report costs of 15%. Orozco (2002) estimates the cost of a Hawala transaction to be less than 2% of the value of the principal. Similarly, remittances through friends, taxi drivers, etc. are also low-cost compared to formal channels.

(19)

Figure 2: Global Average and Global Weighted Average price

Source: RPW (2015)

*The gure excludes all transactions within Soviet Republics as they operate based on the integrated payment systems of the former USSR and are not comparable to RSPs which incur high costs when having

to bridge the national payment systems in two countries.

corridor. 7 _{The weighted average is below the normal average, suggesting that lower prices}

are associated with higher volumes.

One has to bear in mind that the weights do not take into the account the relative market share of each RSP and thus the real average might dier signicantly. Further, Ratha and Riedberg (2005) claim that: ...there is not a single price for a commodity called remit-tance, just as there is no single price for cars. Remittances are dierentiated by a number of characteristics that make them special... [Ratha and Riedberg (2005), p.17]. This further questions the usage of the AGRP for policy purposes. The following section will support this claim by looking at the price variation across several levels- corridor, RSP and product level.

7_{The precise volume and origin of the bilateral remittance ows is not known as in the Balance of Payments,}

countries do not track via which channel or from what country was the payment made. Nevertheless, estimates based on sending and receiving countries GDP and bilateral migrant stock serve as a benchmark. See Ratha and Shaw (2007) for details on the methodology. Figure 19 in the appendix illustrates the bilateral remittance ows for the Netherlands.

(20)

4.3 Remittance Price Worldwide (RPW) Database

4.3.1 The Database

The prices of remittances have been collected by the World Bank on a semi-annual basis from 2009 and on a quarterly basis from 2013 onwards. Information was collected through phone calls to agents (RSPs) on location by posing as customers (mystery-shopping). 8

The database includes information about USD200 and USD500 value of the remittance. World Bank researchers collected the data within each corridor on the same day, in order to control for uctuations in exchange rates and other changes in fee structures. Currently, this Remittance Price Worldwide (RPW) database covers 235 country corridors. The corridors include 32 major remittance sending countries and 94 receiving countries. 9

In each corridor there is a dierent number of RSPs which oer various services. Besides the F ee and the F X margin the database further includes information on the name (Western Union, ING bank) and the type of the RSP (MTO, bank, post oce etc.) , a product (account-to-account, cash-to-cash, online service) and a transfer speed (from less than an hour to more than 6 days). This extensive database consists of close to 50,000 observations for the period 2011 Q1 to 2015 Q2 both for USD 200 and USD 500.10

About 10% of the RSPs within the some corridors are however considered not to be trans-parent as they do not disclose the FX fee. For the purposes of this thesis these observations were excluded from the analysis as the reported fees are likely to be systematically higher. This is also done by RPW (2015) when computing the Global average remittance price. In order to have a balanced panel, 24 country corridors have been excluded as the information is for example available only for the year 2014. Due to data availability of other variables, observations for the 2015 are excluded as well.

Although prices are measured quarterly or semi-annually, most of the explanatory variables are measured annually, at the best. The prices are therefore averaged to obtain annual data. This aggregation leads to a loss in the accuracy but does not bias the estimates.11 _{In the}

end, the total number of corridors analyzed is 201 (31 sending countries and 83 receiving countries) all for the years 2011-2014.

8_{The following questions were asked: (1) I would like to send $200 to my relative in [destination]. How}

much is the remittance cost? (2) When will the money reach the beneciary? (3) How many [local currency units] should I bring for $200 equivalent? (4) How many [local currency units] will my relative receive? (5) Where is your oce or agency located? (6) Can my relative pick up the money from a post oce or a bank with whom you you may have partnership? and others. (Ratha and Shaw, 2007)

9_{This does not capture all the ows as the theoretical total number of corridors can be dened as all}

possible combinations of states multiplied by two as the ow can go both direction. Or around 70,000 possible combinations. Many of these corridors would have however zero ows as for example probably no one sends money from Tonga to Mongolia.

10_{Data for year 2009 and 2010 are currently not publicly available}

11_{How to make optimal use of all available data when some variables are measured at a higher frequency}

than others was rstly introduced by Ghysels et al. (2004) who developed Mixed Data Sampling Models (MIDAS) for this purpose. MIDAS command is however not yet developed for panel data in Stata software. Due to technical limitations, prices across quarters are averaged to obtain annual data.

(21)

4.3.2 Price Variation within Corridors, Countries, RSPs and Products

This subsection aims to illustrate the importance of analyzing the price determinants not only on a corridor level (rather than at the sending or receiving country level) as argued by BP11 but also on the RSP level and the product level. Figure 3 below illustrates the variation in average prices across the 201 corridors, calculated across all surveyed remittance service providers in each corridor. The cheapest transfer is between USA and Honduras where as the most expensive is between South Africa and Malawi.

Figure 3: Average prices across 201 corridors, year 2014

*red line represents the global average

Source: Author`s own analysis based on data from RPW database

There is considerable heterogeneity in prices even within the same sending or remittance receiving country. Figure 4 below shows the variation in price when remittance is sent from the USA. As mentioned already USA-Honduras transfer is the cheapest (less than 2%) whereas the USA - Thailand transfer costs over 12% of the value sent.

(22)

Figure 4: Average Price of Remittances from the USA to 27 Receiving Countries

Figure 5 shows the price variation when Filipinos working abroad wish to send money home. The cheapest transfer is from United Arab Emirates and the most expensive is from Japan.

Figure 5: Average Price of Remittances from 13 sending countries to Philippines

There is also a great price variation even within a single corridor - among individual RSPs. The gure below depicts the price of various individual RSPs operating within Australia-Philippines corridor. GMT Money charges Filipino workers only USD6 (3%) for sending USD200 home. On the other hand, using the ANZ Bank for the same service costs close to USD40 (20%). Averaging the price across all these providers to obtain a single measure, as done by BP11, is likely a risky approach in terms of an information loss.

(23)

Figure 6: Prices by individual RSPs in the Australia-Philippines corridor, year 2014

The last level of the price variation lays within a single provider and a corridor. Each product has its characteristics and its price. Figure below show the variation in price for WU in the United Kingdom-Romania corridor. Less than one hour is more expensive than a next day service (or cash-to-cash is cheaper than credit card service). USD service is slightly more expensive than EUR service. This further strengthens the arguments against a wide use of the global average price.

Figure 7: Price variation across products

(24)

4.3.3 Price Variation across Price Components

Not only the price varies within corridors and RSPs, but further the two price components; the Fee and the FX margin and the two price values; USD200and USD500 vary. Figure 8 below depicts the distributions (kernel densities) of average prices within corridors. The distribution has indeed shifted to the left and thus the price decreased over the past 4 years. Fees over 15% are less and less common. The average price diers greatly when looking at the USD200 and USD500 value. Not surprisingly larger amounts are relatively cheaper to send. There is also a little dierence between the mean and the median price. When looking at the price components, the FX margin is independent of the value send (the two lines are identical) whereas the transaction fee varies greatly for dierent amounts. This suggests that all variation of the total price between the dierent amounts is explained only through the variation in the administration fee.

Figure 8: Comparison of average remittance prices by time, transaction value and price component

*densities are expressed in corridor terms as average prices not as individual RSPs

**The negative total cost does not mean that the RSP is suering a loss in the transaction in general, but rather that the RSP benets from a more favorable exchange rate or that it is a special promotion for the

point in time.

Visual inspection of the aggregate data suggests that MTOs are much cheaper than Banks, likely because they specialize in this service. It is important to note that this dierence is likely explained again only through the administration fee, which is a lot lower within MTOs while the FX margin is similar to both types of RSP with a slightly lower value for the Banks. The high fees on the Bank`s side can be explained through high overhead costs linked to a better paid sta with better benets, security and often a much more representative building in a better neighborhood. Banks also oer remittances only as a side product and get allotted a share of the overall overhead costs. For that reason, many large banks have oered few remittance products, as they see them as much less protable as large loans (Ratha and Riedberg, 2005). Low FX margin can be perhaps explained through larger overall transactions and better connections to the interbank foreign exchange market. 12

12_{Although not shown on the gure, it is also important to realize that the densities depict average price}

(25)

Figure 9: Comparison of average remittance prices by RSP type and price components

Source: Author`s own compilation based on RPW database

Against the intuition, faster service (delivery within a day) appears to be less costly than slower transactions (over 3 days). This is probably because MTOs oer faster services than banks or because online services are cheaper and faster. The most expensive is the use of the account-to-account channel. There is also a great variation in price between dierent sending and receiving regions. The most expensive are the intra-African corridors with a mean price around 20% of the value sent. Intra-Asian corridors density is shifted to the left, suggesting lower prices yet still above the global average. Migrants sending money from North America to Latin America can enjoy average prices close to 5%.

Figure 10: Comparison of average remittance prices by product type and region

5 Empirical Methodology

5.1 The Baseline Model

Following BP11 and FS08, and adding the time dimension, our baseline model is as follows:

there is only one transparent bank that enables transfer from the UK to Ghana. Sending USD200 through this bank and this corridor, cost USD 74 (37%) in 2012. Likely no-one uses this service, as there is over 10 MTOs operating in this corridor whose average price for sending USD 200 is USD20 or 10% only.

(26)

priceijt= β0+ β1corridorijt+ β2receivingjt+ β3sendingit+ εijt (5)

where i = 1, ..., 31 is an individual sending country and j = 1, ..., 83 represents the receiving countries, t = 2011, ...2014 is the time period. i x j = 201 and i x j x t = 804 is the number of observations. price is the price of remittances, corridor is a vector of corridor specic variables such as bilateral migrant stock, receiving and sending is a vector of receiving and sending country specic control variables such as GDP, and ε is the composite error term. The following section will describe the variables, methodology and estimation techniques.

5.2 Dependent Variable

The dependent variable, priceijt, is the average yearly price across all RSPs and products

within a given corridor. As mentioned already, important variable that is missing is the market share of each RSP per corridor. As argued by BP11 this problem is interpreted as a potential case of measurement error in the dependent variable, which should not bias the estimates but would aect their eciency. As a robustness check, median prices will be used besides mean prices as well.

As mentioned in the previous section, estimating an aggregate price can give only a lim-ited insights. An additional estimation will thus be done for MTOs and Banks separately. Moreover, this thesis will look at the price determinants of WU and MG only, because of their major position in the industry and often dierent price structures. Focusing on one nancial institution permits controlling for any bias arising from dierences in institutions across corridors.

Figure 11 below depicts the dierent price developments over time for various groups of RSPs. Whereas WU keeps its prices in the South Africa-Malawi corridor stable over time, MG is more volatile, so is the Banks' average. Aggregation of prices across groups of RSPs can thus lead to a severe loss in accuracy.

(27)

Figure 11: Development of average price over time by RSP type

An important issue to consider is the product heterogeneity. As mentioned above the prices dier systematically for online, mobile, cash or account products but as well as for speed of the transfer. Simple unweighted averaging of prices across RSPs assumes, that this is not true. To account for this heterogeneity, a separate analysis should be done for example for the price for cash-to-cash transactions that take less than an hour. In this way the provider type heterogeneity would not be however accounted for. To decrease the bias to minimum, this thesis will thus nally analyze the cash-to-cash that takes less-than-an-hour product from a single RSP view. 13 _{This is a major contribution compared to BP11 who looked only}

at the aggregate pricing.

The baseline estimation is done for the remittance value of USD200. The USD500 serves as a robustness check. As mentioned already, there is a great variation between the administration fees and FX margins charged by the MTOs and banks. This thesis will thus look at these price components separately.

Lastly, it is important to look at the price from a time series perspective. Question is whether the price follows autoregressive process, e.g. of order one, AR(1). In other words, whether price at time t depends on the price at time t − 1 and so if a lagged dependent variable should be included on the right hand side of the equation (5) to avoid omitted variables bias. AR(1)process can be described as:

pijt = ρpijt−1+ Xβ+εijt (6)

where X is a vector of explanatory variables and the rest is as in eq. (5).

Although the gure 2 above suggests that the Global average price follows a downward trend and that the variance of price is nite and constant, individual corridors exhibit mixed results. 13_{The less than an hour transfer is on average present in over 40% of the observations. Cash-to-cash is}

(28)

See gure 12 below. Thus it is not clear, whether a simple detrending will be sucient for obtaining a stationary series.

Figure 12: Development of price over time, corridor level

However, it is important to realize that the quarterly data is averaged to obtain yearly observations. This in most cases decreases the variance and thus adding a linear trend seems like a sucient solution for obtaining unbiased OLS estimates. The short run dynamics of the price uctuations are nevertheless not that important to the present estimation. Although the inclusion of lags of the dependent variable seems to provide an adequate characterization of many economic dynamic adjustment processes, in panel data analysis with a small number of time periods there often appear to be inference problems, such as small sample bias in coecient estimation and hypothesis testing. See Bun and Saradis (2013) for more detail. To deal with this issue, the series will be de-trended using a linear time trend with not lagged dependent variable.14

5.3 Explanatory variables

Explanatory variables are split into 4 groups. The rst group looks at the demand and supply factors. The demand is proxied by migrant stock and GDP whereas the supply factors include for example a competition proxied by the number of RSPs. The second group looks at variables representing nancial development such as bank account penetration. The third group discusses the factors inuencing costs incurred by the RSPs. The fourth group includes several control variables such as previous colonial relationship or distance between capital cities.

5.3.1 Demand and Supply

A natural candidate for an important variable inuencing the remittance price is the total volume of remittance ows in each corridor, or the demand. Figure 13 below depicts a supply-14_{Preliminary analysis using a GMM estimation shows that even after controlling for other factors, lagged}

price (one period) is signicant and positive only for WU and not for MG nor for MTOs in general. Adding a lagged dependent variable also signicantly decreases the sample making the estimates less precisely estimated as the standard errors become larger.

(29)

demand model where the vertical axis represents a price of a USD200 remittance transaction in corridor A. Horizontal axis represents the volume of remittances send within a corridor A per period. For simplicity it is assumed, that the only possible remittance that can be sent is a USD200. And that all other product characteristics are the same for all transactions. The volume or remittances thus equals the sum of the USD200 transactions. The downward and steep slope of the demand curve is supported by previous studies which nd a cost-elasticity of sending remittances to be lower than -1 (e.g. Gibson et al., 2006).

As mentioned already in section 4.2, the precise volume of bilateral remittances is not known. Instead they are estimated using bilateral migrant stocks, destination country incomes, and source country incomes. Figure 13 illustrates the eect of a positive bilateral migrant-stock shock on the remittance price in a corridor A. This increase in the bilateral migrant stock rstly shifts the demand curve D upwards to D`. With an increase in demand, both the market price and the equilibrium quantity rises to E`. The higher the price and the larger the quantity is achieved as each existing rm in the industry responds to the demand shock. The market, as such, moves along the original market supply curve from equilibrium point E to a higher equilibrium point E`.

However, the higher price leads to above-normal economic prot for existing RSPs. With a freedom of entry and exit, economic prot attracts new RSPs into the industry.15 _{An increase}

in the number of RSPs then causes the market supply curve S to shift to the right, to S`. How far this curve shifts and where it intersects the new demand curve, D', determines if the market is an increasing-cost, constant-cost or decreasing-cost industry. I argue for the latter option.

With more migrant workers, it is easier to attract new customers. The whole process is also supported by a positive shock of technological innovations which decrease the cost of sending remittances. The key point here is that the new equilibrium price E is lower than the original, E. Connecting E and E thus gives a downward sloping industry supply curve. Increase in migrant stock is thus expected to have a negative impact on the price.

15_{With many restrictions in place, the freedom of entry and exit assumption is questionable, but for this}

(30)

Figure 13: The eect of a positive bilateral migrant-stock shock on remittance price in a corridor A

Source: Author`s own compilation

Using the bilateral stock of migrants together with the receiving and sending country GDP, as opposed to using the bilateral remittance ows, has an advantage. As shown by Ahmed and Martínez-Zarzoso (2015), prices also inuence the remittance ows. Including bilateral remittance ows would thus lead to a simultaneity bias. In contrast, the stock of migrants, as argued by BP11, is likely to be independent of the prices. Migrant workers, when deciding where to work, are less likely to take the price of the remittances into consideration. Migrant stock is however not measured annually - data for the year 2010 and 2013 are available. For simplicity, a linear interpolation will be used to construct data points for the years 2011-2014. For details see Meijering (2002).

The share of total migrants to total population might be an important factor which shows the relative importance of the migrant group in the society and thus the economy. To control for the whole market size, the eect of total ow of remittances to all countries from a given country will be estimated. This is likely to be independent of the corridor price and serves as a proxy for the whole market potential. The more migrants (share or total number) or total ows, the lower price is expected.

Competition, in terms of a number of dierent RSPs per corridor, decreases mark-ups as standard micro-economic theory predicts. The competition among RSPs, computed as the number of rms interviewed, is taken from the RPW dataset. Likely, not all RSPs are reported, although the number is as complete as possible. For addition, banking sector competition index obtained from CEPII database will be used instead of the H-statistic for banking sector as used by BP11.

The total number of dierent RSPs interviewed is 528. The highest number of RSPs in receiving country is 131 (the Philippines) and in sending 78 (Australia). Some countries have less than 5 RSPs. Figure 14 shows the situation on the corridor level. Whereas Russia-Uzbekistan corridor is operated only by one RSP, Australia-Philippines market resembles perfect competition with over 25 RSPs.

(31)

Figure 14: RSP competition by corridor and country

It is important to look at the market power of each RSP, which can be from the RPW dataset proxied by the RSP presence in dierent corridors. Figure below shows all MTOs that operate at least in 15 corridors. MG and WU are by far the largest operators which operate in over 95% of corridors. The next largest are Ria and Coinstar which even together cover less than 80% of corridors. 16 _{This is supported by Watkins and Quattri (2014) who}

measured the market power by payout locations in African countries. They argue that the market conditions mostly resemble oligopolistic conditions, where the remittance providers charge high prices. As in any market, limited competition is a barrier to cost reduction and eciency gains.

Figure 15: MTOs by global presence in corridors

16_{Preliminary analysis on a rm level data also shows that with increased market share the price charged}

(32)

Based on the theory, some variables may not have a linear eect. For example number of RSP or migrant stock. An initial small increase in migrant stock will probably drive up the prices to E´. Subsequent inux of a large number of migrants will attract more RSPs and thus decrease the price to E´´ levels. A quadratic term will solve this non-linearity. A monopolistic provider will charge high prices but on the other hand on a perfectly competitive market RSP will not enjoy economies of scale. The following scatter plots show the nonlinear relationship of competition & price and log GDP per capita & price.

Figure 16: Nonlinear relationship of competition, GDP per capita and price

Source: Author`s analysis based on data described in the text

5.3.2 Financial Development

As mentioned above nancial inclusion measures, such as the percentage of adult population with a bank account, are considered as the key driving forces in lowering remittance prices. This data is taken from the recently updated Global FINDEX database. The more people have bank accounts, the cheaper it will be for the RSPs to deliver the amount to the nal client. FINDEX data is however available only for 2011 and 2014. Linear interpolation will be used to construct new data points. BP11 used data on the access to nancial services from 2003, while the remittance prices were dated to 2009.

The nancial literacy of remittance senders can also aect mark-ups as remitters can make better informed choices. Gibson et al. (2014) present a randomized experiment from New Zealand and Australia designed to measure the impact of providing nancial literacy training to migrants. Their results suggest that training appears to increase nancial knowledge and information-seeking behavior and reduces the risk of switching to costlier remittance products. Since nancial literacy of migrants is not recorded, a measure of the education level of migrants will be used as a proxy. For simplicity, a share of medium and highly educated (secondary + tertiary completed education) will be used. This data is available from 2006 and 2011. A simple linear interpolation is used again to construct necessary data points. This is again a major update to BP11, since they used migrants` education all the

(33)

way back from the year 2000. This almost 10 year lag is not discussed in their analysis at all.

The debt of nancial development can be also measured by the Domestic credit to private sector (% of GDP). FS08 also nd that, dollarization, plays an important role in determining the remittance prices. Lower exchange rate volatility should be associated with lower prices as it decreases the uncertainty. This nding is however not conrmed by BP11. Moreover, the dollarization likely inuences only the FX margin and will thus have a little eect on the total costs. In our sample, around 40% of receiving countries are dollarized. 17_.

5.3.3 Cost to RSPs

Thus far, the focused was only on the prices charged to migrant workers. Important question however also is what determines the costs incurred by the RSP when doing their business. Ratha and Riedberg (2005) discuss several factors including sta costs, technology, marketing, foreign exchange risk, location, and transaction costs.

Sta, location and marketing costs mostly increase with country GDP. Cost of technology on the other hand decreases with GDP. The eect of GDP is thus unclear. One easy to measure proxy of technological development can be the mobile cellular subscriptions. Transaction costs can be proxied by corruption levels as measured by Transparency International18_.

Fur-ther controls on the percentage of rural population is used as a more sparsely distributed population might be harder to reach, thus also raising transaction costs for providers. Just as the number of providers, GDP will likely have a non-linear eect as well. GDP itself raises the overall market development and thus should decrease the price. Very high GDP however increases the overhead costs to levels that decrease the benets of the eect of economic development on the price. 19

5.3.4 Controls

As discussed already, product and provider heterogeneity within each corridor are important issues to deal with. Ideally, a multidimensional panel regression would be used which would account for the rm, product as well as year and corridor xed eects. There is however a limited number of providers that are present in all corridors. Many operate only in a single corridor. Further not all corridors have all product types.

In addition, as in the majority of models where bilateral specic variables are used, or in the gravity models, distance between capital (or most populous) cities is used. The greater the 17_{(against USD, EUR, Indian rupee and African rand, pegged exchange rates are also included, as these}

include small risk for currency depreciation. This is also done by BP11)

18_{This thesis does not use the popular Transparency International Index as the country`s ranking itself}

would give limited inference. Instead the Corruption perception scores from Teorell et al. (2015) database are used. The score ranges from 0 (highly clean) to 10 (highly corrupt) and relates to perceptions of corruption as seen by business people or analysts ot the general public.

19_(log(x))2 _{is used instead of log(x}2₎_{, where x refers to GDP per capita or migrant stock. Including a}

(34)

distance, the more expensive the transfer is likely to be. Further, formal colonial relation-ship is often considered to play an important role, as it usually indicates common language and more intensive political and economic ties. Smaller geographic and linguistic distances might also foster the emergence of informal remittance service providers, adding competitive pressure to the formal remittance market.

Some variables report high bivariate correlations and thus are not included in the baseline estimations due to potential cause for multicollinearity bias. This is specically true for log of GDP per capita and share of rural population. The bivariate correlation is over -0.7 and -0.8 for destination and source country respectively, which suggests a strong negative relationship. BP11 nevertheless include these variables alongside each other. I argue that this is not a correct approach and that BP11 estimates might suer a collinearity bias. High correlations are also between s_credit and s_total_RSP (over 0.7). Table 2 reports the correlogram.

(35)

Table 1: Summary statistics of explanatory variables

Abbreviation Min Max Mean Obs. Source Explanatory variables - Supply and Demand

log of number of destination migrants in source country s_ln_mig 4 16 12 783 WB (2013) corridor migrants over total population in source country s_migpop 0 0.2 0.03 804 WDI log of total volume of remittances sent from source country (USD) s_lnREM_out 0.2 0.12 0.24 574 WDI log of total vol. of rem. received by destination country (USD) d_lnREM_in 10 24 21 792 WDI

number of RSPs within corridor nRSP 2 21 8 804 WB, RPW total Number of RSPs in source country s_total_RSP 2 78 31 804 WB, RPW total Number of RSPs in destination country d_total_RSP 2 131 29 804 WB, RPW banking sector competition index in source country1 _{s_bank_comp} ₀ ₄ ₃ ₇₂₀ _CEPII

banking sector competition index in destination country2 _{d_bank_comp} ₀ ₄ ₃ ₈₀₄ _CEPII

Explanatory variables - Financial Development

bank account holders (% of adults) in source country s_account 5 100 84 786 FINDEX bank account holders (% of adults) in destination country d_account 2 97 40 690 FINDEX

domestic credit to private sector (% of GDP) in source s_credit 12 195 115 741 WDI domestic credit to private sector (% of GDP) in destination d_credit 4 158 47 742 WDI dummy for pegged or dollarized exchange rate in destination3 _{d_dollar} ₀ ₁ _0.3 ₈₀₄ _IMF(2013)

share of educated corridor migrants over corridor migrants s_educ 0.16 0.84 0.6 334 OECD Explanatory variables - Costs to RSPs

corruption perception index in destination country4 _{d_corr} _1.7 _6.8 _3.5 ₇₅₆ _{Teorell et al. (2015)}

mobile cellular subscriptions (per 100people) in source s_mobile 55 194 121 804 WDI mobile cellular subscriptions (per 100people) in dest. d_mobile 15 185 94 802 WDI % of rural population in source s_rural 0 76 21 804 WDI % of rural population in destination d_rural 12 85 52 804 WDI

Explanatory variables - Controls

log of distance between capital cities (in km)5 _{ln_dist} ₆ ₁₀ ₈ ₇₇₂ _CEPII

dummy for ever in a colonial relationship6 _colony ₀ ₁ _0.2 ₇₇₂ _CEPII

log of GDP per capita in source country s_lnGDPppp 7 12 10 798 WDI log of GDP per capita in destination country d_lnGDPppp 6 10 8 796 WDI

*1,2,3,4,5,6 _{does not vary over time}

** WDI - World Development Indicators

***Source country is the country from which remittances are sent and which receives migrant workers. Destination country is a country that

(36)

Table 2: Correlogram d_credit 0.4311 0.2023 -0.2083 -0.1955 0.0976 -0.0055 0.0292 0.1741 -0.0321 0.3099 0.5397 0.2383 0.1103 1.0000 s_credit 0.1199 0.3776 -0.1602 -0.3560 0.0517 0.1959 0.7531 0.0857 -0.4592 0.0195 0.0977 0.6579 1.0000 s_account 0.2005 0.6286 -0.1521 -0.5486 -0.1113 0.1222 0.5192 0.1506 -0.2804 0.0293 0.2184 1.0000 d_account 0.6086 0.0787 -0.2686 -0.0423 0.0306 0.0021 0.1002 0.0329 -0.0379 0.2727 1.0000 d_mobile 0.6098 0.0835 -0.4747 -0.0984 0.1309 0.0794 -0.0176 -0.1193 0.0562 1.0000 s_mobile 0.0698 0.2627 0.1171 -0.2259 0.0919 -0.0933 -0.3313 0.0703 1.0000 d_total_RSP -0.0608 0.1681 0.2317 -0.1526 0.0709 0.1881 -0.0070 1.0000 s_total_RSP 0.0111 0.3231 0.0056 -0.2221 0.0686 0.2607 1.0000 nRSP -0.0505 -0.0818 0.0651 0.0682 0.1950 1.0000 s_ln_mig 0.2259 0.2053 -0.1394 -0.0643 1.0000 s_rural -0.2507 -0.8246 0.1952 1.0000 d_rural -0.7199 -0.1338 1.0000 s_lnGDPppp 0.2592 1.0000 d_lnGDPppp 1.0000 d_lnGD~p s_lnGD~p d_rural s_rural s_ln_mig nRSP s_tota~P d_tota~P s_mobile d_mobile d_acco~t s_acco~t s_credit d_credit

Source: Author`s analysis based on data described in the text

(37)

5.4 Methodology

A variety of empirical techniques are employed in the study. The model is rstly estimated using a pooled OLS as a benchmark with standard errors corrected for heteroscedasticity. However, given the panel nature of the dataset, the pooled OLS is only consistent when unobserved xed eect and explanatory variables are uncorrelated (Wooldridge, 2002). In order to take into account the resulting unobserved heterogeneity, a panel data approach i.e. xed and random eects rather than pooled OLS will be used. Hausman test is used to decide for Fixed or Random eect estimation. As proposed by Arellano (1987), standard errors will be clustered on the corridor level making them robust for any type of serial correlation within the times series of a country-pair. Time trend is also added.

The xed eect estimator, however, does not provide a direct estimation of the coecients of time invariant variables. One solution for this is to use the Mundlak approach (Mundlak, 1978) who proposed approximating the country specic eects as a function of the mean of time-variant variables. This is an alternative procedure to the xed eects model, which includes averages of time-varying explanatory variables (Wooldridge, 2002), instead of using dummy variables or the within transformation.

In order to check for the quality of the estimations, several post estimation tests are carried out. The calculation of bivariate correlations between the explanatory variables help us to identify collinearity between the explanatory variables. Variables that are highly correlated are used separately or are dropped from the regression.

As mentioned earlier, there are several dierent dependent variables. Firstly, the determi-nants for the remittance price of one single RSP will be estimated (separately for MG and WU) which control for provider heterogeneity. To control for product heterogeneity, only a certain product of a certain provider will be analyzed. Secondly, the same model will be applied to all MTOs and all banks. Lastly, the price is aggregated across all providers to see if this generalization as done by BP11 holds. As a robustness check, the same analysis is done for USD200 and USD500 remittance value. Thirdly, a separate analysis is done for a fee and FX margin. Results are also compared with those from BP11 and FS08.

6 Empirical Findings

6.1 Baseline Model - Pooled OLS, RE, FE, Mundlak

Table 3 below reports the baseline estimation for average remittance price across all products for one RSP only, specically the MG. This ensures provider homogeneity across corridors. These benchmark estimates provide results for the baseline model using several estimation methods. The rst column shows pooled OLS estimates, the second xed eects (FE) es-timates, the third column presents random eects (RE) eses-timates, and nally the fourth column presents results using the Mundlak approach (MU).

The majority of variables have consistent signs across the dierent estimation methods. Sur-prisingly colonial ties increase the price. Distance is positive yet not signicant. Stock of

(38)

bilateral migrants shows mixed results. Whereas pooled OLS and RE show signicant and negative estimates, FE and MU are positive. This is in line with FS08 but in contrast to BP11 who found negative eect. Destination GDP has a positive eect whereas source GDP has a negative eect. This is again in contrast to BP11 who found consistently positive co-ecients for both the sending and receiving country GDP. FS08 found insignicant results. The inconsistency of these results with previous ndings ask for further investigation. As mentioned earlier pooled OLS does not take into account the unobserved heterogeneity and is thus not an unbiased estimator. As expected, the Hausman test speaks in favor of FE estimator20_.

6.2 Additional Variables

To deal with the unclear eect of GDP and stock of migrants, quadratic terms as well as relative instead of absolute stock of migrants is used. In this section only FE estimates are presented. Column 1 corrects the previous estimation by computing robust standard errors (on corridor level). Time trend is also added and is strongly signicant.

Column 2 adds quadratic terms for GDP and the number of providers. The direction of the eect is the same for both the sending and the destination country. With an increasing GDP, the remittance price decreases as generally markets become mature and development takes place. From a certain threshold, very high GDP per capita results in high overhead costs and the price increases. Both terms are signicant.

The same holds for the competition in the sector. The competition has a negative eect on the price. More RSPs decreases markups. However from a certain point, too many RSP decreases the ability to enjoy economies of scale and the average price starts to slowly increase. The turning point is around 10 RSPs. In reality, the number is however likely higher as not all RSPs are interviewed by the RPW team. Ceteris paribus, a new entrant on the market with 5 already operating RSPs decreases the price on average by 0.2 percentage points.21_The

total number of RSPs in a country is negative but not signicant (column 4). This supports the importance of studying the price on a corridor level rather than on a country level only. Column 3 shows also interesting results. The relative demand, measured as the share of bilateral migrants to total population in a source country, proves to be more important than the absolute number of bilateral migrants in the country. The absolute number of migrants does not play a role - it is insignicant. Results do not change even when using only one of the measures separately. Also when adding the quadratic term of the absolute number of migrants, estimates are insignicant although they conrm the predictions that higher number of migrants rstly increases the price and only after a certain threshold the price starts to decrease. Holding other factors xed, 1 percentage point increase in the share of migrants to total population is on average associated with a 0.85 percentage point decrease in 20_{Hausman test was performed on estimated without clustered standard errors as STATA software does}

not support Hausman command with vce(cluster cvar) option

21_{If the estimated equation is written as y = β}

0+ β1x + β2x2 then the eect of change in x on y can

be approximated as 4y

4x ≈ β1+ 2β2x(Wooldridge, 2002). So with 5 RSPs on the market, the eect of an