Free development aid:
examining the relation
between grants, debt
relief and tax effort
Master Thesis Jan-Tjibbe Steeman Student 6306101 - Amsterdam School of Economics University of Amsterdam - Master program: MSc Economics - Specialisation: International Economics and Globalisation - Supervisor: dr. M. Micevska Scharf Second reader: dr. D.J.M. Veestraeten February 9, 2016
2
Statement of Originality
This document is written by Student Jan-Tjibbe Steeman 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.
3
Free development aid: examining the relation between
grants, debt relief and tax effort.
Jan-Tjibbe SteemanAmsterdam School of Economics, University of Amsterdam (UvA), The Netherlands February 9, 2016
Abstract.
This paper examines the effects of IDA grants and debt relief on tax effort, for a panel of IDA-eligible countries. In addition, this paper studies whether effects differ for HIPC and non-HIPC countries. A VAR model is constructed to represent the fiscal response model by Heller (1975). The panel VAR is estimated with the PVAR package for Stata, made available by Abrigo and Love (2015). Based on the VAR estimations, granger causality tests and impulse-response functions this paper argues that the results for the aggregate panel do not fully suffice, due to the opposing effects of the HIPC and non-HIPC countries within this panel. For the non-HIPC countries, a negative effect is found for IDA grants on tax effort and a positive effect for debt relief on tax effort. For the non-HIPC countries, a positive effect is found for IDA grants on tax effort and a negative effect for debt relief on tax effort.
4
Table of contents
Introduction ... 5
1. Literature review... 7
1.1. Transitions in multilateral development aid ... 7
1.1.1. Debt relief and debt overhang theory. ... 7
1.1.2. IDA pure grants ... 10
1.2. Aid effectiveness and conditionality ... 11
1.3. Development aid and the fiscal behaviour of the recipient country ... 13
1.3.1. The fiscal response model ... 13
1.3.2. Fiscal space ... 14
1.3.3. Fiscal impact of aid instruments ... 15
1.4. Empirical literature ... 16 2. Data description ... 18 2.1. Data sources ... 18 2.2. Variable description ... 19 2.3. Variable statistics ... 20 3. Empirical analysis ... 22 3.1. Hypotheses... 22 3.2. Empirical framework ... 23
3.2.1. Panel VAR and VAR estimation ... 23
3.2.2. Granger causality and impulse-response functions ... 25
3.2.3. Variable stationarity and VAR stability ... 26
3.2.4. Model order selection ... 28
4. Results ... 29
4.1. Aggregate panel ... 29
4.2. HIPC panel ... 34
4.3. Non-HIPC panel ... 38
4.4. Comparing models ... 41
5. Discussion and limitations ... 43
5.1. Discussion ... 43
5.2. Limitations ... 46
6. Conclusion ... 47
References ... 50
Appendices ... 54
Appendix A. Country information ... 54
Appendix B. Variable statistics ... 56
Appendix C. Variable stationarity, VAR stability and model order selection ... 58
Appendix D. Aggregate panel estimates ... 61
Appendix E. HIPC panel estimates ... 65
5
Introduction
On September 25th 2015, the United Nations (UN) agreed on the Sustainable
Development Goals (SDGs). SDG, a development agenda that, among other goals, aims at eradicating poverty by 2030. After adopting this agenda, Secretary-General of the UN, Ban Ki-moon (2015), declared; “we owe this and much more to the vulnerable, the oppressed, the displaced and the forgotten people in our world”. Nevertheless, critics remark that the SDGs are merely an extended deadline of a previous set of goals. In 2000 the UN adopted the Millennium Declaration. Within this declaration the member states and development institutions of the UN obliged themselves to several global development goals, to end world poverty by 2015. Later, these goals became known as the Millennium Development Goals (MDGs) (United Nations, 2015). In conjunction with the implementation of the MDGs, two changes emerged within multilateral development aid. First, an increased emphasis was put on debt reduction, through the Heavily Indebted Poor Countries (HIPC) and Multilateral Debt Relief Initiative (MDRI) initiatives (Cassimon and Van Campenhout, 2008). Second, the World Bank’s International Development Association (IDA) partially switched from loans to grants (Radelet, 2005). However, due to the necessity of a new agenda on global poverty by 2015, one can question the effectiveness of the first.
Aid effectiveness has been subject to a long theoretical and empirical debate. Even the definition of aid effectiveness is unsettled. At first, studies regarding aid effectiveness focused on whether aid had a positive impact on economic growth. However, in recent years, aid is also perceived effective when it is poverty reducing. Furthermore, researchers and policymakers have started to realize that aid effectiveness largely depends on the fiscal reaction of the aid receiving governments. Studies investigating the fiscal behaviour within countries can roughly be divided into two groups. First, fungibility studies assess the relation between aid and the sectoral composition of government spending. Second, fiscal response studies focus on the interdependency between foreign aid and government’s fiscal behaviour (Osei, Morrissey and Lloyd, 2005).
The fiscal response studies focus mostly on the relation between different aid instruments and government taxation. These studies examined whether there is a different impact between grants and loans. Gupta, Clements, Pivovasrky and Tiongson
6
(2003); Odedokun (2003); Morrissey, Islei, and M’Amanja (2007) and Clist and Morrisey (2011) all find evidence indicating that loans have a positive effect on tax revenues, contrary to grants, which have a negative effect on tax revenues. However, the impacts differ when studies made a disaggregation for low and middle income countries, for different time periods or for short- and long run effects. In the last 15 years, debt relief has emerged as an aid instrument. Depetris Chauvin and Kraay (2005) find no relation between debt relief and government domestic revenues. Cassimon and Van Campenhout (2007, 2008) and Cassimon, Ferry, Raffinot and Van Campenhout (2013) studied the effectiveness of the debt relief initiatives. Mostly, evidence is found for a positive effect on domestic revenue. However, Cassimon et al. (2013) recommend to study the impact on tax revenues, instead of total domestic revenues. Moreover, Cassimon and Van Campenhout (2007) write that results of the two debt relief initiatives can hardly be generalized to other non-HIPC countries.
With the amplified emphasis for grants and debt relief, the international community tries to support low-income countries to end poverty and foster economic growth. However, the fiscal behaviour of the recipient country can hinder this aim. It is therefore important to compare the effectiveness of aid instruments. This study augments to this and studies the following research question: What is the effect of IDA
grants and debt relief on the tax effort within the recipient country? And does the effect differ for HIPC and non-HIPC countries?
To answer this research question, annual data for a balanced panel of IDA eligible countries is used. In order to construct the fiscal response model, two databases are required. First, a dataset of fiscal variables is manually constructed from IMF country reports. Second, a dataset of development aid variables is constructed from the OECD DAC and World Bank databases. The variables are potentially dynamically interlinked. Therefore, estimations are performed within a vector autoregressive model (VAR) framework. Moreover, this VAR framework provides the possibility to create impulse-response functions (IRFs). These show the intertemporal relation between aid instruments and tax revenues. The panel VAR is estimated with the use of the PVAR package for Stata, made available by Abrigo and Love (2015), based on the Generalized Method of Moments (GMM) framework.
The rest of this paper is structured as follows. Chapter 1 functions as literature review. Herein the transitions in development aid, the aid effectiveness debate and the
7
theoretical and empirical studies regarding the relation between aid and fiscal behaviour are described. With this knowledge, chapter 2 provides a description of the data sources and variables. In chapter 3 the hypotheses are presented, which are based on the previous theoretical and empirical literature. After this, the empirical framework is presented in order to answer these hypotheses. Here, different panel VAR techniques are described, time series tested for stationarity and the VAR analysed for stability. Chapter 4 presents the results of the VAR estimation. With the use of Granger causality tests and IRFs a coherent answer is created. In chapter 5 the results are discussed. Finally, chapter 6, presents the main conclusions and directions for forthcoming research.
1. Literature review
To answer the research question, several essential topics are addressed in the literature review. First, recent transitions in multilateral development aid are discussed. Second, the concepts of aid effectiveness and conditionality are elucidated. Third, theoretical literature regarding the relation between development aid and government fiscal behaviour is studied. Fourth, empirical literature studying the relation between aid instruments and government taxing behaviour is extensively studied.
1.1. Transitions in multilateral development aid
By the end of the 20th century, the international community increased its effort in
fighting world poverty to help low-income countries (LICs) achieving the MDGs. To accomplish this, the multilateral development institutions agreed on two changes. First, to help developing countries reducing their debt burdens, debt relief programs were implemented. Second, to maintain the lower debt levels, the World Bank decided to fund the poorest countries with pure grants. Both changes are described in the following sections.
1.1.1. Debt relief and debt overhang theory.
In the second half of the 20th century, multilateral development institutions primarily
used concessional loans to support LICs. These loans are a form of development aid, because they partly consist of a grant element. The grants element arises through the
8
below-market interest rates and long grace periods. However, bilateral donors provided loans on non-concessional terms as well. These, sometimes unfavourable, bilateral loan conditions supported by financial shocks and the lack of macroeconomic reforms created debt distress within the LICs. To help these LICs, multilateral institutions, donor countries and recipient countries agreed on several temporary rescheduling agreements (Arnone and Presbitero, 2010). Gueye, Vaugeois, Martin and Johnson (2007) argue that these rescheduling agreements were adopted because donor countries identified it as a short-run payment issue. However, Arnone and Presbitero (2010) argue that donor countries tried to limit moral hazard by the aid receiving governments. Though, by the end of the 80s, creditors realized that the situation was not solely a short-run problem. Therefore, additional steps were necessarily. These steps included several debt cancellation mechanisms installed by the Paris Club (Arnone and Presbitero, 2010). The Paris Club is an organization of creditor countries, which aims at coordinating debt rescheduling and debt cancellation agreements with debtor countries (Weiss, 2013). However, a more comprehensive approach was desired. Moreover, the multilateral development institutions experienced an augmented political pressure to increase their effort due to a growing multilateral debt stock within the LICs. In answer to this, the HIPC initiative was founded in 1996 by the World Bank and the International Monetary Fund (IMF). This initiative combines both multilateral and bilateral debt reduction and has two targets: long-term debt sustainability and poverty reduction (Arnone and Presbitero, 2010).
Cassimon et al. (2013) argue that cancelling debt to sustainable levels was not solely to acknowledge the fact that countries would not repay their debt, it was also an answer to the debt overhang problem. The debt overhang theory, originated by Meyers (1977), is studied by Krugman (1988) and Sachs (1989) regarding middle income countries. They conclude that debt overhang arises when a country’s stock of external debt is exceeding a country’s ability to pay the present value of future debt payments. This burden has a negative impact on domestic and foreign investment. Koeda (2008) studied the debt overhang problem for LICs, which mostly acquire concessional lending. He concludes that LICs have an incentive to use parts of their concessional lending for consumption, instead of using it for profitable investment. Overall, debt overhang theory predicts a lower than optimal investment rate and thus a lower than optimal growth. Cordella, Ricci and Ruiz-Arran (2010) did empirical research on the debt
9
overhang theory. They find no negative effect of high indebtedness on growth and investment, when a country has insufficient institutions. This is due to their already present marginal creditworthiness. Therefore, debt relief should go hand in hand with institutional improvements within highly indebted countries.
To increase the effectiveness of the HIPC initiative, the international community implemented an augmented version in 1999. This HIPC program works in a two-step process. When a country wants to be considered for the initiative, it has to meet four criteria: be part of the IDA, have unsustainable debt, have a track record of policy programs within the IMF and World Bank and write a Poverty Reduction Strategy Paper (PRSP) (IMF Factsheet HIPC, 2014). After completing these conditions, known as the decision point, the country can partly receive debt relief. As a condition, the country has to improve their policies, hold their good performances and implement their PRSP if they wish to reach completion point. After the completion point, the country gets full debt relief agreed on during the decision point (IMF Factsheet HIPC, 2014). Arnone and Presbitero (2010) point out that, despite the positive results of the HIPC initiative, a demand for further debt relief was desired by the international community. Therefore, in July 2005, the G8 proposed full debt relief for post-HIPC countries. The objective was to support countries achieving their MDGs, by 2015. This Multilateral Debt Relief Initiative (MDRI) provides full relief of debt owed by HIPC countries to the African Development Fund (ADF), World Bank’s IDA and IMF (IMF Factsheet MDRI, 2014).
Of the 39 countries that were eligible for the HIPC initiative, 36 countries have reached completion point by November 2015. Three countries are still in their qualifying process. Figure 1.1 on the next page, is taken from the most recent HIPC and MDRI statistical update (2014). It illustrates the effectiveness of the HIPC initiative. Within the figure, the HIPCs’ average poverty reducing expenditures and average debt payments are shown. The debt relief released at the HIPC completion point generates a positive impact on poverty reducing expenditures and a negative effect on debt service payments. Furthermore, the statistical update concludes that debt relief provided under the two initiatives was mostly delivered by multilateral development institutions. They disbursed US $ 75 billion of HIPC and US $ 41 billion of MDRI debt relief, in end-2013 present value (PV) terms. For (Paris Club) bilateral creditors this was more than US $ 21.5 billion for HIPC debt relief, in end-2013 PV terms. And for the non-Paris Club bilateral creditors this was more than US $ 4.9 billion HIPC debt relief, in end-2013 PV
10
terms. Despite these debt relief figures, positive effects have been moderated. Recent data shows that debt service payments will grow in the near future. Moreover, by the end of 2014, most HIPC countries were behind schedule in achieving the MDGs (HIPC and MDRI initiative statistical update, 2014).
Besides the HIPC countries, other low and middle income countries received debt relief or debt rescheduling as well, mainly through the Paris Club. However, studies from Johansson (2010) and Depetris Chauvin and Kraay (2005) show that most debt relief goes to the HIPC countries.
Figure 1.1: Fiscal effects from the HIPC initiative
The figure below illustrates the effect of debt relief on poverty-reducing expenditures and debt service payments of 36 post-decision-point HIPC countries (in % of GDP and with t = completion point).
Source: HIPC and MDRI statistical update (2014).
1.1.2. IDA pure grants
The World Bank operates through two main windows; the International Bank for Reconstruction and Development (IBRD), founded in 1946 and the International Development Association (IDA), founded in 1960. Until the millennium, IDA provided solely concessional loans to LICs. However, due to increasing debt burdens within these LICs, the World Bank decided in 2002 that 20% of their funds should be distributed as pure grant (Radelet, 2005). Radelet (2005) argues that these grants were initially meant for post-conflict reconstruction, natural disasters, HIV/AIDS, education, health, water and sanitation. However, in 2005 the World Bank changed the allocation. Currently, the allocation is by means of the country’s debt risk analysis: the debt sustainability framework. Thus, the IDA lending terms are in line with the possibility of debt distress.
11
When a county has a high risk of debt distress, they get all financial assistance in form of grants. When they have a medium risk of debt distress, the financial assistance is for 50 percent in the form of grants. All other countries receive their financial assistance on regular IDA credit terms, which are concessional loans (IDA lending terms, 2015). Countries, eligible for IDA funding, are obliged to write a Poverty Reduction Strategy Paper (PRSP). They describe the country’s actions to fight domestic poverty, with the objective to increase aid effectiveness (Dijkstra, 2011). In the next section this concept is discussed in more detail.
1.2. Aid effectiveness and conditionality
Development aid is an important source of income for LICs. Svensson (2000) concludes that between 1975-1995, for the 50 most aid dependent countries in the world, aid revenues as share of government expenditures were approximately 54 percent. Moreover, in 2014 the global community provided around US $ 135 billion of development assistance, with US $ 60 billion coming from the World Bank group. Nevertheless, by the end of 2015, more than one billion people are still living in poverty and most MDGs are not met (World Bank, 2015). These opposing facts, give rise to the question whether development aid is effective (Svensson, 2000).
Morrissey (2004) argues that aid effectiveness is generally measured by its effect on economic growth. However, quantifying this effect is difficult because regressions fall short in describing all variables effecting economic growth. The empirical literature regarding the relation of aid and economic growth is inconclusive. Mallik (2008) concludes that there is a positive long-run relationship between aid and growth for a panel of African countries. However, Burke and Ahmadi-Esfahani (2006) find no evidence for this within a panel of Asian countries. Burnside and Dollar (2000) conclude that aid has a positive impact on growth when countries possess good institutions. Contrary to the conclusion of Hansen and Tarp (2001), they find a positive impact on growth, unconditional of the countries’ institutions. Nevertheless, Cassimon and Van Campenhout (2007) argue that government behaviour is important. They state that aid effectiveness will largely depend on the succeeding government’s taxation and expenditures. Studying this fiscal behaviour is thus a more indirect way of analysing aid effectiveness.
12
In response to the aid effectiveness debate, aid disbursements were linked to policy reforms and fiscal outcomes. Morrissey (2004) writes that this ‘conditionality’ covers several areas, such as fiscal and monetary policies and trade liberalization. First, solely the IMF used the instrument. Later, the World Bank and other donors related their aid flows to domestic reforms within the recipient country as well (Montinola, 2010). However, Dijkstra (2011) writes that conditionality was one of the problems in the donor-recipient country relationship. She states that conditionality is ineffective because recipient countries only implement their planned activities. Another problem in the donor-recipient country relationship was the lack of donor coordination. The lack of coordination created institutional fragmentation and an overload of donor administration units. To target these problems the IMF and World Bank introduced the PRSPs (Dijkstra, 2011).
Since 1999, countries pursuing HIPC initiative assistance have to write a PRSP. The PRSP has to be finished before the decision point of the initiative. Later, this framework became dominant in the aid allocation of the World Bank IDA, the low-income facility of the IMF and some of the bilateral donors too (Dijkstra, 2011). The goal of this framework is to gather support of development partners, set clear actions at country level, and bring forth effective and sustainable poverty reduction (IMF and IDA, 1999). Writing the PRSP can be a time consuming activity. However, to make sure aid programs are not delayed, countries can choose to write an interim PRSP. This is a lean version of the normal PRSP, but provides the possibility of getting partial development assistance (Klugman, 2002). Klugman (2002) writes that, contrary to conditionality, PRSPs are more open to the taste of the recipient country. However, the World Bank and IMF still assess whether the PRSP is a suitable path to poverty reduction. The PRSPs were supposed to increase country ownership and decrease conditionality. However, Dijkstra (2011) highlights a number of inconsistencies regarding the PRSP process. In fact, conditionality increases when all countries that want to qualify for development assistance are required to write a PRSP. Moreover, Dijkstra argues that the PRSP can be seen as a shift from content to process conditionality. Therefore, Dijkstra concludes that country ownership did not increase, keeping aid effectiveness unchanged.
13
1.3. Development aid and the fiscal behaviour of the recipient country
This section describes the relationship between development aid and fiscal behaviour within the recipient country. First, the fiscal response model described by Heller (1975) is analysed. This model can be seen as the origin of the fiscal response literature. After this, the concept of fiscal space is described. This section ends with a theoretical explanation on the potential fiscal responses of the recipient government after receiving development aid.
1.3.1. The fiscal response model
In the empirical literature, the relation between development aid and the fiscal behaviour within the recipient country is an ongoing debate. Previous studies can be divided into two groups. First, fungibility studies focus on the sectoral composition of government spending after receiving aid. Aid is fungible when a donor country gives aid for a certain purpose, but the recipient country uses it for another purpose (Osei et al., 2005). Second, fiscal response studies focus on the interdependency between foreign aid and the succeeding government’s fiscal behaviour. Osei et al. (2005) argue that these studies’ analyses go further because they try to incorporate other fiscal dynamics within the recipient country. These fiscal response studies use the fiscal response model (FRM), created by Heller (1975). The model describes the fiscal behaviour of a country, with the use of a maximization problem of the government budget constraint. Within the FRM of Heller, the following utility function is assumed for every period:
𝑼𝑼 = 𝑭𝑭[𝑰𝑰, (𝒀𝒀 − 𝑻𝑻), 𝑪𝑪, 𝑩𝑩, 𝑮𝑮, 𝑳𝑳]
With 𝑰𝑰 being the government investments, (𝒀𝒀 − 𝑻𝑻) is the gross domestic product 𝒀𝒀 minus tax revenues 𝑻𝑻, government consumption is presented as 𝑪𝑪, 𝑩𝑩 is government borrowing from domestic sources, 𝑮𝑮 are the foreign grants to the government and 𝑳𝑳 are the foreign loans to the government.
Heller further divides the government consumption, however this is not vital. The expenditure side consists of government consumption. This includes expenditures with a current character. Government investments are capital expenditures, aiming to create economic growth. The revenue side consists of tax revenues, government
14
domestic borrowing and foreign aid in the form of grants and loans. The function is a budgetary constraint and can therefore be represented as:
𝑰𝑰 + 𝑪𝑪 = 𝑻𝑻 + 𝑩𝑩 + 𝑮𝑮 + 𝑳𝑳
The left-hand side, expenditures, are equal to the right-hand side, revenues. Heller assumes that aid variables are exogenous to the public sector. He argues that the public sector cannot influence the aid flow. Contrary to this is the study from Franco-Rodriguez, Morrissey and McGillivray (1998). In practice, the donor country commits an aid flow to the recipient country. The recipient country then decides how much is actually disbursed. They conclude that the recipient country chooses their own amount of aid and thus classify aid as endogenous. In their FRM, aid is treated like the other revenue categories.
Osei et al. (2005) describe some limitations of the FRM. They argue that the fiscal behaviour can differ between countries. Therefore, case studies are preferred. Moreover, FRMs should be seen as an illustration of how the fiscal variables react to aid instead of providing causal interpretation. Another limitation of these early FRMs is that they do not incorporate debt relief. To understand how debt relief fits in, the concept of fiscal space is described in the next section.
1.3.2. Fiscal space
Heller (2005) argues that for LICs, fiscal space is essential to increase poverty reducing expenditures and thus achieving the MDGs. Fiscal space can be defined as the budgetary room that is available to increase resources, without affecting the sustainability of the government’s fiscal position. Subsequently, these resources can be used to increase government spending. However, they can also be used to reduce taxation. Heller summarizes three ways in which a LIC can create fiscal space. First, increasing domestic revenues can create fiscal space. Several LICs have low tax ratios. Therefore, countries could increase their efforts in augmenting these ratios. Second, decreasing unproductive expenditures can create fiscal space. Mostly recurrent expenditures, such as spending on defences, domestic security and civil services should be evaluated. Third, increasing the grant provision can create fiscal space. Heller argues that creating fiscal space through borrowing (domestic or external) or by printing money is not recommendable.
15
Borrowing temporarily increases the money available for other purposes. However, it increases debt obligations as well. Therefore it does not improve the financial position. Printing money increases uncertainty of monetary policy and increases the possibility of hyperinflation. Heller (2006) argues that debt relief is also a measure in creating fiscal space. Resources, previously used for debt payments, are now available for other purposes. However, Cassimon and Van Campenhout (2007) point out that this is only true when the country was intending to pay the debt. Otherwise, debt relief is solely a clearance of debt stock and does not provide additional resources.
1.3.3. Fiscal impact of aid instruments
Aid instruments can generate different fiscal behaviours of the recipient country’s government. Clements, Gupta, Pivovarsky and Tiongson (2004) argue that it can reduce tax revenues, increase current and investment expenditures or reduce domestic borrowing, or a combination of the three. They describe these scenarios as follows. First, aid flows are transferred to the private sector by reducing tax revenues. Under the most extreme scenario, tax revenues are reduced by the full amount of aid inflow. Second, aid flows are used to increase expenditures. When the full amount is used, with no further increase in expenditures, tax revenues are not affected. In the third scenario the government decides to use the aid flow to reduce domestic borrowing. Therefore, tax revenues and expenditures are not affected.
The central question is whether the composition of aid matters for fiscal behaviour in the recipient country. Morrissey et al. (2007) argue that this discussion is regarding the different degrees of concessionality. Concessional loans, normally used for development aid, have a grant element of at least 25 percent. Grants are 100 percent concessional. Since grants do not have to be repaid, they possibly have a negative effect on fiscal behaviour, contrary to concessional loans. Loans have to be repaid and therefore probably more effectively used (Cohen, Jacquet, and Reisen, 2007). However, Radelet (2005) writes that LICs have a bad track record, with respect to generating growth. Expecting the payback of these loans is therefore naïve and will consequently increase the debt stock of LICs. Grants are therefore preferred.
16
1.4. Empirical literature
The question central in previous empirical literature is whether grants give the recipient country an incentive to reduce tax effort. This incentive arises because the recipient country does not have to repay the grant. Two different empirical frameworks are used in answering this question. First, studies modelling tax effort by means of variables that determine the tax base structure of the economy. Second, studies that use the fiscal response model of Heller (1975). The first models tax effort by means of variables that determine the tax base structure of the economy. This method originates from Lotz and Morss (1967) and is optimized by succeeding studies. Variables commonly used in this model are openness to trade, agricultural and industrial value added and the per capita income. Subsequently, the grant and loan variables are added to the model to estimate their effect on tax effort. The econometric technique used is OLS. For panel data, fixed effects regressions are performed to control for country specific differences.
Within this framework, Gupta et al. (2003) study the effect for a panel of developing countries during the period 1970-1990. They conclude that grants negatively affect tax effort. Moreover, concessional loans are positive related to tax effort. However, Gupta et al. (2003) use total domestic revenues in representing tax effort. Morrissey et al. (2007) question their results. They argue that Gupta et al. (2003) solely study the short run relationship. To allow for medium term effects, they take five years averages as well. With tax revenues as dependent variable they conclude, for the short run, in line with Gupta et al. (2003). However, regarding the medium term relation, no effect of grants on tax effort is found. Clist and Morrisey (2011) study whether this short run relation between aid and tax effort changed between 1970-2005. In line with previous studies, a positive relation between tax effort and loans, and a negative one for grants is found. However, in the period 1985-2005 the relation between grants and tax effort becomes positive. They contribute this to the increased World Bank conditionality, started in the 80s.
The other stream of studies uses the fiscal response model of Heller (1975). With the use of this model, reduced-form equations are derived. Subsequently, the three-stage least squares (3SLS) technique is used to estimate the fiscal impacts of loans and grants. Odedokun (2003) uses such an approach to study the relation between grants
17
and tax effort. He separates the data between low and not so low-income countries. In line with Gupta et al. (2003) and Morrissey et al. (2007), Odedokun (2003) concludes that grants have a negative effect on tax effort within LICs. However, they have a positive effect in the not so low-income countries. One possible explanation for the different effects is found in the study of Teera and Hudson (2004). In their study, they conclude that LICs have generally lower tax revenues ratios and more difficulties increasing these.
Since debt relief programs increased in popularity, studies started to incorporate these figures. Depetris Chauvin and Kraay (2005) do not find a relation between debt relief and tax revenues. However, their empirical framework is weak by taking averages of five years and estimating correlations. Cassimon and Van Campenhout (2007, 2008) study the fiscals impacts of the HIPC debt relief initiatives. Both studies find evidence for a positive effect of debt relief on domestic revenues. However, Cassimon and Van Campenhout (2007) write that these results can hardly be generalized to other non-HIPC countries. Moreover, Cassimon et al. (2013) augment to these studies, by incorporating debt relief from the MDRI initiative. Although their results are less convincing, mainly a positive influence from debt relief on domestic revenues is found. However, they recommend to increase fiscal behaviour precision by studying tax revenues, instead of total domestic revenues. An important feature of the latter three studies is that they use a different econometric technique to estimate the fiscal impact. Instead of using the 3SLS technique, they use a vector autoregressive model (VAR). MacGillivray and Morrissey (2001) write that the 3SLS method is problematic for short time series and that studies using this approach are questionable. Cassimon and Van Campenhout (2007) agree and argue that 3SLS is highly sensitive to starting values and therefore not preferred. Moreover, Cassimon and Van Campenhout (2007, 2008) and Osei et al. (2005) claim that the fiscal response literature fit the characteristics to adopt a VAR approach. First, although different exogenous factors are influencing the fiscal aggregates, variables must stay in relation with each other. A VAR approach will display this relation. Second, a VAR takes into account the different dynamic effects between fiscal variables. Third, these dynamic relations can be graphed with the help of impulse-response functions. These functions give the possibility to investigate what the impulse-response is in one variable, after a shock in another.
18
In this chapter, the literature regarding multilateral debt relief programs and the partly transition from loans to grants by the World Bank IDA were discussed. In addition, the theoretical and empirical studies concerning the fiscal impact of different aid instruments were explained. The next chapter presents the data required in the fiscal response literature to establish the empirical framework.
2. Data description
This chapter provides a verbal and statistical description of the obtained data. First, the source from where the different variable time series are collected is specified. Second, a description of all variables is presented. Third, a statistical depiction is provided. Herein some assumptions are made and transformations performed in order to realize an effective dataset for the empirical analysis.
2.1. Data sources
This study uses annual panel data from 1999 until 2012. To answer the research question, two main datasets are required. First, in line with Cassimon and Van Campenhout (2007, 2008) and Cassimon et al. (2013), a dataset of fiscal variables is manually constructed from IMF Article IV and IMF country reports. A broader time span is preferred. However, these reports generally go back to the year 1999 and solely for a handful of countries recent data years are available. This dataset includes tax revenues (T), government current expenditures (C), government investments (I) and domestic financing (B). Most IMF reports provide these accounts for the central government operations. However, occasionally the general government operations are presented. In general, this is not a problem as long as the same government body is used for a particular country for the entire period. The second dataset contains development aid variables. The IDA grants (G*) are taken from the World Banks’ World Development Indicators database. The remaining grants (G), total loans (L) and total debt relief (D) are from the OECD DAC database.
From the 80 countries that were eligible of receiving IDA grants or HIPC debt relief in the studied period, only 62 actually received one of these two aid instruments. From these 62 countries, 42 are included in the balanced dataset. This makes the aggregate panel dataset N=42 (countries) and T=14 (time periods). There are two reasons for
19
excluding countries. First, missing data due to missing country reports of the IMF. This occurred eleven times. Second, missing data due to the irregular data presentation within the IMF reports. When a country is a significant oil exporter, tax revenues are generally represented in such a way that it is incomparable across the period and to other countries. This occurred nine times. In total 23 HIPC countries (N=23 and T=14) and nineteen non-HIPC countries (N=19 and T=14) are included. Table 2.1 in Appendix
A provides an overview of all countries, whether they are included in the balanced
dataset and the reason of exclusion.
Also presented in Table 2.1 are the fiscal years of the countries included in the balanced dataset. For twelve countries the fiscal year deviates from the calendar year. It is suboptimal to convert fiscal data to calendar year format because governments use a yearly budgetary cycle process. However, development aid data are in calendar year format. Thus, both ways lead to a suboptimal result. The amount of countries with deviating fiscal years is limited and it is preferred to have all variables in the same time span. Therefore, fiscal data of countries with deviating fiscal years are converted to calendar year format. This is manually completed by multiplying the fiscal data with the amount of days that are within that particular calendar year.
2.2. Variable description
In Table 2.2 of Appendix B, an overview and description of all variables is provided. All variables are in percentages of GDP. The variables of the fiscal dataset are provided in this format in IMF country reports. Aid flows are converted to percentages of GDP with the use of the countries’ GDP numbers. The GDP numbers are provided by the World Bank World Development Indicators database. Using variables in percentages of GDP has two advantages. First, it controls for the size of the economy. Second, it controls for omitted factors influencing all countries and variables in the same manner (Cassimon and Van Campenhout, 2007).
To measure tax effort, the tax revenues variable is used. These include all government taxes. This generally contains taxes on ‘income and profits’, ‘domestic goods and services’ and ‘international trade’. The government current expenditures variable is net of interest payments. For government investments, the government capital expenditures account is used. This includes net lending. The fourth variable of the fiscal database is domestic financing. Contrary to previous studies (Osei et al., 2005;
20
Cassimon and Van Campenhout, 2007, 2008; Cassimon et al., 2013), net domestic financing is used, instead of net domestic borrowing. This has two motives. First, data on net domestic financing are better and more reliable displayed in the IMF country reports. Second, it mirrors a more complete domestic government funding.
Data of development aid are provided for disbursements and commitments. However, aid flows, required for this study, are solely presented as disbursements. Fortunately, disbursements are more suitable to use because these depict the real leverage on the budget of the recipient country. The IDA grants variable comprises net disbursements of grants provided by the IDA. Multilateral debt relief is mostly provided as grant (Cassimon and Van Campenhout, 2007). Therefore, in order to acquire the
remaining grants variable without the debt relief part, the following calculations are
completed. First, debt forgiveness is deducted from total net grants. After this, the IDA grant part is subtracted. The remaining number is the net grant flow, without debt forgiveness and IDA grants. The third development aid variable, total loans, contains gross loan disbursements. The last variable in the model is debt relief. Contrary to Cassimon and Van Campenhout (2007, 2008) and Cassimon et al., (2013), total net debt relief is chosen instead of sole debt relief from the HIPC and/or MDRI initiatives. In the studied period, most debt relief is provided through these initiatives. However, non-HIPC countries received debt relief separate from these initiatives as well. Moreover, Cassimon and Van Campenhout (2007) write that collecting data on the HIPC and MDRI initiatives from the country reports is delicate, and consequently less creditable. Therefore, in providing a comprehensive and appropriate representation, total net debt relief from the OECD DAC database is utilized.
2.3. Variable statistics
The amount of observations of the variable time series are presented in Table 2.2. The fiscal dataset contains missing values. At most six observations are missing for a particular fiscal variable. Countries containing these missing values are kept in the balanced dataset on the basis of two criteria. First, not more than one year is missing for a certain country. Second, the year of this missing value is identical for all fiscal variables. Therefore, the negative econometric consequences are limited. However, missing values are undesirable. Therefore, this paper will apply some assumptions in order to deal with these missing values. First, when there are solely future data points
21
available, 𝒕𝒕 + 𝟏𝟏 is used. When there are only previous data points available, 𝒕𝒕 − 𝟏𝟏 is used. When there are previous and future data points available, the value of (𝒕𝒕−𝟏𝟏)+(𝒕𝒕+𝟏𝟏)𝟐𝟐 is taken.
The development aid variables contain missing observations too. For the
remaining grants variable, all missing observations are due to the debt forgiveness
element. Furthermore, the database does not provide a full dataset on debt relief. To proceed, this study assumes that the value of missing observations, from development aid variables, are null. This assumption is based on the following reasoning. The OECD DAC database tries to audit all development aid flows. However, debt relief and forgiveness agreements are sometimes bilateral arrangements apart from the Paris Club. Therefore, the full amount can sometimes be unclear. In contrast, debt relief and forgiveness through the HIPC/MDRI initiatives and Paris Club are well documented. This is exactly what is seen in the data. The missing observations are mainly from non-HIPC countries.
Table 2.2 presents the variables’ mean and standard deviation. Moreover, a disaggregation is made between HIPC and non-HIPC countries. The results indicate that non-HIPC countries collect on average more taxes and spend more on government current expenditures. Contrary to government investments, which are higher for HIPC countries. Domestic financing is, on average, low for both panels (<1% of GDP). Focussing on the standard deviations of the fiscal variables, it shows that these are higher for the non-HIPC panel, except for government investments. This gives reason to assume a more heterogeneous non-HIPC panel. Regarding the aid variable statistics, it shows that a substantial higher disbursement of aid flows goes to HIPC countries. Aside from this, debt relief appears to be marginal for both panels, especially for non-HIPC countries. However, the mean is low due to the one-time character of debt relief. Most non-HIPC countries received debt relief for only one or a few years. The same applies for HIPC countries. However, these amounts are on average higher. Comparing IDA grants to remaining grants, it shows that IDA grants are for both panels only a small part of their total grants received. For the non-HIPC and HIPC countries this is on average 2.6 and 5.44 percent respectively.
Figure 2.1 of Appendix B gives a graphical presentation of all variables in the studied period. Focussing on the fiscal variables, it shows that for both panels, tax revenues and government current expenditures are increasing over time. The
22
government investments increased for both groups as well. The investments are augmented due to an upward shift in 2009. The graph of domestic financing indicates that both panels reduced domestic financing until 2006 and augmented these hereafter. The graph of total loans indicates a decreasing trend for both groups of countries. One possible explanation for this is the substitution of these aid flows by IDA grants and debt relief. For the HIPC countries, debt relief peaks in the year 2003. After, it decreased and stabilised. For non-HIPC countries, the amount of debt relief appears stable. When the graphs of all variables are compared, remaining grants appear most stable. However, after peaking in 2010, a downward trend is seen for both panels.
One common procedure in modelling economic time series is to perform the logarithmic transformation. Stock and Watson (2010) write that this has two advantages. First, many economic time series are approximately exponential. Taking the logarithms of these variables makes the series roughly linear. Second, standard deviations of several time series are often proportional to its levels. Taking the logarithm of a variable makes the standard deviation of the series roughly constant. However, for some time series, performing this transformation can be difficult. These time series contain negative values or values of zero. Performing a logistic transformation on these time series is solely possible when a constant is added. However, this will lead to a large amount of outliers. Therefore, this paper applies the logarithmic transformation solely on the tax revenues, government current expenditures and government investments variables.
3. Empirical analysis
In this chapter the empirical methodology is provided. In the first part the hypotheses are presented. In the second part the empirical framework is constructed, which is used to test these hypotheses.
3.1. Hypotheses
After critically examining previous theoretical and empirical literature, and by taking into consideration the variable statistics, this chapter will specify the hypotheses. The first two hypotheses focus on the relation between IDA grants and tax effort, and debt relief and tax effort.
23
2nd hypothesis: a significant negative effect is expected of IDA grants on tax effort.
These hypotheses are motivated as follows. Debt relief provides fiscal space to the recipient country. Furthermore, as seen in the graphical representation, remaining grants stayed stable in the studied period. Therefore, IDA grants are an augmented grant flow, creating fiscal space too. With the use of PRSPs, this fiscal space should have been directed to poverty reducing expenditures. However, previous literature questions the effectiveness of PRSPs. Therefore, this paper assumes that these PRSPs were indeed not effective. Hence, the hypotheses state that the recipient country used the fiscal space, created through debt relief and IDA grants, to reduce tax effort.
The third hypothesis concerns, whether the effects differ for HIPC and non-HIPC countries.
3rd hypothesis: The effects for HIPC and non-HIPC countries are presumed to be different.
This is due to two reasons. First, aid flows are on average larger for HIPC countries. Therefore, it is expected that these countries are more aid dependent. Hence, changes in aid flows have more impact on tax effort. Second, the non-HIPC panel is more heterogeneous than the HIPC panel. Because the HIPC group is more identical, fiscal responses will be more alike. Hence, providing an more unambiguous result. Thus, the third hypothesis states that the effects will differ for non-HIPC and HIPC countries.
To test the hypotheses this paper will construct an empirical framework, which is presented in the next section.
3.2. Empirical framework
To test the hypotheses, a panel VAR is constructed. First, the panel VAR and the different methods of estimation are discussed. Second, two VAR post estimation concepts are described. These are granger causality and IRFs. Third, the concepts of stationarity and VAR stability are explained and applied. This chapter ends with an explanation and selection of the preferred model order.
3.2.1. Panel VAR and VAR estimation
The variables within the FRM are potentially dynamically related. This implies that current tax effort is related to previous tax effort and previous values of the other variables. This dynamic relation applies to the other variables included in the FRM too.
24
Moreover, no a priori conclusion can be made on which variables are exogenous. These two specifications fit the VAR model. Lütkepohl and Krätzig (2004) write that the VAR model will a priori treat all variables endogenous and potentially put restrictions with the help of econometric techniques. Furthermore, the VAR model assumes large dynamic relations between variables.
Within panel data, it is possible that factors are omitted, which vary across the entities (countries) but are constant over time. This unobserved country heterogeneity is called panel specific fixed-effects. A panel data VAR(p) model with order p and panel specific fixed effects can be represented by the following equation (Abrigo and Love, 2015).
𝒀𝒀𝒊𝒊𝒕𝒕 = 𝒀𝒀𝒊𝒊𝒕𝒕−𝟏𝟏𝑨𝑨𝟏𝟏+ 𝒀𝒀𝒊𝒊𝒕𝒕−𝟐𝟐𝑨𝑨𝟐𝟐+ ⋯ + 𝒀𝒀𝒊𝒊𝒕𝒕−𝒑𝒑𝑨𝑨𝒑𝒑−𝟏𝟏+ 𝒀𝒀𝒊𝒊𝒕𝒕−𝒑𝒑𝑨𝑨𝒑𝒑+ 𝒖𝒖𝒊𝒊𝒕𝒕+ 𝒆𝒆𝒊𝒊𝒕𝒕
𝑖𝑖 𝜖𝜖 {1,2, … , 𝑁𝑁}, 𝑡𝑡 𝜖𝜖 {1,2, … , 𝑇𝑇𝑖𝑖}
With 𝒀𝒀𝒊𝒊𝒕𝒕 being a (1𝑥𝑥𝑥𝑥) vector of dependent variables; 𝒖𝒖𝒊𝒊𝒕𝒕 is a (1𝑥𝑥𝑥𝑥) vector of dependent
variable-specific fixed effects errors; 𝒆𝒆𝒊𝒊𝒕𝒕 is a (1𝑥𝑥𝑥𝑥) vector of idiosyncratic errors and
𝑨𝑨𝟏𝟏,𝟐𝟐,…,𝒑𝒑 are (𝑥𝑥𝑥𝑥𝑥𝑥) parameters to be estimated. Assumed is that 𝒆𝒆𝒊𝒊𝒕𝒕 is an independent
vector described as 𝑬𝑬[𝒆𝒆′
𝒊𝒊𝒕𝒕𝒆𝒆𝒊𝒊𝒕𝒕] = 𝚺𝚺 with 𝑬𝑬[𝒆𝒆𝒊𝒊𝒕𝒕] = 0.
To control for the panel specific fixed-effects, Cagala and Glogowsky (2014) propose to perform fixed effects regressions with the least squares dummy variable estimator (LSDV). However, Abrigo and Love (2015) argue that these fixed effect regressions yield biased estimates, due to the correlation between lagged dependent variables. If N and T are large, this bias is small. However, when T decreases the bias increases. Therefore, Abrigo and Love propose two other estimation methods, based on the generalized method of moments (GMM) framework. The first method uses the forward orthogonal deviation technique (Helmert transformation) to remove panel specific fixed-effects. The second method uses the first-difference transformation. Both are consistent for small T and large N datasets. However, Abrigo and Love recommend the first, therefore this method is applied.1
1 The VAR is estimated with the amount of lags (transformed), selected at the model order selection in section 3.2.4., instrumented by the same lags in level (untransformed).
25
In order to provide an answer on all hypotheses, the empirical analysis will consist of three models. First, to answer whether tax effort was affected by IDA grants and debt relief, a VAR model is estimated with the aggregate dataset. Second, to answer whether the effects differ between HIPC and non-HIPC countries, two separate VAR models are estimated. One with a panel of HIPC countries and one with non-HIPC countries. The equation below is the reduced-form equation of tax revenues within a VAR(p) model, with model order p.2
𝑻𝑻𝒊𝒊𝒕𝒕 = 𝜶𝜶 + 𝑻𝑻𝒊𝒊𝒕𝒕−𝒑𝒑𝜷𝜷𝒑𝒑+ 𝑪𝑪𝒊𝒊𝒕𝒕−𝒑𝒑𝜸𝜸𝒑𝒑+ 𝑰𝑰𝒊𝒊𝒕𝒕−𝒑𝒑𝜹𝜹𝒑𝒑+ 𝑩𝑩𝒊𝒊𝒕𝒕−𝒑𝒑𝜺𝜺𝒑𝒑+ 𝑮𝑮∗𝒊𝒊𝒕𝒕−𝒑𝒑𝜻𝜻𝒑𝒑+ 𝑮𝑮𝒊𝒊𝒕𝒕−𝒑𝒑𝝑𝝑𝒑𝒑+ 𝑳𝑳𝒊𝒊𝒕𝒕−𝒑𝒑𝝀𝝀𝒑𝒑
+ 𝑫𝑫𝒊𝒊𝒕𝒕−𝒑𝒑𝜼𝜼𝒑𝒑+ 𝒖𝒖𝒊𝒊𝒕𝒕+ 𝒆𝒆𝒊𝒊𝒕𝒕
𝑖𝑖 𝜖𝜖 {1,2, … , 𝑁𝑁}, 𝑡𝑡 𝜖𝜖 {1,2, … , 𝑇𝑇𝑖𝑖}
With 𝑻𝑻𝒊𝒊𝒕𝒕 being tax revenues; 𝑪𝑪𝒊𝒊𝒕𝒕, 𝑰𝑰𝒊𝒊𝒕𝒕, 𝑩𝑩𝒊𝒊𝒕𝒕, 𝑮𝑮∗𝒊𝒊𝒕𝒕, 𝑮𝑮𝒊𝒊𝒕𝒕, 𝑳𝑳𝒊𝒊𝒕𝒕, 𝑫𝑫𝒊𝒊𝒕𝒕 being the fiscal and aid flow
variables; 𝜶𝜶 is a constant; 𝜷𝜷𝒑𝒑, 𝜸𝜸𝒑𝒑, 𝜹𝜹𝒑𝒑, 𝜺𝜺𝒑𝒑, 𝜻𝜻𝒑𝒑, 𝝑𝝑𝒑𝒑, 𝝀𝝀𝒑𝒑, 𝜼𝜼𝒑𝒑 are parameters to be estimated;
𝒖𝒖𝒊𝒊𝒕𝒕 is the variable-specific fixed effects error and 𝒆𝒆𝒊𝒊𝒕𝒕 is the idiosyncratic error.
3.2.2. Granger causality and impulse-response functions
Abrigo and Love (2015) argue that a VAR estimation gives the possibility to perform cross-equation hypothesis testing. One regularly used method, is the Granger (1969) causality test. Enders (2004) writes that granger causality tests estimate whether lagged values of one variable enter the equation before another. Stock and Watson (2012) point out that caution is required with granger causality tests. Granger causality does not imply causality. They argue that a more tailored name would have been granger predictability. Hence, a variable is a useful predictor given all other variables in the regression. However, granger causality is a useful tool in providing information on the relation between variables within the VAR. The granger causality test is conducted by estimating the Wald F-statistics. Under the null hypothesis the coefficients of the lagged variables are jointly equal to zero. Under the alternative hypothesis, the coefficients of the lagged variables granger-cause the dependent variable.
2 In a reduced-form equation, a system of equations is separated into single-form equations (Enders, 2004). Here, tax revenues is presented by its own lagged values and lagged values of the other explanatory variables. However, for all variables included in the VAR, a similar reduced-form equation can be presented. For simplicity, solely the equation for tax revenues is presented.
26
Enders (2004) argues that a VAR representation provides the possibility to estimate the time path of shocks within the system. This is done with impulse-response functions (IRFs). These functions estimate the response of one variable, when it is presented with an impulse of another. Normal IRFs estimate the response of a variable after an impulse of another, while keeping other dynamics within the VAR constant. Orthogonalized impulse-response functions (OIRFs) take into account these dynamics. However, to create OIRFs, structural analysis of the VAR is necessary. Hence, ordering the shocks within the VAR with the knowledge of the true economic model. This is beyond the scope of this paper. Therefore, normal IRFs are estimated. However, Abrigo and Love (2015) point out that these IRFs are not suitable for causal interpretation because variables are correlated contemporaneously. Though just as with granger causality, it offers extra information on the relation between variables within the VAR.
One important requirement of having consistent IRF estimations, is stability of the VAR system. Moreover, Enders (2004) writes that in order to perform granger causality tests, the VAR variables should be stationary. The concepts of variable stationarity and VAR stability are explained in the next section.
3.2.3. Variable stationarity and VAR stability
Stock and Watson (2012) argue that an important concept for time series regressions is variable stationarity. A variable is nonstationary when it contains a stochastic trend (unit root). If regressions are performed with nonstationary variables, conclusions are possibly unreliable. One example is spurious regression. This arises when regressions are performed with time series containing a similar stochastic trend. Results appear as if times series are highly related. However, this is due to their trend and not because of any causal relation. Nonstationary variables can be converted into stationary by taken differences. If the variable is stationary in first differences, it is said to be in the first order of integration I(1). Taking first differences has another effect, regarding imperfect multicollinearity. It is possible that a linear correlation exists between explanatory variables. If regressions are performed, this linear correlation can yield imperfect multicollinearity (Stock and Watson, 2012). However, after differencing the variables, this linear trend and potential imperfect multicollinearity is removed.
To detect whether variables are stationary, unit root tests are performed. For panel data, several unit root tests are available. These tests vary in the power they
27
exhibited across panel dimensions. This paper uses three panels. All contain less time periods (T) than countries (N). The most appropriate unit root test for these kinds of panels, is the Harris and Tzavalis unit root test, conducted by Harris and Tzavalis (1999). This test assumes a fixed amount of time periods. With the Harris and Tzavalis unit root test, no pre-determination of the desired lag length is required. The null hypothesis states that all countries have a unit root. The alternative hypothesis states that at least one country is stationary. Hence, it is possible that the null hypothesis is rejected while only one or two countries exhibit stationary data. Enders (2004) write that this alternative hypothesis is one of the limitations of most panel unit root tests.
Another procedure, to indirectly test whether the variables within the VAR are stationary, is by testing the stability of the VAR. Lütkepohl and Krätzig (2004) write that a VAR model is stable, when no VAR variables contain a unit root. Thus, stability of the VAR model implies that all variables are stationary. It is therefore sufficient to solely test for VAR stability. For comprehensiveness, this paper will first apply the Harris and Tzavalis (1999) test. After this, the VAR is tested for stability.
In Appendix C the results of the Harris and Tzavalis unit root tests are presented. Table 3.1 shows the results for the aggregate panel. The p-values in the third column indicate that the null hypothesis regarding tax revenues and government current expenditures consisting a unit root, cannot be rejected at a 0.10 significance level. The same applies to government investments at a 0.01 significance level. With these three variables in first differences, the VAR stability test indicates an unstable VAR. 3 The
remaining grants and IDA grants should be transformed into first differences too, to acquire a stable VAR. 4 Table 3.2 presents the results of the unit root tests for the panel
of HIPC countries. The results indicate that tax revenues, government current expenditures and government investments are non-stationary. However, the VAR stability test shows that IDA grants and remaining grants should be transformed into first differences as well. 5The results of the unit root test for the panel of non-HIPC
countries are presented in Table 3.3. These results indicate that the null hypothesis of time series consisting a unit root cannot be rejected for tax revenues, government current expenditures, government investments and IDA grants at a 0.10 significance
3 Testing VAR stability requires to predetermine the model order. However, at this stage, the desired order is unknown. Different model orders are tested for stability and all indicate VAR instability. 4 Several model orders are tested and all indicate VAR stability.
28
level. However, the VAR stability test indicates that remaining grants should be transformed into first differences as well.6
Enders (2004) points out that there is no agreement on whether variables in a VAR should be stationary. He argues that by taking first differences important information is lost. Furthermore, he points out that the goal of a VAR is to analyse the variables interdependencies, and not to determine the parameter estimates. Moreover, comparing results of time series in levels and first differences is difficult. Lütkepohl and Krätzig (2004) argue that a VAR model is not the most suitable to accommodate I(1) variables. They argue that it ignores a possible co-integrated relationship between a mix of I(0) and I(1) variables. This co-integrated relationship would yield other econometric modelling, such as a vector error correction model (VECM). Though, this goes beyond the scope of this paper. As argument against these statements, this paper deems the problems of multicollinearity and spurious regressions as severe, and prevails them above the loss of information. However, it should be possible to compare estimates and variables. Therefore, two models are estimated. Model A has tax revenues, government current expenditures, government investments, IDA grants and remaining grants in first differences to establish a stable VAR. Model B has all variables in first differences, to improve variable comparability.7
3.2.4. Model order selection
Stock and Watson (2012) argue that the lag length selection is a trade-off between employing too little predictive power and estimation uncertainty. Therefore, the amount of lags (order) must be in balance between the two. Normally, this is estimated by applying one of the information criteria. These information criteria estimate the lag length by measuring the information benefit and estimation cost, when an extra lag is added (Stock and Watson, 2012). Enders (2004) writes that it is possible to allow for different lag lengths within a VAR model. However, symmetry within the model is desired. It is therefore common to apply equal lag lengths for all equations. The model order of the panel VAR is based on Abrigo and Love’ (2015) model selection. They argue
6The stability condition is met when the VAR is estimated with model order two, which correspondents with the VAR model order selection presented in section 3.2.4..
7 When all variables are in first differences the aggregate and HIPC panels satisfy the VAR stability condition. However, the non-HIPC panel does not satisfy the stability condition. This has implications for estimating consistent IRFs and granger causality tests.
29
that the order can be selected by estimating the overall coefficient of determination. This coefficient estimates the fraction of variation explained by the panel VAR.
Table 3.4 presents the results of the model order selection for the aggregate panel. The coefficient of determination is estimated for both VAR models A and B. The coefficients of determination indicates that for both models the order of three is preferred. In Table 3.5, the results for the HIPC panel are presented. Estimating a VAR model is only feasible with more observations than parameters. This applies too, for estimating the coefficient of determination. The dataset of HIPC countries has more observations than parameters, for a maximum of three lags. The results indicate that for model A the order of two is preferred and for model B the order of three. For the non-HIPC dataset, it is only possible to estimate the coefficient of determination for a maximum order of two. With two as maximum, the results in Table 3.6 show that for both models the order of two is desired. However, for both models the coefficient of determination is clearly smaller than for the HIPC and aggregate panels.
The models of the aggregate panel are estimated with model order three. However, the coefficient of determination of model orders two and three are close to each other. Therefore, for robustness, both models are estimated with model order two as well. For comparability, similar model orders are preferred in estimating the models of the HIPC and non-HIPC panels. Therefore, the models are for both panels estimated with model order two. However, for the HIPC dataset, a robustness tests is performed with model order three.
4. Results
In this chapter the results of the VAR estimations, granger causality tests and IRFs are presented. First, the results of the aggregate panel are discussed. After this, the results of the two disaggregate panels are deliberated. This chapter ends with a section where for each panel the results of the two different models are compared.
4.1. Aggregate panel
It is difficult to compare the coefficients of variables in level values and first differences. Therefore, both VAR models are discussed separately. When possible, comparisons are made between the two. Table 4.1A presents the VAR estimation of model A. In Table
30
4.1B, the results of the VAR estimation of model B are presented. In Appendix D, the results of the granger causality test of both models are presented in Table 4.2A and Table 4.2B. As robustness test, the same procedure is applied to both models with lag length two. These results can be found in Tables 4.3A, 4.3B, 4.4A and 4.4B of Appendix D. In general, it can be concluded that there are many significant estimations. This is in line with previous fiscal response studies, using a VAR model representation (Cassimon and Van Campenhout, 2007; 2008 and Cassimon et al., 2013). However, to get a grip on the various estimates, mostly the impact of aid instrument on tax revenues is discussed.
The results of model A are presented first. Concerning the effect of IDA grants on tax revenues, one can see that the coefficients of the VAR estimation are negative. All three lagged values indicate that IDA grants have a negative impact on tax revenues at a 0.10, 0.01 and 0.05 significance level respectively. Moreover, the granger causality test indicates that lagged values of IDA grants contribute in predicting tax revenues at a 0.05 significance level (p=0.028). The effect peaks in the second year. Here, an increase of one unit in the growth of IDA grants as share of GDP decreases the growth of tax revenues as percentage of GDP with 1.9 percent. When model A is estimated with lag length two, similar but smaller coefficients are found for IDA grants. However, the coefficient of determination of model order three was slightly higher. Therefore, these estimates are preferred and used in further estimations. Figure 4.1A depicts the response of tax revenues over time, after an impulse of IDA grants. In the IRF, it can be seen that in the first two years after the impulse the response is negative. 8 However, after three and
four years a marginal but positive response is seen. Focussing on the effect of other development aid variables on tax revenues, no significant coefficients are found in the VAR estimation. Moreover, all granger causality tests indicate that lagged values of remaining grants, total loans and debt relief do not contribute in predicting tax revenues. When model A is estimated with lag length two, similar coefficients appear. Therefore, besides the IDA grant variable, no evidence is found for an effect of aid instruments on tax revenues in model A.
8 Confidence intervals can be estimated with Monte Carlo simulations. However, these yield broad confidence intervals. Moreover, the IRFs are not used for causal interpretation. Therefore, no confidence intervals are presented.
31
Table 4.1A: VAR estimation model A – Aggregate panel
The table below presents the coefficients of the VAR estimation with model order three. Model A has tax revenues, government current expenditures and government investments in log differences. IDA grants and remaining grants are in first differences. All other variables are in levels. The ***, ** and * stand for 1%, 5% or 10% significance level respectively.
Coefficients for Dependent
variable TAX REV. GOV. CUR. EXP. GOV. INVEST. DOM. FIN IDA GRANTS REM. GRANTS TOTAL LOANS DEBT RELIEF
TAX REV. L1 L2 L3 -0.051 -0.089 -0.031 0.154** 0.078 0.122* -0.113 -0.117 -0.285 2.811 1.672 -0.477 -0.430 -0.496 -1.188 -3.571 -1.820 0.263 0.054 -1.530 -1.161 5.743 0.903 0.034 GOV. CUR. EXP. L1 L2 L3 0.077 0.062 0.011 0.044 -0.123** -0.015 -0.049 -0.013 0.085 1.460 2.581** 2.403 1.045* 0.621 0.564 4.172*** 1.583 1.509 1.037 -1.044 2.374** -1.635 -5.358* 4.977* GOV. INVEST. L1 L2 L3 0.004 0.027** 0.011 0.006 0.020 0.006 0.206** -0.023 0.070 0.541 0.010 -0.043 0.352* 0.207 0.180 2.545** 1.305** -0.135 0.489 0.296 -0.180 1.412* -1.657 -0.643 DOM. FIN. L1 L2 L3 0.001 -0.001 0.009*** -0.005 -0.004* 0.004* -0.009 0.002 0.011 0.249*** 0.052 -0.044 -0.006 0.008 -0.005 -0.006 0.042 0.067 0.012 -0.023 -0.004 -0.090 0.061 0.013 IDA GRANTS L1 L2 L3 -0.012* -0.019*** -0.010** -0.006 -0.016 -0.010 -0.045* -0.032 -0.050*** -0.268 -0.235 -0.100 -0.573*** -0.337*** -0.118 0.211 0.474** 0.203 0.099 -0.098 0.057 0.135 -0.114 0.059 REM. GRANTS L1 L2 L3 -0.001 0.002 0.003 0.000 0.000 0.000 0.016** 0.009 0.003 0.093 -0.012 0.171** 0.042* 0.049 0.026 -0.224* -0.022 0.006 0.093 0.097 0.017 -0.018 -0.070 -0.112 TOTAL LOANS L1 L2 L3 0.003 -0.002 -0.002 0.004 -0.012*** -0.001 -0.012 -0.011 0.028** -0.001 -0.117 0.063 -0.013 0.074* 0.045 0.043 -0.022 0.270** 0.478*** 0.207*** 0.107** 0.718** -0.138 -0.168 DEBT RELIEF L1 L2 L3 0.001 0.003 0.000 0.003 0.006*** 0.003* -0.001 -0.004 -0.006* 0.003 0.096** -0.089** -0.020 -0.003 -0.017** -0.045 -0.030 -0.005 -0.043 0.026 -0.046** 0.081 0.071 0.005
Figure 4.1A: Impulse-response function
