Grants, Loans and Tax Revenues
What are the effects of grants and loans on tax revenues of recipient
countries and to what extent are these effects conditional on the
institutional quality of recipient countries?
Master Thesis MSc Economics Specialization International Economics & Globalisation Faculty of Economics and Business University of Amsterdam July 15, 2018 Author: A.M. Kingma Student number: 10574360 Email: annemkingma@gmail.com Supervisor: Ms. N.J. Leefmans Second reader: Dr. D.J.M. Veestraeten Number of words: 14960
2 STATEMENT OF ORGINALITY This document is written by Anne M. Kingma, who declares to take full responsibility for the content of this document. I declare that the text and 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 solely responsible for the supervision of completion of the work, not for the content.
3
Abstract
In recent times international interest in the effect of aid on the tax-to-GDP ratio of recipient countries has increased, as demonstrated by the Addis Tax Initiative of 2015. However, foreign donors fear that grants may have a discouraging effect on the tax revenues of recipient countries and are used for tax relief. This study investigates the effects of grants and loans on tax revenues in recipient countries and whether the impact of aid is conditional on the quality of the institutions of recipient countries. For the empirical research, a panel dataset consisting of 102 countries for the period 1980 to 2015 is used and regressions are performed with fixed effects ordinary least squares (FE OLS) and system generalized method of moments (GMM). In an attempt to capture the behavioural effects of aid occurring in the medium term and to reduce the potential endogeneity caused by simultaneous causality, different lagged values of grants and loans are used.
The main finding of this thesis is that the seemingly opposing results of previous research regarding the effect of grants on tax revenues have been brought closer. The reason for this is that the results depend on the period and methodology used. Despite the differences in the FE OLS and system GMM results, both estimation techniques provide evidence that grants increase tax revenues in the medium term, favouring a medium term of three years. Furthermore, this research finds only weak evidence that loans increase tax revenues. However, loans appear to positively affect the tax-to-GDP ratio in politically stable countries. By contrast, there is no strong evidence that the effects of loans and grants on tax revenues are conditional on corruption or government effectiveness.
4
Table of Contents
1. Introduction 5 2. Literature review 8 2.1. Theoretical literature 8 2.1.1. The behavioural effects of aid on tax revenues 8 2.1.2. Effects conditional on institutional quality 10 2.1.3. Summary 11 2.2. Empirical evidence 11 2.2.1 Impact of loans and grants on tax revenues 11 2.2.2. Governance variables, aid and tax revenues 14 2.2.3. Summary 14 3. Methodology 16 3.1. Baseline regression 16 3.1.1. Medium-term effect 16 3.1.2. Control variables 18 3.1.3. Conditionality on institutional quality 19 3.2. Robustness tests 21 3.3. Estimation techniques 22 3.3.1. Fixed effects OLS 22 3.3.2. System GMM 23 4. Data 4.1. Data description 27 4.2. Sample selection 28 5. Results 32 5.1. Results of baseline regressions 32 5.1.1. Main results 32 5.1.2. Results of different lagged values 35 5.1.3. Results conditional on the institutional quality 38 5.2. Results of robustness tests 40 6. Conclusion 43 References 46 Appendices 49 Appendix 1: Variables 49 Appendix 2: Countries 50 Appendix 3: Statistics 52 Appendix 4: Tests 54
5
1. Introduction
One debate about foreign aid is focused on the differences in effectiveness of the components of aid: grants and loans. It is argued that issuing foreign loans leads to debt accumulation in developing countries and those countries therefore fail to achieve development objectives (Gupta et al., 2003). The Meltzer Commission, established by the United States Congress in 2000, for this reason argued that development associations should only provide aid in the form of grants. However, some donor countries have doubts about increasing the share of grants, since for some recipient countries that share would reach up to three times the level of domestic tax revenues (Gupta et al., 2003). Figure 1 for example illustrates that in Malawi the grants-to-GDP ratio is almost four times the tax-to-GDP ratio in 1994. Figure 1: Malawi’s tax-to-GDP and grants-to-GDP ratios per year in percentages. Source: Graph constructed by the author based on data from the OECD (2018). When grants are given to recipient governments to increase their general budget or to finance certain areas of government expenditure, it has a direct effect on their government expenditures. Furthermore, when the amount of grants received is large compared to the level of domestic tax revenues, it might also affect domestic borrowing and taxation (Clist & Morrissey, 2011). The fear of donors is that grants will, for example, act as a substitute for tax revenues to privileged political interest groups or to stimulate investments. In the case of issuing loans, policymakers need to repay the
0 5 10 15 20 25 30 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 Tax-to-GDP ratio Grants-to-GDP ratio
6
amount in the future and therefore have an incentive to keep the amount of tax collection at least constant (Yohou et al., 2016).
To avoid tax relief being financed by grants, the International Development Committee (2014) recently stated that the United Kingdom should increase the share of bilateral loans for middle-income countries at the expense of grants. The reason is that in the UK an important objective of aid is mobilizing the domestic revenues of recipient countries. However, other countries have also become more aware of the importance of domestic resource mobilization in developing countries. For that reason, those countries joined forces in the Addis Tax Initiative of 2015, which stipulates that 19 donor countries will double their spending on aid used as support for domestic revenue mobilization before 2020 (International Tax Compact, 2015). This initiative proves that the relationship between aid and the tax revenues of recipient countries is increasingly coming to the attention of the world community.
The question as to whether an increase in tax revenues is achieved by providing grants or loans to recipient countries is therefore becoming ever more important. However, not only is there disagreement between donors about the different impacts of grants and loans, but also in the academic field there is no consensus. Regarding the impact of grants, Gupta et al. (2003) and Benedek et al. (2012) find that grants have a negative effect on tax revenues, whereas Clist and Morrissey (2011) and Prichard et al. (2014) argue that grants have a positive impact. The effect of loans on tax revenues is less ambiguous but still under discussion: Gupta et al. (2003) and Clist and Morrissey (2011) provide results indicating a significant positive impact, whereas Benedek et al. (2012) and Prichard et al. (2014) find no significant effect.
International awareness of the relationship between aid and domestic revenue mobilization and the lack of academic consensus about the impact of grants and loans on tax-to-GDP ratios make it relevant to research the effects of loans and grants on tax revenues. This thesis will therefore aim to answer the question: what are the effects of grants and loans on tax revenues of recipient countries for the period 1980 to 2015? Also, this thesis will investigate whether the effect of aid on tax revenues is conditional on the institutional quality of recipient countries. Three variables will measure this institutional quality: corruption, political stability and government effectiveness. In previous research regarding the relation between aid and tax revenues only corruption is used as conditionality. This thesis will further contribute to the existing literature by
7
including five additional years and using multiple estimation techniques and lagged variables, to research whether the choice of variables and estimation techniques affects the empirical outcomes.
The central question of this thesis will be answered by using annual data from 102 countries. The data are drawn from datasets of the World Bank, International Centre for Tax and Development (ICTD) and Organization for Economic Cooperation and Development (OECD). The empirical analysis will be conducted using second lagged variables of grants and loans, in an attempt to capture the assumed medium-term effect of aid on tax revenues and to avoid the potential endogeneity caused by simultaneous causality. In further attempts, I will also experiment with first, third and averages of first to fourth lagged variables. The methodology of this thesis is based on that of Gupta et al. (2003), and elaborates on that of Clist and Morrissey (2011) and Benedek et al. (2012). Gupta et al. (2003) were the first to introduce the baseline regression regressing aid on tax revenues and used fixed effects (FE) ordinary least squares (OLS) to perform the regression. Besides using FE OLS, I will also use the system generalized method of moments (GMM) estimators for the regression, as done by Benedek et al. (2012). To investigate whether the relation between aid and tax revenues is conditional on institutional quality, I augment the baseline regression with interaction variables consisting of aid components and dummy variables based on governance variables. This thesis is ordered as follows. The next chapter, provides an overview of the existing literature, which is divided into an empirical part and a theoretical part. Chapter 3 describes the methodology and estimation techniques used to answer the central question. Chapter 4 elaborates on the data used in this thesis. Chapter 5 reveals the outcomes of the empirical models and chapter 6 presents the conclusion and discussion.
8 2. Literature review This chapter includes an overview of the existing literature about the relation between aid and tax revenues. Section 2.1 will focus on the possible behavioural effects of aid on tax revenues from a theoretical perspective, while section 2.2 consists of an overview of empirical evidence found by previous research. 2.1. Theoretical literature 2.1.1. The behavioural effects of aid on tax revenues The definition of ‘aid’ used in this thesis is that of official development assistance (ODA) consisting of concessional loans and grants that aim to support the economic development of recipient countries (OECD, 2018). Morrissey (2004) argues that in the majority of the recipient countries aid is spent on the provision of public services or used for general budget support. Therefore, aid affects fiscal behaviour and the relationship between foreign aid and tax revenues can thus be illustrated using the government budget constraint, as Gupta et al. (2003) did. Here the budget constraint is incorporated into the theoretical framework to analyse the relations between grants and loans and tax revenues: 𝐺 = 𝑇 + 𝐺𝑟 + 𝐿 + 𝐵
where 𝐺 denotes government spending, 𝑇 the tax revenues, 𝐺𝑟 the amount of grants, 𝐿 the amount of loans, and 𝐵 domestic net borrowing. The above equation implies that when governments receive grants or loans, policymakers decide either to increase government spending, reduce tax effort, lower domestic borrowing, or combine the multiple options. In extreme cases, government reduces tax effort by the full amount of the received grants and keeps domestic borrowing and government spending constant. Financing tax relief with grants and loans is also called the ‘substitution effect’. (Gupta et al., 2003, p. 4).
There are several reasons for recipient governments to use aid to finance tax relief, some of which can be related to the economic policies of the ruling administration. For example, tax relief can occur when it is the government’s objective to increase investments and economic growth. This objective can be achieved through a lowering of taxation on capital (Heller, 1975). Aid may therefore enable an increase in investments
9
and economic growth without lowering government expenditure, thus without a reduction in the provided public services (Hisali & Ddumba-Ssentamu, 2013). Other reasons that support the substitution effect are based on the personal incentives of politicians. One example is that ruling politicians can desire to increase their popularity to remain in power or to win an election, which can be achieved by tax relief. This reduction in tax revenues can be financed by aid (Knack, 2001).
The above-mentioned reasons have a direct negative effect on tax revenues, but there are also indirect behavioural effects of aid that may result in a decrease in tax revenues. For example, it is argued that aid lowers tax revenues through undermining the development of governmental institutions (Bräutigam & Knack, 2004). Receiving a continuing influx of grants can reduce the incentives for rent-seeking governments to increase their spending on institutions responsible for tax collection because keeping the amount of tax revenues low might ensure that the influx of aid will continue in the future (Gupta et al., 2003). Bräutigam and Knack argue that in the 1990s several African countries were unwilling to increase tax efforts to be able to finance public services themselves, because the recipient governments had ‘developed a cozy accommodation with dependency’ (2004, p. 257). Dependence on aid resulted in the fact that in the 1990s almost 40% of government expenditures in Ghana, Malawi and Zambia were financed with aid.
The above arguments imply that grants and loans may have a negative effect on tax revenues. From a theoretical point of view the opposite can be true as well. As previously mentioned, receiving aid may lead to an increase in government spending instead of a reduction in domestic borrowing or tax effort. Gambaro et al. (2007) argue, for example, that recipient governments may use aid in order to strengthen the capacities of government institutions, including those responsible for tax collection. The increase in capacity may result in raising tax revenues. The same may hold for loans. Indeed, loans need to be repaid to the lender and therefore policymakers may be induced to strengthen the collection of tax revenues (Gupta et al., 2003).
In addition, aid does not only have a positive effect on tax revenues by influencing the government behaviour, it may also have a positive and indirect effect on tax revenues through its impact on the economy. Yohou et al. (2016) argue that when aid results in a raise in government expenditures or improves the effectiveness of public services, it may increase the level of human development and the rate of economic
10
growth and therefore enhance future tax performances. Also, when the government succeeds in stimulating investments and economic growth through tax relief, it may generate more future tax revenues through an increase in national income (Yohou et al., 2016).
2.1.2 Effects conditional on institutional quality
The previous section described several behavioural effects of aid on tax revenues. In addition, the impact of aid could be conditional on the quality of the institutions of the recipient country in question. Also, the institutional quality may interact with the presence of aid and therefore affect the relationship between aid and tax revenues. For example, according to Knack (2001) the substitution effect, where tax relief is financed by aid, mostly occurs in dictatorships or corrupt countries to strengthen the power of the dictator or ruling elite, as tax relief might favour certain political interest groups that keep the government in power. Thus, when countries are highly corrupt, aid may lead to lower tax revenues compared to countries that have a low degree of corruption. In addition, Gambaro et al. (2007) argue that the impact of aid on tax revenues is stronger in less corrupt countries, since the governments in those countries suffer less from a lack of accountability making it more difficult for the governments to enrich themselves instead of efficiently spending aid.
The effect of aid on tax revenues might also be conditional on the degree of political stability. Yohou et al. (2016) argue that when it is not certain the ruling government will survive in office, the citizens’ trust that the government will provide public services is lowered and therefore citizens are less willing to pay taxes. Thus, aid might be ineffective in increasing tax revenues when countries are politically unstable. Also, the influx of aid might worsen the political instability, leading to less efficient tax collection. It is argued that the presence of extensive food aid in Somalia led to an increase in the value of Somalia’s governmental institutions, resulting in more competition to take control of the institutions, which in the end led to civil war (Knack, 2001).
Besides corruption and political stability, the quality of public services might affect the relationship between aid and tax revenues. Knack argues (2001) that when the quality of public services is low it can be weakened further when receiving aid, because talented public officials are lured away from the government sector by donor
11
organizations such as nongovernmental organizations. These organizations settle in recipient countries and persuade public officials to work for them by offering higher salaries. The decrease in talented public officials may result in less effective state bureaucracies responsible for tax collection and therefore in a decline in tax revenues (Knack, 2001).
2.1.3 Summary
In summary, it appears that there are multiple behavioural effects that may occur when countries receive loans and grants. Several arguments related to economic policy and politicians’ incentives support the substitution effect, suggesting that receiving aid discourages the tax efforts of recipient countries. However, one can also argue that aid strengthens the capacity of domestic institutions and stimulates investments and economic growth, resulting in an increase of tax revenues. The ambiguous effect of aid on tax revenues indicates a relationship that is theoretically not straightforward and therefore the topic of this thesis is rather an empirical question.
2.2. Empirical Evidence
2.2.1 Impact of grants and loans on tax revenues
Heller (1975) is the first who performed an empirical cross-country study on the effect of aid on public fiscal behaviour, including 11 developing African countries in his research. Previous research has focused on the effect of aid on the tax efforts of individual countries. The author developed an econometric model including multiple utility and constraint functions and is also the first who assessed within his model the various effects of loans and grants. He finds that grants increase consumption whereas foreign loans tend to stimulate investments. Khan and Hoshino (1992) also examined the effect of aid on tax revenues in an empirical cross-country research including five Southeast Asian countries and used a similar model as Heller (1975). Their results indicate that grants and loans have opposing effects on tax revenues because loans increase the tax burden, and therefore have a positive effect, while grants lessen it. However, the results of the empirical cross-country research by Otim (1996) indicate different conclusions, namely that grants and loans both positively affect tax efforts in
Southeast Asian countries.
12
Odedokun (2003) is the first to carry out empirical research into a larger number of recipient countries, namely 22 donor and 72 recipient countries. Also, instead of building an extensive empirical model with multiple equations, as in previous research, he created an empirical model including one regression in which the variable tax revenues is a function of GDP per capita and first lagged grants. Odedokun (2003) finds that grants reduce tax effort. Gupta et al. (2003) are the first to include as many recipient and donor countries as possible in their research and estimated the effect of total grants and loans on the tax revenues of 107 recipient countries during the period 1970 to 2000. Their regression model is used as a starting point by subsequent research studies and, compared with the research of Odedokun (2003), is much broader because of more control and explanatory variables. The authors are also the first to add variables that represent the sectoral composition of countries’ economies to their regression model and to allow for nonlinearities by including squared variables of grants and loans. The control variables they include are trade openness, gross domestic product (GDP) per capita, and agricultural and industry value added. With the help of fixed effects ordinary least squares estimators, Gupta et al. (2003) estimate that tax revenues decrease by 28 cents for each grant dollar received, while tax revenues increase by 35 cents for each additional dollar of aid in the form of loans. Thus, grants have a negative effect on tax effort whereas loans have a positive effect.
Although Gupta et al. (2003) provide a thorough analysis of the relationship between aid and tax revenues, they are criticized by Clist and Morrissey (2011), who decided to change the previously used methodology in several ways. For example, they replaced the trade openness variable for separate import and export variables. Clist and Morrissey (2011) argue that during the 1980s export taxes were reduced and imports and exports were thus not taxed at the same rate. Therefore imports and exports should be incorporated separately. Second, they started using second lagged variables of aid as explanatory variables instead of current aid variables. The reasons for this are to reduce the potential problem of endogeneity and to give aid time to affect the tax-to-GDP ratio or to capture the behavioural effect of aid. The endogeneity problem arises because the inflow of aid can also be determined by the current tax-to-GDP ratio, rather than the other way around, suggesting that there is simultaneous causality between aid and tax revenues. The authors divided their sample into two different time periods. They conclude that for the period 1970 to 2005 their outcomes for 82 countries coincided
13
with the results of Gupta et al. (2003). For the period 1985 to 2005, however, they find that, when using lagged variables of loans and grants, both have a positive and significant effect on tax revenues but there is no significant impact when using current variables. Therefore, Clist and Morrissey (2011) claim that before and during the beginning of the 1980s grants had a negative effect on tax revenues but that this effect has changed during the 1980s to become a positive effect afterwards.
Benedek et al. (2012) respond to Clist and Morrissey (2011) with a paper introducing a more elaborate estimation technique, namely the general method of moments (GMM) estimators. Benedek et al. (2012) used both the difference GMM and system GMM estimators to regress the relationship between aid and tax revenues. They proposed using this technique to remove the problem of potential endogeneity and the possibility of serial correlation because of the persistence of the tax-to-GDP ratio. For the period 1980 to 2009 they find that loans do not affect tax revenues and that grants have a negative effect on tax revenues but that this effect weakens over time. Clist (2016) argues that using GMM estimators in the context of Benedek et al. (2012) is inappropriate, because difference GMM instruments would highlight differences that are mainly caused by the use of multiple data sources. For example, Benedek et al. (2012) used multiple data sources to construct the variable tax revenues because of the many missing data points in each dataset. Clist (2016), however, state that the use of multiple data sources to replace the missing observations of tax revenue data is the main problem of econometric research into the impact of aid on tax revenues. Clist (2016) supports his argument by proving that the empirical outcomes depend on the data sources of tax revenues used and suggests using only one data source in the form of a more comprehensive dataset, such as the Government Revenue Dataset (GRD) of the International Centre for Tax and Development (ICTD).
The studies of Prichard et al. (2014) and Clist (2016) both used the GRD dataset. The latter investigated the period 1980 to 2009 and included current and first and second lagged values of grants and loans. Clist (2016) finds for both lag values that grants and loans have insignificant effects on tax revenues. Although he finds evidence that current grants have a negative effect on tax revenues like Gupta et al. (2003), Clist (2016) argues that this result is probably due to donors compensating recipient governments for the shortfall of tax revenues to finance government expenditures. Prichard et al. (2014) did not use second lagged values or system GMM estimators to
14
avoid simultaneous causality but experimented with averages of four to eight lag lengths for the aid variables. They find that grants variables consisting of an average of four lag lengths have a positive effect on tax revenues, whereas grants variables constructed from an average of eight lag lengths have no significant effects.
2.2.1 Governance variables, aid and tax revenues
Variables related to government institutions are increasingly incorporated in research into the determinants of tax revenue. Gupta (2007) examined whether different institutional variables affect tax revenues. The results are mixed: government stability and law and order do not have significant effects, while political stability, economic stability and corruption have been proven to be determinants of tax revenues.
When good governance came to the attention of donors during the 1990s, institutionally related variables also became part of research investigating the relationship between aid and tax revenues (Gambaro et al., 2007). For example, Ghura (1998) demonstrates for 39 Sub-Saharan countries that tax collection is affected by corruption and the quality of public services. When the level of corruption is high, aid affects the tax-to-GDP ratio negatively. The opposite effect holds true for the quality of public services. Gupta et al. (2003) also include corruption as a control variable and find that corruption increases the negative effect of grants on tax revenues, while in highly corrupt countries the positive effect of loans offsets the negative effect of corruption. While governance-related variables could be determinants of tax revenues, the variables could also interact with aid having a different impact on tax revenues. Gambaro et al. (2007) are the first to investigate the interaction effects between corruption and aid on tax revenues and to use interaction variables. They estimate the impact of aid on tax revenues conditional on corruption for 65 countries during the period 1990 to 2004 by using a simple model without lags or squared variables of aid. The authors divided the countries into less corrupt and highly corrupt countries and generalized this for the entire period. They find no evidence that grants and loans have a different impact on tax revenues when countries are highly or less corrupt.
2.2.2 Summary
The inconclusiveness of the theoretical literature is also present in the empirical literature. There is no agreement about the sign or significance of the effects of grants
15
and loans on tax revenues. Empirical models differ with regard to the inclusion of control and lagged variables and econometric approach. For example, GMM and OLS FE estimators are both used, import and export are exchanged for trade openness and different numbers of lags of grants and loans are included. However, there is consensus about the inclusion of some variables. GDP per capita, the shares of agricultural and industrial values added, and the squared variables of grants and loans are included in all previous research starting from Gupta et al. (2003).
16
3. Methodology
The aim of this thesis is to answer the question as to whether there is a substitution effect between grants and loans and tax revenues. Since the relationship between aid and tax revenues is in fact an empirical question, this study will be empirical as well. This chapter explains the methodology that is used for the empirical research. The first section explains the content of the baseline regressions including the regressions estimating the impact of aid on tax revenues conditional on the quality of the domestic institutions. The second section describes the robustness tests and the third section explains the two estimation techniques used in the empirical research.
3.1. Baseline regression
3.1.3. Medium-term effect
To analyse the effect of grants and loans on tax revenues for the period 1980 to 2015, an unbalanced panel dataset with data from 102 countries is used. The dataset contains data of five more years and a higher number of observations compared with the most recent previous research, thanks to the increased availability of data of tax revenues and aid. Further information about the data of this thesis can be found in Chapter 4.
The baseline regressions of this thesis are based on the regression of the first researchers to investigate the relationship between the composition of total aid and tax revenues across multiple countries and regions, Gupta et al. (2003). Furthermore, the regressions are elaborated upon those of Clist and Morrissey (2011) and Benedek et al. (2012). The regressions include multiple control variables that represent the composition of the economy and the tax base structure of a country (Tanzi, 1992). To see whether aid affects tax revenues, the standard regression for tax effort is augmented by including the variables grants and loans and their squared equivalents. This results in the following baseline regression: 1 ln 𝑇𝑎𝑥!" = 𝛽! + 𝛽!𝐴𝑔𝑟 !"+ 𝛽!𝐼𝑛𝑑!"+ 𝛽!𝐺𝐷𝑃!"+ 𝛽!𝐸𝑥𝑝!"+ 𝛽!𝐼𝑚𝑝!"+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!! + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!! + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!! + 𝛼! + 𝛾!+ 𝜀
where 𝑇𝑎𝑥!" represents the tax revenue in percentage of GDP of country 𝑖 in year 𝑡. The dependent variable is transformed into a logarithm because the variable is non-negative and positively skewed. The explanatory variables of interest are the grants (𝐺𝑟𝑎𝑛𝑡𝑠!"!!)
17
and loans (𝐿𝑜𝑎𝑛𝑠!"!!), both in percentages of GDP. To allow for nonlinearities in the relationships between aid and tax revenues, squared variables of grants 𝐺𝑟𝑎𝑛𝑡𝑠!"!!! and loans (𝐿𝑜𝑎𝑛𝑠!"!!! ) are included in the regression. When the variables grants and loans have a positive sign, it indicates that an increase in grants and loans results in more tax revenues, and when the variables have a negative sign it provides evidence of a substitution effect. When the squared grants and loans variables have a negative sign, it proves that, after a certain threshold, more loans or grants result in less additional tax revenues. To control for country-specific and time-invariant shocks, country-specific dummies are included, summarized in 𝛼!. Furthermore, year dummies are included to correct for the annual and common shocks occurring in all countries, denoted by 𝛾!. To prevent multicollinearity, one dummy is omitted. For the variables grants and loans lagged values will be used, as indicated by the subscription. The reason for using lagged values is that it is expected that aid will have behavioural effects on tax revenues in the medium term. It takes time for the amount of aid received to influence the decisions of politicians to raise more or less taxes (Prichard et al., 2014). Changes in tax effort are therefore only likely to happen in the medium term instead of in the short term, certainly in low-income countries where improvements in tax effort are hard to accomplish. Also, the fraction of aid given to improve tax revenues immediately to meet the current government expenditure is less than 1.7%, suggesting that a contemporaneous relationship between aid and tax revenues is very small (Prichard et al., 2014).
The primary lagged value of interest is the second lag of aid variables. The reason for this is that Clist and Morrissey (2011), being the first to use two-year lagged values of grants and loans, argue that using second lags may capture the behavioural effects and could reduce potential endogeneity depending on the persistence of the tax-to-GDP ratio. Nevertheless, defining ‘the medium term’ remains ambiguous. Prichard et al. (2014) have stated that in a further attempt to find behavioural effects, deeper lag lengths could be helpful and therefore included averages of four lagged values of grants and loans in their regression. This thesis will follow Prichard et al. (2014) and also include the four-year lagged averages of the aid variables in the regression. In addition, the first and third lags of grants and loans will be used to investigate whether the second lag is indeed the best lagged value to capture the medium-term effect. The first lag has
18 already been employed in previous research, whereas the third lag will be introduced in this thesis. 3.1.2. Control variables The control variables of equation (1) are the proxies of the tax base structure and the composition of the economy that determine tax revenues. These control variables include variables that have been used in previous research investigating the relationship between aid and tax revenues and have been identified as having a significant effect on tax revenues.
To control for the sectorial composition of the economy, the agricultural (𝐴𝑔𝑟 !") and industrial (𝐼𝑛𝑑!") shares of value added are included in the regression and presented as fractions of GDP. The agricultural variable is expected to have a negative effect on tax revenues, because if a large part of the agricultural sector is subsistence agriculture then the sector might be difficult to tax (𝛽! < 0). Also, from a demand point of view, a large agricultural sector may lower the need to spend government tax revenues on public goods and services, which tend to be more urban-based (Gupta, 2009). By contrast, a large industrial sector might be easier to monitor, making it less complicated for policymakers to tax activities in the industrial sector. The industrial sector could also act as proxy for the size of the formal private sector (Clist & Morrissey, 2011). A positive sign for the coefficient of the industrial sector (𝛽! > 0) is therefore expected.
The control variable 𝐺𝐷𝑃!" represents the GDP per capita and is transformed into a logarithm because the variable is non-negative and positively skewed. Furthermore, the transformation makes the variable stationary. Evidence that this and other variables are stationary is presented in Appendix 4. The variable is used as proxy for a country’s economic development and is expected to have a positive effect on tax revenues (𝛽! > 0). Tanzi (1992) has illustrated that the development of tax bases increases more than proportionately to the growth of the level of GDP per capita. Moreover, GDP per capita is a good indicator of a country’s economic structure and ability to collect taxes (Gambaro et al., 2007).
Another important determinant of tax revenues is trade taxes (Tanzi, 1992). One explanation for the importance of trade taxes is that they are easier to charge than income taxes because trading goods enter the country at specific border posts.
19
Therefore developing countries typically also collect a higher share of their tax revenues from trade taxes than developed countries (Gupta, 2007). In the first equation, imports
(𝐼𝑚𝑝!") and exports (𝐸𝑥𝑝!") are included separately in the regression, following Clist and
Morrissey (2011). The authors argue that imports and exports generate different tax revenues because they are not taxed at the same rate and in the 1980s export taxes were almost completely removed. The expectation here is that exports have a positive effect on tax revenues because when exports increase and are taxed, they will result in more tax revenues (𝛽! > 0). Due to the reduction in export taxes and the importance of import taxes for developing countries, a larger positive effect from the control variable imports (𝛽! > 0) is expected. According to Moutos (2001), the reason why developing countries in particular rely on import taxes is that the majority of the citizens of such countries consume domestically produced goods instead of imported goods, and therefore the domestic citizens have a preference for higher import taxes over higher income or export taxes.
In the regressions of other researchers, trade openness acts as a proxy of trade tax revenues instead of imports and exports. Gupta et al. (2003) and Benedek et al. (2012) include trade openness in their regressions instead of imports and exports. Trade openness is defined as the sum of imports and exports as a percentage of GDP. Since Clist and Morrissey (2011) and Benedek et al. (2012) disagree about including trade openness or imports and exports and also record different empirical outcomes for each, the present research will estimate a second baseline regression to check whether including trade openness indeed provides significantly different results. It is expected that trade openness, like exports and imports, will have a positive effect on tax revenues (𝛽! > 0). The second baseline regression is as follows: 2 ln 𝑇𝑎𝑥!" = 𝛽! + 𝛽!𝐴𝑔𝑟 !"+ 𝛽!𝐼𝑛𝑑!"+ 𝛽!𝐺𝐷𝑃!"+ 𝛽!𝑇𝑟𝑎𝑑𝑒!"+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!! + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!! + 𝛽 !𝐿𝑜𝑎𝑛𝑠!"!!! + 𝛼! + 𝛾!+ 𝜀 3.1.3. Conditionality on institutional quality
In this thesis, the impact of grants and loans on tax revenues will also be estimated conditional on the quality of the government institutions of recipient countries. Three governance variables will represent this institutional quality: corruption, government effectiveness and political stability. In previous research by Gambaro et al. (2007), the
20
impact of aid was only investigated conditional on corruption. However, since Gupta (2007) found that political stability affects tax revenues significantly, it might be that the impact of aid on tax revenues is conditional on the degree of political stability. In addition, there is theoretical evidence that the quality of public services might affect the relationships between aid and tax revenues (Knack, 2014).
It might be incorrect to assume that corruption, political instability and government effectiveness have the same impact on the relationship between aid and tax revenues since corruption does not need to be a correct proxy for government effectiveness or political stability. For example, Turkey is politically unstable but scores high on government effectiveness and is in the middle range in terms of corruption. Therefore, this study includes three governance variables representing the quality of domestic institutions separately and capturing the interaction effects with aid. Since the three political variables are correlated (Kaufman et al., 2010), the variables are separately augmented to the first baseline regression (1) using second lagged values, resulting in three new regressions: 3 ln 𝑇𝑎𝑥!" = 𝛽!+ 𝛽!𝑋 + 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!+ 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!𝐿𝑜𝑤𝐶𝑜𝑟𝑟 + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!𝐿𝑜𝑤𝐶𝑜𝑟𝑟 + 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!! + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!! + 𝛼!+ 𝛾!+ 𝜀 4 ln 𝑇𝑎𝑥!" = 𝛽!+ 𝛽!𝑋 + 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!+ 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!𝐻𝑖𝑔ℎ𝑆𝑡𝑎𝑏 + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!𝐻𝑖𝑔ℎ𝑆𝑡𝑎𝑏 + 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!! + 𝛽 !𝐿𝑜𝑎𝑛𝑠!"!!! + 𝛼!+ 𝛾!+ 𝜀 5 ln 𝑇𝑎𝑥!" = 𝛽! + 𝛽!𝑋 + 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!+ 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!+ 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!𝐻𝑖𝑔ℎ𝐺𝑜𝑣 + 𝛽!𝐿𝑜𝑎𝑛𝑠!"!!𝐻𝑖𝑔ℎ𝐺𝑜𝑣 + 𝛽!𝐺𝑟𝑎𝑛𝑡𝑠!"!!! + 𝛽 !𝐿𝑜𝑎𝑛𝑠!"!!! + 𝛼!+ 𝛾!+ 𝜀
The variable X represents the vector of control variables as identified in the previous section. In regression (3) the interaction variables consist of a dummy variable 𝐿𝑜𝑤𝐶𝑜𝑟𝑟 that includes only countries with a low level of corruption, and a second lagged aid variable. This researcher has chosen to use current values of governance variables and not to include second lagged values because it would remove two data points, resulting in an even smaller sample since only annual figures of governance variables for the period 2002 to 2015 are available. Furthermore, the governance variables are used as dummy variables so countries are either less corrupt or highly corrupt, meaning that the
21
precise value of corruption is less relevant and, finally, the average values of the governance variables remain roughly the same over time, as can be seen in Appendix 3. The corruption variable reflects the extent to which public power is exercised for private gain and how the state is controlled by elites and their interests (Kaufman et al., 2010). One could expect that a corrupt administration would invest aid less efficiently or would choose to spend grants for the purpose of self-enrichment. Also, governments in highly corrupt countries might more easily favour political interest groups by granting them tax relief financed with aid (Knack, 2001). It is therefore expected that grants and loans have positive effects on tax revenues in less corrupt countries (𝛽! = 𝛽! > 0). Regression (4) includes the dummy variable 𝐻𝑖𝑔ℎ𝑆𝑡𝑎𝑏, representing politically stable countries. The original variable captures the likelihood that a country’s
government will be overthrown by violence or illegitimate means (Kaufman et al., 2010). The expectation is that aid received in politically unstable countries is ineffective in increasing tax revenues because then the citizens mistrust that the government will provide public services and are therefore less likely to pay taxes. By contrast, the sign of the coefficient of the interaction variable including a high degree of political stability will be positive (𝛽! = 𝛽! > 0).
Finally, government effectiveness can be affected by aid; the variable of government effectiveness is captured by 𝐻𝑖𝑔ℎ𝐺𝑜𝑣 in equation (5), representing countries with a high degree of government effectiveness. The variable is measured by the quality of public and civil services (Kaufman et al., 2010). The expectation is that if aid interacts with a low degree of government effectiveness, it will have a negative impact on the collection of tax revenues, and vice versa if aid interacts with a high degree of government effectiveness (𝛽! = 𝛽! > 0). The reason for this is that public officials in countries characterized by a low-quality public sector might be lured away from the low-quality public sector because of the higher payment or securities offered by donor organizations that settle in the recipient countries (Knack, 2001). This draining of talent leads to less effective public institutions, implying a negative effect on tax revenues. 3.2. Robustness tests Clist and Morrissey (2011) argue in their paper that the relationship between grants and tax revenues changed during the latter part of the 1980s, since they find that grants had
22 a positive impact for the period 1985 to 2005 but no effect for the period 1970 to 1985. To explain these empirical results, the authors suggest that the positive effect during the period 1985 to 2005 is associated with the policy reforms introduced by the structural adjustment programmes (SAPs) of the Bretton Woods institutions and implemented in recipient countries during the 1980s (Prichard et al., 2014). They state that the policy reforms, together with the technical assistance accompanying aid, turned the effect of grants on tax revenues in a positive direction. To investigate whether the policy reforms of the SAPs indeed resulted in a different effect of grants on tax revenues after the 1980s, the sample of the present research is divided into two subsamples. Following Prichard et al. (2014), the first period includes data from the 1980s, 1980 to 1989, and the second period includes data from 1990 to 2015. The first subsample includes the period of implementation of the policy reforms and the second subsample represents the post-1980 period in which the effect of grants on tax revenues should be positive, if the hypothesis of Clist and Morrissey (2011) is correct.
Furthermore, the sample will be divided into two subsamples based on income classification to investigate whether including high-income countries, leads to different empirical results compared to excluding them. In the sample of this thesis, all countries that received aid somewhere during the period 1980 to 2015 are included, thus also current high-income countries that received aid before becoming advanced economies. The results of the comprehensive sample will be compared with the results of the subsample that excludes the ten current high-income countries and accordingly only contains developing countries. The reason for separating countries is that previous researchers did not include current high-income countries and therefore may have found different empirical outcomes. For example, Prichard et al. (2014) did not include European high-income countries, such as Cyprus and Malta, in his sample. However, those countries received aid during the period 1980 to 2015 and are therefore included in this thesis’ sample. The classification of countries based on level of income can be found in Appendix 2.
3.3 Estimation techniques
3.3.1. Fixed effects OLS
The fixed effects OLS model is a method to control for omitted variables when those variables vary across entities, countries in this case, but do not change over time. The
23
model deals with heterogeneity that may (not) be observed and examines whether the intercept differs per country (Park, 2011). The main reason for using fixed effects in this thesis is that the data probably will include heterogeneity, because it is likely that the effects of grants and loans on tax revenues will be affected by unobserved country-specific factors that are constant over time (Prichard et al., 2014). Examples of such factors are the number of inhabitants or national characteristics regarding taxation.
Although most researchers perform fixed effects OLS regressions to estimate the effects of aid components on tax revenues, Gupta et al. (2003) also used random effects (RE) regressions as methodology. The main difference between RE and FE regressions is the place of the dummy variable in the regression. The parameter estimate of the dummy variable in the RE model is an error component and in the FE model a component of the intercept (Park, 2011). To check which model is preferred, the Hausman test was performed, which compares the FE model with its RE equivalent. The null hypothesis was rejected, suggesting that using RE regressions would be problematic and thus that FE regressions are preferred (Park, 2011).
However, FE models with solely country fixed effects have their shortcomings. First, those FE models do not include time-specific unobserved effects, that is, effects which are constant across countries but differ over time. Financial crises and extreme weather are examples of time fixed effects and could affect the relationship between aid and tax revenues (Stock & Watson, 2012, p. 407). To control for time-specific unobserved effects, time dummies are included in the regression and a two-way FE model is created (Park, 2011). The second shortcoming is that FE OLS regressions do not correct for potential endogeneity, meaning that the explanatory variables can be correlated with the error term and are therefore endogenous variables (Stock & Watson, 2012, p. 471). A reason for potential endogeneity would be simultaneous causality, because of the possibility that the current inflow of aid is also determined by the current tax revenues of countries, rather than the other way around (Prichard et al. 2014).
3.3.2. System GMM
Whereas previous researchers used lagged variables to correct for potential endogeneity between aid and tax revenues, Benedek et al. (2012) were the first to use generalized method of moments estimators (GMM estimators) to overcome the potential problems of endogeneity and the possibility of serial correlation. In this
24 context, the use of GMM estimators seems valid. First, there is theoretical evidence that tax revenues might determine the amount of grants and loans. Therefore the variables grants and loans may be endogenous. Second, fixed effects may be correlated with the explanatory variables (Mileva, 2007). Third, the persistence of the tax-to-GDP ratio may give rise to serial correlation, which could be prevented by using GMM estimators
(Benedek et al., 2012).
Although it may seem that GMM estimators solve all potential econometric problems and perhaps invalidate the use of FE estimators, this is not necessarily the case. The GMM estimation also has its limitations; Roodman (2009) has argued that GMM estimators should be used in samples with a high number of observations across a much lower number of entities, because otherwise it will lead to insignificant coefficients. The data sample of this thesis consists of a high number of observations but also of a high number of entities, which may lead to insignificant results. Roodman (2009) has recommended in that case using FE OLS estimators. Therefore, the emphasis of this thesis is on both estimation techniques.
The GMM estimation technique includes two estimators: the difference generalized method of moments (difference GMM) estimator and the system generalized method of moments (system GMM) estimator. The difference GMM estimation technique transforms the equation to a first-difference equation, eliminating the fixed effects. Subsequently, the differenced lagged dependent variable is instrumented with its past levels to remove the endogeneity (Mileva, 2009). The two-step system GMM estimation technique consists of a difference and a level equation, estimating at the same time the system of the two equations. In the difference equation lagged levels of the endogenous variables are used as instruments and in the level equation the lagged differences of the endogenous variables serve as instruments (Roodman, 2009). A difference between the two estimators is that the difference GMM estimator has poor finite-sample properties if series are highly persistent, which might apply to tax revenue data (Benedek et al., 2012). In that situation, the lagged levels are weak instruments for the difference equation because they are weakly correlated with the first differences. Also, in the system GMM estimation, the level equation is included, allowing for cross-country information. Thus, the system GMM estimator is more efficient than the difference GMM estimator, certainly in the case of weak instruments (Roodman, 2009). Therefore, for the empirical research of this thesis the second-step system GMM estimator will be used.
25
When using the system GMM estimation, one needs to determine first which variables are exogenous, which are predetermined, and which are endogenous. The classification of the variables needs to be based on economic theory (Roodman, 2009). This research study follows Benedek et al. (2012) and classifies the variables grants, loans, their squared equivalents, and tax revenues as being endogenous. Furthermore, the variables GDP per capita, imports, exports, trade openness, and agricultural and industrial value added are assumed to be exogenous and not determined by variables within the model. Limiting the number of endogenous and predetermined variables reduces the amount of GMM instruments, which leads to a lower probability of over-identification (Roodman, 2009).
The second step is to decide how many lags will be used for the GMM instruments. The number of lags chosen is based on the results of statistical tests. Using second lagged variables for the difference equation, with or without deeper lags, results in significant second-order serial correlation and low p-values for the Hansen test. A potential explanation for the significant second-order serial correlation could be the high persistence of the tax-to-GDP ratio. Hence, in this thesis third lagged variables are chosen to act as instruments for the difference equation and second lagged differenced variables to serve as instruments for the level equation. One could also choose to use a lag depth of three to four instead of only three. However, using more and deeper lags as instruments reduces the sample size and increases the probability of over-identification of instruments. A rule of thumb regarding the number of instruments is that they should not exceed the number of entities in the sample, meaning that the number of instruments in this context should not exceed 102 which can be prevented by using the
collapse command (Roodman, 2009).
The final step in using system GMM estimators is to perform two tests in order to check the validity of the GMM instruments. The Arellano-Bond test is used for the difference equation to investigate whether there is first-, second- and third-order serial correlation (Mileva, 2009). No first-order correlation is expected in the difference equation, since a lagged dependent variable on the right-hand side of the baseline regression is not included. Including a lagged dependent variable is not necessary when using the xtabond2 command, since it allows for zero lags (Roodman, 2009). Prichard et al. (2014) also argue that using a dynamic model, thus including a lagged dependent variable, is not necessary because of the high persistence of the tax-to-GDP ratio.
26 Besides no first-order serial correlation, there should also be no second- and third-order serial correlation included in the regression, because otherwise system GMM estimators are inconsistent (Roodman, 2009). Thus, the null hypotheses of the Arellano-Bond test regarding autocorrelation of order 2 and order 3, respectively AR(2) and AR(3), should not be rejected (Mileva, 2009). The second test that will be employed is the Hansen test for over-identifying restrictions when using robust errors (Mileva, 2009). The test is used to examine whether the excluded instruments are correctly excluded and whether the instruments are not correlated with the error term. The p-level should be as high as possible, but should not be 1.00 since that would imply that the empirical model is incorrect (Roodman, 2009).
27
4. Data
This chapter elaborates on the data used for the empirical research of this thesis. The first section contains the data sources and the definitions of the variables. The second section on the sample selection describes which countries are included in the sample and why data might be excluded.
4.1. Data description
Data of tax revenues is drawn from the Government Revenue Dataset of the International Centre for Tax and Development. This dataset contains data about government revenues in percentages of GDP for the period 1980 to 2015. The GRD includes data from major databases such as the Governmental Finance Statistics (GFS) database. Before the creation of the GRD database in 2014, the GFS database was used as primary data source for empirical research on tax revenues and aid (Clist, 2016). However, many data is missing from the GFS database, and it was attempted to resolve this by using total government revenues as proxy for tax revenues or by using data from multiple data sources, each of which use a different definition of tax revenues (Benedek et al., 2012). Clist (2016) disagreed with both methods and suggested using the GRD dataset for future research. This dataset contains 70% more observations of developing countries for the period 1990 to 2000 than the GFS, and obviates the need to use proxies or different variables of tax revenues to create a substantial dataset. Another advantage of the GRD database, compared with the GFS database, is that it does not include tax revenues from natural resources. The exclusion of tax revenues from natural resources is important for this thesis, since aid is expected to have behavioural effects and thus only affect non-resource tax revenues (Prichard et al, 2014).
Data about grants and loans are collected from the OECD’s Development Assistance Committee (DAC) statistics (2018), which includes data on the volume and types of aid flows to 150 countries and territories. The variable grants includes the total amount of donations classified as being Official Development Assistance (ODA) received per country per year. The variable loans of this thesis contains the net ODA-loans, that consist of a grant element of at least 25% and are extended by the official sector for the
promotion of economic development.
Data of the control variables of the baseline regressions are drawn from the World Bank’s World Development Indicators dataset (World Bank, 2018a). The
28
variables imports and exports of goods and services, and shares of agriculture and industry value added are already noted in percentages of GDP. The variable GDP per capita is measured in current US dollars for each year and country. Data of the governance variables are also drawn form the World Bank (2018b) and include annual figures from 2002 for six dimensions of governance for each country. For this thesis, three dimensions are used: political stability, government effectiveness, and corruption. The governance variables are constructed by using 31 data sources including surveys of firms and households and evaluations of private companies and public sector organizations (Kaufman et al., 2010). A summary of the descriptions about the variables used in this thesis can be found in Appendix 1.
4.2. Sample selection
To provide an answer to the question whether aid affects tax revenues, a panel dataset of 102 countries for the period 1980 to 2015 is constructed. The sample contains 21 low-income countries, 38 lower-middle-income countries, 32 upper-middle-income countries and 10 high-income countries. Countries were included if they have least one complete data point. Thus, countries were omitted from the sample if there is no data available for one or more of the time series. The absence of data also applies to high-income countries that did not receive any aid during the period 1980 to 2015. However, some the current high-income countries, such as Cyprus, did receive aid until recently and are therefore included in this thesis’ sample. Because of the recent improvements in data measurement and the availability of data on tax revenues and aid, the majority of the countries have complete time series. Another consequence of the increased availability of data on tax revenues is that the sample of this thesis has a much higher number of observations than previous empirical research on aid and tax revenues. The complete list of countries included in this sample can be found in Appendix 2.
One could argue that it is preferable only to include countries that at least have half of the data points available in order to create a more balanced dataset. However, excluding countries on the basis of an arbitrary condition could lead to sample selection bias, since for example low-income countries have fewer complete datasets and excluding those specific countries may affect the empirical outcome of this research. Furthermore, to control whether the data of the sample selection match previous research, it was determined whether the averages of the variables are in line with the
29 averages of the same variables of recent research. This indeed appears to be the case. For example, Prichard et al. (2014) also finds an average grants-to-GDP ratio around 5% and an average loans-to-GDP ratio below 1%. The statistics of the main variables of this thesis are summarized in figure 4.1. From this figure it appears that the average tax-to-GDP ratio is increasing over time while the grants-to-thesis are summarized in figure 4.1. From this figure it appears that the average tax-to-GDP ratio initially increased but decreased from 1991 onwards. The average loans-to-GDP ratio likewise increased first but decreased after 1993 whilst remaining constant since 2008 at a ratio of around 0.35%. More statistics of the sample can be found in Appendix 3.
Figure 4.1: The average tax-to-GDP, grants-to-GDP and loans-to-GDP ratios per year in percentages.
Source: Graph constructed by the author based on data from the ICTD/UNU-WIDER (2018a) and OECD (2018).
For some countries, data was available but needed to be excluded because of measurement inconsistencies and errors. Measurement inconsistencies are found in the tax revenue data of countries that underwent transitions from planned economies with state-owned enterprises to free market economies without those enterprises (ICTD/UNU-WIDER, 2018a). During the period of the planned economy, the revenues of state-owned enterprises were incorporated as tax revenues. Thus, the sudden absence of state-owned enterprises during the transition period is the reason tax revenues in Mongolia decreased by 50% between 1990 and 1991, as can be seen in Figure 4.2. Therefore, the years 1986 to 1990 will be omitted because of measurement inconsistencies. For Albania, there was a noticeable drop in tax revenue during the years 1989 to 1990, as illustrated in Figure 4.2. The other previous state-owned economies for 0 2 4 6 8 10 12 14 16 18 20 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 2012 2014 Tax-to-GDP ratio Grants-to-GDP ratio Net loans-to-GDP ratio
30 which the data prior to and during transition was omitted are Armenia, Azerbaijan, the Kyrgyz Republic, Mongolia, Ukraine and Uzbekistan. Figure 4.2: Tax-to-GDP ratios of Albania and Mongolia per year in percentages. Source: Graph constructed by the author based on data from the ICTD/UNU-WIDER (2018a).
Another reason for removing data is because of measurement errors. In this sample, Zimbabwe and Benin have very volatile data for tax revenues during the periods 1999 to 2015 and 1980 to 1984, respectively (Figure 4.3). Researchers of ICTD/UNU-WIDER (2018a) mention that data are probably not correctly measured in these countries and advice to remove these data from econometric analyses. This advice was heeded for the current research. Figure 4.3: Tax-to-GDP ratios of Zimbabwe and Benin per year in percentages. Source: Graph constructed by the author based on data from the ICTD/UNU-WIDER (2018a). 0 5 10 15 20 25 30 35 Mongolia Ukraine 0 5 10 15 20 25 30 35 Benin Zimbabwe
31
The last remark about data measurement errors is related to data for grants. In 2006 several countries had a substantially high grants-to-GDP ratio compared to other years (figure 4.4). This is due to a sudden increase in the absolute amounts of grants and not related to GDP (OECD, 2018). However, no evidence of the existence of a measurement error could be found, nor evidence that there was a sudden increase in the total number of grants allocated because of drought or other circumstances. Despite the reason for the sudden increase in the grants-to-GDP ratio being unknown, the data points are clearly outliers within their time series and were thus removed from the dataset. Figure 4.4: Grants-to-GDP ratios per year of several countries in percentages.1 1 Excluding the 2006 data points of these thirteen countries did not result in different outcomes 3 To construct the variable grants in percentages of GDP, data of GDP are drawn from the World Bank 0 10 20 30 40 50 60 70 80 Bolivia Burkina Faso Cameroon Ghana Madagascar Malawi Mauritanie Nigeria Senegal Sierrea Leone Tanzania Uganda Zambia
