Faculty of Economics and Business
Master’s Thesis
The Effectiveness of Multilateral Health Aid
The Case of the Global Fund
Tim Ngoc Hai Nguyen (S4072677) Email: t.n.h.nguyen.1@student.rug.nl
Abstract
Multilateral health aid is a thoroughly discussed topic in the studies of economic development. This paper builds upon the literature on the Global Fund’s health aid programme by outlining the general issues concerning health aid paradigms. Our quantitative framework uses a fixed-effects panel-data regression model to measure the impact of health aid on disease burden and deaths related to tuberculosis, HIV/AIDS, and malaria. Among other things, this paper discovers some moderate reducing effects of health aid on the risk of HIV/AIDS and malaria infections, whilst also showing significant reducing effects on deaths related to tuberculosis. Likewise, country-specific characteristics such as income levels play a substantial role in successful grant implementation, whilst government expenditure has no correlation.
Keywords: Health Outcome, Multilateral Health Aid, Global Fund, Disease Burden
2
1. Introduction
In recent past, studies in economic development shifted its focus from monetary goals to utilitarian objectives including improvements in health levels across developing countries. There is a unanimous consensus that better health levels result impact economic growth in numerous ways. For example, it reduces production losses due to worker illness and increases the productivity of adults due to better nutrition (Lustig, 2007). Thus, concepts of multilateral health aid present contributions to this development goal. Yet, the Paris Declaration outlined the goal of improving the effectiveness of health aid, stating that “while the volumes of aid and other development resources must increase to achieve [the Millennium Development Goals], aid effectiveness must increase significantly as well to support partner country efforts to strengthen governance and improve development performance” (OECD, 2005).
In this paper, I will evaluate the effectiveness of multilateral health aid by using the case study of ‘The Global Fund to fight AIDS, tuberculosis and malaria’, or in this paper shortly referred to as ‘the Global Fund’. Hereby, the empirical analysis is based on health and aid data of 128 low- and middle-income countries. Therefore, my paper contributes to the research on multilateral health aid and its effects on health levels in developing countries. More specifically, this paper builds upon a wide range of literature investigating the outcomes of multilateral health aid programmes and adds the Global Fund as a viable case study for our broader understanding on this topic. Based on my research, there are no extensive studies on the impact of the Global Fund on actual health outcomes. Although, plenty of studies show particular country-specific characteristics to be the central determinants for health aid success (Lu et al., 2010). Hence, our analysis adds to this knowledge by recognising the success of grant implementation to be substantially affected by country-specific characteristics. Thus, I state the hypotheses of this paper based on the discussed literature, arguing that the Global Fund’s multilateral health aid positively affects disease burden and deaths related to tuberculosis, HIV/AIDS, and malaria in developing countries. Furthermore, the likelihood of successful grant implementation significantly depends on specific country-specific characteristics.
3
health aid remains weak, mostly revealing correlation rather than causation. Yet, there is no evidence for grant fungibility in our linear probability model.
The remainder of this paper is structured as follows: Chapter 2 outlines the existing literature on the effectiveness of health aid whilst drawing a sensible argumentation line to conclude with the main hypotheses of this paper. Based on the literature, Chapter 3 presents the quantitative framework of this paper. Hereby, the data sources and collection methods are discussed in relation to the quantitative framework. This chapter discusses the background and data collection, grant eligibility and grant disbursement, and country-specific characteristics. I also specify the applied regression model. Subsequently, I discuss the results in Chapter 4 where I apply the findings of our empirical model to provide evidence for our hypotheses. Lastly, Chapter 5 draws the conclusion of the found results.
4
2. Literature Review
The global average incidence rates for HIV, tuberculosis and malaria have substantially declined since the year 2000 (see Figure 1), meaning that the risk of infection for those diseases decreased throughout this century. This can be considered as a testimony to successful development over the past decades. Nonetheless, it can be argued that improvements in health levels can generally be attributed to the significant improvements in global wealth and income (Easterlin, 2000). In contrast, this paper argues multilateral health aid to be a main contributor to this development. Hence, the objective of this paper is to offer an extension to the literature by gaining evidence on the impact of multilateral health aid. Hereby, I use the case study of the Global Fund, an international financing and partnership organisation that aims to “attract, leverage and invest additional resources to end the epidemics of HIV/AIDS, tuberculosis and malaria […]” (The Global Fund, 2016).
Figure 1
Health Outcomes
5
Nonetheless, more recent studies on different aid programmes present more feasible approaches. For instance, there are various studies examining the health impact of Gavi, an aid consortium aiming to expand immunisation across developing countries. Jaupart et al. (2018) apply a quasi-experimental difference-in-differences strategy by evaluating the development of immunisation rates in Gavi-eligible countries compared to non-eligible countries. There are significant, positive effects for Gavi aid on multiple vaccines, although concerns arise in terms of its biasness. In fact, the found causalities may possibly be unrelated or cannot be attributed to Gavi aid because the empirical analysis mainly focuses on countries located at the lowest baseline of vaccination levels (Jaupart et al., 2018). Consequently, the findings display cases with greater potential for coverage expansion. In contrast, Lu et al. (2006a) apply a GMM model for Gavi support, showing significant, positive effects for countries below 65% DPT1
vaccination coverage. Also, an extension to their model includes additional years of data and finds significant, positive effects for countries between 65% and 80% DPT vaccination coverage, although evidence for countries below 65% DPT vaccination coverage is not found (Hulls et al., 2010).
Lu et al. (2006a) and Hulls et al. (2010) examine the health outcomes for aid recipients compared to countries who failed to negotiate a Gavi aid package for mostly unknown reasons. Therefore, whether a country actually receives health aid is largely driven by unobserved country-specific characteristics (Dykstra et al., 2019). In addition, the research done by Lu et al. (2006a) and Hulls et al. (2010) examines only one vaccine supported by Gavi. In comparison, Dykstra et al. (2019) presents a more extensive study using a regression discontinuity analysis and examining a larger number of Gavi-funded vaccines as well as infant and child mortality. Notably, Dykstra et al. (2019) regress a Gavi treatment dummy2 on an indicator of Gavi eligibility in their first stage model, their results measure the effects of country-specific characteristics, such as income levels, on the probability of a country receiving Gavi aid. Furthermore, Gavi aid for cheap, existing vaccines like hepatitis B and DPT moderately raise vaccination rates by single digit margins whilst vaccination rates increasing substantially over time (Dykstra et al., 2019). Likewise, newer and more expensive vaccines show positive effects on vaccination coverage. As their empirical analysis observes the effects for countries on the aid income eligibility threshold, positive effects are likely to be even larger in cases when countries are further located from the threshold, assuming that the effects are smallest near the eligibility threshold (Dykstra et al., 2019).3
A major concern of aid is that allocated funds may not be used appropriately by its recipients; in short, aid fungibility describes the issue whereby health funds are displaced to non-health purposes. Hence, fungibility presents a considerable issue concerning the effectiveness of health aid. Several studies investigate the issue by primarily examining government expenditures in response to received health aid. The issue of fungibility exists in a case study on vaccination aid, although fungibility does not automatically imply welfare loss as they find no evidence for outright wastage, at least from a public finance perspective (Dykstra et al.,
1 Diphtheria, pertussis, and tetanus
6
2019). Instead, aid waste can be separately measured as the excess of aid over and above the total vaccination rate in a given country. This methodology, however, essentially requires specific data on aid flows and vaccine prices.
In contrast, Lu et al. (2010) discover high levels of health aid fungibility by using a systematic analysis of all data sources available for government expenditures on health as agent in developing countries. In fact, health-related foreign aid to governmental sectors has significant, negative effects on domestic health spending but significant, positive effects for health-related aid to non-governmental sectors (Lu et al., 2010). For every US$1 development assistance for health, government health expenditures were reduced by $0.43 to $1.12. The model uses a panel-data model to estimate the association between health spending and GDP, government size, HIV prevalence, debt relief, and development assistance for health to governmental and non-governmental sectors. This implies that health aid could have been entirely displaced into non-health uses, although this finding is contested based on incorrect model identification (Roodman, 2012) and model misspecification problems (Van de Sijpe, 2013). Nevertheless, other case studies show further support for fungibility of development aid (Dejavaran, 1999; Feyzioglu, 1998).
The Global Fund regularly evaluates the aid programme’s performance in order to improve the effectiveness of their aid scheme. Radelet & Siddiqi (2007) investigates the Global Fund’s evaluation mechanism by applying an ordered probit multivariate analysis to link evaluation scores to different characteristics of grant implementation. The Global Fund decides on the continuation of grants based on the country’s evaluation scores (Radelet & Siddiqi, 2007). It is found that, among other things, countries with higher density of doctors as well as higher disease-prevalence score higher in evaluation scores, implying that country-specific characteristics play an important role for grant implementation. Although the results show solely association, not causality, and focus on evaluation scores rather than actual health outcomes. Similarly, Katz et al. (2010) examines the Global Fund’s tuberculosis grants in 88 countries by using stepwise regression models in order to investigate the correlation of grant performance with a range of grant and country characteristics. Notably, a positive correlation is evident between evaluation scores and political stability as well as disease burden (Katz et al., 2010). Yet, another study finds health systems, risk management, and general organisational efficiencies being an important factor for successful grant implementation (Macro International, 2009).
7
the poorest countries would be able to effectively use their resources in order to improve their health structures (Lu et al., 2006b).
Hypotheses
The discussed literature underscores the motivation for our conceptual framework. Based on the literature on the after-effects of Gavi aid, there is a strong argument for multilateral health aid to have a largely positive effects on health outcomes since recipient countries perform usually better in vaccination coverage rates than non-recipient countries. Likewise, parallels can be drawn between Gavi and the Global Fund eligibility criteria. For once, the Global Fund applies similar mechanisms for aid implementation like Gavi. This includes a “focus on ‘national ownership and respect [for] country-led formulation and implementation’; the evaluation of ‘proposals through independent review processes based on the most appropriate scientific and technical standards’; giving ‘due priority to the most affected countries’; and ‘linking resources to the achievement of clear, measurable and sustainable results’”(The Global Fund, 2001; cited in Schmidt-Traub, 2018). In fact, Gavi and the Global Fund are the only institutions, which were reviewed by the UK Multilateral Development Review (DFID, 2016), to operate with an independent review panel, underlining both institution’s integrity, commitment, and transparency.
Still, the differences in health outcome measures must be taken into consideration. Whilst Gavi targets the increase of vaccination coverages in developing countries, the Global Fund focuses on fighting the disease burden of AIDS, tuberculosis, and malaria. Thus, to determine the effectiveness of the Global Fund’s programme, one must examine the impact on the risk of infection for each disease. In addition, the Global Fund also aims to reduce deaths related to AIDS, tuberculosis, and malaria. Therefore, our main hypotheses are as follows:
H1: The Global Fund’s aid programme reduces the risk of infection for HIV/AIDS, tuberculosis, and malaria.
H2: The Global Fund’s aid programme reduces the deaths related to tuberculosis, HIV/AIDS, and malaria.
Yet, country-specific characteristics affecting the Global Fund’s eligibility criteria risks endogeneity in the recipient variable. Therefore, our model controls for income levels, government expenditure as well as various governance indicators. Alternatively, models measuring fungibility presented in the literature are not feasible for our empirical model as aid flows cannot be examined due to data availability. Instead, our model investigates the Global Fund’s aid programme independent from its grant size. Taking this into consideration, fungibility is partially explored since the probability of grant disbursement is regressed on government expenditure. Consequently, our model tests the correlation between government expenditure and the likelihood of receiving grant disbursements. This allows the estimates to detect potential differences in grant disbursements in response to changing government expenditures, in which a negative correlation implies potential fungibility. As a result, our subsidiary hypotheses states as follows:
8
9
3. Data and Methodology
The core premise of our empirical analysis is to uncover the effects of grant reception on disease burden. Notably, I separate disease burden into two components: incidence rates as well as mortality rates. This chapter specifies the quantitative framework and investigates the applied panel datasets. Moreover, the Global Fund’s mechanism on grant implementation is examined at which the eligibility policies are critically reviewed. Due to limited availability on aid flow data, this paper suggests an alternative variable to measure the effect on disease burden whilst also considering potential risks of endogeneity. Hence, I specify a two-stage least squares model in order to control for model identification problems. Furthermore, our framework uses panel data on country-specific characteristics to keep other variables constant for our result. Lastly, this chapter also discusses potential problems concerning fixed effects as well as time-delayed outcomes.
Health Outcomes
The Global Fund actively invests in both preventative measures (e.g. mosquito nets) as well as medical treatments (e.g. antiretroviral therapy) to fight tuberculosis, HIV/AIDS, and malaria. It is therefore sensible to reduce the level of risk for new infections as well as deaths related to each disease. Thereby, I obtain panel data on incidence as well as mortality rates for tuberculosis, HIV/AIDS, and malaria from 2003 onwards by using a database offered by the World Bank. Unlike the model presented by Lu et al. (2010), I use incidence rates instead of prevalence rates for measuring the risk of infection. This is because incidence rates measure
Table 1: Summary Statistics
Variable Observations Mean Std. Dev. Min Max
Incidence Rates
Tuberculosis per 100,000 people 2,252 176.3572 202.4029 0 1280
HIV per 1,000 uninfected population 1,680 1.134696 2.443588 0.01 17.35 Malaria per 1,000 population at risk 1,632 107.1556 153.3027 0 743.5143
Mortality Rates
Tuberculosis per 100,000 people 2,252 21.33549 25.1086 0 157
HIV/AIDS per 100,000 people 2,100 70.79295 167.5762 0.031152 1174.156
Malaria per 100,000 people 2,100 21.33736 44.34047 0 230.7879
Control Variables
ln (GNI per capita) 2,190 7.74443 1.072115 4.70048 9.859013
Government Expenditure in % of GDP 1,928 16.00597 9.974429 0.951747 115.9325
Control of Corruption 2,276 -0.48335 0.650367 -1.86871 1.724581
Rule of Law 2,280 -0.49241 0.681522 -2.60645 1.555118
10
the proportion of persons who develop conditions at a certain time, whilst prevalence rates refer to the proportion of persons who suffer from conditions within a certain time period. In other words, incidence rates represent the risk of newly contracting the disease, whereas prevalence rates indicate how widespread the disease is. Hence, prevalence rates become more impractical when measuring the risk of newly infected patients as it would also include people who suffer from pre-existing conditions.
Notably, tuberculosis incidence rates were measured relative to a population of 100,000 whilst HIV and malaria are measured relative to a population of 1,000 people. The data was acquired by the WHO, UNAIDS, and OECD. The summary statistics are listed in Table 1. At first glance, most data observations are available for tuberculosis which suggests wider global reach of the disease. Moreover, data on HIV and malaria are mostly measured in affected countries, hence fewer observations are available. As a result, the average incidence rate of malaria exceeds the rate of tuberculosis and HIV. This can be figured when considering the unit differences for each disease, hence close attention must be paid when analysing the data. Also, the substantially high standard deviations indicate higher disparities between countries’ health levels.
The World Health Organisation provide a comprehensive dataset on the mortality rates of tuberculosis4. In contrast, I find no comprehensive dataset by the same source on mortality rates for both HIV/AIDS or malaria. Hence, our empirical analysis has to use data provided by the Global Burden of Disease Collaborative Network for HIV/AIDS and malaria5. Note that some might suggest selection bias in the data, however, both institutions collect mortality data through monitoring country civil registration systems. Yet, all mortality data is measured relative to a population of 100,000 people. Similar to incidence rates, the standard deviations for mortality rates are considerably large, thus possibly indicating country-specific differences affecting health treatments and therefore deaths. It might be worth considering, following a similar argument like Jaupart et al. (2018), that countries with substantially high disease burden may naturally display greater improvements due to their, by far, greater potential.
The Global Fund: Eligibility and Disbursement
The Global Fund provides datasets which show all 146 countries with at least one eligible component from 2003 until 2020. Since the Global Fund’s first board meeting, the Technical Review Panel (TRP) evaluates programme performance and sets grant eligibility policies as a result. (The Global Fund, n.d.). The TRP determines a country to be eligible for funding by considering two components: economic capability and disease burden. The country’s economic
capability categorises countries based on their GNI per capita6 which are determined by the World Bank and it is the primary indicator for eligibility. For instance, all lower income and lower-middle income countries automatically qualify for fund eligibility regardless of the disease burden component; although prior to 2014, lower-middle income countries had to demonstrate ‘high’ disease burden to qualify for fund eligibility. Since the most recent changes,
11
only upper-lower middle income countries must show a ‘high’ disease burden to be eligible for grant eligibility. The available funds are distributed through an allocation formula which is further explained in Appendix A. To be classified with ‘high’ disease burden, the Global Fund looks at multiple criteria including prevalence, incidence, and mortality rates, depending on disease. The detailed requirements to be classified as ‘high’ disease burden are illustrated in Appendix B.
Yet, eligibility does not automatically guarantee allocation of funds. The Global Fund maintained their right to refuse an allocation to a country’s component when they have not received existing grants7, when they have never received Global Fund grants, or when a country has successfully transitioned and/or when commitments have been made to ensure domestic financing of the programme (The Global Fund, 2018). In reality, despite Myanmar’s eligibility the Global Fund had withdrawn their grant in 2005 due to newly imposed travel restrictions which limited the grants to be implemented8. Likewise, North Korea faced similar measures to their grant programmes as they did not receive any disbursements until 2010 despite being eligible throughout the time period. Other cases suggest that countries may have refused to apply for funds as they had suffered from relatively low disease burden. Algeria, for example, was eligible for malaria funds from 2003 to 2010, however, did not receive any disbursements in 2003, 2009 and 2010. Similarly, Cabo Verde did not receive malaria disbursements from 2003 to 2009, despite being eligible throughout that time period. Both countries’ malaria disease burden were relatively low throughout that time period and likely qualified for the grant due to their income levels, hence it is a reasonable assumption that those countries refused grants as they appear to be unnecessary.
Although the Global Fund also provides some data on their grant disbursements, they do not offer a comprehensive annual payment dataset. Hence, I use the dataset provided by the OECD which reveals the Global Fund’s total annual payment to each country, showing the total value
of disbursements for each year. Note that I assume any grant disbursement to be equivalent to grant implementation. After merging both datasets, I categorise countries on their recipient status. I make another assumption that if a country was grant eligible for a specific disease at a given year and received a disbursement then I label the country as a recipient for the specific disease in the given year. For example, the Republic of Congo in 2012 was eligible for tuberculosis and malaria grants but was not eligible for HIV/AIDS. According to the data provided by the OECD, Congo had received disbursements in 2012. Therefore, it is concluded that the Republic of Congo was a malaria and tuberculosis grant recipient but was not a HIV/AIDS grant recipient in 2012.
7 Grants which have been signed but not yet disbursed
12
As illustrated in Figure 2, I find 1,478 out of 1,857 eligible cases (80%) for HIV/AIDS funds where the country received the actual grant. Furthermore, 1,308 out of the 1,585 eligible cases (83%) for malaria as well as 1,389 of the 1,761 eligible cases (79%) for tuberculosis had received grant disbursements.
Note that our found disbursement dummy is a binary variable; therefore, it only distinguishes between recipients and non-recipients of the grant. This limits the model to further investigate the effects of differences in the value of aid. Nonetheless, using a simplified binary indicator of grant disbursement instead of the actual value helps our empirical model to provide a more universal answer to whether aid, in general, adds beneficially to health outcomes.
Regression specification
As previously mentioned, I apply a grant disbursement dummy variable instead of specific aid flow data. Primarily, this is due to the limited available data on the Global Fund’s aid flows for each disease. In fact, the Global Fund only provides explicit aid flow data from 2017 onwards. Secondly, aid flow data are significantly more endogenous than a simple dummy variable. This is because grant size highly depend on a country’s disease burden. In essence, highly affected countries receive more aid than marginally affected countries.
Thus, to create a grant disbursement dummy, I generate a variable which takes the value 1 if country 𝑐 has received a grant by the Global Fund, and 0 otherwise. I use time-lagged disbursement dummies to observe effects of grant disbursements considering time-lagged effects. In other words, I assume that grants require a certain time period to take full effect since implementation, application, and health outcomes do not occur immediately at the time of disbursement. Hence, I use 𝑡 − 1 and 𝑡 − 2 in order to investigate the effects of aid after 12 months and after 24 months. I refer to the 1-year lagged disbursement dummy as 𝑅𝑐(𝑡−1) in
the equation.
As previously explained, our model uses a fixed-effects within-estimator to control for country-specific, time-invariant characteristics. This allows our model to remove omitted variable bias as time-invariant variables such as geography, climate, and ethnicity can be accounted for. To test whether fixed-effects within-estimators are more appropriate than a random-effects model,
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Tuberculosis HIV/AIDS Malaria
Figure 2: Grant Recipients vs. Eligible Recipients
13
I test both regressions estimators using a Hausman test (see Appendix C)9, consequently
rejecting the null hypothesis. Therefore, fixed-effects within-estimators are more suitable for this model.
In addition, our model controls for GNI per capita as I expect a negative relationship between disease burden and income levels, meaning that higher income countries have comparatively lower disease burden than lower income countries, and vice versa. The expectation is based on richer countries having larger budgets for health services. Furthermore, the empirical model uses natural log-transformed GNI per capita data as it becomes more convenient for the interpretation of our results. I denote the variable as 𝐼𝑛𝑐𝐿𝑒𝑣𝑒𝑙̈ in the equation, indicating its log transformation.
Likewise, our regression model controls for government expenditure on final consumption, referred to as 𝐺𝑜𝑣𝑆𝑝𝑒𝑛𝑑, which is measured in % of total GDP. This variable allows the model to take efforts of national governments to improve health facilities into account. Similarly, government expenditure is also expected to have reducing effects on disease burden as high government expenditure usually indicates larger efforts of the government to improve their public services which include health services but also indirect factors such as infrastructure, logistics, and education. To avoid multicollinearity between income levels and government spending, I measure government expenditure in percentage relative to total GDP in order to control for effects caused by the size of the economy. This holds the model accountable for overidentification.
Based on the findings by Lu et al. (2006b) on the impact of governance indicators on grant implementation, our empirical analysis also controls for differences in level of corruptions, regulatory quality, and the enforcement of the rule of law. To keep the research feasible, the empirical model solely focuses on control of corruption, regulatory quality, and rule of law as measurements of governance quality. Datasets on governance are available from the World Bank’s Worldwide Governance Indicators database. These indicators measure country-specific governance differences. They approximately range from -2.5 (low) to 2.5 (high) and are further described by Kaufmann et al. (2010). Yet, the indicators seem highly correlated at first sight, therefore potential multicollinearity issues cannot be eliminated. Hence, the interpretation of the results will take this issue into account. The error term is marked as 𝜀.
𝑌𝑐𝑡 = 𝛽0+ 𝛽1𝑅𝑐(𝑡−1)+ 𝛽2𝐼𝑛𝑐𝐿𝑒𝑣𝑒𝑙̈ 𝑐𝑡+ 𝛽3𝐺𝑜𝑣𝑆𝑝𝑒𝑛𝑑𝑐𝑡 + 𝛽4𝐶𝑜𝑟𝑐𝑡
+ 𝛽5𝑅𝑒𝑔𝑄𝑢𝑎𝑙𝑐𝑡+𝛽6𝐿𝑎𝑤𝑐𝑡+ 𝜀𝑐𝑡 (1)
Likewise, I control for reverse causality issues in order to prevent the model to violate the OLS assumption which states that 𝑐𝑜𝑣(𝑥𝑖, 𝜀𝑖) = 0. Keeping this in mind, I previously discussed the hypothesis that countries might refuse grant disbursements if their disease burden is low.
14
Hence, I expect disease burden to reversely affect the disbursement dummy 𝑅𝑐(𝑡−1). Resembling the model presented by Dykstra et al. (2019), I estimate the first stage regression model in Equation (2) using the disbursement dummy as the instrumented variable and an eligibility dummy as the exogenous instrument. Like the disbursement dummy, I also use a time-lags to control for time delays in the outcomes. The eligibility dummy, which I denote as 𝐸, takes the value 1 if country 𝑐 has met the grant requirements set by the Global Fund’s eligibility policy at year 𝑡 − 1 (and 𝑡 − 2), and 0 otherwise. To test whether the model is correctly identified, I look for a high degree of joint significance as well as a high value of R-squared to give more confidence about our first stage model. Furthermore, I apply a Hausman test between the suggested two-stage least squares model and a regular OLS model.
The Global Fund primarily looks at income threshold to determine country’s eligibility rather than levels of disease burden. This emphasises that disease burden only plays a role in certain cases. Therefore, I assume the instrument to be largely exogenous. In other words, eligibility is assumed to be uncorrelated with the error term of the dependent variable. Moreover, it is reasonable to expect the eligibility dummy to have a sufficiently strong correlation with the endogenous disbursement variable and, therefore, it fulfils the relevance criteria.
Note that the first stage regression model uses a binary dependent variable, indicating a linear probability model. Therefore, the first stage results can be separately evaluated as they find evidence the relationship of successful grant implementation on country-specific characteristic; thus, becoming important for our third hypothesis. A linear probability model is preferred over a probit and logit model due to its interpretability. For instance, the results of a linear probability model can be simply interpreted as the change in probability of successful grant implementation, whilst holding all other explanatory variables constant.
𝑅𝑐(𝑡−1) = 𝛼0 + 𝛼1𝐸𝑐(𝑡−1)+ 𝛼2𝐼𝑛𝑐𝐿𝑒𝑣𝑒𝑙̈ 𝑐𝑡+ 𝛼3𝐺𝑜𝑣𝑆𝑝𝑒𝑛𝑑𝑐𝑡+ 𝛼4𝐶𝑜𝑟𝑐𝑡
15
4. Results & Discussion
In this section, the results from our regression models are discussed. Take into account that I present six different columns as I regress for the three different diseases as well as two different time-lagged dummy variables (𝑡 − 1 and 𝑡 − 2). Nevertheless, the relevance of the instruments are explored by using the findings of our first-stage OLS model. Besides from that, the findings suggest a reducing effects of income on the likelihood of receiving a grant. Yet, our model provide limited evidence for grants affecting incidence rates HIV/AIDS and malaria. Despite of that, our empirical model show expected impacts for almost all control variables such as income levels, government spending, and governance indicators. Although control of corruption as well as regulatory quality demonstrate increasing effects on both incidence and mortality rates, opening up gaps for further studies on that matter.
Eligibility and disbursements
The relationship between eligibility and disbursement variable reveal significant estimates for our used exogenous instrument variable as can be observed in Table 2. The F-test values for each column show substantially high values which gives us confidence for overall joint-significance of our estimated first stage model. In addition, the R-squared values in each column suggest an appropriate goodness-of-fit measure for our first stage regression, hinting that our model has been correctly specified (Hill et al., 2012). However, after running a Hausman test to choose between our suggested 2SLS within-estimator model and a fixed-effects OLS model, I cannot reject the null hypothesis (see Appendix C). The test results have been almost identical for each model. Therefore, the differences in coefficients are not systematic and our disbursement dummy variable is exogenous. I change my fixed-effects regression model from a two-stage least squares to a conventional fixed-effects OLS regression model.
Nonetheless, the linear probability model applied in the first stage regression offers extended knowledge about the relationship between successful grant disbursement and grant eligibility. Therefore, I interpret the estimates found from the linear probability model where I find positive, moderate effects of changing income levels on the disbursement dummy variable at a 1% significance level, implying that a 1 percent increase in GNI per capita affects the likelihood of receiving grants by 0.13 to 0.15 percent for tuberculosis, 0.14 to 0.17 percent for HIV/AIDS, and 0.11 to 0.15 percent for malaria. This opposes the contention of low-income countries being more likely to implement health aid funds than richer countries. As previously discussed, viable reasons for this could be based on either the Global Fund refusing grants to low-income countries, as was seen by the example of North Korea and Myanmar, or lower income countries may refuse grants by the Global Fund, although this seems to be more unlikely from a realistic perspective. Nonetheless, the found estimates suggest country-specific differences playing a major factor in successful grant implementations.
16
The findings meet the expectations as all grant recipients were naturally grant eligible. More notably, government expenditure appears to almost no impact on the likelihood of successful grant implementation with only column (1) and (5) showing little significance. This finding adds to the understanding of aid fungibility, suggesting that government expenditure does not have a relationship with grant implementation. In fact, the findings do not provide sufficient evidence for aid fungibility, therefore additional research on this finding can further contribute to this topic.
A two-stage least squares model does not identify to be the most viable regression model, therefore I continue by applying a regular panel-data OLS regression model to estimate the impact of recipients to avoid larger standard errors. However, the linear probability model applied in the first stage provides a feasible model to investigate the likelihood of successful
Table 2: Eligibility and Disbursements
(1) (2) (3) (4) (5) (6)
VARIABLES 𝑅𝑐(𝑡−1)Tuberculosis 𝑅𝑐(𝑡−2)Tuberculosis 𝑅𝑐(𝑡−1)HIV/AIDS 𝑅𝑐(𝑡−2)HIV/AIDS 𝑅𝑐(𝑡−1)Malaria 𝑅𝑐(𝑡−2)Malaria
Ec(t−1)Tuberculosis 0.777*** (0.018) Ec(t−2)Tuberculosis 0.776*** (0.018) Ec(t−1)HIV/AIDS 0.729*** (0.020) Ec(t−2)HIV/AIDS 0.737*** (0.021) Ec(t−1)Malaria 0.805*** (0.015) Ec(t−2)Malaria 0.807*** (0.016) ln (GNI per capita) 0.155*** 0.128*** 0.167*** 0.137*** 0.148*** 0.114*** (0.015) (0.016) (0.017) (0.018) (0.014) (0.015) Government Expenditure in % of GDP -0.004** 0.001 -0.003 0.001 -0.004** 0.000 (0.002) (0.002) (0.002) (0.002) (0.002) (0.002) Control of Corruption 0.049 0.043 0.021 0.007 0.065* 0.065 (0.041) (0.044) (0.044) (0.047) (0.038) (0.040) Rule of Law -0.029 -0.028 -0.024 0.003 -0.029 -0.003 (0.041) (0.044) (0.045) (0.047) (0.039) (0.041) Regulatory Quality -0.062* -0.008 -0.042 0.017 -0.068** -0.043 (0.035) (0.037) (0.037) (0.040) (0.033) (0.034) Constant -1.094*** -0.943*** -1.154*** -0.971*** -1.043*** -0.821*** (0.129) (0.136) (0.141) (0.148) (0.121) (0.125) Observations 1,891 1,891 1,891 1,891 1,891 1,891 R-squared 0.524 0.519 0.434 0.427 0.613 0.616 F test 321.8 315.3 224.8 218 464.4 469.3 Prob>F 0 0 0 0 0 0
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
17
grant implementation. In fact, I find statistically significant evidence for a positive relationship between income level and grant implementation, suggesting that increasing income levels are likely to cause successful grant implementation. However, there is no substantial evidence for government expenditure to influence grant implementation; hence, one might suggest that there is no evidence for aid fungibility as no significant correlation is found between government expenditure and grant disbursement.
Effects of grant disbursement on incidence rates
In Table 3, the findings for our panel-data OLS model are presented where incidence rates are used as our dependent variables. Yet again, the measurement units must be taken into account to correctly interpret the findings; incidence rates of tuberculosis were measured per 100,000 people whilst HIV and malaria incidence rates account for infections per 1,000 people. The estimates of our regression model indicate that countries which received a grant did reduce incidence rates, although the results only find statistically significant estimates for HIV in column (3) as well as weak significance for malaria in column (5). Also, each column shows
Table 3: Fixed-Effects OLS Regression - Incidence Rates (𝒀𝒄𝒕)
(1) (2) (3) (4) (5) (6)
VARIABLES Tuberculosis Tuberculosis HIV HIV Malaria Malaria
Rc(t−1)Tuberculosis -5.184 (3.749) Rc(t−2)Tuberculosis -4.888 (3.534) Rc(t−1)HIV/AIDS -0.139** (0.058) Rc(t−2)HIV/AIDS -0.079 (0.055) Rc(t−1)Malaria -5.453* (2.966) Rc(t−2)Malaria -1.729 (2.849) ln (GNI per capita) -34.685*** -34.622*** -0.395*** -0.403*** -24.280*** -24.520***
18
negative estimates, suggesting reducing effects of grants. However, HIV in column (4) as well as malaria in column (6) show relatively large standard errors for the estimates, implying notable irregularities for the actual coefficients. Additionally, neither estimates for tuberculosis are statistically significant. Therefore, the regression models find only partial support for our initial hypothesis.
However, there are various potential reasons which may explain our findings in Table 3. First of all, it is likely that structures of grant implementation differ between receiving countries, thus outcomes differ significantly from observations. This would also explain the shown large standard errors in our estimates. For example, differences in health systems have not sufficiently been controlled for despite the implementation of governance indicators in our regression model. This is because it is likely that unobserved structural advantages in countries’ health systems play a role in the success of health aid programmes. Therefore, further studies on individual country cases of grant implementation may help to broaden our understanding of the impact of development grants on health outcomes. Secondly, despite using time-lagged variables, the Global Fund’s aid programme may entail a longer timeframe to produce substantial effects on health outcomes. This was, however, not possible as the Global Fund is still a young institution, thus data availability only reaches back to the year 2003. Therefore, I argue that continuous research on the impact of the Global Fund’s aid programme is needed to provide better results. As can also be observed, the standard errors for 𝑡 − 2 lagged dummy variables are smaller than for 𝑡 − 1 time-lagged variables, possibly indicating that true effects become more visible after a longer timeframe. Again, this recommends continuous studies on long-term effects of grants as it may be beneficial for supporting our hypothesis.
Despite this, the control variables provide additional results for our empirical model. As expected, there is significant evidence for income levels to have reducing effects on incidence rates for each disease. The estimated coefficients suggest that for every 1 percent increase in GNI per capita, the incidence rate of tuberculosis decreases by around 0.35 people relative to a 100,000 population. Furthermore, incidence rates for malaria decrease by almost 0.25 people relative to a 1,000 population for every 1 percent increase in GNI per capita, whilst HIV incidence rates decrease by 0.004 people relative to a 1,000 population. The magnitude of the estimates are moderate relative to the average incidence number (see Table 1), however, the estimates are significant at a 1% level and correspond with the initial assumption of the negative relationship between income levels and disease burden. Also, I find that government spending reduces incidence rates for tuberculosis and HIV. This confirms the expectation of health outcomes being significantly affected by government’s efforts. However, the effects of government spending are relatively small, suggesting that other factors play a greater role for reducing the incidence rates.
R-squared 0.100 0.100 0.095 0.093 0.068 0.066
# of obs. countries 128 128 101 101 98 98
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
19
The model finds strong evidence for rule of law reducing incidence rates for tuberculosis and HIV. Remarkably, other governance variables show positive effects on incidence rates. For example, estimates for control of corruption reveal increasing impacts on incidence rates; meaning that more corrupt countries account for lower incidence rates. Similarly, the model also finds regulatory quality positively affecting incidence rates for tuberculosis and HIV. This implies that the ability of a government to formulate and implement beneficial policies leads to an increase in incidence rates for tuberculosis and HIV. The results are statistically significant at a 1% level and leave large gaps for further studies to investigate governance impact on disease burden. One might argue that control of corruption as well as regulatory quality are detrimental to incidence rates due to higher levels of bureaucracy which as a result puts up barriers for effective policy making. Other reasons may be sampling biases in more corrupt countries which lead to underestimated data for disease burden. Further insight about the relationship between disease burden and governance quality might complement this model. Nonetheless, the results could, at least partially, be attributed to high levels of collinearity between governance indicators. As illustrated in Figure 3, it becomes evident that the used governance indicators have strong linear correlation, hence implying a multicollinearity issue. In general, multicollinearity does not affect the fit of the model but it undermines the statistical significance of the particular variables. Hence, I continue to evaluate the results of the model, however, neglect the findings for the governance indicators.
Effects of grant disbursement on mortality rates
20
people relative to a 100,000 population which may seem small but is still a considerable number of people. Although with large standard errors, it is worth mentioning that the effect of grants in each column shows negative coefficients. Again, structural differences in health systems as well as insufficient timeframes may explain our results. In addition, the small number of observations may certainly play a role in the limited findings on the Global Fund’s effectiveness. Hence, studying the impacts of health aid by the Global Fund over a longer time period may certainly help to prove our hypothesis.
Table 4: Fixed-Effects OLS Regression - Mortality Rates (𝐘𝐜𝐭)
(1) (2) (3) (4) (5) (6)
VARIABLES Tuberculosis Tuberculosis HIV/AIDS HIV/AIDS Malaria Malaria
Rc(t−1)Tuberculosis -0.743* (0.383) Rc(t−2)Tuberculosis -0.954*** (0.361) Rc(t−1)HIV/AIDS -9.138* (5.375) Rc(t−2)HIV/AIDS -6.031 (5.105) Rc(t−1)Malaria -1.080 (0.761) Rc(t−2)Malaria -1.170 (0.722) ln (GNI per capita) -6.040*** -6.005*** -42.176*** -42.471*** -9.077*** -8.983***
(0.353) (0.353) (4.956) (4.955) (0.727) (0.732) Government Expenditure in % of GDP -0.274*** -0.272*** -3.021*** -2.984*** -0.226*** -0.225*** (0.040) (0.040) (0.565) (0.565) (0.083) (0.083) Control of Corruption -0.433 -0.366 46.243*** 45.947*** 5.412*** 5.475*** (0.940) (0.940) (13.416) (13.423) (1.974) (1.975) Rule of Law -4.272*** -4.356*** -71.660*** -71.301*** -6.653*** -6.714*** (0.955) (0.955) (13.872) (13.879) (2.045) (2.046) Regulatory Quality 2.710*** 2.755*** 28.619** 28.870** 0.034 0.051 (0.799) (0.798) (11.532) (11.539) (1.697) (1.696) Constant 71.199*** 71.041*** 451.412*** 451.064*** 96.461*** 95.779*** (2.904) (2.902) (40.848) (40.867) (6.010) (6.016) Observations 1,869 1,869 1,755 1,755 1,755 1,755 R-squared 0.196 0.197 0.086 0.085 0.114 0.114 # of country 128 128 127 127 127 127
Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
21
Nonetheless, the estimates for the control variables provide supplementary findings. The model finds strong evidence for a negative relationship between income and mortality rates for each disease at a 1% significance level. More specifically, a 1 percent increase in GNI per capita reduces tuberculosis mortality rates by 0.06 people, HIV/AIDS mortality rates by 0.42 people, and malaria mortality rates by 0.09 people relative to a 100,000 population. Again, this is consistent with the assumption of a reverse relationship between income levels and disease burden. In addition, Table 4 indicates a significant reverse relationship between government spending and mortality rates for each disease. In essence, tuberculosis mortality decreases by 0.3 people for every added percentage point of government spending relative to GDP. Similarly, malaria mortality decreases by 0.2 people and HIV/AIDS mortality decreases by around 3 people for every added percentage point of government spending. This meets our expectations as increases in government spending often link with national efforts to improve health levels.
22
5. Conclusion
This paper aimed to find statistically significant evidence to show the effectiveness of the Global Fund’s multilateral health aid programme to support the claimed hypotheses which were based on the academic literature on the impacts of health aid. In fact, I argued that multilateral health aid reduces the disease burden of tuberculosis, HIV/AIDS, and malaria. Then I specified further, arguing that health aid also reduces the deaths related to each of the three diseases. I expected to find statistically significant evidence showing recipient countries to perform better in disease burden than non-recipient countries. This was argued based on the literature (Dykstra et al., 2019, Jaupart et al., 2018; Lu et al., 2006b) which analysed similar health aid programmes with positive results.
This paper’s findings makes contributions to the research on health aid programmes. Also, it provides some evidence for our initially stated hypotheses. Firstly, our empirical model offers a handful of evidence which describe the relationship between successful grant implementation and grant eligibility as well as income levels. However, more noteworthy, I find no correlation statistical relationship between government spending and grant implementation, thus suggesting no evidence for aid fungibility. On the other hand, the Global Fund’s health aid is found to likely reduce the risks of HIV/AIDS and malaria infections, although statistical significance was relatively weak. Furthermore, our empirical model finds health aid to significantly reduce deaths related to tuberculosis, whereby both findings show relatively moderate magnitude of their impact. Besides, the results for the applied control variables matches the expectations as income levels as well as government spending substantially reduce the risk of infections (incidence rate) as well as deaths (mortality rate) related to the diseases. In contrast, the results for control of corruption, rule of law, and regulatory quality are rather inconclusive. For instance, control of corruption as well as regulatory quality show increasing effects on disease burden whilst rule of law show reducing effects. Hence, I suggest multicollinearity to be a main contributor to these findings.
In terms of our empirical analysis, the initially expected endogeneity between eligibility and disbursement variable is revealed to be not correctly identified. Nonetheless, the linear probability model contributes to the understanding of the third hypothesis. For example, I find income levels to be a significant factor of successful grant implementation, suggesting that higher income countries are more likely to receive aid funds than lower income countries. In fact, the results were statistically significant at a 1% level for each disease. In addition, the model does not find statistical support for government expenditure to affect grant implementation, although some values show slight statistical significance. Nevertheless, the estimates show neglectable coefficients for the effects of government spending.
23
24
6. Bibliography
Begovac, J., Gedike, K., Lukas, D., & Lepej, S. Z. (2008). Late presentation to care for hiv infection in croatia and the effect of interventions during the croatian global fund
project. Aids and Behavior, 12(Suppl. 1), 53. https://doi.org/10.1007/s10461-008-9398-9 Car, J., Paljärvi Tapio, Car, M., Kazeem, A., Majeed, A., & Atun, R. (2012). Negative health system effects of global fund's investments in aids, tuberculosis and malaria from 2002 to 2009: systematic review. Jrsm Short Reports, 3(10), 1–14.
https://doi.org/10.1258/shorts.2012.012062
De Jongh, T. E., Harnmeijer, J. H., Atun, R., Korenromp, E. L., Zhao, J., Puvimanasinghe, J., & Baltussen, R. (2014). Health impact of external funding for hiv, tuberculosis and malaria: systematic review. Health Policy and Planning, 29(5), 650–662.
https://doi.org/10.1093/heapol/czt051
Devarajan, S., Rajkumar, A. S., & Swaroop, V. (1999). What does aid to Africa finance?. Washington (DC): World Bank.
DFID. (2016). Raising the Standard: The Multilateral Development Review 2016. Department for International Development, London.
Dykstra, S., Glassman, A., Kenny, C., & Sandefur, J. (2019). Regression discontinuity analysis of gavi's impact on vaccination rates. Journal of Development Economics, 140, 12– 25. https://doi.org/10.1016/j.jdeveco.2019.04.005
Easterlin, R. (2000). The Worldwide Standard of Living since 1800. The Journal of
Economic Perspectives, 14(1), 7-26. Retrieved January 4, 2021, from
http://www.jstor.org/stable/2647048
Feyzioglu, T., Vinaya, S., & Zhu, M. (1998). A panel data analysis of the fungibility of foreign aid. The World Bank Economic Review, 12(1), 29–58.
Hill, R. C., Griffiths, W. E., & Lim, G. C. (2012). Principles of econometrics. John Wiley & Sons
Hulls, D., Venkatachalam, P., Kumar, K., Cochrane, T., Kaur, H., Gulati, N., Jones, D. (2010). Second GAVI Evaluation Report. CEPA, LLP, London.
Jahn, A., Floyd, S., Crampin, A. C., Mwaungulu, F., Mvula, H., Munthali, F., … Glynn, J. R. (2008). Population-level effect of hiv on adult mortality and early evidence of reversal after introduction of antiretroviral therapy in malawi. Lancet (London, England), 371(9624), 1603–11. https://doi.org/10.1016/S0140-6736(08)60693-5
Jaupart, P., Dipple, L., & Dercon, S. (2019). Has gavi lived up to its promise? quasi-experimental evidence on country immunisation rates and child mortality. Bmj Global
Health, 4(6), 001789. https://doi.org/10.1136/bmjgh-2019-001789
25
International Journal of Tuberculosis and Lung Disease : The Official Journal of the International Union against Tuberculosis and Lung Disease, 14(9), 1097–103.
Kaufmann, Daniel, Aart Kraay and Massimo Mastruzzi (2010). "The Worldwide Governance Indicators: Methodology and Analytical Issues". World Bank Policy Research Working Paper No. 5430 (http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1682130)
Lu, C., Michaud, C. M., Gakidou, E., Khan, K., & Murray, C. J. (2006a). Effect of the global alliance for vaccines and immunisation on diphtheria, tetanus, and pertussis vaccine
coverage: an independent assessment. Lancet (London, England), 368(9541), 1088–95. Lu, C., Michaud, C. M., Khan, K., & Murray, C. J. (2006b). Absorptive capacity and disbursements by the global fund to fight aids, tuberculosis and malaria: analysis of grant implementation. Lancet (London, England), 368(9534), 483–8.
Lu, C., Schneider, M. T., Gubbins, P., Leach-Kemon, K., Jamison, D., & Murray, C. J. L. (2010). Public financing of health in developing countries: a cross-national systematic analysis. The Lancet, 375(9723), 1375–1387. https://doi.org/10.1016/S0140-6736(10)60233-4
Lustig, N. (2007). Investing in health for economic development: The case of Mexico. In Advancing Development (pp. 168-182). Palgrave Macmillan, London.
OECD (2005), Paris Declaration on Aid Effectiveness, OECD Publishing, Paris, https://doi.org/10.1787/9789264098084-en.
Radelet, S., & Siddiqi, B. (2007). Global fund grant programmes: an analysis of evaluation scores. Lancet (London, England), 369(9575), 1807–1813. https://doi.org/10.1016/S0140-6736(07)60818-6
Roodman, D. (2012). Doubts about the evidence that foreign aid for health is displaced into non-health uses. The Lancet, 380(9846), 972–973.
https://doi.org/10.1016/S0140-6736(12)61529-3
Samb, B., Evans, T., Dybul, M., Atun, R., Moatti, J. P., Nishtar, S., … Polman, K. (2009). An assessment of interactions between global health initiatives and country health
systems. Lancet, 373(9681), 2137–2169.
Schmidt-Traub, G. (2018). The role of the technical review panel of the global fund to fight hiv/aids, tuberculosis and malaria: an analysis of grant recommendations. Health Policy and
Planning, 33(3), 335–344. https://doi.org/10.1093/heapol/czx186
The Global Fund to Fight AIDS, Tuberculosis, and Malaria (2016). Bylaws of the Global
Fund to fight AIDS, tuberculosis & malaria. The Global Fund. Retrieved January 14, 2020,
from https://www.theglobalfund.org/media/6007/core_globalfund_bylaws_en.pdf The Global Fund to Fight AIDS, Tuberculosis, and Malaria (2017). The Global Fund
Strategy 2017-2022 – Investing to end epidemics. The Global Fund. Retrieved September 5,
26
The Global Fund to Fight AIDS, Tuberculosis, and Malaria (2018). The Global Fund
Eligibility Policy. The Global Fund. Retrieved September 23, 2020, from https://data-service.theglobalfund.org/downloads
The Global Fund to Fight AIDS, Tuberculosis, and Malaria. (n.d.). Technical Panel Review. The Global Fund. Retrieved November 27, 2020, from
https://www.theglobalfund.org/en/technical-review-panel/
Van de Sijpe, N. (2013). The fungibility of health aid reconsidered. Journal of Development
Studies, 49(12), 1746–1754. https://doi.org/10.1080/00220388.2013.819424
WHO (2020). Global Tuberculosis Report. World Health Organization. Retrieved January 19, 2021, from
27
7. Appendix
Appendix A
28
Appendix C
Appendix D
(V_b-V_B is not positive definite) Prob>chi2 = 0.0000
= 31.29
chi2(6) = (b-B)'[(V_b-V_B)^(-1)](b-B) Test: Ho: difference in coefficients not systematic
