The Effects of Development Aid on Infant Mortality and Life Expectancy

(1)

1

The Effects of Development Aid on

Infant Mortality and Life Expectancy

Does Governance Quality Matter?

Author : Steven Hommes Student Number : 1286528

Date : August, 27th, 2007

Email-address : stevenhommes@hotmail.com

(2)

2

ABSTRACT

(3)

3

1 INTRODUCTION

1.1 Background and Context of the Thesis

The purpose of development aid is to permanently improve a recipient country’s economic growth and “quality of life”, which distinguishes it from humanitarian aid which merely aims at a temporarily relief of human suffering. The idea behind development aid has been around for quite a long time. The first major project that can be seen as development aid funded, was the well known Marshall Plan (named after U.S. Secretary of State George Marshall) which aimed at the economic reconstruction of Western Europe after World War II. According to McGillivray et al. (2006), the Marshall Plan was widely considered to be a great success, and therefore it was a promising first assessment regarding the effectiveness of development aid.

Subsequently, the focus of development aid switched to the so called “third world” or

“developing” countries (at first often for strategic purposes during the Cold War) where it has stayed up till this day. Substantial amounts of aid have flown for many years to these

countries, but positive effects with respect to economic growth and “quality of life” have often been much more difficult to detect. With respect to the effectiveness of development aid, Doucouliagos & Paldam (2005d) describe two particular pieces of evidence that seem illogical:

1. Contrasting country stories: There seem to be massive differences in the effectiveness of aid throughout aid receiving countries. South Korea, for instance, received substantial aid for about a decade whereupon its economy went into a high growth phase. On the

contrary, Tanzania has been a major aid recipient for 40 years, and has showed little economic growth throughout the entire period. Likewise, 30 years of considerable aid receipts in Zambia go together with an unusually poor economic performance.

(6)

6

absolute aid effectiveness. Since Mosley (1986), the contrast has been known as the micro-macro paradox.

One of the most often proposed reasons for the apparent lack of absolute aid effectiveness is the argument that development aid may be fungible, as found by Mansoob Murshed & Sen (1995), Feyzioglu et al. (1998) and McGillivray & Morrissey (1999). Among donors of development aid there is a widespread fear that aid intended for poverty reduction may in fact be financing the military, as found by Collier & Hoeffler (2007), or enriching the political elite. The apparent lack of absolute aid effectiveness has stimulated the amount of research and debate enormously, which has resulted in a large collection of empirical macro studies on the effectiveness of development aid.

The large majority of the research that has been done on the effectiveness of development aid has focused on the effects of aid on growth in per capita income, either direct or indirect through the effect of aid on another factor which in turn influences growth in per capita

income (usually savings or investment). As Kosack (2003) argues, rising income undoubtedly is an important feature of development, but it should certainly not be the only criterion by which the effectiveness of aid is judged. He emphasizes that the effectiveness of aid should also be judged by its ability to improve the “quality of life”.

Only recently have researchers begun to determine the effectiveness of aid by means of its effect on the “quality of life”. This approach was pioneered by Boone (1996), who among other things assessed the effects of development aid on the “quality of life” indicators infant mortality, life expectancy and primary school enrollment.

1.2 Statement of Purpose and Contributions of the Thesis

(7)

7

to identify the direct effects of aid as well as the effects of “the interaction between aid and governance quality” on infant mortality and life expectancy. In order to achieve this, a

regression model has been developed with which relevant data on 47 developing countries for the 1990-2001 time period has been analyzed.

The effectiveness of aid by means of its impact on infant mortality and life expectancy has not often been examined before. The effects of development aid on infant mortality have, based on my literature research, previously only been examined in Boone (1996), Mosley et al. (2004), Gomanee et al. (2005), Gomanee, Girma & Morrissey (2005) and Verschoor & Kalwij (2006). The effects of development aid on life expectancy have, based on my

literature research, so far only been assessed in Boone (1996). What is particularly innovative about this study is that it not only examines the direct effect of development aid on infant mortality and life expectancy, but also tries to determine whether a country’s “quality of governance” has an influence on the outcome. The degree to which development aid is fungible, and therefore less effective, is likely to be influenced by a developing country’s “quality of governance”. In this study, a country’s “quality of governance” will be determined by a constructed governance quality index which captures aspects of democracy, stability and institutional quality. The importance of the latter is emphasized in Rodrik et al. (2004).

1.3 Structure of the Thesis

The remainder of this thesis is organized as follows. In section 2, the existing literature on the effectiveness of development aid is discussed and divided into three groups. The literature on the determinants of a country’s health level is also explored in this section. Section 3

(8)

8

2 LITERATURE REVIEW

2.1 Different Types of Aid Effectiveness Studies

In order to develop an appropriate model for this research with which I am capable of achieving the goals stated in the introduction, I explored the different models that have been used in the aid effectiveness literature. Several different types of studies on the effectiveness of development aid can be identified. In order to bring some structure to the landscape of different studies in the aid effectiveness literature, I adopt a classification that was proposed by Doucouliagos & Paldam (2005d).

Doucouliagos & Paldam (2005d) thoroughly investigate the aid effectiveness literature and argue that at that time the aid effectiveness literature had reached approximately 100 studies of which the empirical models used can be divided into the following three families:

1. Accumulation studies: The ones where aid is expected to cause an accumulation in savings and investment which in turn increases growth.

2. Growth studies: The ones where aid has a direct effect on growth.

3. Conditional growth studies: The ones where the effect of aid on growth depends on certain conditions.

Some of the studies in the aid effectiveness literature use only models from one of the three families, but several other use models from more than one of these three families.

2.1.1 Family 1: Accumulation Studies

According to Doucouliagos & Paldam (2005d), the accumulation studies of the impact of aid on savings or investment is the oldest type of the three families, with the first paper by Rahman (1968), but it has continued to this day. According to Doucouliagos & Paldam (2005a) and McGillivray et al. (2006), the basic reasoning behind this family of models comes from the ”two gap” theory which states that investment which leads to economic growth is restricted by both domestic saving and the import capacity. Doucouliagos & Paldam (2005a) mention that in the studies where the impact of aid on investment is examined, development aid is seen as a device that allows the foreign goods and services balance to turn negative (and thus increases the import capacity) in order to allow investment, and

(9)

9

that in the studies where the impact of aid on savings is examined, development aid is seen as a device that may increase savings which in turn boost investment, and economic growth as a consequence.

According to Doucouliagos & Paldam (2005a), the studies in this family are based on models of the following types:

(a) Iit = β0 + βaidAIDit +

∑

=

m j1

βcon,jCONit,j + εit for investment models

(b) Sit = β0 + βaidAIDit +

∑

=

m j1

βcon,jCONit,j + εit for savings models

Where (it) is the index to countries and time, (Iit) is the investment proportion of GDP, (Sit) is

the savings proportion of GDP, (AIDit) is the variable that measures the amount of received

aid and (εit) indicates the random disturbances. For a set of m control variables, (CONit,j)

indicates the jth control variable. All β’s are the coefficients to be estimated.

This family of the aid effectiveness literature has produced several different results. Virtually all of the early studies including Rahman (1968), Griffin (1970), Griffin & Enos (1970), Gupta (1970) and Weiskopf (1972) find that there exists either no relationship or a negative relation between aid and the domestic saving rate (Sit). This peculiar result is explained by

Boone (1996) when he finds that aid is ineffective in increasing investment (Iit) because aid

increases public consumption by the same amount (which thus leaves nothing for savings). Controversely, more recent studies such as Lensink & Morrissey (2000), McGillivray (2000) and Collier & Dollar (2004) find that aid does have a robust positive effect on investment (Iit)

which in turn stimulates economic growth.

2.1.2 Family 2: Growth Studies

(10)

10

direct investment). Most of the literature in this family uses GDP per capita growth as the dependent variable but, as will be discussed extensively in section 2.2, more recently, just as in this study, researchers have begun to switch the dependent variable with several quality of life indicators.

According to Doucouliagos & Paldam (2005b) the studies in this family are based on models of the following type:

(c) Git = β0 + βaidAIDit +

∑

=

m j1

βcon,jCONit,j + εit

Where (Git) indicates the (economic) growth rate.

Also this family of the aid effectiveness literature has produced different results. The different papers conclude that there is either a positive effect of aid on growth or that there is no effect of aid on growth. Some of the papers that belong to the group that finds aid to have a positive an significant direct effect on growth are Papanek (1973), Dowling & Hiemenz (1982), Gupta & Islam (1983) and Economides et al. (2004). Some of the papers that belong to the group that finds aid to have no effect on growth are Mosley (1980), Voivodas (1973), Mosley et al. (1987), Boone (1996) and Ram (2003). Most & van den Berg (1996) are somewhat in between and find aid to have a direct effect on growth in several countries, but also find it to be ineffective in others.

2.1.3 Family 3: Conditional Growth Studies

This family of literature differs from the growth studies family by adding certain conditions to the relationship between aid and growth. In this family, not only can the direct effect of aid on growth be assessed (as in the growth studies), but the effect of aid on growth also depends upon an additional factor (Zit). When this additional factor is favorable, aid has a positive

(11)

11

Generally, the studies in this family are based on interaction models of the following type:

(d) Git = β0 + βaidAIDit + βzZit + β1AIDitZit +

∑

=

m j1

Where (Zit) is the variable on which the effectiveness of aid is conditional.

In the above described interaction type model, (Zit) is the variable on which the effectiveness

of aid is conditional. Several different variables for (Zit) have already been used throughout

this family of the aid effectiveness literature.

By far the most used measure for the variable on which the effectiveness of aid is conditional (Zit), is a measure for a country’s policy environment, leading to the so called “good policy

model” which was popularized in the landmark World Bank policy research report:

“Assessing Aid, What Works, What Doesn’t, and Why” (1998). The “good policy model” is described in detail in Burnside & Dollar (2000). They use a policy index, which is a weighted sum of the budget balance, the inflation rate and trade openness, as the variable on which the effectiveness of aid is conditional. Among the studies in which aid is conditional on policy are Collier & Dehn (2001), Dalgaard & Hansen (2001), Hansen & Tarp (2001), Guillaumont & Chauvet (2001), Lensink & White (2001), Hudson & Mosley (2001), Collier & Dollar (2002), Islam (2002), Ovaska (2003), Dalgaard et al. (2004), Burnside & Dollar (2004), Easterly et al. (2004), Ram (2004) and Tashrifov (2005). Results among these studies are quite mixed, although the majority of these studies (just over 2/3) rejects the hypothesis that aid effectiveness is conditional on good policies.

Other measures for the variable on which the effectiveness of aid is conditional (Zit), have

(12)

12

A special measure for (Zit), the variable on which the effectiveness of aid is conditional, is aid

itself. When aid itself is used as the condition, the model reduces to:

(e) Git = β0 + βaidAIDit + β1AIDit2 +

∑

=

m j1

This model is known as the “medicine model”. Just as medicine works if it is given in

moderation and harms if it is taken in excess, the model assumes diminishing marginal returns to aid which become negative at some point. According to Doucouliagos and Paldam (2005c), this model was first proposed in Hadjimichael et al. (1995). Studies in which aid is

conditional on itself include Dalgaard & Hansen (2001), Hansen & Tarp (2001), Hudson & Mosley (2001), Lensink & White (2001), Lu & Ram (2001), Collier & Dollar (2001), Collier & Dollar (2002), Islam (2002), Ovaska (2003) and Dalgaard et al. (2004). The majority of these studies (about 80%) support the hypothesis that aid is effective with diminishing marginal returns.

2.2 Quality of Life Indicators as the Dependent Variable

As mentioned, the majority of the aid effectiveness literature has focused on the effects of aid on growth in GDP per capita, either directly (as in the growth studies), through increased savings and investment (as in the accumulation studies) or when the effect of aid is

conditional on something else (as in the conditional growth studies). Kosack (2003) argues that rising income is certainly an important feature of development, but that it is not the only important feature, and is therefore not the only criterion by which aid effectiveness should be judged. He argues that aid effectiveness should be judged by its ability to improve overall life quality. For this reason, researchers, using both growth and conditional growth type models, have more recently started to change the dependent variable for several “quality of life” indicators. Because this study looks at the effects of development aid on two “quality of life” indicators, infant mortality and life expectancy, I explored this literature.

(13)

13

wealthy political elite. He also notes however that these results may be biased since he is aggregating a range of political regimes. Therefore he conditions for political regimes and tests whether different political regimes use aid differently. His regression results reject his hypothesis that democratic/liberal regimes use aid differently, and indicate that under each regime there is no significant impact of aid on “quality of life” indicators. Boone (1996) does conclude however that liberal political regimes, ceteris paribus, have approximately 30% lower infant mortality than the least free regimes.

Kosack (2003) switches the dependent variable to the change in the Human Development Index (HDI). The HDI is an index on the “quality of life” which has been developed by the United Nations, and incorporates life expectancy, infant mortality, school enrollment and literacy. Kosack (2003) judges aid by its ability to improve a country’s “quality of life”, as measured by the HDI and assumes the effectiveness of aid to be conditional on

democratization. He argues that there still is a considerable correlation between GDP per capita and “quality of life” as measured by the HDI. In his data, the correlation coefficient between GDP per capita and the HDI is 0,69 and between GDP per capita and growth in the HDI it is 0,53. Kosack (2003) argues however that there are also considerable differences between the two, the GDP per capita measure may be unlikely to capture the full effect of aid (the purpose of which, after all, is often not only to increase economic growth, but also to improve health and education). The major result of his paper is that aid improves “quality of life”, as measured by the change in the HDI, in democratic countries and is ineffective (and possibly harmful) in autocracies. Furthermore, he finds that the results are not the same when GDP per capita growth is used as the dependent variable instead.

(14)

14

Recently, a collection of papers has emerged which emphasize that there are two different ways in which aid can have an influence on “quality of life”, or welfare as they call it.

Gomanee et al. (2005) argues that aid can have both a direct and an indirect effect on welfare. The direct effect is that aid can generate income-earning opportunities or provide social services which can both increase aggregate welfare. The indirect effect is that aid is directed through government spending, and aid may increase government spending on social sectors which increases aggregate welfare.

Gomanee et al. (2005) use both the infant mortality rate and the Human Development Index as the dependent variable (and thus as a measure of aid effectiveness). In their model they also add a measure of government expenditure on social sectors to the independents. The primary result of their study is that the direct effect of aid on welfare is indeed present. A significant coefficient on their aid variable signals that aid has either a direct effect on welfare, or that it increases welfare through economic growth. Gomanee et al. (2005) do not find any evidence for the indirect channel of aid effectiveness which assumes aid to be effective by increasing government spending on social sectors.

On the contrary, both Mosley et al. (2004) and Gomanee, Girma & Morrissey (2005) find that the indirect effect, that aid impacts on “quality of life” through the effect of aid on

government spending in social sectors, is significant. Furthermore, neither study finds that aid impacts directly on the “quality of life”. Mosley et al. (2004) uses both the infant mortality rate and the poverty headcount ratio, as estimated by the percentage of a country’s population living on less than one dollar per day, as the dependent variable. Gomanee, Girma &

Morrissey (2005) use either infant mortality or the HDI as the dependent variable. Using quantile regressions they find that aid is associated with higher human development (as measured by the HDI) and lower infant mortality and that aid is more effective in countries with a low “quality of life”.

(15)

15

2.3 Factors Influencing Health Levels

A country’s health level is a major indicator for a country’s “quality of life”. Infant mortality and life expectancy are two measures of a country’s health level. Because this study looks at the effects of development aid on infant mortality and life expectancy, I examined the literature on the determinants of a country’s health level in order to find out which factor’s other than aid potentially affected my dependent variable (that is infant mortality or life expectancy).

Gbesemete & Johnson (1992) argue that differences in health levels can be explained as the outcome of a composite of factors, namely socio-economic, demographic, medical,

environmental and political factors. They focus attention on factors that have featured most prominently in the literature and sum up these factors are as follows:

A) Socio-economic factors:

• Education: In several studies education is pointed out as one of the most powerful predictors of mortality. An educated population seems to be more receptive to medical, sanitary and nutritional information which leads to increased longevity. The educational level of females is particularly vital to the health of infants.

• Water supply and sanitation: Improved water supply and sanitation is very effective against bacillary dysentery, cholera and other diarrhoeal diseases. Hence, it is assumed that the higher the proportion of the population with an acceptable standard of water supply, the lower will be the level of mortality.

• GNI per capita: The most general measure of resource availability is the gross national income (GNI) per capita. Numerous studies have found that as incomes decrease, health levels tend to deteriorate.

• Percentage of the population with access to health care: A substantial number of scholars have analyzed the relationship between various measures of health and medical services and found positive effects.

(16)

16

• Economic social standing: Economic social standing is determined by Gbesemete & Johnson (1992) as the average of ranks for GNP per capita, educational status of the population and level of health. It is supposed to have a positive effect on the health level. • Health care expenditure: The level of health care expenditures determines the level of

health care services and the purchase of medical equipment. Therefore health care expenditure has a positive effect on the health level.

• Diet and nutrition: Malnutrition is one of the most prominent causes of mortality and diet and nutrition is thus an important determinant of the health level.

• Food aid: Food aid can help to compensate for the loss of food in times of war and drought and thus contributes to the health level.

B) Demographic factors:

• Fertility: It is argued that infant deaths increase with ascending number of births and shorter birth intervals. With shorter birth intervals, less care is given to the children. Furthermore, high fertility may lead to competition for the limited resources available to a household.

• Density: High population density is deleterious to health insofar as it aggravates problems of sanitation and facilitates disease transmission. On the other hand, it is argued that the effectiveness of health service inputs varies inversely with the dispersion of the

population.

C) Medical factors:

• Physicians, nurses and hospital beds: When the stock of physicians increase and the workload decreases doctors may spend more time on each patient than otherwise. A similar argument holds for nurses. An increased number of hospital beds may lead to more admissions in hospitals and thereby the diagnosis and cure, provided there are drugs and equipment in the hospital, of potential killers.

D) Environmental and Political factors:

(17)

17

After identifying the factors that have featured most prominently in the literature on health levels, Gbesemete & Johnson (1992) re-evaluate two models in which infant mortality is explained by socio-economic, demographic, medical, environmental and political factors. They find that the GNP per capita, school population as a percentage of the population under 19, population density and the percentage of the population with access to health care together explain 80% of the variations in infant mortality in their sample study.

The literature on the determinants of infant mortality (one of the measures of a country’s health level) in developing countries, produces several different findings. In contrast to Gbesemete & Johnson (1992), Younger (2001) finds that only two variables, primary school enrolments and DPT vaccination rates for infants show any consistent correlation with declining infant mortality, and even those correlations are not robust to the inclusion of fixed effects. Frey & Field (2000) find that only economic disarticulation (which is present when the various sectors of a country’s economy are disconnected and unevenly developed), female education, debt dependence, and Sub-Saharan African status have the expected effects on infant mortality. Hojman (1996) finds that to reduce infant mortality, the most influential variables are the vaccination coverage, the incidence of low weight at birth, female education, and the number of nurses in relation to the population size.

2.4 Literature Review Conclusions

1. From the existing literature, it can be concluded that it is not possible to make a definitive statement as to whether development aid is effective or not. The landscape of different aid effectiveness studies produces several conflicting results.

2. With respect to the different aid effectiveness models, Doucouliagos & Paldam (2005d) demonstrate that they can generally be divided in three broad categories: accumulation models, growth models and conditional growth models.

(18)

18

(19)

19

3 RESEARCH QUESTIONS

From the literature on the effectiveness of development aid follows that the effectiveness of development aid may either be unconditional (as assumed in the growth studies) or

conditional (as assumed in the conditional growth studies). As mentioned in the introduction, this study tries to identify a relationship between development aid receipts and two health related “quality of life” indicators, infant mortality and life expectancy. Furthermore this study tries to find out whether a country’s “quality of governance”, through the interaction of aid and governance quality, influences the outcome. In other words, this study looks at both the unconditional and the conditional effect (where aid effectiveness is conditional on “quality of governance”) of development aid on infant mortality and life expectancy. This resulted in the following general and specific research questions.

3.1 General

Does development aid and/or “the interaction of development aid and quality of governance” have an influence on infant mortality growth and/or life expectancy growth?

3.2 Specific

1a. Can a causal relationship between development aid and infant mortality growth be identified?

1b. Can a causal relationship between “the interaction of development aid and quality of governance” and infant mortality growth be identified?

2a. Can a causal relationship between development aid and life expectancy growth be identified?

(20)

20

4 THEORETICAL MODEL

4.1 Model Construction

4.1.1 General

In order to answer the general and specific research questions, a regression model will be developed. With respect to the three different families of aid effectiveness models, as identified by Doucouliagos & Paldam (2005d), the most promising type of model for this study seems to be the conditional growth model. This type of model will be most effective in answering my research questions, because it allows for both the direct effect of aid as well as the interaction effect of aid and governance quality on infant mortality and life expectancy to be assessed. Therefore, I develop a conditional growth type regression model, in which the variable on which the effectiveness of aid is conditional (Zit) is a constructed “quality of

governance” index.

The model building process is divided into two steps. First, a basic conditional growth type model, including governance quality as the variable on which the effectiveness of aid is conditional (Zit), will be developed. Subsequently, a set of 10 control variables, of which 7 are

related to aid effectiveness in general and 3 are specific for health related “quality of life” indicators as the dependent variable, will be added to the model.

4.1.2 Basic Model

The model building process starts at the most basic interaction type regression model that is used in the family of conditional growth studies on the effectiveness of development aid:

(1) AEit = β0 + βaidAIDit + βzZit + β1AIDitZit +

∑

= 10

1

j

Where (it) is the index to countries and time, (AEit) is the dependent variable that is used as a

proxy for aid effectiveness, (AIDit) is the variable that measures the amount of received aid,

(Zit) is the variable on which the effectiveness of aid is conditional and (εit) indicates the

random disturbances. For the set of 10 control variables, (CONit,j) indicates the jth control

(21)

21

In this study the aid effectiveness variable (AEit) will be either infant mortality growth

(INFMit) or life expectancy growth (LIFEXit). The variable on which the effectiveness of aid

is conditional (Zit), will be “quality of governance” (GQit). This variable is measured by an

index which was constructed from a set of governance indicators that have been collected from the World Bank, and will capture six key dimensions of governance (voice and accountability, political stability, government effectiveness, regulatory quality, rule of law and control of corruption). In including the governance quality variable (GQit) as the variable

on which the effectiveness of aid is conditional (Zit), the model now becomes as follows:

(2) AEit = β0 + βaidAIDit + βgqGQit + β1AIDitGQit +

∑

= 10

1

j

Where (GQit) is the “governance quality” variable and (AIDit*GQit) is the aid and

“governance quality” interaction variable.

4.1.3 Control Variables

Following Burnside & Dollar (2000) and Kosack (2003), I assume that the effectiveness of aid is influenced by macroeconomic policy. Burnside and Dollar (2000) develop a policy index that is used as the interacting variable in their model, leading to the described “good policy model”. Kosack (2003) treats several macroeconomic policy variables as controls in his model however, and this is the approach I will follow in this study.

The influence of macroeconomic policy on aid effectiveness is estimated through three variables in Burnside & Dollar (2000) and Kosack (2003). The three variables are as follows: 1. A dummy variable measuring the openness to international trade, developed by Sachs and

Warner (1995). Closed economies are ones that have average tariffs on machinery and materials above 40 percent, or a black market premium above 20 percent, or pervasive government control of key tradables.

2. The inflation rate as a measure of monetary policy, as proposed by Fischer (1993).

(22)

22

Following Kosack (2003), it was my initial intention to include all these variables as controls in the model, but unfortunately the datasets on countries’ budget balances that were available to me (mostly from the World Bank and the IMF) had too many missing values which would result in dropping too many countries from the data set. Therefore the budget surplus/deficit control variable has not been added to my model. The openness to trade measure and the inflation rate have been added to the set of control variables however.

I also follow Kosack (2003) by including a terms of trade measure to the controls. The terms of trade variable is calculated as the change in the ratio of the export price index to the import price index and is used to control for changes in the international economy (particularly changes in primary commodity prices, on which many of the economies of developing countries are based).

I also decided to include a dummy variable for sub-Saharan Africa into my regression equation to capture some of the influence of geography and natural endowments on the effectiveness of foreign aid. I chose to use a sub-Saharan dummy because both Burnside & Dollar (2000) and Kosack (2003) have shown aid to be less effective in sub-Saharan Africa than elsewhere in the world.

Following Burnside & Dollar (2000) and Kosack (2003), I also assume that the initial value (at the beginning of the period of which the growth rate is measured) of the aid effectiveness indicator is influential on the observed growth rate, and therefore this variable has to be included in the regression equation.

It is possible that specific circumstances in specific time periods had an influence on the aid effectiveness indicator (AEit) in my model. In order to capture the possible impact of specific

time periods on the effectiveness of aid, fixed time effects dummy variables are added to the model. As will be explained in section 5.1, this study uses three time periods, and therefore there will be two fixed time effects dummy variables added to the regression equation, one for the first and one for the second time period.

Since the dependent variable (AEit) in my model is either infant mortality growth (INFMit) or

life expectancy (at birth) growth (LIFEXit), the model is also extended with some health level

(23)

23

country’s health level has produced a whole range of different findings as to which variables are the most significant. Based on this literature, by using variables that have proved to be significant in these studies, and on data availability, I included a variable for DPT

immunization rates, a variable for population density and a variable for female education to the regression model.

Adding the control variables to the model yields the final model for infant mortality growth and life expectancy (at birth) growth as a proxy for aid effectiveness, which now looks as follows:

(3) AEit = β0 + βiIVit + βaidAIDit + βgqGQit + β1AIDitGQit + βotOTit + βinfINFit + βtotTOTit +

βssdSSDit + βimmIMMit + βdsyDSYit + βfedFEDit + βtd1TD1it + βtd2TD2it + εit

Where (OTit) is the openness to trade dummy variable, (INFit) is the inflation rate variable,

(TOTit) is the terms of trade variable, (SSDit) is the sub-Saharan dummy variable, (IMMit) is

the DPT immunization variable, (DSYit) is the population density variable, (FEDit) is the

female education variable and (TD1it) and (TD2it) are time dummy variables for the first and

second time period. Finally, (IVit) is the initial value of the variable that is used as a proxy for

aid effectiveness, this is either (IVINFMit) for the infant mortality model or (IVLIFEXit) for

(24)

24

4.2 Hypotheses

Based on the literature review, research questions were developed in section 3. In this section, hypothesis corresponding to the proposed research questions will be developed.

It is my assumption that development aid as well as the interaction of aid and governance quality have the capacity to improve health related “quality of life” indicators such as infant mortality and life expectancy. Therefore infant mortality growth and life expectancy (at birth) growth, should display corresponding relationships with these variables. Assuming that these relationships are indeed present, the following hypotheses corresponding to the proposed research questions have been developed:

1a. H0: There exists no relationship between development aid and infant mortality growth.

Ha: There exists a negative relationship between development aid and infant mortality

growth.

1b. H0: There exists no relationship between “the interaction of development aid and

quality of governance” and infant mortality growth.

Ha: There exists a negative relationship between “the interaction of development aid

and quality of governance” and infant mortality growth.

2a. H0: There exists no relationship between development aid and life expectancy growth.

Ha: There exists a positive relationship between development aid and life expectancy

growth.

2b. H0: There exists no relationship between “the interaction of development aid and

quality of governance” and life expectancy growth.

Ha: There exists a positive relationship between “the interaction of development aid

(25)

25

5 METHODS AND DATA

5.1 General

In order to estimate my regression model I decided to use the ordinary least squares (OLS) approach. A problem which is associated with this approach in the literature, is a potential correlation between the aid variable and the error term. That is, it is possible that aid is correlated with a number of other variables which also have an influence on the aid

effectiveness measures in the model, or that the aid effectiveness measures have an influence on received aid. This potential problem can be resolved by using a two stage least squares (2SLS) or instrumental variable procedure instead, but this methodology is beyond the scope of this research.

Several studies have attempted to determine the extent to which aid is correlated with the error term, by using both OLS as well as 2SLS techniques and subsequently comparing the results. To this end, the outcomes of these studies have been mixed. Hansen & Tarp (2001), Lensink & White (2001) and Kosack (2003) find no evidence for a correlation between aid and the error term while other studies such as Guillaumont & Chauvet (2001), Dalgaard & Hansen (2001) and Easterly et al. (2004) find small but significant differences between the OLS and 2SLS estimates for the aid variable or its composite variables.

In order to estimate the model using OLS regressions, data on the model variables had to be collected from several sources. Data sources that have been used are the World Bank’s World Development Indicators (WDI) database, the World Bank’s Governance Indicators database and the updated Sachs & Warner (1995) openness to trade dataset, which has been updated by Roodman (2007) and has been provided by the Center for Global Development. This resulted in a dataset1 composed of time series data for a sample of 47 developing countries. The 1990-2001 time period has been used for model estimation. The twelve year period has been divided into three 4-year periods: 1990-1993 (period 1), 1994-1997 (period 2) and 1998-2001 (period 3). Therefore, this leaves 3 × 47 = 141 observations for model estimation.

1

(26)

26

5.2 Sample Construction

In order to construct the sample of developing countries that will be used to estimate my model, the following steps have been taken.

I started off with all 147 countries in the world that the World Bank classifies as developing countries. Subsequently, I excluded all countries from the sample for which aid receipts data were not or only partially available for the 1990-2001 period which is used in model

estimation. This decreased the sample to 100 countries. Then I excluded all countries that on average received very low amounts of aid (under 1%) relative to their GNI during the 1990-2001 time period that is used for estimation. This is based on the logic that aid needs to have at least a reasonable impact on a country’s GNI per capita if it is going to be of any effect. By the exclusion the sample was reduced to 78 countries.

Several other countries had to be excluded from the data set, because data on either the openness to trade, the inflation rate or the terms of trade variable were not sufficiently available. This resulted in the final sample of 47 developing countries that has been used in order to estimate the models. In the final sample, 30 countries are sub-Saharan whereas there are only 48 sub-Saharan countries in the original 147 countries which the World Bank classifies as developing countries. This contrast is caused by missing data and the fact that sub-Saharan developing countries on average received much more aid relative to their GNI than developing countries elsewhere in the world. A list of the developing countries that are included in the final sample can be found in appendix A.

5.3 Variable Construction

5.3.1 Development Aid Receipts

The data for the development aid receipts variable (AIDit) has been collected from the World

(27)

27

my estimate for a country’s received aid, because differences between the two indicators will be small and data availability for the “aid receipts as a percentage of GNI” indicator was better. The aid receipts variable in my model is constructed by averaging a country’s aid receipts as a percentage of GNI over the years of a time period that is used for the model estimation. So for instance, for the first time period (1990-1993), the aid receipts variable is calculated as the average aid as a percentage of GNI for the years 1990, 1991, 1992 and 1993.

5.3.2 Governance Quality

The governance quality variable (GQit) is measured by a “quality of governance” index which

has been constructed from a set of governance indicators that have been collected from the World Bank. Its worldwide governance indicators capture six key dimensions of governance in over 200 countries and have been measured biannually since 1996 and annually since 2002. The worldwide governance indicators project defines governance as the set of traditions and institutions by which authority in a country is exercised. The political, economic and

institutional dimensions of governance are captured by the following six aggregate indicators:

1. Voice and Accountability: the extent to which a country’s citizens are able to participate in selecting their government, as well as freedom of expression, freedom of association, and a free media.

2. Political Stability: perceptions of the likelihood that the government will be destabilized or overthrown by unconstitutional or violent means, including domestic violence and terrorism.

3. Government Effectiveness: the quality of public services, the quality of the civil service and the degree of its independence from political pressures, the quality of policy

formulation and implementation, and the credibility of the government’s commitment to such policies.

4. Regulatory Quality: the ability of the government to formulate and implement sound policies and regulations that permit and promote private sector development.

(28)

28

6. Control of Corruption: the extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as “capture” of the state by elites and private interests.

All of the six governance indicators are indexed on a scale from -2,5 (very poor governance) to +2,5 (very good governance) which, like Kosack (2003) did with his democratization index, I converted to a scale from 0 to +5. Subsequently, I added the values for the six

indicators for each country in my sample to determine its overall “quality of governance” (this means that I consider all six indicators equally important). This combined value is used as a measure for the governance quality variable in the model. In order to calculate the values for the interaction variable (AIDit*GQit), the “quality of governance” measures were equated with

the average aid receipts as a percentage of GNI for each of the time periods used in estimation.

The governance indicators have only been measured since 1996, which is unfortunate since I’m using the 1990-2001 period in order to estimate my model. By looking at the governance indicators dataset, it is found however that the value for the governance index for a certain country does not change by a lot over time. I calculated the amount of change in a sample country’s value for the governance index between 2000 and 2006, this because it is a 6-year period, just like the 1990-1996 period for which data are missing. Between 2000 and 2006 the average absolute percentage change in a sample country’s value for the governance index was only 4,1% relative to the index, with a standard deviation of 3,2%. Therefore I assume that the value for the governance index for a sample country does not change by a lot over time. For this reason I decided to use the 1998 values throughout the entire 1990-2001 time period, which I use for estimation.

5.3.3 Infant Mortality and Life Expectancy

Data for the infant mortality variable (INFMit) and the life expectancy variable (LIFEXit) has

(29)

29

measure is for 2004, since the 2005 data were not yet available in the version of the database that was available.

With respect to infant mortality or life expectancy as the dependent variable (AEit), I look at

the effect that the independent variables have on the percentage change in these measures. In other words, the dependent variable is either the infant mortality growth rate (INFMit), or the

life expectancy growth rate (LIFEXit). In order to determine the infant mortality growth rate

and the life expectancy growth rate, I calculate the percentage change in these indicators during a five year period (or the four year period 2000-2004, since 2005 data were not yet available). For instance, for the 1990-1995 time period, the infant mortality variable is calculated as follows:

(i) INFMi1 = (Inf. Mort. 1995 – Inf. Mort. 1990) / (Inf. Mort. 1990)) * 100%

With respect to the construction of the dependent variable in my model, like Kosack (2003), I use near-future infant mortality growth and life expectancy growth. This is based on the logic that aid spending will take some time to have a noticeable effect on infant mortality and life expectancy, because it takes some time for development aid receipts to contribute to

improvements in vital underlying factors, such as hygiene, health facilities and the amount of physicians and nurses. Therefore, for period 1 (1990-1993) the infant mortality or the life expectancy growth rate for 1990-1995 is used as the dependent variable. Likewise, for period 2 (1994-1997) the infant mortality or the life expectancy growth rate for 1995-2000 is used as the dependent variable and for period 3 (1998-2001) the infant mortality or the life

expectancy growth rate for 2000-2004 is used as the dependent variable. By using this methodology, a time lag of 2 years is assumed for period 1 whereas a time lag of 3 years is assumed for period 2 and 3. Furthermore, a 5-year change in infant mortality and life expectancy is used as the dependent variable for period 1 and 2, whereas a 4-year change is used for period 3. Therefore, this methodology is not ideal, but it had to be used because of non-perfect data availability.

(30)

30

measure at the beginning of a time period. So for instance, the 1990 infant mortality value is used as the value for the initial value variable (IVit) for the first time period.

5.3.4 Control Variables

5.3.4.1 Openness to Trade

Data for the openness to trade dummy variable (OTit) has been collected from the updated

Sachs & Warner (1995) dataset which has been provided by the Center for Global

Development. In the Sachs & Warner (1995) dataset, closed economies are ones that have average tariffs on machinery and materials above 40 percent, or a black market premium above 20 percent, or pervasive government control of key tradables. The Sachs & Warner (1995) paper is written in 1995, so it contains no data on the openness to trade for 1996-2001. However, the Sachs & Warner (1995) dataset has been updated by Easterly et al. (2004) to include years up to 1998, and has been updated once more by Roodman (2007) to include years up to 2001. The dummy variable values for the countries in my sample have been collected from the Roodman (2007) dataset, which has been collected from the Center for Global Development. If a country was open to international trade during one of the time periods used in model estimation, it receives the value 1, and if it was closed the value 0 was given.

5.3.4.2 Inflation Rate

Data for the inflation rate variable (INFit) has been collected from the World Bank’s World

Development Indicators (WDI) database. I used the GDP deflator estimates as a measure of the inflation rate for the countries in my sample. The inflation rate variable (INFit) is

constructed by averaging a country’s inflation rate (as estimated by the GDP deflator) over the years of a time period that is used for the model estimation.

5.3.4.3 Terms of Trade

Data for the terms of trade variable (TOTit) has been collected from the World Bank’s World

(31)

31

Base year for the time series data is 2000, which gets the value 100 in every country. The terms of trade values for all other years are relative to this base year.

In order to be of any potential value for this research we need to look at the change in the terms of trade indicator. The terms of trade variable (TOTit) in my model is therefore

calculated as the average percentage change in the terms of trade indicator over the time period to be estimated. For instance, for the first time period (1990-1993), the terms of trade variable is calculated as the percentage change in the 4-year average value as follows:

(ii) TOTi1 = ((Av. value 1990-93 – Av. value 1986-89) / (Av. value 1986-89)) * 100%

I chose to use the change in the year average terms of trade instead of the change in the 4-year point estimates. I chose this method because using point estimates has the potential of under- or overestimating the 4-year terms of trade change in the case of high yearly volatility in a country’s terms of trade. Using the 4-year average change instead, reduces this

undesirable effect.

There were 2 cases in which data was not available for all of the four years of a specific time period. In those instances I took the average terms of trade indicator value of the years within that period that were available to me.

5.3.4.4 Sub-Saharan Dummy

As mentioned, the Sub-Saharan dummy variable (SSDit) simply measures whether a country

is geographically positioned south of the Sahara desert. On the World Bank’s website I found a list of developing countries that are considered to be south of the Sahara. All countries in my sample that are sub-Saharan receive the value 1, whereas all other countries in the sample receive the value 0 for this dummy variable.

5.3.4.5 DPT Immunization

(32)

32

model is constructed by averaging a country’s DPT immunization rate over the years of a 4-year time period that is used for model estimation.

5.3.4.6 Population Density

Data on population density has also been collected from the World Bank’s World

Development Indicators (WDI) database. For a country, it measures the average number of people per square kilometer. The population density variable (DSYit) in my model is

constructed by averaging a countries population density over the years of a 4-year time period that is used in model estimation.

There was 1 case in one country in which data was not available for all of the four years of a specific time period. For that case, I took the average population density for the years within that period that were available to me.

5.3.4.7 Female Education

The construction of the female education variable (FEDit) was more difficult since data was

not widely available and a lot of observations were missing. Likewise, the data on female education has been collected from the World Bank’s World Development Indicators (WDI) database. I decided to use the ratio of female to male in primary school enrollment (in %) as a measure for female education because the dataset for that indicator, as compared to other female education indicators, had the least missing values.

(33)

33

Altogether, after constructing my dataset on female education, 8 values were still missing. This means that by including the female education variable (FEDit) in the regressions, the

total number of observations is reduced from 141 to 133.

5.4 Descriptive Statistics

In order to explore the model’s variables before starting with the regressions, I first examined some descriptive statistics on the dataset that will be used in model estimation. The

descriptive statistics can be found in table 1.

Table 1: Descriptive statistics on the dataset

INFM LIFEX IVINFM IVLIFEX AID GQ AIDxGQ

Mean -7,1361 0,2712 85,4085 54,2903 12,1375 11,4901 132,8720 Std. Deviation 10,61021 5,53382 38,30897 10,42632 11,44491 3,32773 124,7484 Minimum -34,00 -24,81 16,10 31,17 0,61 1,38 5,06 Maximum 48,00 29,14 191,00 73,59 62,47 19,61 750,77 25 -14,4858 -1,3712 52,5000 45,3868 3,2365 9,7485 34,2430 50 -6,6667 1,5883 85,0000 53,1426 10,3445 12,5871 114,5436 Percentiles 75 0,0000 3,0039 114,0000 64,2005 16,1074 13,4540 173,8804

OT INF TOT SSD IMM DSY FED

Mean 0,5745 117,5490 -4,7300 0,6383 65,7979 90,9857 84,8481 Std. Deviation 0,49619 671,5708 16,50797 0,48221 21,73094 145,3174 15,17274 Minimum 0,00 -2,16 -44,52 0,00 15,25 2,05 44,66 Maximum 1,00 7014,81 56,15 1,00 97,75 980,79 107,08 25 0,0000 4,1090 -14,2307 0,0000 49,0000 13,9248 72,2191 50 1,0000 9,3869 -4,1551 1,0000 70,0000 47,0591 89,4956 Percentiles 75 1,0000 17,8144 3,1432 1,0000 84,5000 99,8514 97,5291

Number of Observations (N) is 133 for FED and 141 for all other variables INFM = Change in the Infant Mortality Rate (in %)

LIFEX = Change in Life Expectancy (in %) IVINFM = Initial Value of Infant Mortality IVLIFEX = Initial Value of Life Expectancy

AID = Aid as a percentage of GNI

GQ = Quality of Governance

AIDxGQ = Aid and Quality of Governance Interaction

OT = Openness to Trade Dummy

INF = Inflation Rate (in %)

TOT = Change in the Terms of Trade (in %)

SSD = Sub-Saharan Dummy

IMM = DPT Immunization Rate (in %)

DSY = Population Density

FED = Rate of Female to Male in Primary School Enrollment (in %)

(34)

34

Since the mean of the change in infant mortality (in %) is substantially negative (-7,1361), it can be concluded that on average infant mortality in the sample of countries has decreased during the three time periods that are examined. Furthermore, 75 % of al observations demonstrate a decrease in infant mortality as can be seed from the 75th percentile. Life expectancy in the sample of countries has increased on average, although only slightly, by 0,2712 percent (as can be seen from its mean) during the three time periods.

On average, the developing countries in the sample demonstrated aid receipts which were 12,1 % of their GNI during the three time periods, but this ranges from 0,61 % in Indonesia in period 2 to 62,5 % in Guinea-Bissau in period 2. With respect to the governance quality index (which has a scale from 0 to 30), the country which scores the lowest is Zaire (1,38), and the country which scores the highest is Botswana (19,61). The mean of the openness to trade dummy variable is 0,57, which indicates that in just over half of the observations a sample country was open to international trade. The mean of the percentage change in the terms of trade is negative (-4,73), which indicates that on average the terms of trade has deteriorated in the sample countries during the three time periods. DPT immunization rates for infants were 65,8 % on average in the sample countries during the three time periods, but they range from 15,25% in Chad in period 1 to 97,75 % in Sri Lanka in period 3.

It can also be seen from table 1 that most variables have a mild standard deviation relative to their interquartile range (which is the difference between the 25th and the 75th percentile). Exceptions are the population density variable and especially the inflation rate variable. The relatively large standard deviation of the population density variable is caused by large

(35)

35

6 RESULTS

The procedure that has been used in order to test my hypotheses is as follows. I first estimated both the infant mortality model and the life expectancy model while using the full dataset. Subsequently, in order to determine whether the obtained regression results are sensible, a robustness analysis has been done. In this robustness analysis, the following topics have been treated: multicollinearity, outliers, leverage points and influential observations, normality of the regression residuals and homoscedasticity. This leads to adjusted regressions of which the results are used to test my hypotheses.

6.1 Infant Mortality

6.1.1 Initial Estimation

I started by estimating the infant mortality model using the full dataset. The regression results of this estimation can be found in table 2. As can be seen from the results, the model has substantial explanatory power, R squared is 0,616. The significant coefficients are the ones for the aid variable (significant at the 10 % level), the governance quality variable (significant at the 10 % level), the interaction of aid and governance quality variable (significant at the 1 % level), the openness to trade variable (significant at the 5 % level), the sub-Saharan dummy variable (significant at the 1 % level) and the female education variable (significant at the 10 % level).

The coefficient for the aid variable is positive, which is the opposite sign from the one

expected. Although barely significant, the coefficient for the governance quality variable also carries the opposite sign from the one expected. These are odd results which therefore

emphasize the necessity of a robustness analysis. The coefficient for the interaction between aid and “quality of governance” variable is negative, which is the expected sign (the

(36)

36

variable has the opposite sign from the one expected. It is strange that the initial regression results indicate that increased female to male primary school enrollment has a positive effect on infant mortality growth.

Table 2: Initial infant mortality model regression results

Variable Coefficients t –Statistic Significance

(Constant) _-30,135 _-3,732 _0,000 IVINFM -0,040 -1,114 0,268 AID 0,340 1,933 0,056 GQ 0,578 1,679 0,096 AIDxGQ -0,048 -2,970 0,004 OT -3,748 -2,270 0,025 INF 0,001 0,750 0,455 TOT 0,015 0,389 0,698 SSD 20,379 9,559 0,000 IMM 0,031 0,672 0,503 DSY -0,007 -1,148 0,253 FED 0,113 1,945 0,054 TD1 0,102 0,062 0,951 TD2 -0,830 -0,527 0,599

Dependent Variable: INFM

Number of observations (N): 133

R-Squared: 0,616

6.1.2 Robustness analysis

Several topics will be discussed in this robustness analysis. These are multicollinearity,

outliers, leverage points and influential observations, normality of the regression residuals and homoscedasticity. I will first discuss all these topics before re-estimating the infant mortality model in section 6.1.3.

6.1.2.1 Multicollinearity

For the purpose of this robustness analysis, it was first investigated whether multicollinearity among the independent variables is potentially distorting the regression results. In order to determine if correlations among the independent variables do indeed exist I have developed a correlation matrix for the infant mortality and life expectancy models which can be found in appendix B. When looking at the correlation matrix, it is easily discovered that the aid

(37)

37

In order to determine whether multicollinearity is indeed present and potentially distorting the regression results, I looked at two collinearity diagnostics, the tolerance and the variance inflation factor, which according to Cohen et al. (2003) can be used to determine the degree of multicollinearity. Both indicators are calculated by first treating the specific independent variable as the dependent variable and executing a regression on all other independent variables. The tolerance is calculated subsequently by taking 1-R2 for the regression results. The tolerance is thus a measure of to which extent an independent variable’s variation is independent from the variation in the other independent variables. A low value of a variable’s tolerance indicates that the independent variable under consideration is to a large extent a linear combination of the other independent variables. The variance inflation factor is simply the reciprocal of the tolerance. Therefore, when it is high there is high multicollinearity and possible instability of the model coefficients. According to Cohen et al. (2003) a commonly used rule of thumb is that a tolerance value of 0,10 or less (or correspondingly, a variance inflation factor of 10 or more) provides evidence of serious multicollinearity involving the corresponding independent variable.

It is found that the aid variable and the interaction variable both show a low tolerance (0,92 and 0,93 respectively) and a high variance inflation factor (10,868 and 10,728 respectively). When applying the mentioned rule of thumb, these values indicate the presence of serious multicollinearity. Although the multicollinearity between the aid variable and the interaction variable was somewhat expected because of the construction of the interaction variable, it may be wise also to estimate the model after excluding one of the two variables. This means that in order to test the hypotheses, the regression model has to be estimated three times for every dependent variable, the first time including both variables, the second time excluding the interaction variable from the model and the third time excluding the aid variable from the model. There appear to be no other cases of high multicollinearity among the independent variables in the infant mortality model.

6.1.2.2 Outliers, Leverage Points and Influential Observations

(38)

38

substantially. Outliers are observations of which the value for the dependent variable is unusually distant from the value which is predicted by the model. The standardized residual value of an observation is a good measure to assess if an observation is a potential outlier. The leverage of an observation is determined by how far the values of the independent variables are from their mean values. The leverage of an observation can be determined by the so-called leverage statistic. Finally, influential observations are observations that change the regression coefficients substantially. The degree of influence of an observation is determined by both the leverage and the difference between a dependent’s actual and predicted value. Therefore it can be seen as the product of leverage and outlierness. According to Cohen et al. (2003) two types of measures of influence are commonly used. First, global measures of influence provide information about how a particular observation affects overall characteristics of the regression equation. Second, specific measures of influence provide information about how a particular observation affects each individual regression coefficient.

Cohen et al. (2003) mention that a good measure to determine whether an observation is influential in a global sense is Cook’s distance measure (or Cook’s D). Cook’s distance measure determines which effect deleting a particular observation has on the total set of regression coefficients. According to Fox (1991), observations with a Cook’s D value over 1 can be considered extremely influential, and observations of which the Cook’s D value exceeds 4/(n-k-1), where n denotes the number of observations and k denotes the number of independent variables in the model, can be considered potentially influential.

(39)

39

I found one observation that could be identified as an outlier. For the purpose of this analysis I consider all observations with a standardized residual value of over +3 or under –3 to be a potential outlier. I found one observation beyond these borders which had a standardized residual of +5,613. Upon closer inspection I found that the outlier was caused by a 48 % increase in the infant mortality rate in Botswana between 1995 and 2000. A 48 % increase in infant mortality in five years is indeed unusually large. According to Foster (1998), the escalation of the AIDS crisis in Africa is directly responsible for this extreme result. Foster (1998) argues that parental death caused by AIDS is the most prominent reason for the deterioration of the infant mortality rate. Furthermore, it is argued that HIV infection

increases susceptibility to tuberculosis which led to a large outbreak of the disease in the area and infants are particularly vulnerable to tuberculosis. The reason that the effects of the AIDS epidemic escalation were much more visible in the change in the infant mortality rate in Botswana than in other countries in the area is that Botswana was a relatively prosperous country. Life expectancy was initially much higher and infant mortality was initially much lower than in other sample countries in the area. GDP per capita in Botswana was in 1995 on average almost 7 times higher than in sample countries in the area such as Angola, Zaire, Mozambique, Zambia and Zimbabwe. The escalation of the AIDS crisis brought infant mortality rates and life expectancy in Botswana more in line with the area. This sudden change caused the unusually large increase in infant mortality in Botswana during the 1995-2000 time period. Therefore I decided to exclude this observation from the data set.

For the purpose of identifying high leverage and influential observations, first of all Cook’s distance measure has been calculated for all observations. The results showed no cases of extreme influential observations (where Cook’s D > 1). There were 7 cases however where the value of Cook’s distance measure exceeded the absolute value 4/(n-k-1) = 4/(133-13-1) = 0,0336 which indicates a possible problem. The Cook’s D values for these 7 cases can be found in table 1 of appendix C. Furthermore, I found 2 cases with undue leverage (I consider leverage statistic values of over 0,5 to be excessive), both corresponding to one of the high Cook’s D values.

(40)

40

the mean. This explains their influential status at least to a certain extent. For Rwanda (period 1), this was accompanied with an exceptionally large change in the terms of trade (30,2%) and a high population density (265,7 people per square km), for Bangladesh (period 3) this was accompanied by an exceptionally high population density ( 980,8 people per square km) and for Botswana (period 1 and 3) this was accompanied by an exceptionally high governance quality (the highest in the sample) and a very low population density (2,6 and 3,1 people per square km respectively). These factors may also contribute to the influential status of these observations. The influential status of Pakistan (period 3) is more difficult to explain but it may be caused by a low female to male primary education rate (only 62,7 %), low aid receipts (1,7% of GNI) and a relatively high population density (177,0 people per square km).

In order to finalize the search for potential influential observations I calculated standardized dfbeta values for all observations of the aid variable and the interaction variable in the dataset, and identified those which exceeded the absolute value 2/√N = 2/√133 = 0,173. Using this technique, I found 4 cases of possible high influence for the aid variable and 4 cases of possible high influence for the interaction variable (of which 3 correspond to cases for the aid variable). The standardized dfbeta values for these cases can also be found in table 1 of appendix C.

The observations with standardized dfbeta values which exceed the absolute value 2/√N for both the aid variable and the interaction variable are Botswana (period 1 and 3) and Rwanda (period 2). Rwanda (period 2) had extremely large aid receipts (49,0 % of GNI) and as a result also large values for the interaction variable (334,0 and 370,5 respectively). The influential status of Botswana (period 1 and 3) can be explained to an extent by large increases in infant mortality combined with low aid (3,2 % and 1,1% respectively). The observation for which only the standardized dfbeta value for the aid variable exceeds this threshold is Guinea-Bissau (period 2). The influential status of this observation can be explained to a certain extent by extremely large aid receipts (62,5 % of GNI). The observation for which only the

The Effects of Development Aid on Infant Mortality and Life Expectancy