1 – Introduction
1Malaria ranks among the foremost public health and development issues facing tropical countries. The numbers are staggering: an estimated 41% of the world population lives in malarial areas, and 300 to 500 million clinical cases are recorded every year. Malaria kills between 700,000 and 2.7 million people per year, 75% of whom are African children. Around 30% of external medical checks and 20-40% of hospitalisations in Africa are due to malaria.
2Why are malarious countries poorer? Why do they develop slower than non-malarious countries? A series of papers have explored the link between economic development and malaria.  Contrary to other tropical diseases, since the failure of eradication efforts in the 1980s, malaria is widely described as an unavoidable effect of tropical location and natural forces (heavy rains, flooding). Recent empirical studies have used malaria as an exogenous variable in regression analyses of economic growth and income level on malaria endemicity. Nevertheless, even today, little is known on the transmission channels.
3Sachs and Gallup (2001), using a malaria exposure index defined as the population fraction at risk of contracting falciparum malaria in a country, show that poverty is not a leading cause of malaria ceteris paribus. The intensity and distribution of the disease are determined by the ecological conditions enabling the reproduction and development of malarial vectors. Working within a cross-country regression framework, the authors found that countries with intensive malaria are poorer and grew 1.3% less per person per year. These results point to a causal link from malaria to poverty — not vice versa. In contrast, Acemoglu (2002) asserts that malaria “is unlikely to be the reason why many countries in Africa and Asia are very poor today.”  Whatever conclusions are drawn, malaria plays a now well-acknowledged role as a “killer”, a “weakener”, and an element in this “vicious circle which makes the poor malarious and the malarious poor” (Watts, 1999).
4Using WHO estimates of malaria morbidity, McCarthy et al. (2000) suggest that economic development could influence malaria control. Although confirming the dominant role of climate in determining malaria intensity and the negative correlation between malaria and growth, they also found that access to rural health care and income equality influence malaria morbidity after controlling for climate. This raises the problem of the endogeneity of malaria in respect to growth and the robustness of results in this field of research.
5A natural way to study growth is to identify and quantify all the possible transmission channels (Rodrik and Rodriguez, 2000). As human capital formation and accumulation have been shown to be key variables in nation-level growth and development, this paper explores the impact of malaria on human capital development in children. As basic education and health are ends in themselves, this paper contributes to understand the suffering of individuals and communities in affected areas. We investigate this issue using a dataset from Samer Al-Samarrai (2006) to build a cross-country regression framework for assessing the impact of a 1994 index of malaria endemicity on repetition and completion rates in 1996. We control for other variables already used in the literature on the determinants of primary education quality. We add indicators of general health conditions, governance and climate to the model, and check the robustness of our results via a series of tests. Our results suggest that high levels of malaria endemicity increase repetition rates and decrease completion rates, ceteris paribus. These findings fill part of the gap between microeconomic assessments of the impact of malaria and macroeconomic assessments identifying one of the macroeconomic channels through which the disease impacts development.
The next section gives the conceptual framework of the study and a summary of previous literature on the link between malaria and school performances. Section 3 describes the variables and methodology used in the study. Section 4 presents the results of our cross-country regressions, while section 5 goes on to provide robustness checks. Finally, section 6 discusses the results, and draws our conclusions.
2 – Background and Previous Literature
The Link Between Malaria and School Performances: Medical and Empirical Evidence
6Malaria is a parasitic disease transmitted by anopheles mosquitoes, and which results from biological developments of protozoal parasites. The present study focuses exclusively on P. falciparum malaria, which is far more severe than the other types of malaria. The incidence and severity of P. falciparum malaria depends on various entomological, environmental and human factors. There are numerous potential combinations between factors. The clinical presentation of malaria varies, among other factors such as personal behavior or genetic characteristics, according to immune status. Populations exposed to frequent infections acquire partially labile protective immunity. Level of immunity affects not only the mortality and severity of malaria but also non-complicated malarial attacks. Therefore, paradoxically, it is more difficult to assess the real burden of the pathology in highly endemic areas where people are more exposed to infected-mosquito bites, since parasitization is not directly and proportionally associated with apparent clinical symptoms. Conversely, diagnosis is supposed to be easier in low-endemic areas where, in the absence of immunity, symptoms are caused directly by infection. In children, acquired immunity does not play an efficient protective role until age 5-6 years, even in highly endemic areas. This is part of the reason why malaria is a major threat to child survival. In 2000-2003, malaria accounted for 8% of the 10.6 million deaths recorded in children under 5 (WHO-CHERG), making it the third leading cause of mortality in under-5s. 94% of global child deaths attributable to the malaria occur in Africa (Bryce et al., 2005). In holo-endemic areas of Africa (where transmission is said to be stable), the disease represents 20 to 25% of all-causes mortality in the 0-5-years age band and 25% of paediatric hospitalizations, and lethality is around 15%. The depression of acquired immunity during pregnancy means that malaria also causes acute prenatal disease, with severe health consequences for both mother and child. 
7A number of studies have emphasized the consequences of malaria on children’s behavioral and cognitive development, learning performances, and school attendance. Nevertheless, the relationship is only beginning to be explored. Figure 1 schematically illustrates the direct links between malaria and school performances. Other socioeconomic externalities may contribute to amplify the impact of malaria on school outcomes.
8In 1917, the Texas State Health Department, the University of Texas Extension Department and the International Health Board made a first attempt to measure the effects of malaria on physical and mental development (Kelley, 1917). One interesting conclusion was that being cured from malaria leads to an immediate increase in physical and mental development, but the article does not give any explanation on the causal relationship. Moreover, the method and data used in this study remain crude at best. McDonald (1950) also underlines the consequences of malaria on child education through absenteeism and chronic infections. From there until the 1980s, research in this field did not progress further, and the indirect as well as direct consequences of malaria on child development remained neglected.
9In a recent unpublished paper taking a historical perspective, Bleakley (2006) looks at the malaria eradication campaigns led in the USA (1920) and in Brazil, Colombia and Mexico (1950) to assess the impact of childhood exposure to malaria on labor productivity. Using a cohort-level dataset based on microeconomic data, Bleakley found that cohorts born after the eradication campaign had higher literacy at adulthood than the previous generation. Childhood malaria exposure had similar effects on adults across the four countries studied.
10Holding and Snow (2001) reviewed the evidence on the impacts of P. falciparum malaria on performance and learning through neurological impairment after cerebral malaria, brain insults during malaria, complications of clinical cases, malaria induced-anaemia, nutritional effects, foetal exposure to malaria, low birth-weight, and prematurity. For instance, on a total of 1,854 Kenyan children with cerebral malaria, they found that 302 died (16.3%) and 248 (16% of survivors) were reported as presenting neurological sequelae at discharge. These figures do not mean that the sequelae will be accompanied by widespread impairment of cognitive functions or learning capacities, but the association seems to exist. Our study attempts to quantify this impact of malaria on school performance at macroeconomic level. Another review of evidence (Kihara et al., 2006) found similar conclusions and gave further details on the cognitive impacts of malaria. Malaria infection is known to lead to deficits in attention, memory, visuo-spatial skills, language, and executive functions. These deficits are not only caused by cerebral falciparum malaria, but also appear to occur in less severe infections. Finally, Holding and Kitsao-Wekulo (2004) outlined priorities for future research in this field. The full range of potential pathways to impaired development needs to be investigated through different adapted methods. Understanding the source of the variability in outcome (differences in socioeconomic status, presence of other infections, age, environment) could also help identify malaria-specific impact, children most at risk, and modalities for intervention.
11Two case studies realized in Sri Lanka (Fernando et al., 2003) revealed that initial and repeated malaria infections have an adverse impact on school performances (measured by test scores in mathematics and language after controlling for socioeconomic variables such as the education level of parents, household income, habitation types and nutritional status.) For instance, in 571 schoolchildren aged 6-14 years, a child who had experienced more than five attacks of malaria would score roughly 15% less than a child who had experienced less than three attacks of malaria. The cognitive performance at school entry of 325 grade-1 schoolchildren aged 5-6 years in two endemic districts decreases as the number of malaria infections increases. These studies suggest that malaria in malaria-endemic areas adversely impacts cognitive capacities not only at school entry but also during primary schooling.
Malaria can have non-cognitive consequences through school absenteeism, general health conditions, or other socioeconomic conditions and behaviors associated with the disease. For instance, in a case study led in Kenya, Brooker et al. (2000) attributed 13-50% of medically-related school absences to malaria. Leighton and Foster (1993) also provide evidence on the number of school days lost due to malaria in Kenya and Nigeria. In Kenya, primary school students were considered to miss 11% of the school year (20 school days missed per child per year) due to malaria, while in Nigeria the figure varied between 2% and 6% of the school year (3 to 12 days per year per student). Kimbi et al. (2005) estimated that in the Muea area in Cameroon, 53 out of 144 (36.8%) malaria-infected children lose 0.5 to 14 days of school (averaging 1.53 schooldays). The social and economic costs of the disease can amplify absences: parents can retain children at home to take care of younger siblings who are sick or perform other household or productive tasks. This may be especially true in agricultural economies due to coping mechanisms and the fact that farming activities are a family concern (Chima et al., 2003). Nevertheless, these conclusions remain speculative, since we still know very little about the impact of malaria on school attendance. Absenteeism in developing countries stems from numerous factors, many unrelated to malaria.
Malaria is associated with a number of neglected tropical diseases, including hookworm, schistosomiasis, onchocerciasis, filariasis, dengue fever and trypanosomiasis. Brooker et al. (2006) suggest that controls on malaria and parasitic helminths in school children could be viewed as essential co-contributors to promoting health among schoolchildren. The incidence of malaria can be increased by co-infection with the other tropical diseases.
Taken together, these studies tend to emerge a complex relationship between malaria infection, human capital formation, and a large number of risk factors which themselves may be associated with performance at school. Our purpose here is not to exhaustively review the literature on the micro-impacts of malaria on the cognitive and learning capacities of children. We simply underline that the few case studies exploring the question conclude that the link is strong. Microeconomic literature on health also recognizes that physical and mental health problems can impede children’s human capital accumulation (see, for instance, Currie and Stabile, 2006). Hence, it would be surprising not to encounter the impact of malaria at the macroeconomic stage, especially since the disease creates other negative external effects that can amplify the phenomenon.
Direct Impacts of Malaria on School Performance (Source: Author)
Direct Impacts of Malaria on School Performance (Source: Author)
Conceptual Framework of the Study
12Cross-country and panel literature on the determinants of schooling quality tends to focus on education resources. Increasing investments in basic education has been seen as a key strategy for achieving the Education For All (EFA) goals (World Education Forum, 2000). Nevertheless, cross-country comparisons consistently show the link between resources and educational outputs is weak (Al-Samarrai, 2006). Increases in expenditure in primary education alone are unlikely to be sufficient to achieve the education millennium development goals. Other variables such as the effectiveness of the public expenditure management systems, household spending, and the composition of public education spending may explain this weak link. This study explores another potentially important factor that influences schooling quality in developing countries and that may contribute to rendering expenditure on primary education ineffective. We use different datasets in a cross-country and panel regression framework to estimate the impact of an index of malaria endemicity on repetition and completion rates. A novel feature of our work is that we introduce a proxy for general child health conditions, climate variables and governance indicators into the model.
The relationship between school outputs and inputs is generally analyzed with an education production function, particularly at macroeconomic level. However, health conditions and diseases like malaria that alter cognitive and learning abilities are not taken into account in macroeconomic studies of the determinants of education quality. Following Barro and Lee (2001), we define a highly simplified education production function taking health into account:
13Where Q denotes schooling quality, F, family factors (principally family incomes), R, public resources used for primary education, H, child health conditions, and ?, unmeasured factors influencing schooling quality.
3 – Data and Methodology
Dependant Variable: Primary School Repetition and Completion Rates
14Conceptually, education has two main qualitative goals. The first is to develop the cognitive capacities of students. The degree to which systems actually achieve this goal is one indicator of their quality. The second element is “education’s role in encouraging learners’ creative and emotional development, in supporting objectives of peace, citizenship and security, in promoting equality, and in passing global and local cultural values down to future generations” (EFA, Global Monitoring Report, 2005). It is difficult to measure education quality for a broad number of countries, especially if the second objective is factored in (Barro and Lee, 2000). A handful number of papers have tackled this problem. Barro and Lee (2001) provide a panel dataset that includes output and input measures of schooling quality from 1965 to 1990. Hanushek and Kimko (2000) also develop cross-country measures of labor force quality based on information on international differences in mathematics and science knowledge for 1991. Altinok and Murseli (2007) followed a different methodology to obtain qualitative indicators of human capital (QIHC). First, they compiled recent surveys led between 1995 and 2003 in a cross-country database level covering 105 countries. Second, they updated the Barro and Lee (2001) dataset. Third, they gathered all the international surveys on children’s achievement into a panel dataset spanning the period 1964 to 2005. However, the dataset published by Altinok and Murseli (2007) has not yet been intensively used, making it difficult to claim it as a benchmark for comparing different cross-country studies. Moreover, the panel dataset pooling scores on international tests is highly unbalanced, and the cross-sectional indicator compiled different years, again making interpretation difficult. Al-Samarrai (2006) proposed to use the simple measures of education quality available through the UNESCO Institute for Statistics website (primary repetition rates, primary completion rates, primary survival rates) which are widely used to compare school outcomes across countries in international monitoring reports. Since one of our objectives is to compare our results to other studies, we have opted for these measures.
15As we are particularly interested in measuring the specific impact of malaria on children’s cognitive and learning abilities, we elected to start with primary school repetition rates which reflect this particular aspect of education quality. Indeed, despite the fact that repetition rates are sometimes affected by school and educational policies (such as national regulations), they are more directly determined by school results. The choice of this variable combined with our large database will allow us to work with a reasonably-sized sample with good coverage of developing countries (particularly in Africa) for our regressions (from 80 to 84 observations in table 1). Other alternative indicators are test scores. However, although some cross-country studies have been undertaken using students’ scores on internationally comparable tests of achievement in knowledge, the samples are often downsized due to limited availability of internationally comparable data. Recent efforts to define a standardized indicator, through MLA (Monitoring Learning Achievement – UNESCO/UNICEF), SACMEQ (Consortium for Monitoring Educational Quality) and PASEC (“Programme d’Analyse des Systèmes Éducatifs de la CONFEMEN”), still include a small sample of countries. Primary repetition rates are defined as the proportion of pupils enrolled in a given grade at a given school-year at primary level and who study in the same grade in the following school-year (UNESCO).
16In order to take into account other effects of malaria on school results, we also use primary completion rates as an alternative dependent variable.  The choice of this variable gives us a smaller size sample due to the lack of data for both the explained and explanatory variables in the models, but still with good coverage of developing countries (plus Africa) (from 54 to 55 observations in table 2). Primary completion rates are defined as the total number of students (of any age) in the last grade of primary school, minus the number of repeaters in that grade, divided by the total number of children of official graduation age (World Bank).
Repetition and completion rates have been widely used to compare school outcomes across countries in international monitoring reports.
The P. Falciparum Malaria Index
17A major problem when assessing the economic cost of malaria — at the mac-roeconomic level (Sachs and Gallup, 2001) as well as at the microeconomic one (Audibert et al., 2003) — is how to measure exposure to malaria risk and severity of the disease. This is clearly due to the lack of malaria incidence statistics. As stated in the general introduction, there are different methodological and technical barriers hampering accurate measurement of the number of malaria cases.
18As a proxy for child malaria morbidity, we first used the index described in Sachs and Gallup (2001), defined as the fraction of population at risk of contracting falciparum malaria in a country. This index uses historical maps of the geographic extent of malaria prevalence combined with detailed data on world population distribution and the fraction of cases of malaria that are falciparum malaria (for 1990). We computed this indicator for four different years (1946, 1966, 1982, 1994). This assumes that the relative proportion of P. falciparum cases did not change substantially over the period 1946 to 1994. As there are so many problems relative to the measurement of the number of malaria cases (even more so for historical data), we consider these indexes as the best measures of falciparum malaria risk for these particular years.
19Second, we extend these indicators with more recent data available from the Malaria Atlas Project. The main objective of this project is to develop a detailed model of the spatial limits of P. falciparum and P. vivax malaria at a global scale, and its endemicity within this range (Guerra et al., 2008). The MAP identified 87 countries at risk of P. falciparum transmission between 2002 and 2006. Figures for raw estimated population at risk are projections for 2007 made from the Global Rural Mapping Project (GRUMP) alpha population surface (adjusted to United Nations national country estimates for 2000) and national medium variant population growth rate by country from the United Nation Population Database (UNPD). To create the percentage of population at risk for 2007 by country, we used the UNPD estimates for 2000 as a baseline and the national population growth rate by country. This index has two main advantages. First, it enables us to update and improve the Sachs and Gallup (2001) database on falciparum malaria risk and to make cross-sectional regressions with recent years. Indeed, the MAP estimates are calculated differently to the Sachs and Gallup (2001) Malaria index, and therefore this new index is not strictly comparable to the first one. Nevertheless, it does provide a differently-constructed index that can be used for robustness analysis. Our assumption is that the quality of the Malaria Index is improved and upgraded. Second, the MAP database also provides projections of populations at risk within areas of unstable and stable P. falciparum transmission. Therefore, among people at risk, it is possible to have the proportion of those who live in stable or unstable areas and to dissociate the effects of malaria in both cases.
20Costs in terms of education are expected to differ with level of endemicity and percentage of areas in stable/unstable areas. Malaria is said to be stable if it is transmitted throughout the year by long-lived, anthropophilic vector anopheles mosquitoes (Kiszewski et al., 2004). MAP defines stability as an annual P. falciparum parasite prevalence ? 0.1 per 1,000 people per year. We can distinguish three main situations that will have different costs in terms of human capital and basic education at macroeconomic level in our study. In highly endemic regions (where the fraction of the population at risk of contracting falciparum malaria is high), mortality and morbidity mainly occurs among children aged 4 months to 5 years and in pregnant women. Consequently, malaria would be expected to have a bigger impact on primary school quality and school performances in these areas through the channels detailed in the general introduction (figure 1), and due to the other external effects associated with the disease. By contrast, in regions where malaria transmission is less stable and where herd immunity is lower, malaria and severe malaria can affect people of all ages, but less frequently (mainly during epidemics or seasonal variations). Consequently, the educational impacts are expected to be smaller. In malaria-free area, there will obviously be no impact.
In this study, we consider that a country is highly endemic when the P. falciparum malaria index is greater than 0.5 for both index measures. We also attempt to assess the impact in stable and unstable areas.
Other Independent Variables
21Three resource variables that the literature has previously identified as potential determinants of education outcomes are used to take into account the impact of public spending on repetition rates: primary current expenditure per pupil (PPP), public current expenditure in primary education as a percentage of GNP, and primary-school pupil-teacher ratio.
22Income per capita is also included in the regressions, for two main reasons. On the one hand, many studies have shown that countries with higher income have better primary education quality. Indeed, income per capita may be interpreted as a proxy for parents’ income (Barro and Lee, 2000). On the other hand, malaria and income levels are intimately connected, and our falciparum malaria index could just serve as a proxy for poverty.
23Level of urbanization, another explanatory variable that is a potentially influential factor in primary repetition and completion rates, is also included in the model. Indeed, school results are expected to be higher in urban areas since the travel costs tied to school attendance may be lower than in rural areas, and children can get better access to educational services or better general conditions to work in (electricity for instance) especially in developing countries.
24As stressed in the literature, the effectiveness of the public expenditure management system could explain the weak link between resources and education quality (see for instance Reinikka and Svensson, 2004). In order to control for this factor, we include in the regressions a measure of governance quality developed by Kaufmann et al. (2006): “government effectiveness 1996”. The Worldwide Governance Indicators are a compilation of information and perceptions from diverse groups of respondents but have to be seen as proxies of the complex phenomenon of governance. We do not pretend here to answer the specific question of the impact of management system quality on education outcomes, but rather to capture the influential effects of these factors on our dependant variables, ceteris paribus.
Malaria could also be a proxy for a range of other variables, principally general health conditions, notably child health conditions, plus climate and geographical location. Therefore, the regressions given in tables 1 and 2 include the under-5 mortality rate (as a proxy for child health conditions), and the percent of land area in subtropics and tropics (Sachs and Gallup, 2001), which allow us to control for climate, ecological conditions and diseases associated with tropical and subtropical location that could drive our coefficient estimates on the falciparum malaria index. Alternative measures of geographical location (regional dummies) have also been used in the robustness analysis.
25Descriptive statistics, sources and definitions of the variables used in the main cross-sectional analysis (1996) presented in this paper can be found in table A1. Simple correlations between variables are given in table A2. Panel data were drawn from the Barro and Lee (2001) dataset completed with World Development Indicators (World Bank 2006). Data for robustness analysis were also borrowed from the World Development Indicators (World Bank, several years) plus the UNESCO database.
As an example, take the 125 observations available for both the primary repetition rate for 1996 and the falciparum malaria index for 1994. 36 of the 125 countries (29%) are host to intensive malaria (29 are African countries). Ranking the 125 countries by primary repetition rate, 25 of the 36 countries with severe malaria (69%) are in the bottom half of the ranking. A first look at the data would however suggest that malaria could only be a proxy for geographic location or other variables. In the next subsection, we detail the methodology used in the study to explore the link between malaria and school results, and address possible sources of errors. We underline that we only present our identification strategy for the cross-sectional regression analysis (1996), which is comparable to other studies.
26Our modelling approach is based on previous research (Barro and Lee, 2000; Al-Samarrai, 2006), and the relevance of including malaria in the model has been tested using a stepwise selection procedure and the usual tests.
27We begin by estimating the regressions using Ordinary Least Squares (OLS). Standard errors are adjusted for heteroscedasticity in all cases. In this first step, the full set of available observations is considered. Outlying observations are not automatically mistakes in data entry and can provide important economic information by increasing variation in the explanatory variables. In addition to these regressions, we also analyze how robust the OLS results given in tables 1 and 2 are to different estimation techniques.
28Firstly, this study assesses whether the coefficient estimate on the malaria index suffers from endogeneity bias (principally due to measurement errors or two-way causation, but also to omitted variables). The falciparum malaria index is clearly exposed to measurement errors. It is also possible that low education levels have a negative impact on malaria eradication, either directly or through other variables such as income level or general health conditions. It is possible that countries with bad educational outcomes face greater difficulties in controlling malaria, but this argument remains speculative as it is not clear whether or not the general level of formal education in a country will improve malaria control. For instance, it can be argued that some MENA (Middle East and North Africa) countries have recently succeeded in controlling malaria but still have relatively high rates of illiteracy. Indeed, the Malaria Report 2008 lists Algeria, Egypt, Iraq and Saudi Arabia among 10 countries worldwide that have successfully reached the elimination phase for malaria. The United Arab Emirates was malaria-free in January 2007. Much of this success in combating malaria may be attributed to improvements in anti-malaria interventions. The educational achievement of MENA countries remains below other countries at similar levels of economic development. The 2007 TIMSS test on 8th-grade Math and Science capabilities resulted in none of the 12 participating MENA countries reaching the average scale.  However, in this part of the world, malaria epidemiology differs widely from Sub-Saharan Africa, with a higher proportion of P. vivax malaria cases and different predominant anopheles vectors. Moreover, despite low-level education performance, MENA countries have almost reached full primary education enrolment and have halved illiteracy rates within the space of the past 20 years. Mother’s education is a key determinant of access to health services and use of ITNs in Africa. Hence, a bidirectional relationship cannot be completely excluded.
29To correct for potential endogeneity bias, we use the spatial index of the stability of malaria transmission based on the interaction of climate with the dominant biological properties of the anopheles vector of malaria (biting activity, proportion of blood meal taken from human hosts, daily survival of the vector, duration of the transmission season and of extrinsic incubation). This interaction determines vectorial capacity and explains a large part of the strength, stability and regional variation of malaria transmission (Kiszewski et al., 2004). This index is measured on a highly disaggregated level and then averaged for the entire country, weighted by population. The total number of countries in which malaria is endemic or potentially endemic has been divided into 260 different regions based on their particular characteristics (to represent habitat diversity) and 34 anopheles vectors were considered as dominant. Because it is built on climatological and vector characteristics, “Malaria Ecology” is said to be exogenous to public health interventions and economic conditions. There is no reason to think that the ecology-based distribution of malarial mosquito vectors and the variation in their biological properties are causes of bad educational outcomes apart from the direct influence of malaria or through observable variables such as tropical location already included in our models. This variable is therefore a good candidate instrument for the falciparum malaria index. As we said, Sachs and Gallup (2001) had already used a similar instrument. The first stage regression of the falciparum malaria index on the instrument (“Malaria Ecology”) has an R2 of 0.56.
30Secondly, the sample size dictates that we can only introduce a limited number of control variables to isolate the causal effect of the falciparum malaria index. The specification models presented in tables 1 and 2 already include a relatively high number of independent variables. Introducing more variables may lead to a “small sample size problem” and multicollinearity can increase the variance of our unbiased estimators. We therefore performed different specifications of the regression model with a restricted number of independent variables and a higher-size sample (from 85 to 117 observations for regressions with repetition rates and from 56 to 73 observations for regressions with completion rates). We separately control for different factors by successively including in each regression a set of variables for geographical and climate characteristics, variables for health conditions, governance indicators, and education resources (data available on request). Using these specification models does not alter the main conclusions of this paper. The regional dummy variable that appears to play a significant role on the Two-Stage Least Squares (2SLS) regressions with repetition rates is Latin America and the Caribbean. These countries, on average, have higher primary repetition rates than average, after controlling for the model-integrated variables. No dummy variable appears to play significant role on the 2SLS regressions with repetition rates. Malaria could also be a proxy for a range of tropical diseases that are not adequately controlled for by under-5 child mortality (1995) and tropical and/or subtropical location. To assess whether other diseases were responsible for the correlation of malaria with primary repetition and completion rates, we jointly include two relevant indicators for the time period under study here: an average of the proportion of the population living in areas with dengue fever from 1975 to 1995, and an estimated fraction of the population at risk of contracting yellow fever for 1996 (Sachs and Gallup, 2001). Descriptions and sources of these variables are given in table A1. Unfortunately, we do not have macroeconomic-level indicators for other tropical diseases or for AIDS for 1996. Since AIDS has spread so quickly, the regressions did not include recent data on the AIDS burden. However, we do not think our conclusions would change with the inclusion of an AIDS indicator since AIDS affects education by different channels to malaria (for instance through its impact on teachers). The estimated impact of malaria remains relatively stable and statistically significant at the 5% level for regression (2) and the 1% level for regression (6). The two tropical diseases have no significant correlation with repetition and completion rates.
Thirdly, OLS regressions assume that the residuals are independent. It is eminently possible that primary repetition rates within each subcontinent may not be independent, and this could lead to residuals that are not independent within subcontinents. We thus allow regression observations to be clustered into subcontinents to take into account the probable spatial correlation of observations within subcontinents (not presented here but available on request). The 2SLS results are almost identical to those presented in tables 1 and 2. Therefore, we cannot reject the assumption that the residuals are independent.
Fourthly, we explore whether or not our OLS and 2SLS results are driven by influential observations. OLS is vulnerable to outlying observations since it minimizes the sum of squared residuals. Individual countries with large residuals and high leverage are identified and dropped for each regression.  From 8 to 11 countries were identified as outliers in the regressions and consequently dropped (data available upon request). The size and significance of the coefficient estimates on the falciparum malaria index show only slight change.
4 – Results
31The results from the OLS and 2SLS regression analysis with the Samer Al-Samarrai dataset for primary repetition and completion rates are presented in table 1 and table 2. Different models are reported so as to take into account different options chosen from the literature. There appears to be a very strong relationship between the falciparum malaria index and primary repetition and completion rates. The positive coefficient on the malaria index (0.096; se = 0.039) in regression (1) of table 1 indicates that countries with intensive malaria have 9%-higher repetition rates in primary education than countries without malaria. A one standard deviation increase in the falciparum malaria index is associated with a 0.43 standard deviation increase in primary repetition rates. The negative coefficient on the malaria index (-0.295; se = 0.087) in regression (1) of table 2 indicates that countries with intensive malaria have 29%-lower primary completion rates than countries without malaria. A one standard deviation increase in the falciparum malaria index is associated with a 0.47 standard deviation decrease in primary completion rates.
32The coefficient estimates on the malaria index are significantly different from 0 (at the 5% level) in regressions (1) and (2) of table 1. In regression (3) of table 1, the p value of the coefficient estimate is 0.13, and the value itself is lower than in the previous regressions. This may be due to the high correlation between the falciparum malaria index and the pupil-teacher ratio within the sample (r = 0.73). However, the corresponding Instrumental Variable estimation, given in regression (6) of table 1, shows that the impact of malaria is still significant, positive, and high. In all the OLS regressions of table 2, the coefficient estimates on the malaria index are significantly different from zero, negative, and relatively stable.
33Section 3 suggested that the falciparum malaria index may be endogenous, and went on to explain that the instrumental variable estimation can be used to account for this problem. 2SLS regressions are given in columns (4) to (6) of tables 1 and 2. We use the previously-described “Malaria Ecology” instrument for the falciparum malaria index. If there were a reverse causation between the malaria index and primary repetition and completion rates, the coefficient estimates on the malaria index reported in columns (1) to (3) would be positively biased in table 1 and negatively biased in table 2.
After correcting for the possible endogeneity of the falciparum malaria index, the absolute value of the estimated effect of malaria on repetition and completion rates is higher than our previous results, and still statistically significant in all regressions. This generally occurs when the endogenous variable suffers from measurement errors, and the falciparum malaria index remains relatively crude. Therefore, the 2SLS regression analysis confirms that the link between falciparum malaria and primary repetition and completion rates is strong. The positive coefficient on the malaria index (0.180; se = 0.060) in regression (4) of table 1 indicates that countries with intensive malaria have 18%-higher repetition rates in primary education than countries without malaria. The negative coefficient on the malaria index (-0.543; se = 0.184) in regression (4) of table 2 indicates that countries with intensive malaria have 54%-lower primary completion rates than countries without malaria. As 2SLS and OLS estimates on the falciparum malaria index differ significantly (due to measurement errors on the malaria index), the 2SLS estimator is more efficient than the OLS estimator. Therefore, we will now focus on regressions (4) to (6) of table 1 and 2 for the continuing analysis.
Cross-country OLS and 2SLS Estimates of the Effects of Malaria on Primary Repetition rates
Cross-country OLS and 2SLS Estimates of the Effects of Malaria on Primary Repetition rates
34GNP per capita is only significant in regression (6) of table 1, and its impact on primary repetition rates, though very small, is positive. Coefficient estimates of GNP per capita in regressions for completion rates are not significantly different from zero (table 2). These results suggest that income per capita does not have a strong impact on primary repetition and completion rates.
Cross-country OLS and 2SLS Estimates of the Effects of Malaria on Primary Completion rates
Cross-country OLS and 2SLS Estimates of the Effects of Malaria on Primary Completion rates
35Tropical and subtropical location does not appear to have a significant impact on primary repetition and completion rates. It could be argued that the impact of malaria on educational outcomes reflects the influence of a given region and not the isolated effect of malaria. For example, Sub-Saharan Africa has, on average, high levels of malaria, high repetition rates and relatively low completion rates. Hence, the malaria index could just be a proxy for Sub-Saharan Africa. However, results are globally unchanged when regional dummies are included (South Asia, East Asia, Sub-Saharan Africa, Middle East and North Africa, and Latin America and the Caribbean).
36Although under-5 mortality rate was positively and significantly associated with repetition rates in the OLS regressions, it does not appear to have a significant impact on repetition or completion rates in 2SLS regressions. In all the 2SLS regressions of table 1 and 2, the coefficient estimates on under-5 mortality are not significantly different from zero.
37Level of urbanization and the governance indicator have no significant impact on the educational outcomes under study here.
38The coefficient estimates on current public primary expenditure (% of GNP) is negative in all the regressions of table 1. It is positive in regressions (4) and (6) and negative in regression (5) of table 2, but never significantly different from 0. This insignificant effect of total educational spending on repetition/completion rates suggests that an increase in total school resources may not itself lead to an increase in student achievement. This conclusion is in accordance with Hanushek (1995) and Samer Al-Samarrai (2006). There is no systematic relationship between school performance and general measures of school resources.
Where regressions include two school variables (primary expenditure per pupil and current public primary expenditure as a proportion of GNP), the results are not dramatically different. They suggest that primary expenditure per pupil is also an insignificant determinant of repetition and completion rates, ceteris paribus.
Nevertheless, when the regressions include the three school variables, the pupil-teacher ratio is positively and significantly (at 1% level) associated with repetition rates in regression (6) of table 1. This result is coherent with the conclusions of Barro and Lee (2001) on the link between repetition rates and pupil-teacher ratio. It is estimated that a lower pupil-teacher ratio improves repetition rates and test scores. Components of school expenditure allocated to lower pupil-teacher ratio will reduce primary repetition rates, whereas the other-resource variables still have no significant impact. Nevertheless, the literature remains ambiguous on this point (Samer Al-Samarrai, 2006). Moreover, pupil-teacher ratio is not significant in the primary completion rate regression (regression (6) of table 2). Therefore, as in the literature, there seems to be an ambiguous relationship between educational outcomes and commonly-measured school resources. Nevertheless, as this question is not the primary focus of our study, we will refocus our efforts on testing the robustness of the coefficient estimates on the malaria index.
5 – Other Robustness Tests
Panel Dataset: Sachs and Gallup Indices (2001)
39The panel data models used to estimate the effect of malaria on primary repetition rates are random effects models (table 3). This is justified by the fact that 1) our time dimension is reduced, 2) our between-group variation is higher than our within-group variation.
Panel Data Estimates of the Effects of Malaria on Primary Repetition Rates (1960-1995)
Panel Data Estimates of the Effects of Malaria on Primary Repetition Rates (1960-1995)
40Since Sachs and Gallup (2001) gives malaria measures for 1946, 1966, 1982, 1994 and the output measures for 5-year intervals from 1960 to 1995, we matched the input measures with outputs in the nearest year for which the malaria measure is available. Completion rates are not available for this period. The results show similar patterns, but the coefficients appear lower than in our cross-country analysis. This may be due to the fact that we have no time-variant instrument for malaria for this period of analysis, meaning the coefficients are almost certainly biased downward.
Cross-country Robustness Analysis: Using an Updated Malaria Index from MAP (Malaria Atlas Project)
41Table 4 below presents the results with the updated malaria index from MAP (2007). Since the dataset is more recent, some of the variables are still incomplete and count very few observations, and have consequently been dropped from the analysis. This was the case for current primary expenditure per pupil (PPP) and current public expenditure in primary education as a percentage of GNP. As these variables had no significant effect in the previous analysis, we suppose that they are controlled through the well-documented primary school pupil-teacher ratio. All other variables have been kept in the models. The malaria index used here and the Malaria Atlas Project have been described earlier. The malaria index is the percentage of the population at risk of falciparum malaria in 2007 (figure 1). This index is then broken down into two sub-indexes:
%Tot Pop at risk of falciparum malaria in 2007 = % Tot Pop at risk of falciparum malaria in 2007 living in stable transmission areas + % Tot Pop at risk of falciparum malaria in 2007 living in unstable transmission areas
OLS and 2SLS Regressions of Primary Repetition and Completion Rates on the Malaria—MAP Index
OLS and 2SLS Regressions of Primary Repetition and Completion Rates on the Malaria—MAP Index
42According to Guerra et al. (2008), in 2007, 1 billion people were living under unstable or extremely low malaria risk while 1.37 billion were living under stable and high malaria risk. Hence, we first use the global index of malaria separately and next test the robustness of this index by introducing into the regression the % Tot Population at risk of falciparum malaria in 2007 living in stable transmission areas. This allows us to test the relative contribution of stability of the malaria burden.
We find similar patterns in these results to our previous testing on the falciparum malaria global index for both OLS and 2SLS analysis. However, it seems that the % of population at risk of falciparum malaria living in stable transmission areas in 2007 is a driver of our estimates of the coefficient associated with the global falciparum malaria index. Indeed, the partial effect of the falciparum malaria index is no longer statistically significant once associated with the “falciparum malaria index – stable area" (regressions (2), (4), (6) and (8)), whereas the partial effect of this stable-area falciparum malaria index remains statistically significant. When holding other model-integrated factors fixed, particularly the percent of population at risk of falciparum malaria in a country, the countries with a high proportion of population living in stable transmission areas have 13%-higher primary repetition rates than countries without stable transmission areas. This means that countries with relatively higher proportions of population living in stable transmission areas (versus unstable areas) are more affected by the effects of malaria on human capital accumulation.
6 – Discussion and Conclusions
43Our cross-country panel regression analysis has shown that the link between the level of falciparum malaria endemicity and primary repetition rates is strong and positive, while the link between the level of falciparum malaria endemicity and primary completion rates is strong and negative. These results suggest that malaria contributes to impede children’s human capital accumulation at macroeconomic level.
44Nevertheless, the results presented in the previous section should be interpreted with caution. Why is there such a strong link? Malaria specifically affects children’s cognitive and learning abilities, and consequently their school results. Our conjecture was that high repetition rates and low completion rates reflect particularly bad school results. Therefore, the strong link between malaria and primary repetition and completion rates reflects the impact of malaria on school results, and does not imply that there will systematically be a similar relation between malaria and other educational outcomes. We did not explore the relationship between falciparum malaria and other educational indicators.
45Moreover, malaria is a complex economic and social phenomenon. It is very difficult to dissociate the direct medical impact of the disease from its numerous effects on society and human organization. In addition to its medical effects, the behavioral and social aspects linked with malaria may explain part of the estimated impact of malaria on repetition and completion rates reported in this paper.
46A major problem hampering assessments of the impact of malaria on education is the lack of good data. Our conclusions are valid within the limits of data precision and availability. Some variables that may be expected to influence education outcomes or bias the coefficient values of the malaria index were not included due to the lack of cross-country data. We attempted to account for this problem by using different specifications of the regression model and different estimation techniques, but there may still be variables correlated with the malaria index that are missing. Furthermore, the conclusions drawn are only valid at the macroeconomic stage, and are not a substitute for detailed per-country analysis or case studies. Microeconomic evaluations of the impact of malaria on school performances may differ from macroeconomic projections.
47Some variables included in the regression models have ambiguous effects. Note that GNP per capita has a positive effect on primary repetition rates in some regressions. However, this effect is not robust to different model specifications nor to the use of re-updated data. In regressions where GNP per capita is not associated with under-5 mortality rate or pupil-teacher ratio, its effect is insignificant or negative (see regression (4) of table 3 for instance). This may be due to the high correlation between GNP per capita, the malaria index, under-5 mortality rate and pupil-teacher ratio. However, the results do not change when these variables are removed from the model (data available on request). Recent empirical literature on this topic suggests unstable effects of GNP or GDP per capita on education quality at the macroeconomic level. Al-Samarrai (2006), for instance, reported that this variable has a negative but non-significant effect on primary completion rates. In Barro and Lee (2001), the GDP variable proves insignificantly associated with mathematics and science test scores but significantly associated with reading scores (positively) and repetition rates (negatively). Therefore, our analysis casts serious doubts on the relevance of using macroeconomic data to study such important development issues due to the levels of data heterogeneity and quality involved. The same scepticism and criticism addressed at Sachs and Gallup (2001) or Acemoglu et al. (2002) with respect to data quality hold true for our study.
However, our results on the link between falciparum malaria and primary repetition rates suggest that the achievement of the UN millennium development goals for education (particularly the aspects concerning education quality) will require more than just focusing on expenditure in primary education. These results have been shown to be robust to a series of tests (different specification models, controls for influential observations, different estimation techniques, different datasets, and different years). They appear to provide a good idea of the malaria burden at macroeconomic stage. Regressions with completion rates also suggest that there are other externalities linked with the disease, as the impact is even higher. Hence, our macroeconomic results remain useful for a first-look analysis of the relationship. Nevertheless, the questionable level of macroeconomic data quality discussed previously means these results are obviously insufficient to conclude that a causal relationship exists between these variables. It is possible that we are asking questions that are too subtle for the available data to answer with any meaningful precision.
48The author is grateful for comments from M. Audibert, J.-C. Berthélémy, S. Poncet, C. Bros, A. Tatem, all the participants of the Séminaire d’économie CES (Paris), and particularly S. Al-Samarrai and J. Sachs who accepted to provide part of the database for this study. The author remains solely responsible for any errors.
Definitions, Sources and Descriptive Statistics for Main Cross-country Analysis (1996)
Definitions, Sources and Descriptive Statistics for Main Cross-country Analysis (1996)
Other Collinearity Diagnostics
Other Collinearity Diagnostics
Post-Doc. École des Hautes Études en Santé Publique – Département SHS-CS, CAPPS – avenue du Professeur-Léon-Bernard – CS 74312 – 35043 Rennes Cedex. Josselin.Thuilliez@ehesp.fr
See for instance McDonald (1950), Barlow (1967), Newman (1968), Gomes (1993) and Audibert et al. (1999).
Note that the relationship between malaria and growth is not the primary focus of this study.
For a historical perspective of childhood malaria mortality in Africa, see Snow et al. (2001). For a review of the economic effects of malaria in pregnancy, see Worrall et al. (2007).
See Samer Al-Samarrai (2006) for a discussion on conceptual problems associated with the use of completion rates.
http://nces.ed.gov/timss/results07_math07.asp (accessed: November 2009)
Studentized residuals and leverages were examined as an initial approach for identifying outliers and observations that have a potentially strong influence on regression coefficient estimates. Next, DFIT statistics were used as an overall measure of influence.