wPS w149 POLICY RESEARCH WORKING PAPER 2798 Fever and Its Treatment among the More and Less Poor in Sub-Saharan Africa Deon Filmer The World Bank Development Research Group Public Services March 2002 | POLICY RESEARCH WORKING PAPER 2798 Abstract Filmer empirically explores the relationship between an episode of fever are significantly associated: wealthier household poverty and the incidence and treatment of households are substantially more likely to seek care in fever-as an indicator of malaria-among children in the modern health sector. In Central and Western Africa Sub-Saharan Africa. He uses household Demographic those from richer households are more likely to seek care and Health Survey data collected in the 1990s from 22 from all types of sources: government hospitals, lower- countries in which malaria is prevalent. level public facilities such as health clinics, as well as The analysis reveals a positive, but weak, association private sources. In Eastern and Southern Africa the rich between reported fever and poverty. The geographic are primarily more likely to seek care from private association becomes insignificant, however, after facilities. In both regions there is substantial use of controlling for the mother's education. There is some private facilities-use that increases with wealth. Like the evidence that higher levels of wealth in other households incidence of fever, treatment-seeking behavior is strongly in the cluster in which the household lives are associated associated with the level of wealth in the cluster in which with lower levels of reported fever in Eastern and the child lives. Southern Africa. Poverty and the type of care sought for This paper-a product of Public Services, Development Research Grou4p-is part of a larger effort in the group to understand the relationship between poverty and health. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Hedy Sladovich, mail stop MC3-3 11, telephone 202-473-7698, fax 202-522-1154, email address hsladovich@worldbank.org. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at dfilmer@worldbank.org. March 2002. (37 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of idie.as about development issues. An objective of the series is toget the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Produced by the Research Advisory Staff Fever and its treatment among the more and less poor in Sub-Saharan Africa Deon Filmer Development Research Group The World Bank 1818 H Street NW Washington, DC 20433 USA +1-202-473-1303 dfilmer(di.worldbank.org March 2002 The author thanks Davidson Gwatkin, Menno Pradhan, and Adam Wagstaff, as well as participants at a session of the international Health Economics Association 2001 and at a seminar at the World Bank's Health Nutrition and Population Thematic Group Seminar series for comments on earlier drafts. Partial funding for this paper was received from World Bank Research Support Budget RPO 683-32. 1. Introduction There are over 300 million new cases a year of malaria in the world, resulting in over 1 million deaths. Of these, 90 percent of the cases, and 97 percent of the deaths occur in Sub-Saharan Africa (World Health Organization 2002). Reducing the burden of malaria figures high on the agenda as the poorest nations around the world commit themselves to reaching the Millennium Development Goal of halving of child mortality by 2015. Malaria is frequently characterized as being intimately linked with poverty. For example, the World Health Organization's recent Report on Infectious Diseases argues forcefully that these diseases, and malaria in particular, are both a consequence of poverty as well as an obstacle that keeps people in poverty (World Health Organization 2002). Much of the evidence on economic aspects of malaria focus on the relationship between malaria and GDP growth as derived from cross-national data. Additional work focuses on the loss of productivity associated with malaria among adults and the resulting loss in welfare. This paper uses individual and household level data to empirically explore the associations between household wealth and the incidence and treatment of fever-as an indicator of malaria-among children in Sub-Saharan Africa. The data used are from Demographic and Health Surveys collected in the 1990s from over 20 countries in which malaria is prevalent. The approach used is to assess first the broad geographic associations - and test whether these are significant across two broad regions: Western and Central Africa, and Eastern and Southern Africa. Next, the approach is to explore the association between fever (and treatment) of a particular child and the level of wealth in the household and cluster in which he or she lives. The analysis reveals a positive, but weak, association between reported fever and poverty. The geographic association becomes insignificant, however, after controlling for mother's education. There is some evidence that higher levels of wealth of other households in the cluster in which the household lives are associated with lower levels of reported fever in Eastern and Southern Africa. Poverty and the type of care sought-or not sought-for an episode of fever are significantly associated: wealthier households are substantially more likely to seek care in the modern health sector. In Western and Central Africa those from richer households are more likely to seek care from all types of sources: government hospitals, lower-level public facilities such as health clinics, as well as private sources. In Eastern and Southern Africa the rich are primarily more likely to seek care from private facilities. In both regions there is substantial use of private facilities-usage that increases with wealth. Like the incidence of fever, treatment seeking behavior is strongly associated with the level of wealth in the cluster in which the child lives. I Section 2 discusses issues related to measuring fever, using fever as a proxy for malaria, and measuring poverty using DHS data. Section 3 describes the results of ani analysis of the relationship of the incidence of fever and poverty. Section 4 describes the results of the analysis of the correlates of the care seeking behavior for the subset of countries where these data are available. 2. Measuring fever and poverty The data used in this paper are Demographic and Health Survey (DHS) data collected in Sub-Saharan African countries in the 1990s. These data are nationally representative household surveys with large sample sizes ranging from 2,252 households in Comoros to 9,282 in Mozambique. Table 1 reports the countries, the survey years, and the number of households from each country included in the analysis. All the countries in the study are poor, with nationally defined headcount poverty rates ranging from 26 percent in Zimbabwe to 86 percent in Zambia (poverty data from World Bank 1999). In a cross-country average of the countries for which the data are available the median percentage of the population living on under $1 a day in purchasing power parity terms is 38 percent, and the median living on under $2 a day is 84 percent. Under-five mortality is high in these countries, with a cross-country average of 171 under-five deaths per 1000 births (as derived from the DHS data) ranging from 90 per 1000 in Zimbabwe to 318 per 1000 in Niger. Note that the selection of countries into this analysis is driven by data available. All countries that had the relevant questions in the survey instrument were included. Measuring the incidence offever and its relationship to malaria In general the DHS in Sub-Saharan Africa interview all women 15-49 about a variety of issues relating to their fertility preferences, contraceptive behavior, and reproductive and child health. The data analyzed here on the incidence of fever are derived from questions asked of mothers of all children born in the past three or five years, depending on the survey. For consistency, the analysis here is restricted to children under the age of three. The exact formulation of the question about fever varies somewhat across countries, but the typical questionnaire will ask whether the child had an episode of fever in the past 2 weeks.' Subsequently, the respondent will be asked whether the episode of fever was accompanied by a cough and shortness of breath. The focus here on self-reported episodes of fever is completely data driven. The DHS data offer a unique large database in order to analyze patterns of incidence and ' While there are differences across countries, a major advantage of the DHS is the consistency in survey instrument and implementation across countries. 2 treatment in a consistent way across countries. This benefit comes at the limitations of using fever as a proxy for malaria. Table 1. Summary and background information of data used from DHS surveys GNP per Popula- Popula- Popula- Number capita-1990s tion below tion tion Analysis Analysis Year of of average Under-5 the poverty below $1 below of of survey households (PPP, $1995) mortality line Year a day $2 a day Year incidence treatment Western and Central Africa Benin 1996 4,499 798 184 33 1995 V Burkina Faso 1992/3 5,143 813 187 61 86 1994 - Burkina Faso 1999 4,812 813 224 -I Cameroon 1991 3,358 1,412 126 40 1984 -I Cameroon 1998 4,697 1,412 146 V C.A.R. 1994/5 5,551 1,085 157 67 84 1993 V Chad 1996 6,840 829 201 64 1995/6 V C6ted'lvoire 1994 5,935 1,372 150 12 49 1995 V Vf Ghana 1993 5,822 1,646 119 31 1992 V Ghana 1998 6003 1,646 110 V Mali 1995/6 8,716 653 252 73 91 1994 V Niger 1992 5,242 723 318 63 1989/93 61 85 1995 V Niger 1997 5,242 723 303 V Nigeria 1999 7,647 737 133 34 1992/3 70 91 1997 V V Senegal 1992/3 3,528 1,223 131 33 1991 26 68 1995 V V Togo 1998 7,517 1,337 144 32 1987/9 V Eastern and Southern Africa Comoros 1996 2,252 1,586 113 V Kenya 1998 8,380 948 105 42 1992 27 62 1994 V V Madagascar 1992 5,944 751 163 70 1993/4 60 89 1993 V Madagascar 1997 7,171 751 164 V Malawi 1996 2,798 526 234 54 1990/1 - V Mozambique 1997 9,282 612 219 38 78 1996 V Rwanda 1992 6,252 150 51 1993 36 85 1983/5 Vf v Tanzania 1991/2 8,327 446 141 51 1991 20 60 1993 Vf Tanzania 1996 7,969 446 145 V Uganda 1995 7,550 934 156 55 1993 37 77 1992 V Zambia 1992 6,209 703 191 86 1993 73 92 1996 V Zambia 1996/7 7,286 703 192 V Zimbabwe 1999 6,369 2,460 90 26 1990/1 36 64 1990/1 V V Average* 6,081 1,003 171 48 46 77 Std. dev.* 1,785 461 56 17 21 14 Median* 6,003 813 156 47 38 84 Maximum 9,282 2,460 318 86 73 92 Minimum 2,252 446 90 26 12 49 * Unweighted. Source: GNP per capita from World Development Indicators (World Bank 1999). Poverty rates from World Bank (2000). DHS information from DHS Final reports and updates from www.measuredhs.com (Macro International, various years). Focusing on fever as a marker for malaria in areas or seasons of high malaria endemicity is not without justification, however. A review of the guidelines for Integrated Management of Childhood Illness (IMCI) (Gove 1997) summarizes the 2 While there are differences across countries, a major advantage of the DHS is the consistency in survey instrument and implementation across countries. 3 recommended sequence leading to the delivery of oral antimalarials in the presence of fever that is not classified as very severe febrile disease. The recommendations (reproduced in Appendix Table 1) are to treat presumptively for malaria if (1) in high malaria-risk areas, the patient presents with fever, without any general danger sign or stiff neck and without cough with fast breathing and (2) in low malaria-risk areas, the patient presents with fever, without any general danger sign or stiff neck, and without runny nose, measles, other known cause of fever, no cough with fast breathing. Typically, in areas of high endemicity the recommendation is that all patients with fever or history of fever be treated with antimalarials. As the figure (reproduced from MVARA 1998) in Annex 1 shows, the areas of Sub-Saharan Africa under study here are virtually all in areas which are suitable to stable malaria).3 In order to be encompassing, this paper uses all episodes of fever as the outcome measure. In some cases results for fever without a cough are presented as well. While there is a difference in the incidence of fever and fever without a cough, the patterns across groups are not substantively different. Although the use of fever to identify malaria is in line with the current INMCI recommendations, it is certainly not perfect. All fever is not malaria. Brinkmann and Brinkmann (1991), after reviewing a substantial volume of literature, estimate that malaria is responsible for 40 percent of episodes of all fever in Africa. A validation of the IMCI protocol for minimally trained health workers in a high malaria transmission area of Kenya (Perkins and others 1997) found that 96 percent of 1,674 patients presenting with fever were classified as having malaria. A follow-up physicians assessment (based on measurement of temperature, hemoglobin determination, blood smear for malarial parasites and chest X-ray) determined that 456 (27 percent) were unlikely to have malaria, and that as many as 16 percent of cases had "fever requiring referral." Others argue that better algorithms for diagnosing fever could be used (for example Redd and others 1996) or that it is not sufficient to rule out relative alternative treatments (Redd and others 1992) - a limitation that the holistic IMCI approach ainied to remedy. The DHS rely not on third-party health worker or doctor assessments of health status, but on the report of mothers on the fever episodes of their children. This introduces two additional potential sources of error. Do mothers recognize fever in their children, and is there a systematic bias to which mothers recognize fever-and in particular, is this related to household wealth? 3 Stable malaria describes areas with year-round transmission, which may be low or high intensity. Northern regions of Chad, Mali, and Niger are not suitable to stable malaria but in countries covered by the DHS. Northern Mali is excluded by virtue of DHS sample design, and northern Chad and Niger are excluded in the geographic analysis in this paper. In the multivariate analysis, dummy variables for national subregion will adjust for this. 4 First, do mothers (or care-givers) recognize fever in their children? A recent exchange in The Lancet set off by Einterz and Bates comes to no firm conclusion (Einterz and Bates 1997, Dunyo, Doram, and Nkrumah 1997, Verhoef and others 1998, Koefoed and others 1998), and this paper will certainly not resolve the issue. Table 2 summarizes the main results pertaining to children to come out of the exchange. While the studies are not entirely comparable (e.g., differences in study methodology and differences in overall percentage of population with fever) they are consistent in finding that among children with measured fever (actual temperature of 37.5°C or higher) parents (or care-givers) tend to report fever accurately most of the time, in the cases here, between 78 and 98 percent of the time. On the other hand, among children without fever, anywhere from almost none to almost all of the children are reported to have a fever. It is somewhat reassuring that the study that most resembles the DHS-the community survey in Ghana-has reasonably high sensitivity (78 percent) and rarely found reports of fever when it was not present (0.8 percent). Note that the findings are consistent with a bias towards reporting fever, even when it isn't true. Since such a bias would tend to reduce the variation in reported fever, this would tend to dampen any subsequent study of the correlation of fever with other factors. Table 2. Summary of results on sensitivity and specificity of reported and measured fever in three studies Percentage with Percentage without Overall measuredfever measuredfever who Number percentage who are reported are reported to have of with measured to have fever fever (100- children fever ("sensitivity") "specificity') Patients, district hospital in northern Cameroon, children 494 34 92 56 under 5 Two communities in southern Ghana, children under 5 1714 3.5 78 0.8 Health Center in the outskirts of Bissau, Guinea-Bissau, "children" with symptoms compatible with malaria 203 81 98 97 Source: Einterz and Bates (1997), Dunyo, Doram, and Nkrumah (1997), Verhoef and others (1998), Koefoed and others (1998). An additional complication is the fact that individuals of different socioeconomic backgrounds might report a true episode of fever differently, and that a subsequent analysis of the correlation of fever with these socioeconomic variables would be biased in the direction of the selective reporting. While the issue of self-selective reporting is frequently cited as a cause for worry, there has been little systematic investigation of how severe a problem this might be. Studies typically conclude that more easily observed symptoms are less likely to suffer from self-selection (for examples of studies addressing this issue see Butler and others 1987 for an example from the United States and 5 Deolalikar 1998, Sindelar and Thomas 1991, and Strauss and Thomas 1996 for discussions relating to poor countries). The DHS data cannot be used to test or correct for whether fever is recognized more systematically by mothers from richer households which would dampen a relationship between fever and poverty. The multivariate analysis, however, will control for mother's education, which would likely capture a large part of the self-selective nature of reporting. The percentage of children in the DHS reporting any fever in the past two weeks is high (Table 3).4 The average for Western and Central African countries is 35 percent with a rural-urban differential on the order of 6 percentage points. Among these countries, Ghana has the lowest reported level of fever (29 percent) and Benin has the highest (55 percent). In the Eastern and Southern African countries the overall level is slightly higher, with a smaller rural-urban difference. The range is narrower, with Zimbabwe at the low end (31 percent) and Uganda at the high end (50 percent). While the average rural-urban differential is smaller than that in Western and Central Africa it is still large in some countries. For example in Malawi and Uganda it is about ten percentage points. In Western and Central Africa roughly half of all fevers are unaccompanied by a cough, and this ratio does not vary substantially across urban and rural areas. There is some variation in this ratio across countries, mostly between 35 percent in Burkina Faso and 59 percent in Niger and Ghana. The exception is C.A.R. where only 14 percent of all 5 fevers are reported to be without a cough. In Eastern and Southern Africa roughly one- third of all fevers are unaccompanied by a cough, ranging from 24 percent in Rwancla to 43 percent in Mozambique. Measuring household wealth rankings Studies analyzing inequalities in outcomes typically use consumption expenditures as a measure of long-run income (see discussion in Deaton 1997). Although the DHS include detailed health information, data on consumption expenditures are not collected. This paper uses an approach based on an index of assets owned by household members as well as housing characteristics advocated and applied in Filmer and Pritchett (1999, 2001) for the analysis of inequalities in education outcomes. A similar asset index approach has also been used by others to analyze health outcomes in DHS data, for 4. DHS surveys frequently use a weighting scheme to adjust sample to population averages. These weights are used in this analysis. See discussion below on cross-country weights. 5. C.A.R. has a slightly different questionnaire design. If treatment was sought then respondents were asked how much was spent on treatment. It is unlikely though that this would have affected the percentage reporting no cough with a fever. 6 example child mortality in Bonilla-Chacin and Hammer (1999), child survival in Uganda in Stecklov, Bommier, and Boenna (1999), child anthropometric outcomes in Wagstaff and Watanbe (1999), and to document inequalities in a variety of health outcomes and behaviors in Gwatkin and others (2000). Sahn and Stifel (2000) use a similar approach to analyze poverty directly. Table 3. Percentage of children under 3 who are reported to have had an episode of fever (and fever without a cough) in the past two weeks Any fever Fever without a cough Country Year Total Rural Urban Total Rural Urban Western and Central Africa Benin 1996 54.5 57.8 47.3 26.1 27.9 22.4 Burkina Faso 1999 41.0 41.8 34.1 14.4 14.5 13.2 C.A.R. 1994-95 35.7 36.3 34.8 5.0 6.0 3.4 C6te d'Ivoire 1994 44.2 46.3 39.8 19.5 20.8 16.8 Cameroon 1998 30.8 31.1 30.2 11.7 12.1 10.5 Ghana 1998 29.0 29.8 26.4 14.9 15.4 13.5 Mali 1995-96 39.3 41.2 34.1 19.8 20.8 17.0 Niger 1997 49.4 51.1 41.2 28.9 30.1 22.9 Nigeria 1999 31.2 32.6 27.3 17.6 18.5 15.3 Senegal 1992-93 45.7 48.3 40.8 17.8 18.4 16.8 Chad 1996 36.9 36.9 36.6 19.9 20.0 19.5 Togo 1998 37.7 37.7 37.8 15.9 16.5 14.0 All 35.3 36.8 31.3 17.8 18.7 15.5 Eastern and Southern Comoros 1996 48.7 48.8 48.4 17.9 18.5 16.0 Kenya 1998 42.9 42.9 43.2 16.8 16.6 17.5 Madagascar 1997 32.7 33.1 31.0 9.8 10.2 8.1 Mozambique 1997 44.7 43.2 50.4 19.2 21.2 11.8 Malawi 1998 47.6 48.8 37.5 12.0 12.1 11.2 Rwanda 1992 48.6 49.0 39.2 11.8 11.7 14.2 Tanzania 1996 35.8 35.7 36.1 12.8 13.0 12.0 Uganda 1995 50.3 51.5 40.6 14.6 14.8 13.0 Zambia 1996-97 46.4 47.6 44.5 17.2 19.1 14.2 Zimbabwe 1999 31.1 32.5 28.2 8.3 7.4 10.2 All 41.9 42.4 39.3 14.1 14.4 12.7 Source: Author's calculations from DHS data. Notes: Data are weighted. See next section for discussion of weights. 6. DHS surveys frequently use a weighting scheme to adjust sample to population averages. These weights are used in this analysis. See discussion below on cross-country weights. 7. C.A.R. has a slightly different questionnaire design. If treatment was sought then respondents were asked how much was spent on treatment. It is unlikely though that this would have affected the percentage reporting no cough with a fever. 7 The DHS typically collect information on whether or not any household members own each of a set of basic assets (radio, refrigerator, television, bicycle, motorcycle, car) and basic characteristics of the dwelling in which the household lives (whether or not the house has electricity, the number of rooms per person, and whether or not the dwel:ling has floors made of a "finished" surface). In order to use these variables to rank households by their economic status, they need to be aggregated into an index, and a major problem in constructing such an index is choosing appropriate weights.8 This is done here using the statistical technique of principal components. Principal components is a technique for summarizing the information contained in a large number of variables to a smaller number by creating a set of mutually uncorrelated components of the data. Intuitively, the first principal component is that linear index of the underlying variables that captures the most common variation among them. The details of the methodology are in Filmer and Pritchett (2001). That paper also describes how in three datasets (from Indonesia, Nepal, and Pakistan) where there was both consumption expenditure and asset data, Spearman rank correlation coefficients between ranking households by the asset index and ranking them by expenditures adjusted for household size were 0.64 in Nepal, 0.56 in Indonesia, and 0.43 in Pakistan. In the context of education outcomes, Filmer and Pritchett (2001) argue that in the three countries studied the wealth index performs as well as household-size-adjusted consumption expenditures, in predicting educational enrollment and attainment.9 Table 4 reports the "factor scores" for the first principal component in the analysis of the Western and Central Africa and Eastern and Southern Africa data. In both regions, the first principal component captures about 30 percent of the variation in the data. The factor scores, which are in essence the weights in the aggregation of the assets into the index, all have the expected sign (except for owning a bicycle in Western and Central Africa, which could reflect the fact that poorer households are more likely to own a bicycle whereas richer households may be more likely to own a motorcycle or a car instead). Although the results imply that the second eigenvalue is greater than one (the usual value used as a cutoff for "relevant" components) the difference between the first and second eigenvalues is large, lending support to the notion that the first is capturing a significantly larger part of the important information in the asset and housing variables.10 8. If these assets were only to be used to examine the impact of some other factor (e.g., maternal education) as a "control" for wealth in a multivariate regression we would not need to aggregate the variables (see Montgomery and others 2000). 9. Unlike Filmer and Pritchett (2001) water and sanitation variables have not been included in the set of variables used to derive the index as these likely have direct effects on health status. They are included as individual variables in the multivariate analysis. 10. Principal components in typically derived for continuous variables. Monte Carlo simulations comparing the first principal component derived from (1) continuous variables to (2) dummy variables with the same mean (derived from the ranking of the continuous variables) yields very similar factor loaclings 8 Table 4. Factor scores and other summary statistics for first principal component from analysis of pooled Western and Central Africa and Eastern and Southern Africa data Western and Central Eastern and Southern Africa Africa Factor scores Own a radio 0.346 0.351 Own a refrigerator 0.431 0.461 Own a television 0.479 Own a bicycle -0.047 0.072 Own a motorcycle 0.169 0.185 Own a car 0.292 0.396 House has electricity 0.453 0.507 Number of rooms per person 0.046 Floor is made of "finished" surface 0.376 0.459 Summary statistics Proportion of variation explained by first principal component 33.1 34.7 Value of first eigenvalue 2.98 2.43 Difference between first and second eigenvalue 1.75 1.28 Number of households 71,673 65,309 Source: Author's calculation from pooled and weighted DHS data. In order to define quintiles, individuals are sorted by the wealth index within each region, and cutoff values for the quintiles of the population are derived. Households are then assigned to each of these groups on the basis of their value of the asset index.'4 The interpretation is therefore that the poorest quintile is the group in which the poorest 20 percent of the population live. Note that the use of the term "poor" here differs from the across the different types of variables, and the pairwise correlation (across replications) between the loading from a continuous variable and weight on the corresponding dichotomous variable are very high as well, on the order of .9. 11. Principal components in typically derived for continuous variables. Monte Carlo simulations comparing the first principal component derived from (1) continuous variables to (2) dummy variables with the same mean (derived from the ranking of the continuous variables) yields very similar factor loadings across the different types of variables, and the pairwise correlation (across replications) between the loading from a continuous variable and weight on the corresponding dichotomous variable are very high as well, on the order of .9. 12. This method of ranking households is analogous to fairly standard approaches which use consumption expenditure quintiles. 13. Principal components in typically derived for continuous variables. Monte Carlo simulations cornparing the first principal component derived from (1) continuous variables to (2) dummy variables with the same mean (derived from the ranking of the continuous variables) yields very similar factor loadings across the different types of variables, and the pairwise correlation (across replications) between the loading from a continuous variable and weight on the corresponding dichotomous variable are very high as well, on the order of .9. 14. This method of ranking households is analogous to fairly standard approaches which use consumption expenditure quintiles. 9 usual notion derived from being below a poverty line. In this analysis, it refers to the, population that lives in households with low values of the asset index. Pooling data across regions The analysis in this paper is undertaken for "Western and Central Africa" and "Eastern and Southern Africa" as regions-pooling the data across countries-and in order to do so the data needs to be weighted appropriately. Many DHS collect weights in order for sample averages to represent nationwide averages. For example in some countries urban areas are over sampled and weights will adjust for this. In pooling the data across countries one needs to adjust, in addition, for the fact that the sample size relative to the population varies in the different countries. Observations from a country where the sample is only a very small percentage of the population need to be inflated relative to those in a country where a relatively large percentage of the population was sampled, and vice versa. For example, the Nigeria 1999 sample corresponded to 0.031 percent of Nigeria's population whereas the Togo 1998 sample equaled 1.016 percent of Togo's population. Annex 2 describes the derivation of weights in more detail. 3. Fever and poverty: Bivariate and multivariate analysis This analysis uses both a bivariate and a multivariate approach to analyzing the links between poverty and fever (and subsequently its treatment). The bivariate approach allows us to investigate to what extent fever and its treatment "move" with poverty, that is investigate the proposition that the rich suffer less from illness, or seek treatment more if ill. The multivariate approach allows one to disentangle the partial association of fever and its treatment with poverty after controlling for factors which may be correlated with household wealth and have independent relationships to fever. This approach allows one to investigate the degree to which the observed association is determined by other variables-such as mother's education. Geographic distribution offever and poverty At the global level, there is no doubt that malaria and poverty move together. Based on country-level data, Gwatkin and Guillot (2000) have recently estimated that 57.9 percent of deaths due to malaria occurred among the poorest 20 percent of the world's population in 1990. McCarthy, Wolf and Wu (2000) and Gallup and Sachs (1998) investigate the extent to which the national level association between malaria and low GNP per capita or low growth in GNP per capita is causal. Both studies conclude that malaria causes accounts for a substantial percentage reduction in GDP per capita growth. More generally, the countries where malaria is transmitted are also poor countries and while the relationship is not perfect, a focus on malaria is a focus on the world's poor. 10 Turning back to the DHS data, now that we have an indicator of fever and a measure of wealth, we can look at the relationship between the two within the two large regions of Sub-Saharan Africa. First, consider the geographic distribution of fever and poverty. Overall the numbers do not support the notion that there is a strong association between the incidence of fever and the wealth of the area in which the child lives. Figure 1 shows the incidence of fever in the past two weeks among children under three and of the proportion of the population in the by poorest quintile by geographic subregions. 15 While there are some subregions that are both poor and have a lot of fever, there is not an overwhelming resemblance between the figures in either of the regions. This impression is confirmed by examining the correlation between the two variables in each region (Table 5). Figure 1. Fever and poverty in Sub-Saharan Africa: Proportion of children under 3 who had a fever in the past two weeks and proportion of population in the poorest quintile Fever Poverty Western and Central Africa E10.168 - 0.287 LI0.017 - 0.136 _ 0.287 - 0.368 0.136 - 0.244 _ 0.368 - 0.426 0.244 - 0.368 _ 0.426 - 0.577 _ 0.368 - 0.568 _ 0.577 - 0.729 _ 0.568 - 0.782 = Nodata = No daa 15. Geographic subregions are generally administrative provinces or regions of countries reported in the surveys. The breakdown is reported in Annex 3. Eastern and Southern Africa ..6 - z ~~~~~~~~ ~~~~~~~A w ~ U- U i 0.127 - 0.275 0- 0.043 0.275 - 0.365 0.043 - 0.138 0.365 - 0.441 0.138 - 0.234 0.441 - 0.522 0.234 - 0.331 0.522 - 0.649 0.331 - 0.444 LiiiNodata Li Nodata Table 5. Correlation coefficients between percentage poor and incidence of fever (and fever without a cough) in children under three in subregion Any fever Fever without a cough Correlation Correlation coefficient P-value coefficient P-value Western and Central Africa .035 .760 .217 .052* Eastern and Southern Africa .218 .059* .215 .062* Excluding Madagascar .300 .012** .345 .003"* **(*) Significantly different from zero at 5(10) percent level. Source: Author's calculations from pooled and weighted DHS data. The bivariate relationship between fever and poverty is insignificant in Western and Central Africa and is significant at the 10 percent level in Eastern and Southern Africa, but small in magnitude. Since Madagascar stands out on the map as an area with a low level of fever but a high level of poverty, it was temporarily removed from the sample on the theory that it may be capturing a different relationship than on the continent. When it is removed the magnitude of the correlation increases, and the relationship becomes significant at the 5 percent level. Of course, this bivariate relationship is picking up many confounding factors. For example, in a simple regression that includes the average years of schooling of the mothers in the sample the association 12 becomes negative and significant in Western and Central Africa and insignificant in Eastern and Southern Africa (Table 6). After excluding Madagascar, the relationship is positive but insignificant in Eastern and Southern Africa. The relationship with average education of mothers is statistically significarnt in the expected direction: higher education is associated with a reduction in the reported incidence of fever. Table 6. OLS regressions of average incidence of fever in sub-region on average poverty and average mothers' education Fever Fever without a cough Western Eastern Eastern and Western Eastern Eastern and and and Southern and and Southern Central Southern Africa (excL Central Southern Africa (excL. Africa Africa Madagascar) Africa Africa Madagascar) Proportion in poorest quintile -0.134 -0.058 0.059 0.008 -0.019 0.083 (2.25)* (0.35) (0.33) (0.20) (0.22) (0.88) Average years of schooling of mothers -0.027 -0.021 -0.020 -0.011 -0.010 -0.009 (4.19)** (2.70)** (2.5 1)* (2.38)* (2.46)* (2.16)* Constant 0.479 0.504 0.484 0.196 0.189 0.169 (16.91)** (8.17)** (7.67)** (9.96)** (5.77)** (5.08)** Observations 81 76 70 81 76 70 R-squared 0.18 0.13 0.17 0.11 0.12 0.18 Note: Absolute value of t-statistics in parentheses; * significant at 10%; ** significant at 5%. Source: Author's calculations from pooled, and weighted and aggregated DHS data. Bivariate relationship between fever and poverty at the individual level Abstracting from geography, Table 7 reports the percentage of children under three who have had a fever in the past two weeks by quintile. This bivariate relationship conforms to the geographic results: although the incidence is generally larger in the poorest quintile than in the richest quintile, there is not a strong relationship between fever and poverty in either region. In Western and Central Africa, 28 percent of children in the richest quintile are reported to have had a fever in the past two weeks, whereas almost 3 8 percent of children in the poorest quintile reported fever. In Eastern and Southern Africa the percentage is 36 percent for the richest quintile and 42 percent for the poorest. 13 Table 7. Percentage of children under 3 who are reported to have had a fever (and fever withoiut a cough) in the past two weeks by quintile ranking of household. Poorest- Poorest Quintile 2 Quintile 3 Quintile 4 Richest Richest Western and Central Africa Any fever 37.6 39.5 37.7 32.1 27.9 9.7 Fever without a cough 19.1 19.8 18.2 17.2 13.9 5.2 Eastern and Southern Africa Any fever 41.7 42.6 43.5 43.9 36.4 5.3 Fever without a cough 13.4 14.0 14.3 15.5 12.7 0.7 Source: Author's calculation from pooled and weighted DHS data. The change does not appear to be uniform across the distribution. In Western and Central Africa the percentage with a fever goes from 38 to 32 percent between the third and fourth quintiles, and in Eastern and Southern Africa it goes from 44 to 36 percent between the fourth and fifth (richest) quintiles. Across the remainder of the distribution the level is almost flat-with a slight increase between the first and second quintile. Because differences in absolute levels of wealth can be small between one quintile and the next, quintiles can distort one's impression of the relationship. Figure 2 shows a non- parametric estimate of the incidence of fever (and fever without a cough) for each value of the wealth index. 16 The visual impression of these graphs is that of a weak downward slope suggesting that the quintiles may indeed lead one to understate the differential across the distribution. 16. The estimate is a moving average estimate of the percentage who report fever across the wealth distribution and includes about 5 percent of the sample for each estimate (2,501 observations in Western and Central Africa and 2,001 observations in Eastem and Southern Africa). Point estimates are then connected with a cubic spline. 14 Figure 2. Percentage of children under three who were reported to have a fever (and fever without a cough in the dotted line) in the previous two weeks: moving average estimates across the wealth distribution. Western and Central Africa Eastern and Southern Africa Fever Fever without a cough Fever Fever without a cough 50 - 50 40 -0 40- a)0 30 - 30- X 20- - X 20- 10 - 10 __ i ' 6 6 1 7 4 6f Index Gf wealth Index of weafth Multivariate relationship between fever and poverty at the individual level The results so far indicate a weak (although sometimes statistically significant) relationship between the geographic distribution of fever and poverty and in the bivariate relationship across the wealth distribution. The analysis has not, however, controlled for background characteristics and location of residence: that is, given a set of individual and househ old background characteristics, is being from a richer or poorer household associated with a lower the incidence of fever for a given child? In order to investigate this proposition, the following model was estimated for each of the regions: F icr* = b, x Wicr + b2 X W2icr + + a x Xicr + Xk=2,R dr x Dr + Uicr (1) where F ir* is an unobserved variable whose observed counterpart, whether or not a child -indexed by i, in cluster c, in sub-region r - had an episode of fever, is defined as F icr = 1 if F icr* >=° = 0 otherwise. Wealth effects are specified by including the child's household wealth index (W) and its square (W2). An additional specification includes the average level of wealth of 1S other households (i.e. the non-self mean) in the cluster in which the child lives (CW) and its square (CW2) as well as the average level of fever among children from other households in the cluster (CF) to control for the overall local level of fever. 17 Child and household background variables (X) include the child's age in months, age in months squared, the child's gender, the child's mother's years of schooling, her husband's (or partner's) years of schooling, a dummy variable equal to one if the household gets its drinking water from a covered source (such as a tap or a pipe as opposed to a well or a streamn), a dummy variable equal to one if the household has a flush toilet or a pit latrine (as opposed to no toilet facilities), and a dummy variable equal to one if the child resides in an urban area. Last, in order to control for district level variables, a set of dummy variables Dr for each of the R subregions (the areas shown in Figure 1) is included in the regression as well (one excluded sub-region constitutes the reference category). These will capture general area effects, and the resulting partial relationships for the other variables are therefore estimated conditional on the sub-region of residence. The results, reported in Table 8, again show a weak association between the reported incidence of fever and the household's wealth, conditional on the control variables. In the models without cluster variables the relationship is small and negative, and the coefficients on the wealth variables are jointly significant at the 5 percent level in Eastern and Southem Africa. With cluster variables the association is negative in Western and Central Africa and positive in Eastern and Southern Africa, but in both regions the wealth variables are jointly insignificantly different from zero.'8 The wealth of other households in the cluster is significantly negatively associated with fever in Eastern and Southern Africa, but insignificantly so in Western and Central Africa. The models consistently find a significant relationship between fever and the incidence of fever in other households in the cluster. While this may not appear to be surprising, recall that this is conditional on many other individual, household, cluster, and regional characteristics- suggesting a strong underlying geographic concentration of episodes of fever. Male children have statistically significant higher reported incidence of fever in Western and Central Africa, but the differential is small with the average predicted probability going from 34 for girls to 36 for boys. In Eastern and Southern Africa the relationship is insignificant, although it is positive as well. Age has a statistically significant inverse-U shaped relationship with incidence in both regions, with the highest incidence occurring at 16 months. Mother's years of schooling is significantly negatively related to the reported incidence of any fever in Eastern and Southern Africa, but the 17. Clusters are the lowest level from which a sample of households is drawn, i.e., these are typically the primary sampling unit in the data with about 20 households in a cluster. 18. An alternative specification that includes wealth as dummy variables for quintile produces results that are qualitatively similar. 16 magnitude of the effect is not large. The average predicted probability of fever setting the education of all mothers to zero is 44 percent whereas it is 41 percent setting the education of all mothers to 6 years, i.e., only about a three percentage point differential for 6 years of schooling. The mother's husband's (or partner's) education is insignificantly related to the incidence of fever. Table 8. Probit estimates of the relationship between fever, wealth and background characteristics Western and Central Africa Eastern and Southern Africa (1) (2) (3) (4) Wealth index -0.024 -0.029 -0.011 0.004 (1.33) (1.60) (0.62) (0.22) Wealth index squared -0.000 0.004 -0.003 -0.002 (0.01) (0.54) (0.91) (0.61) Cluster mean wealth index 0.017 -0.057 (0.72) (2.17)** Cluster wealth index squared -0.016 0.001 (1.61) (0.10) Cluster proportion with fever 0.794 0.585 (9.07)** (8.68)** 1 = Male 0.055 0.055 0.030 0.028 (2.48)** (2.46)** (1.56) (1.45) Age (months) 0.065 0.066 0.065 0.066 (12.35)** (12.55)** (16.88)** (16.93)** Age squared (months) -0.002 -0.002 -0.002 -0.002 (12.33)** (12.51)** (16.97)** (17.02)** Mother's schooling (yrs) -0.001 -0.001 -0.012 -0.010 (0.18) (0.15) (3.03)** (2.64)** Husband's schooling (yrs) -0.000 -0.001 0.001 0.001 (0.04) (0.22) (0.38) (0.37) 1= Drinking water covered -0.003 0.010 0.024 0.024 (0.08) (0.28) (0.93) (1.03) 1= Toilet flush or pit latrine 0.057 0.055 0.004 0.011 (1.44) (1.65)* (0.13) (0.38) 1 = Urban -0.085 -0.069 0.007 0.057 (1.63) (1.55) (0. 19) (1.65)* Constant -0.781 -1.084 -0.207 -0.582 (7.08)** (11.27)** (1.21) (4.00)** Model includes dummy variables for sub-region and constant term (not reported) Observations 38,916 38,828 29,035 28984 Joint tests (p-values) Own wealth 0.13 0.20 0.04** 0.76 Cluster wealth 0.27 0.01** Cluster variables (wealth and 0.00** 0.00** fever) Note: Model includes dummy variables for region (not shown) and a dummy variable for husband's data available. Robust z-statistics in parentheses. * significant at 10%; ** significant at 5%. Source: Author's calculations from pooled and weighted DHS. T-statistics, adjusted for clustering are reported in parentheses. 17 Country by country results It is possible that there are too many confounding factors that would be obscuring a significant relationship. In particular, country-to-country differences in the exposure to malaria, in the support of the health care system, or relative position in the wealth distribution might not be well captured in the pooled model above (despite the dummy variables for subregions). In addition, the pooled approach to estimating wealth might overemphasize differences between countries compared to smaller differences within countries. Therefore, these country-by-country regressions used a wealth index that was recalculated country by country. Table 9 reports the selected p-values of the joint tests of significance of the wealth and cluster-level wealth and fever variables in a model that also includes all the other control variables (see Table 8) and that allow all the coefficients to differ across countries. Table 9. P-values of tests on wealth, average cluster wealth, and average cluster fever variables in country-by-country probit regressions of fever Joint test Joint test Joint on cluster Joint test on cluster test on wealth on wealth wealth and wealth and and cluster Average and cluster Average wealth wealth cluster wealth wealth cluster Country squared squared fever Country squared squared fever Benin 1993 0.403 0.421 0.072** Comoros 1996 0.567 0.848 0.097* Burkina Faso 1999 0.701 0.677 0.000** Kenya 1998 0.730 0.044** 0.001** C.A.R. 1994-95 0.313 0.525 0.025* Madagascar 1997 0.824 0.438 0.001** Chad 1996 0.082* 0.307 0.000** Mozambique 1997 0.978 0.450 0.900 C6te d'lvoire 1994 0.195 0.133 0.245 Malawi 1996 0.948 0.996 0.067* Cameroon 1998 0.069* 0.388 0.000** Rwanda 1992 0.559 0.685 0.001 ** Ghana 1998 0.431 0.137 0.172 Tanzania 1996 0.475 0.022** 0.054* Mali 1995-96 0.773 0.160 0.000** Uganda 1995 0.253 0.902 0.000** Niger 1997 0.604 0.685 0.000** Zambia 1996-97 0.175 0.276 0.000** Nigeria 1999 0.529 0.055 0.000** Zimbabwe 1999 0.869 0.583 0.125 Senegal 1992-93 0.423 0.166 0.000** Togo 1998 0.910 0.848 0.000** Note: *(**) indicates underlying variables that are significantly different from zero at the 10(5) percent level. Source: Authors' calculations from DHS data. Model includes individual, household and cluster variables as well as dummy variables for sub-regions (not reported, see Table 8 for list of variables). P-values are calculated adjusting for clustering. Again, the results do not provide evidence for a strong relationship between reported fever and poverty. Wealth-both at the household and cluster levels-are typically not statistically significantly associated with fever. Chad and Cameroon are exceptions where household wealth is significant, and Kenya and Tanzania are 19. An alternative specification that includes wealth as dunmmy variables for quintile produces results that are qualitatively similar. 18 exceptions where cluster wealth and fever are significantly related. On the other hand, the average level of fever in the cluster is significantly related to wealth (with the exception of CMte d'Ivoire, Ghana, Mozambique, and Zimbabwe). 4. Treatment seeking behavior and poverty Seeking carefor malaria The focus on fever here is driven by its availability in the DHS data. Typically, however, episodes of fever are what prompt caregivers to seek treatment, and most often patients are treated presumptively (both by parents and medical personnel) for malaria. Based on their review of the literature, Brinkmann and Brinkmann (1991) conclude that between 8 and 25 percent of persons with malaria visit health services, with self- treatment being more common in urban than in rural areas (more than 60 percent versus between 2 and 25 percent). McCombie (1996) reviews the literature on treatment seeking for malaria and finds a substantial variation across countries. On average, close to 50 percent of cases rely exclusively on self-treatment-usually with antimalarials. Most episodes involve some form of self-treatment, which in general involves the purchase of drugs. The use the official health sector-hospitals, clinics, dispensaries, private practitioners, and village health workers-for treatment varied from 10 to 99 percent, depending on the country and the type of study (with about half the studies finding more than 50 percent). Very few cases rely exclusively on traditional methods (or not even at all for uncomplicated malaria). The review also identified urban-rural differences as the most common source of variation across studies. McCombie (1996) also observes that the community prevalence of malaria reduced the probability of seeking care from a doctor. In general, "[...] experience with malaria affects treatment seeking behavior and leads to diffusion of information on how to treat it (op. cit. p. 941). Other more recent household survey-based studies conform to these results. In western Kenya 60 percent of fever episodes were treated at home with only 18 percent resulting in a visit to a health clinic or a hospital, with the remainder seeking no treatment (Ruebush and others 1995). In coastal Kenya 23 percent of mothers reporting that a child had malaria in the prior two weeks had taken the child to a health facility (Mwenesi, Harpham, and Snow 1995). Fifty-four percent had given over the counter drugs to the sick child and 24 percent had given no treatment, or had given a home remedy. In southern Ghana, fever is mostly treated at home with commonly available drugs and herbal remedies and a visit to a health center was the last resort after failure of home treatment (Ahorlu and others 1997). On the other hand, in Malawi a higher share of episodes, 52 percent, resulted in a visit to a clinic (Slutsker and others 1994). In that study higher socioeconomic status was found to be positively correlated with clinic attendance. 19 DHS data on the treatment offever In a subset of the DHS questionnaires, mothers were asked to report what, if any, action was taken if they responded that their child had a fever in the past two weeks. Analyzing this data can be done for Burkina Faso 1992-93, Cameroon 1991, C6te d'Ivoire 1994, Ghana 1993, Niger 1992, Nigeria 1999, and Senegal 1992-93 in Western and Central Africa, and for Kenya 1998, Malawi 1996, Rwanda 1992, Madagascar 1992, Tanzania 1992, Zambia 1992 and Zimbabwe 1999 in Eastern and Southern Africa. Note that Niger, Madagascar, Tanzania, and Zambia refer to a survey from a different year from that used in the incidence analysis. Typically, the surveys will ask "did you seek advice or treatment for the fever" for a child who is reported to have had a fever in the past two weeks. In some cases, the question is asked whether the child is reported to have had a fever or a cough with rapid breathing in the past two weeks. In those cases where the child is reported to have had both a fever and a cough it is impossible to know whether the advice or treatment was sought for the fever and not for the cough. In this analysis we ignore the problem and include the advice/treatment seeking behavior as long as the child is reported to have had at least a fever. The types of modern sector facilities/persons that the mother can report having visited are grouped into: "higher-level public" (i.e., government hospital); "lower-level public" (e.g., government health center, government health post, mobile clinic, community health worker); "private medical" (e.g., private hospital/clinic, private doctor, private mobile clinic); "private commercial" (e.g., pharmacy, shop); and traditional healers.20 In addition to these generic options, country specific options (for example a nurse's practice, public health post, and a pharmaceutical depot) have been mapped to the basic classification. Among the responses included in the "no modern sector" category are "no treatment or advice" and "advice from friends or family." Table 10 shows the basic results in on treatment seeking behavior in the study countries. There is about a ten-percentage point difference between the two regions in the overall level of modern sector use. In Western and Central Africa about 56 percent of cases of fever resulted in a visit to the modern health sector, whereas in Eastern and Southern Africa about 65 percent of cases did. While the focus here is on regional averages, the data do show wide variation across countries within each region. For example, 43 percent of cases of fever among children in Western and Central Africa did not result in any medical advice sought, but this ranges from 15 percent in Cote d'Ivoire to about 75 percent in Burkina Faso and 20.Private medical facilities may be "commercial" in nature. This terminology is used purely to distinguish the two types of private services here. 20 Niger. The type of treatment sought clearly depends on country characteristics and policies. For example, 29 percent of cases of fever in children resulted in a visit to a lower-level public facility in Eastern and Southern Africa but this masks a range of 16 percent in Malawi to 41 percent in Zambia. The data are consistent in showing a very small degree of treatment or advice sought from traditional healers: on average 1.2 percent of cases in Western and Central Africa and 1.6 percent of cases in Eastern and Southern Africa. Table 10. Treatment/Advice sought as a result of a child under 3 having a fever in the past two weeks No Public, Treatment / higher Public, Private, Private, Advice level lower level medical commercial Traditional Total Western and Central Africa Burkina Faso 1992/3 76.6 1.7 16.5 1.7 0.0 3.5 100 C6te d'Ivoire 1994/5 15.0 26.8 32.1 17.5 5.4 3.3 100 Cameroon 1991 52.1 8.1 19.2 14.5 2.3 3.8 100 Ghana 1993 31.3 22.3 15.6 12.5 14.3 4.1 100 Niger 1992 74.4 0.1 10.7 8.7 3.8 2.4 100 Nigeria 1999 38.8 13.1 14.2 11.5 22.2 0.3 100 Senegal 1992/3 60.7 4.0 25.7 4.8 2.1 2.6 100 Total 43.0 12.2 15.7 10.9 17.0 1.2 100 Eastern and Southern Africa Kenya 1998 23.4 12.2 24.7 22.5 16.5 0.8 100 Madagascar 1992 49.4 13.1 18.5 13.7 2.6 2.7 100 Malawi 1996 31.1 3.1 16.0 17.2 31.1 1.5 100 Rwanda 1992 54.6 5.4 26.3 7.5 3.0 3.2 100 Tanzania 1991/2 35.7 10.7 40.2 9.1 2.6 1.7 100 Zambia 1992 22.1 9.4 40.9 19.1 6.2 2.3 100 Zimbabwe 1999 35.2 6.2 31.2 15.2 12.0 0.3 100 Total 33.2 9.5 29.2 15.6 10.9 1.6 100 Source: Author's calculations from pooled and weighted DHS data. The classification is: "higher-level public" is government hospital, "lower-level public" is government health center, government health post, mobile clinic, community health worker, "private medical" is private hospital/clinic, private doctor, private mobile clinic, and "private commercial" is pharmacy or shop. In some countries additional options have been mapped to this classification. The "no modern sector" category includes no treatment or "professional" advice and advice from friends or family. 21 Treatment offever and household wealth. Figure 3 and Table 11 show the association between treatment seeking and poverty as shown in.21 The first striking result is that wealthier households are substantially more likely to seek treatment or advice in the modem sector in response to an episode of fever. In Western and Central Africa the percent who seek no modem sector care is 2.8 times as high in the richest than in the poorest quintile, 64 versus 23 percent, and in Eastern and Southern Africa it is 1.8 times as high, 41 versus 23 percent. There is quite a bit of differentiation across countries underlying these regional averages. In Western and Central Africa the smallest differential is in Senegal (1.4 times) and the largest is in C6te d'Ivoire (3.7 times). In Eastern and Southern Africa the ratio ranges from 1 in Zimbabwe to 2.6 in Zambia. The second feature to emerge from Table 11 is the high degree of unequal usage of higher-level public health facilities by the rich and the poor. In Western and Central Africa, 25 percent of fever cases involving children from the richest quintile result in a trip to a government hospital, whereas among the poorest quintile the number is only 5.2 percent. In Eastern and Southern Africa the percentage among the rich is about 17 percent, but is 8.4 percent for the poor. In both regions there is a substantial increase going from the fourth to the fifth (richest) quintile. Figure 3. Type of treatment sought as a result of fever in the past two weeks Western and Central Africa Eastern and Southern Africa 100 - 100 80 - 80 60- 60 - ffi 40- 1' !. 40 20 -| _ 20 0 0 Poorest Quintile 2 Quintile 3 Quintile 4 Richest Poorest Quintile 2 Quintile 3 Quintile 4 Richest * Public, higher level E Public, lower level 0 Private, medical E Private, commercial * Traditional 21. Since the data sets are different from the incidence analysis, the wealth index was recalculated on the pooled and reweighted observations in the new set of countries/years. The principal components yields a very similar set of "weights" for the index. These are not reported here, but are available on request from the author. 22 Table 11. Advice/Treatment sought as a result of a child having a fever in the past two weeks Public, No /Self higher Public, Private, Private, treatment level lower level medical commercial Traditional Total Western and Central Africa Quintile I (poorest) 64.2 5.2 9.7 8.6 11.1 1.4 100 Quintile 2 50.2 8.6 16.2 6.9 15.9 2.2 100 Quintile 3 36.5 12.7 20.5 9.7 19.9 0.7 100 Quintile4 27.5 16.6 18.3 17.5 19.4 0.8 100 Quintile 5 (richest) 23.0 24.5 14.0 16.6 21.8 0.1 100 Total 43.0 12.2 15.7 10.9 17.0 1.2 100 Eastern and Southern Africa Quintile I (poorest) 40.5 8.4 28.6 13.1 7.2 2.2 100 Quintile 2 37.3 5.8 32.1 11.2 11.2 2.4 100 Quintile3 34.4 8.1 28.7 16.5 11.9 1.3 100 Quintile 4 28.9 10.5 30.6 14.9 13.9 1.3 100 Quintile 5 (richest) 21.3 17.3 25.4 25.0 10.6 0.5 100 Total 33.2 9.5 29.2 15.6 10.9 1.6 100 Source: Author's calculations from pooled and weighted DHS data. The use of lower-level public facilities is greater than that of higher-level facilities, especially for the poorest groups in both regions. Perhaps surprisingly, in Western and Central Africa the poorest use lower-level public facilities substantially less than people from the upper quintiles-the differential is on the order of 4 percentage points. This is not the pattern in Eastern and Southern Africa where the use of lower-level public facilities is fairly constant at around 30 percent, falling slightly in the richest quintile. Seeking treatment from private sources, either medical or commercial, is about 25 percent. In both regions the use of private facilities increases over most of the wealth distribution, although it falls for commercial sources at the upper quintiles in Eastern and Southern Africa. This fall is (more than) made up for by an increase in the use of private medical facilities and government hospitals however. Overall in this region the use of all private facilities for treatment and advice increases from 17 percent in the poorest quintile to 28 percent in the richest. In Western and Central Africa the use of private facilities increases from 20 percent in the poorest quintile to 38 in the richest quintile. Multivariate analysis Again, multivariate analysis can help sort out some of the confounding factors. The approach used here is to estimate a Multinomial Logit (MNL) model for the sample that reports having had fever in the past two weeks, i.e. an estimate of the correlates of treatment choice conditioned on the sample who were ill. The approach allows one to investigate the partial association between treatment choices and household wealth, after conditioning on the same set of variables as in the incidence analysis, i.e., observed 23 individual, household characteristics, and cluster variables as well as controlling for thLe subregion in which the individual lives (again through the inclusion of dummy variables). Multinomial Logit estimates can be derived from what is known as the "randoim utility" model. For example, the utility from choicej for individual i am given as Uij = ,j x Zij + vij (2) where Z refers to all the regressors in the model (see equation 1). Under the assumption that the error terms (v) are identically distributed with a specific distribution, one can derive the expressions for the coefficients for each outcome, specifically: Prob(Y1 =j) = (e 31 X zij) / (Y l , e p x Zij) (3) The model is unidentified (i.e., there are many solutions yielding the same set of probabilities) and the usual way of estimating it is under the restriction that the coefficients for the "reference choice" are all equal to zero. The resulting coefficient estimates are therefore only interpretable relative to this base category. In this analysis the reference category has been set to those who sought no modem sector advice or treatment in response to an episode of fever.22 There are two major assumptions that are being made in this estimation. First, since the model is estimated conditional on the set of those reporting fever, the estimates will be biased if unobserved factors that determine fever (u s from equation 1) are correlated with unobserved factors determining treatment choice (v s from equation 2). In addition, the model assumes that the error terms from the different choices are uncorrelated (i.e., the v s from the different choices in equation 2), also known as the Independence of Irrelevant Alternatives (IIA) assumption. Therefore the Multinomial Logit model is estimated under the assumptions of no sample selectivity and independence of the errors in the choice model. Table 12 reports the results from the MNL estimation for the specification that excludes average cluster wealth and fever. The results confirm that the relationship between being in the richer quintiles in Western and Central Africa and the use of public care is significantly different from zero. In the MNL results, higher wealth is significantly associated with more public facility use-substantially more so for higher-level services. 22. Note that seeking care from traditional healers is grouped with no treatment in this part of the analysis. This is mostly because the multivariate choice analysis is difficult to identify when only few cases choose one particular choice. 24 In Eastern and Southern Africa, wealth is significantly positively associated with more public higher-level care, and private care from both medical and commercial sources. It is unrelated to public lower level care. Table 12. Multinornial logit estimates of treatment choice for children under 3 reporting fever in the past two weeks (no modern sector treatment is reference choice) Western and Central Africa Eastern and Southern Africa Public, Public, Private, Public, Public, Private, higher lower Private, commer higher lower Private, commer level level medical -cial level level medical -cial Wealth index 0.325 0.190 0.165 0.161 0.129 -0.016 0.121 0.146 (2.58)** (1.71)* (1.24) (1.29) (1.75)* (0.31) (2.03)** (1.83)* Wealth index squared 0.040 -0.093 0.006 -0.087 -0.012 0.000 0.014 -0.038 (0.61) (1.60) (0.08) (1.21) (0.59) (0.02) (0.79) (1.59) I = Male 0.091 0.053 0.175 0.222 0.043 -0.004 0.023 -0.127 (0.34) (0.26) (0.64) (0.89) (0.39) (0.05) (0.26) (1.13) Age (months) 0.047 0.034 -0.006 -0.064 0.048 0.054 0.068 0.011 (0.71) (0.78) (0.10) (1.17) (1.91)* (3.36)** (3.44)** (0.44) Age squared (months) -0.001 -0.000 0.000 0.002 -0.002 -0.002 -0.002 -0.000 (0.33) (0.38) (0.06) (1.52) (2.52)** (3.84)** (3.51)** (0.22) Mother's schooling (yrs) -0.017 0.055 0.113 -0.037 0.050 0.053 0.060 -0.003 (0.37) (1.49) (2.69)** (0.77) (2.53)** (4.08)** (3.50)** (0.15) Husband's schooling (yrs) 0.025 0.033 0.100 0.065 0.035 0.033 0.024 0.016 (0.72) (1.05) (2.68)** (1.74)* (1.89)* (2.54)** (1.46) (0.82) 1= Drinking water covered -0.516 0.235 0.101 -0.197 0.099 0.276 0.206 0.209 (1.48) (0.76) (0.30) (0.64) (0.75) (3.34)** (1.88)* (1.51) 1= Toilet flush or pit latrine 0.449 -0.056 0.001 0.235 0.362 0.139 0.272 0.229 (1.21) (0.20) (0.00) (0.73) (2.29)** (1.45) (2.20)** (1.52) I = Urban 1.624 0.294 0.018 0.804 1.164 -0.445 0.068 0.001 (4.29)** (0.98) (0.05) (2.30)** (7.22)** (3.01)** (0.44) (0.00) Constant -5.454 -1.840 -1.518 -2.049 -1.418 -0.163 -1.334 -1.811 (4.27)** (2.96)** (1.93)* (2.87)** (4.41)** (0.74) (4.64)** (4.82)** Observations 6406 6406 6406 6406 8176 8176 8176 8176 Joint tests (p-values) Own wealth 0.018* 0.065* 0.410 0.223 0.168 0.939 0.004** 0.157 Note: Model includes dummy variables for region (not shown) and a dummy variable for husband's data available. T- statistics, adjusted for clustering are reported in parentheses. * significant at 10%; ** significant at 5%. Source: Author's calculations from pooled and weighted DHS. The magnitude of these conditional associations can be assessed from the predicted probabilities summarized in Table 13. The table shows the percent probability of seeking each type of care conditional on having a fever, setting all observations to have the samne wealth (chosen to be the means of each quintile) and averaged across all observations. Even controlling for other characteristics, the use of public higher level services goes from 9.2 percent for the average wealth of the poorest quintile to 19 percent for the average wealth of the richest quintile in Western and Central Africa, with the magnitude increasing substantially between the wealth level of those in the fourth quintile to those in the fifth. The use of public lower-level care in Western and Central Africa 25 increases fastest at lower levels of wealth, and then stabilizes at around 18 percent at the wealth level of those in the third and fourth quintiles, and then decreases slightly to 16 percent at the richest level of wealth evaluated. In Eastern and Southern Africa while the associations are significant, they are typically not as large. The percentage who use no modem care falls from 36 to 32 percent for the average wealth of the poorest to richest quintiles. It is only the use of private medical facilities that increases appreciably with wealth: from 14 percent for those with the wealth of the poorest quintile to 19 percent to those with the wealth of the richest quintile. Table 13. Average predicted probabilities from Multinomial Logit Estimation of Advice/Treatment sought as a result of a child under 3 having a fever in the past two weeks Public, Public, No / Self higher lower Private, Private, treatment level level medical commercial Total Western and Central Africa Probabilities evaluated at: Mean wealth in the poorest quintile 53.4 9.2 12.0 10.7 14.7 100 Mean wealth in quintile 2 46.1 9.4 16.0 10.3 18.1 100 Mean wealth in quintile 3 41.0 10.7 18.2 10.5 19.6 100 Meanwealthinquintile4 38.0 13.8 18.1 11.2 18.9 100 Mean wealth in the richest quintile 36.6 19.1 15.6 12.4 16.3 100 Eastern and Southern Africa Probabilities evaluated at: Mean wealth in the poorest quintile 36.0 8.8 30.9 14.3 10.0 100 Meanwealthinquintile2 35.9 8.8 30.8 14.4 10.1 100 Mean wealth in quintile 3 34.9 9.5 29.6 14.9 11.1 100 Meanwealthinquintile4 34.1 10.0 28.6 15.6 11.7 100 Mean wealth in the richest quintile 32.0 11.0 26.0 19.1 11.8 100 Source: Author's calculations from pooled and weighted DHS data. See Table 10 for definition of categories. Predictions are the average probability averaging over all individuals with their observed characteristics but substituting all observations to have the average wealth of the specified quintile. The MNL results conform to the expectation formed by the bivariate analysis in Table 11. A comparison between Table 11 (which doesn't control for background characteristics) and Table 12 (which does) implies that although a substantial amount of the rich-poor differential is explained by characteristics other than household wealth, wealth still plays a substantial role in determining whether to treat using modem methods versus home or no care, as well as treatment choice. Although the predicted differential is reduced after controlling for other characteristics it still exists. For example, the average predicted probability of seeking no modem sector care ranges from 53 percent for the poorest quintile to 37 the richest quintile in Western and Central Africa and from 36 to 32 percent in Eastern and Southern Afrca. The unadjusted ranges are 64 to 23 percent and 41 to 21 percent in each region respectively. Clearly other factors (including subregional fixed attributes) go a long way in explaining wealth differentials. 26 None of the other variables included in the model have a clear impact on the choice of service. In Western and Central Africa mother's education is associated with an increase in the probability of private medical facilities. In Eastern and Southern Africa mother's years of schooling is statistically significantly positively associated with seeking advice or treatment from all sources except private commercial. Table 14 reports the average predicted probability of seeking various types of care setting mother's education to 0 and then to 6 and averaging across all observations. Public lower-level facility use is about 4 percentage points higher in both regions for children of mothers with 6 years, as opposed to zero years, of schooling. Similarly, the use of private medical facilities is only about 3 or 4 percentage points higher among the more educated mothers. While the relationship with education is statistically significant, it is not large. Table 14. Average predicted probabilities from Multinomial Logit Estimation of Advice/Treatment sought as a result of a child under 3 having a fever in the past two weeks Public, Public, No / Self higher lower Private, Private, treatment level level medical commercial Total Western and Central Africa Mother has 0 years of schooling 45.4 13.8 13.4 6.6 20.8 100 Mother has 6 years of schooling 42.5 11.9 17.2 11.9 16.5 100 Eastern and Southern Africa Mother has 0 years of schooling 39.0 8.6 26.4 13.3 12.6 100 Mother has 6 years of schooling 33.4 9.7 30.3 16.0 10.6 100 Source: Author's calculations from pooled and weighted DHS data. Predictions are the average probability averaging over all individuals with their observed characteristics but substituting all observations to have the years of schooling of the mother in question. One surprising result is that having a flush toilet or a pit latrine is positively associated with an increase in the likelihood of seeking treatment or advice from many of the modern sector choices listed. One might expect that better sanitation is a proxy for the general health environment in the household and would therefore affect the probability of being reported as sick (which is not true according to Table 8), but it is surprising that it should affect the choice of care conditional on illness. It is possible that toilet facilities are picking up a component of wealth. A second specification estimated for each region includes the average wealth among other households in the cluster (and its square) as well as the incidence of fever among children from other households in the cluster. This specification will allow a test of whether it is a household's own wealth that matters or whether it is the general wealth of the surrounding households that matters. Moreover, it will allow a test of McCombie's (1996) observation that experience with malaria affects treatment choices. 27 Including the cluster variables, as reported in Table 15, changes neither the significance nor the magnitude of the control variables. On the other hand, household wealth in this model is almost always insignificant-the exception being increased public lower-level services among wealthier households in Westem and Central Africa. By contrast, cluster wealth is significantly related to many of the treatment choices analyzed: public higher-level facilities and private facilities in Western and Central Africa, and all types of services in Eastem and Southem Africa. Table 16 summarizes the magnitudLes of the estimated effects. The table shows the predicted probability of choosing a given type of care, setting the cluster average wealth variable to the mean level of wealth in the poorest 20 quintile of clusters, the second quintile of clusters, and so on; and then averaging those predicted probabilities across all observations for each of these levels. In both regions the probability of using no modem care falls substantially with increased cluster wealth: from 56 to 36 percent seeking no modem care among the poorest and richest clusters in Western and Central Africa, and from 39 to 27 seeking no modern care in Eastern and Southern Africa. In Westem and Central Africa the results imply large and statistically significant differentials between richer and poorer clusters: public higher level care goes from 7.2 to 20 percent, private medical care goes from 7.5 to 17, and seeking care from a private commercial source increases and then decreases among the richest clusters (with the largest differential being between the poorest clusters at 5 percent and the clusters with the wealth of those in the fourth quintile at 27 percent). Likewise in Eastem and Southem Africa the results imply large differentials for all but private commercial sources: use of public higher-level facilities increase from 7.9 to 11 percent, the use of public lower-level facilities decreases from 31 to 21 percent, and the use of private medical facilities increases from 12 to 31 between the poorest and richest clusters. 28 Table 15. Multinoniial logit estimates of treatment choice for children under three reporting fever in the past two weeks (no modern sector treatment is reference choice). Model including cluster wealth and fever Western and Central Africa Eastern and Southern Africa Public, Public, Private, Public, Public, Private, higher lower Private, commer higher lower Private, commer level level medical -cial level level medical -cial Wealth index 0.195 0.211 0.044 0.057 0.067 -0.016 0.015 0.114 (1.40) (1.71)* (0.29) (0.42) (0.77) (0.30) (0.24) (1.38) Wealth index squared 0.058 -0.086 0.031 -0.018 0.004 0.012 0.019 -0.021 (0.87) (1.42) (0.45) (0.24) (0.17) (0.64) (1.09) (0.84) Cluster mean wealth index 0.596 -0.135 0.518 0.407 0.421 0.074 0.532 0.273 (2.42)** (0.59) (1.99)** (1.55) (2.13)** (0.81) (4.92)** (1.81)* Cluster wealth index squared -0.013 -0.106 0.036 -0.698 -0.137 -0.074 -0.057 -0.110 (0.09) (0.72) (0.24) (3.72)** (1.79)* (1.76)* (1.20) (1.76)* Cluster prop. with fever 1.142 -0.684 -0.297 -0.096 -0.699 0.482 0.037 0.560 (1.48) (0.98) (0.39) (0.15) (2.22)** (2.19)** (0.15) (1.75)* 1 = Male 0.089 0.065 0.182 0.257 0.060 -0.008 0.029 -0.127 (0.33) (0.31) (0.65) (1.01) (0.55) (0.11) (0.32) (1.12) Age (months) 0.048 0.030 0.004 -0.051 0.046 0.053 0.063 0.009 (0.70) (0.68) (0.06) (0.90) (1.84)* (3.34)** (3.18)** (0.38) Age squared (months) -0.001 -0.000 -0.000 0.002 -0.002 -0.002 -0.002 -0.000 (0.33) (0.25) (0.02) (1.23) (2.43)** (3.80)** (3.22)** (0.15) Mother's schooling (yrs) -0.039 0.063 0.106 -0.028 0.046 0.052 0.055 -0.004 (0.83) (1.71)* (2.42)** (0.58) (2.30)** (4.04)** (3.16)** (0.21) Husband's schooling (yrs) 0.014 0.031 0.084 0.047 0.035 0.032 0.022 0.015 (0.42) (0.98) (2.17)** (1.25) (1.86)* (2.50)** (1.30) (0.76) 1= Drinking water covered -0.564 0.202 -0.043 -0.132 0.041 0.260 0.136 0.170 (1.65)* (0.67) (0.12) (0.40) (0.30) (3.12)** (1.23) (1.22) 1= Toilet flush or pit latrine 0.324 -0.144 -0.242 -0.105 0.313 0.138 0.237 0.209 (0.84) (0.49) (0.65) (0.30) (2.03)** (1.44) (1.90)* (1.39) 1 = Urban 1.102 0.315 -0.522 0.430 0.871 -0.482 -0.409 -0.177 (2.41)** (0.89) (1.16) (1.10) (3.91)** (2.93)** (2.37)** (0.74) Constant -5.259 -1.420 -0.845 -0.365 -0.719 -0.353 -1.024 -1.898 (3.64)** (1.86)* (0.82) (0.42) (1.71)* (1.32) (3.06)** (4.26)** Observations 6394 6394 6394 6394 8170 8170 8170 8170 Joint tests (p-values) Own wealth 0.260 0.083* 0.820 0.893 0.523 0.814 0.289 0.386 Cluster wealth 0.037** 0.698 0.134 0.000** 0.097* 0.209 0.000** 0.131 Cluster variables (wealth and fever) 0.040** 0.683 0.252 0.001** 0.044** 0.050* 0.000** 0.070* Note: Model includes dummy variables for region (not shown) and a dummy variable for husband's data available. T- statistics, adjusted for clustering are reported in parentheses. * significant at 10%; ** significant at 5%. Source: Author's calculations from pooled and weighted DHS. 29 Table 16. Average predicted probabilities from Multinomnial Logit Estimation of Advice/Treatment sought as a result of a child under 3 having a fever in the past two weeks No / Self Public, Public, Private Private treat- higher lower medi- com- ment level level cal mercial Total Western and Central Africa Probabilities evaluated at: Mean cluster wealth in the poorest quintile of clusters 56.2 7.2 23.6 7.5 5.4 100 Mean cluster wealth in second quintile of clusters 46.9 8.5 20.9 8.0 15.8 100 Mean cluster wealth in third quintile of clusters 39.0 9.8 17.0 8.6 25.6 lO1 Mean cluster wealth in fourth quintile of clusters 35.1 13.4 13.3 11.3 26.9 100 Mean cluster wealth in the richest quintile of clusters 36.2 19.8 10.9 17.1 16.0 100 Eastern and Southern Africa Probabilities evaluated at: Mean cluster wealth in the poorest quintileofclusters 38.7 7.9 31.4 11.8 10.2 100 Mean cluster wealth in second quintile of clusters 36.3 8.8 31.0 13.2 10.8 100 Mean cluster wealth in third quintile of clusters 33.7 9.9 30.2 15.0 11.3 100 Mean cluster wealth in fourth quintile of clusters 30.4 11.2 28.2 18.5 11.7 100 Mean cluster wealth in the richest quintile of clusters 27.0 11.3 21.0 30.8 10.0 100 Source: Author's calculations from pooled and weighted DHS data. See Table 10 for definition of categories. Predictions are the average probability averaging over all individuals with their observed characteristics but substituting all observations to have the average cluster wealth of the specified quintile of clusters. Last, the results support the notion that experience with fever affects treatment choice, but only in Eastern and Southern Africa. The coefficient on the average incidence of fever in other households of the cluster is significantly negatively related to the use of public higher-level facilities; positively related to the use of lower-level public facilities; and (weakly) positively related to seeking care from private commercial sources. In Western and Central Africa, however, the coefficient is always insignificantly different from zero. 30 5. Conclusions There are serious caveats to using DHS data to analyze the relationship between malaria, its treatment, and poverty in Sub-Saharan Africa: there is not a one-to-one correspondence between fever and malaria, even in areas with stable malaria; fever is not always recognized in children, and it may be recognized in a selective way that potentially biases results; the DHS do not collect the data typically used for poverty analysis. Recognizing these caveats, this analysis explores the existing DHS data for patterns between fever and poverty. The results show a positive, but weak, relationship between reported fever and poverty across geographic space. Some areas have both high levels of fever and poverty and the correlation between the two is positive. However, this correlation, while statistically significant, is small in magnitude. Moreover, it becomes insignificantly different from zero after controlling for the education of mothers in the area. The relationship between fever and wealth across households is also insignificant, although in Eastern and Southern Africa a higher level of wealth in the cluster in which a household lives is associated with a lower incidence of fever. Fever is highly geographically concentrated: even after controlling for a variety of individual, household, and cluster attributes (including wealth), the average level of fever in the cluster in which a child lives is a strong determinant of their own probability of reporting fever. Treatment or advice from the modern sector is more likely in Eastern and Southern Africa than in Western and Central Africa. In both regions, the percent who seek care in the modern sector is substantially larger for the rich than for the poor, even after controlling for individual, household, and cluster attributes. While the type of treatment sought is related to the wealth of the household, the average level of wealth in the cluster in which the household lives supercedes this relationship. In Western and Central Africa children with fever in wealthier communities are substantially more likely to seek care from government hospitals and from private sources than those from poorer households. Care for children from wealthier households, regardless of the wealth in the cluster, is significantly more likely to be from public lower-level sources in this region. In Eastern and Southern Africa care is significantly more likely to be sought from government hospitals and from private medical facilities in richer than in poorer communities, whereas in this region the use lower level public services falls in wealthier communities. Seeking care from private medical sources increases substantially in wealthier communities, although the use of private commercial sources is unrelated to wealth. The experience that a cluster has with malaria is significantly related to the type of treatment sought in Eastern and Southern Africa, but not in Western and Central Africa. Unlike cross-national results which show a strong association between malaria and poverty, the results of this analysis of DHS data across these 22 countries in Sub- 31 Saharan Africa do not support the notion that fever and poverty are closely related. There is some support for the idea that levels of wealth in the community might affect the incidence of fever in Eastern and Southern Africa. The results suggest that poverty affects the type of treatment sought as a result of an episode of fever, although the patterns differ between the two broad regions analyzed here. Moreover, the results suggest that it is levels of wealth in the community in which a household lives that influences treatment the most, rather that wealth of the household itself. References The word "processed" describes informally reproduced works that may not be commonly available through library systems. Ahorlu, Collins K., Samuel K. Dunyo, Edwin A. Afari, Kwadwo A. Koram and Francis Nkrumah. 1997. "Malaria-Related Beliefs and Behaviour in Southern Ghana: Implications for Treatment, Prevention, and Control." Tropical Medicine and International Health 2(5). Bonilla-Chacin, Maria, and Jeffrey S. Hammer, 1999. "Life and Death among the Poorest." The Johns Hopkins University and Development Research Group, World Bank. Processed. Brinkmann, Uwe, and Agnes Brinkmann. 1991. "Malaria and Health in Africa: The Present Situation and Epidemiological Trends." Tropical Medicine and Parasitology 42(3). Butler, J. S., Richard V. Burkhauser, Jean M. Mitchell, and Theodore P. Pincus. 1987. "Measurement Error in Self-Reported Health Variables." The Review of Economics and Statistics 69(4). Deaton, Angus. 1997. The analysis of household surveys: a microeconometric approach to development policy. Baltimore, Md. : Johns Hopkins University Press. Deolalikar, Anil. 1998. "The Demand for Health Services in a Developing Country: The Role of Prices, Service Quality and Reporting of Illness." In Aman Ullah and David E. A. Giles eds. ,Handbook of Applied Economic Statistics. New York: Michael Dekker, Inc. Dunyo, Samuel K., Kwadwo A. Koram, and Francis K. Nkrumah, 1997. "Correspondence: Fever in Africa and WHO recommendation." The Lancet 350(9090): 1549. Einterz, Ellen M., and Myra E. Bates, 1997. "Fever in Africa: do patients know when they are hot?" The Lancet 350(9080): 781. Filmer, Deon, and Lant Pritchett, 2001. "Estimating wealth effects without income of expenditure data - or tears: Educational enrollment in India." Demography 38(1). Filmer, Deon, and Lant Pritchett, 1999. "The Effect of Household Wealth on Educational Attainment: Evidence from 35 Countries." Population and Development Review 25(1). 32 Gallup, John Luke, and Jeffrey D. Sachs. 2000. "The Economic Burden of Malaria." Center for International Development Working Paper No. 52. Harvard University: Cambridge. Gove, S. 1997. "Integrated management of childhood illness by outpatient health workers: technical basis and overview." Bulletin of the World Health Organization 75(Supplement 1): 7-24. Gwatkin, Davidon and Michel Guillot. 2000. "The Burden of Disease among the World's Global Poor: Current Situation, Future Trends, and Implications for Strategy." Human Development Network: Health Nutrition and Population Discussion Paper. The World Bank: Washington, DC. Gwatkin, Davidson R., Shea Rutstein, Kiersten Johnson, Rohini Pande and Adam Wagstaff. 2000. "Socio-Economic Differences in Health, Nutrition, and Population." World Bank, HNP/Poverty Thematic Group, Washington, D.C. Kofoed, Poul-Erik Lund, Francisco Dias, Francisco Lopes, and Lars Rombo. 1998. "Correspondence: Diagnosis of Fever in Africa." The Lancet 351(9099): 372. MARA. 1998. "Towards and Atlas of Malaria Risk in Africa: First Technical Report of the MARA/ARMA Collaboration." South African Medical Research Council, Durban South Africa. Available at http://www.mara.org.za (January 10, 2002). McCarthy, F. Desmond, Holger Wolf, and Yi Wu. 2000. "Malaria and Growth." World Bank Policy Research Working Paper No. 2303. The World Bank. Washington DC. Montgomery, Mark R., Michele Gragnolati, Kathleen A. Burke, and Edmundo Paredes. 2000. "Measuring Living Standards with Proxy Variables." Demography 37(2): 155-74. McCombie, S.C. 1996. "Treatment Seeking for Malaria: A Review of Recent Research." Social Science and Medicine 43(6). Mwenesi, Halima, Trudy Harpham, and Robert W. Snow. 1995. "Child Malaria Treatment Practices among Mothers in Kenya." Social Science and Medicine 40(9). Perkins, B.A., J. R. Zucker, J. Otieno, H.S. Jafari, L. Paxton, S. C. Redd, B. L. Nahlen, B. Schwartz, A. J. Oloo, C. Olango, S. Gove, and C. C. Campbell. 1997. "Evaluation of an algorithm for integrated management of childhood illness in an area of Kenya with high malaria transmission." Bulletin of the World Health Organization 75(Supplement 1): 33-42. Redd, Stephen C., Peter B. Bloland, Peter N. Kazembe, Ellen Patrick, Rodney Tembenu, and Carlos C. Campbell. 1992. "Usefulness of Clinical Case-Definitions in Guiding Therapy for African Children with Malaria or Pneumonia." The Lancet 340(Nov 7). Redd, Stephen C., Peter N. Kazembe, Stephen P. Luby, Okey Nwanyanwu, Allen W. Hightower, Charles Ziba, Jack. J. Wirima, Lester Chitsulo, and Michael Olivar. 1996. "Clinical Algorithm for Treatment of Plasmodium Falciparum Malaria in Children." The Lancet 347(Jan 27). 33 Ruebush, T. K., M. K. Kern, C. C. Campbell, and A. J. Oloo. 1995. "Self-treatment of Malaria in a Rural area of Western Kenya." Bulletin of the World Health Organization 73(2). Sahn, David E., and D.C. Stifel. 2000. "Poverty Comparisons Over Time and Across Countries in Africa." World Development 28(12). Sindelar, Jody, and Duncan Thomas. 1991. "Measurement of Child Health: Maternal Response Bias." Discussion Paper No. 633. Yale University Economic Growth Center, Connecticut. Slutsker, L., L. Chitsulo, A. Macheso, R.W. Steketee. 1994. "Treatment of Malaria Fever Episodes among Children in Malawi: Results from a KAP Survey." Tropical Medicine and Parasitology 45. Stecklov, Guy, Antoine Bormmier, and Ties Boerma, 1999. "Trends in Equity in Child Survival in Developing Countries: A Illustrative Example using Ugandan Data." MEASURE Evaluation Project/Carolina Population Center, INED. Processed. Strauss, John, and Duncan Thomas. 1996. "Measurement and Mismeasurement of Social Indicators." American Economic Review Papers and Proceedings 86(2). Verhoef, Hans, Elsa Hodgins, Clive E. West, Jane Y. Carter, and Frans J. Kok. 1998. "Correspondence: Diagnosis of Fever in Africa." The Lancet 351(9099): 372. Wagstaff, Adam, and N. Watanbe, 1999. "Inequalities in child malnutrition in the developing world." World Bank, Development Research Group, Washington, DC. Processed. World Bank. 1999. World Development Indicators 1999. Washington, DC: The World Bank. World Bank. 2000. World Development Report 2000/2001: Attacking Poverty. New York: Oxford University Press. World Health Organization. 2002. World Health Organization Report on Infectious Diseases. Geneva: WHO. Available at http://www.who.int/infectious-disease- report/2002/ (February 4 2002). 34 Annex 1: Table Al. Classification table for management of childhood fever Signs Classify as Treatment High malaria risk area -Any general Very severe -Give quinine for severe malaria (first dose) danger sign febrile disease -Give fist dose of an appropriate antibiotic -Stiff neck -Treat the child to prevent low blood sugar -Give one dose of paracetamol in clinic for high fever -refer urgently to hospital -Fever (by history Malaria -If no cough with fast breathing treat with oral antimalarial or feels hot or or if cough with fast breathing, treat with cotrimoxazole for 5 days temperature 2 -Advise mother when to return immediately 37.50C) -Follow-up in 2 days if fever persists -If fever is present every day for more than 7 days, refer for reassessment Low malaria risk area -Any general Very severe -Give quinine for severe malaria (first dose) unless no malaria risk danger sign febrile disease -Give fist dose of an appropriate antibiotic -Stiff neck -Treat the child to prevent low blood sugar -Give one dose of paracetamol in clinic for high fever -refer urgently to hospital -No runny nose Malaria -If no cough with fast breathing treat with oral antimalarial and no measles or if cough with fast breathing, treat with cotrimoxazole for 5 days and no other -Advise mother when to return immediately cause of fever -Follow-up in 2 days if fever persists -If fever is present every day for more than 7 days, refer for reassessment - Runny nose Fever - malaria -Give one dose of paracetamol in clinic for high fever (38. 5°C or above) present or unlikely -Advise mother when to return immediately measles present -Follow-up in 2 days in fever persists or other cause of -If fever is present every day for more than 7 days, refer for reassessment fever present Source: Adapted from Gove (1997). Figure Al. Climate suitability for stable malaria 0 I ,mdorlob(. o oeF. .. Source: Adapted from MARA (1998). 35 Annex 2 : Derivation of country level weights. Assume, for example, that two countries A and B have the same population, NA=NB. In country A 1 percent of the population (SA people) were sampled and in country B 2 percent of the population was sampled (SB people). In order for the mean of the pooled sample to be a valid estimate of the mean of the pooled populations, one needs to weight the sample from country A by a factor of 2. The specific weight we use here is a relative weight such that the sum of the weighted samples within each region equals the actual regional sample size. In particular, the weight for each country is equal to [SC/ST]/[PC/PT] where SC is the sample size in the country, ST is the total regional sample size, PC is the country's population and PT is the population of the region as a whole. Annex Table 2 reports the relative weights derived for each country in the two regions. In the Western and Central Africa region Nigeria gets a large weight (5.836) as the sample is a small percentage of the country's population yet the country constitutes a large part of the regions population. On the other hand, C.A.R. gets a small weight (0.211) as a large percentage of the population was sampled, but the country only contributes a small share of the region's population. In the Eastern and Southern Africa region Comoros gets a very low weight (0.077), whereas Kenya and Tanzania receive large weights (1.65). 36 Table A2.1 Derivation of country level weights for analysis of incidence Population in DHS Population in country Percentage ofpopulation sample (in thousands) sampled Relative weight (1) (11) (I) as a percentage of (11) Benin 1996 27,892 5,632 0.495 0.367 Burkina Faso 1998 32,181 10,996 0.293 0.621 C.A.R. 1994-95 28,050 3,254 0.862 0.211 Cameroon 1998 26,523 14,303 0.185 0.979 Chad 1996 37,213 6,937 0.536 0.339 C6ted'lvoire 1994 38,783 13,132 0.295 0.615 Ghana 1998 22,625 18,460 0.123 1.482 Mali 1995-96 50,159 9,849 0.509 0.357 Niger 1997 36,722 9,799 0.375 0.485 Nigeria 1999 38,558 123,897 0.031 5.836 Senegal 1992-93 31,966 7,800 0.410 0.443 Togo 1998 44,157 4,345 1.016 0.179 Western and Central Africa 414,829 228,404 0.182 1.000 Comoros 1996 14,297 504 2.837 0.077 Kenya 1998 37,705 28,612 0.132 1.656 Madagascar 1997 35,059 14,148 0.248 0.881 Malawi 1996 12,597 8,986 0.140 1.557 Mozambique 1997 44,822 16,630 0.270 0.810 Rwanda 1992 31,881 7,350 0.434 0.503 Tanzania 1996 40,220 30,488 0.132 1.654 Uganda 1995 36,026 19,168 0.188 1.161 Zambia 1996-97 39,721 9,214 0.431 0.506 Zimbabwe 1999 28,523 11,904 0.240 0.911 Eastern and Southern Africa 320,851 147,004 0.218 1.000 Source: Population data from World Development Indicators (World Bank, 1999). Table A2.2 Derivation of country level weights for analysis of treatment Population in DHS Population in country Percentage ofpopulation sample (in thousands) sampled Relative weight (I) (11) (I) as a percentage of (11) Burkina Faso 1992/3 34,222 9,198 0.372 0.312 Cameroon 1991 20,724 11,797 0.176 0.660 C6te d'lvoire 1994 38,783 13,132 0.295 0.393 Ghana 1993 22,139 16,200 0.137 0.849 Niger 1992 34,297 8,261 0.415 0.279 Nigeria 1999 38,558 123,897 0.031 13.938 Senegal 1992/3 31,966 7,800 0.410 0.283 Western and Central Africa 220,689 190,285 0.116 1.000 Kenya 1998 37,705 29,295 0.129 1.661 Malawi 1996 12,597 8,987 0.140 1.525 Rwanda 1992 31,881 7,350 0.434 0.493 Madagascar 1992 31,423 12,202 0.258 0.830 Tanzania 1991/2 46,733 26,691 0.175 1.221 Zambia 1992 34,943 8,262 0.423 0.505 Zimbabwe 1999 28,523 11,904 0.240 0.892 Eastern and Southern Africa 223,805 104,690 0.214 1.000 Source: Population data from World Development Indicators (World Bank, 1999) 37 Policy Research Working Paper Series Contact Title Author Date for paper WPS2778 Technology and Firm Performance Gladys L6pez-Acevedo February 2002 M. Geller in Mexico 85155 WPS2779 Technology and Skill Demand Gladys L6pez-Acevedo February 2002 M. Geller in Mexico 85155 WPS2780 Determinants of Technology Adoption Gladys L6pez-Acevedo February 2002 M. Geller in Mexico 85155 WPS2781 Maritime Transport Costs and Port Ximena Clark February 2002 E. Khine Efficiency David Dollar 37471 Alejandro Micco WPS2782 Global Capital Flows and Financing Ann E. Harrison February 2002 K. Labrie Constraints Inessa Love 31001 Margaret S. McMillan WPS2783 Ownership, Competition, and Geroge R. G. Clarke February 2002 P. Sintim-Aboagye Corruption: Bribe Takers versus Lixin Colin Xu 37644 Bribe Payers WPS2784 Financial and Legal Constraints to Thorsten Beck February 2002 A. Yaptenco Firm Growth: Does Size Matter? Asli DemirgOu-Kunt 38526 Vojislav Maksimovic WPS2785 Improving Air Quality in Metropolitan The Mexico Air Quality February 2002 G. Lage Mexico City: An Economic Valuation Management Team 31099 WPS2786 The Composition of Foreign Direct Beata K. Smarzynska February 2002 P. Flewitt Investment and Protection of 32724 Intellectual Property Rights: Evidence from Transition Economies WPS2787 Do Farmers Choose to Be Inefficient? Donald F. Larson Febrnary 2002 P. Kokila Evidence from Bohol, Philippines Frank Plessmann 33716 WPS2788 Macroeconomic Adjustment and the Pierre-Richard Agenor February 2002 M. Gosiengfiao Poor: Analytical Issues and Cross- 33363 Country Evidence WPS2789 "Learning by Dining" Informal Somik V. Lall Febru<,. 9002 Y. D'Souza Networks and Productivity in Sudeshna rGhosh 31449 Mexican Industry WPS2790 Estimating the Poverty Impacts of Jeffrey J. Reimer February 2002 P. Flewiti Trade Liberalization 32724 Policy Research Working Paper Series Contact Title Author Date for paper WPS2791 The Static and Dynamic Incidence of Dominique van de Walle February 2002 H. Sladovich Vietnam's Public Safety Net 37698 WPS2792 Determinants of Life Insurance Thorsten Beck February 2002 A. Yaptenco Consumption across Countries Ian Webb 31823 WPS2793 Agricultural Markets and Risks: Panos Varangis February 2002 P. Kokila Management of the Latter, Not the Donald Larson 33716 Former Jock R. Anderson WPS2794 Land Policies and Evolving Farm Zvi Lerman February 2002 M. Fernandez Structures in Transition Countries Csaba Csaki 33766 Gershon Feder WPS2795 Inequalities in Health in Developing Adam Wagstaff February 2002 H. Sladovich Countries: Swimming against the Tide? 37698 WPS2796 Do Rural Infrastructure Investments Jocelyn A. Songco February 2002 H. Sutrisna Benefit the Poor? Evaluating Linkages: 88032 A Global View, A Focus on Vietnam WPS2797 Regional Integration and Development Maurice Schiff February 2002 P. Flewitt in Small States 32724