- Split View
-
Views
-
Cite
Cite
Deon Filmer, Fever and its treatment among the more and less poor in sub-Saharan Africa, Health Policy and Planning, Volume 20, Issue 6, November 2005, Pages 337–346, https://doi.org/10.1093/heapol/czi043
- Share Icon Share
Abstract
This paper uses individual and household level data to explore empirically 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 22 countries where malaria is prevalent. The results suggest that the incidence of fever and its treatment are related to poverty in sub-Saharan Africa. Incidence is typically lower at the very top of the wealth distribution. The relationship, however, is not strong, especially after controlling for potentially confounding factors. Treatment patterns are strongly related to poverty as wealthier households are more likely to seek care or advice. While it is perhaps unsurprising that treatment from private sources increases with household wealth, government services – despite their public nature – are typically also used more by wealthier households. While general results hold for many of the countries, there is sufficient variation across countries that any policy seeking to reform the health sector in order to better cater to the poor needs to be informed by country-specific work.
Introduction
Each year, there are more than 300 million new cases of malaria in the world, resulting in over 1 million deaths. Of these, 90% of the cases and 97% of the deaths occur in sub-Saharan Africa (World Health Organization 2002). Reducing the burden of malaria is high on the agenda as the poorest nations around the world commit themselves to reaching the Millennium Development Goal of halving child mortality by 2015. Malaria is frequently characterized as being intimately linked with poverty. For example, the World Health Organization's Report on Infectious Diseases argues forcefully that infectious diseases, and malaria in particular, are a consequence of poverty as well as an obstacle that keeps people in poverty (World Health Organization 2002). More generally, this paper contributes to a growing literature that documents income-related inequalities in health in poor countries; for example, see Gwatkin (2000) and accompanying papers in a special issue of the Bulletin of the World Health Organization.
Much of the evidence on economic aspects of malaria focuses on the relationship between malaria and GDP growth as derived from cross-national data (for example, Gallup and Sachs 2000; McCarthy et al. 2000). Additional work focuses on the loss of productivity associated with malaria among adults and the resulting loss in welfare (for a review see Chima et al. 2003). 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 where malaria is prevalent.
The paper is organized as follows: data and methods are described in the next section, followed by results on the incidence of fever, results on treatment-seeking behaviour, and a final section containing a brief discussion and conclusion.
Data and methods
The data used in this paper are Demographic and Health Survey (DHS) data collected in sub-Saharan African countries in the 1990s. These were nationally representative household surveys with large sample sizes ranging from 2252 households in Comoros to 9282 in Mozambique (summary and background information for the data are in an Annex table available from the author). All the countries in the study are poor, with poverty rates ranging from 26% in Zimbabwe to 86% in Zambia (World Bank 2000). Under-five mortality rates are high, ranging from 90 per 1000 births in Zimbabwe to 318 per 1000 births in Niger (World Bank 1999). The selection of countries into this analysis is driven by data availability. All countries that had the relevant questions in the survey instrument were included.1
1Sampling weights are used to adjust for over- and under-sampling within countries.
The data on fever incidence are derived from questions asked of mothers of all children under 3 years old. The exact formulation of questions varied somewhat across countries, but the typical questionnaire asked whether the child had an episode of fever in the 2 weeks prior to the interview.
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 2 weeks. Analysis of this information can be done for 14 of the 22 countries.2 The surveys typically asked ‘did you seek advice or treatment for the fever’ for a child reported to have had a fever in the past 2 weeks. In some cases, the question was asked whether the child had a fever or a cough with rapid breathing in the 2 weeks prior to the interview. In cases where the child 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. This analysis includes the advice/treatment seeking behaviour as long as the child is reported to have had at least a fever.
2In Madagascar, Niger, Tanzania and Zambia, data on treatment refer to a survey from a different year than that used in the incidence analysis.
The types of facilities or persons that respondents could report having visited are grouped into: public hospitals; ‘lower-level public’ (e.g. government health centre, government health post, mobile clinic, community health worker); ‘private medical’ (e.g. private hospital/clinic, private doctor, private mobile clinic); or a commercial establishment such as a pharmacy or a shop. In addition to these generic options, country-specific options (for example, a nurse's practice, public health post or a pharmaceutical depot) have been mapped to the basic classification. ‘Advice from friends or family’ is among the responses included in the ‘no treatment or advice’ category. Advice or treatment from traditional healers is treated as a separate category in parts of the analysis. Visits refer to the first facility or person sought since this is the only information collected through the DHS.
Studies analyzing inequalities in outcomes typically use consumption expenditures as a measure of long-run income (see discussion in Deaton 1997) but this information is not available in DHS data. This paper uses an approach based on an index of household-owned assets and housing characteristics advocated and applied in Filmer and Pritchett (1999, 2001) for the analysis of inequalities in education outcomes.3 Principal components analysis is applied to the variables measuring assets and housing characteristics, and the first principal component is retained and defines the wealth index used to define wealth quintiles.4
3A similar asset index approach has also been used by others to analyze health outcomes in DHS data; for example, child mortality in Bonilla-Chacin and Hammer (1999), child survival in Uganda in Stecklov et al. (1999), child anthropometric outcomes in Wagstaff and Watanabe (1999), and to document inequalities in a variety of health outcomes and behaviours in Gwatkin et al. (2000). Sahn and Stifel (2000) use a similar approach to analyze poverty directly.
4In order to define quintiles, individuals are sorted by the wealth index within each country, 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. The interpretation is, therefore, that the poorest quintile is the group in which the poorest 20% of the population live. Note that the use of the term ‘poor’ here differs from the 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 wealth index.
Some of the results in this paper pool the data from all surveys. In order to do so, a wealth index is defined on the pooled dataset (although country-specific indexes are used in the country-by-country analysis) and the survey sampling weights are adjusted to reflect different population to survey-size ratios in the different countries.5
5That is, despite the fact that smaller countries will have larger sample sizes relative to their populations, the weights will adjust for this, so the results for the aggregated data can be interpreted as population weighted averages.
Poverty and the incidence of fever
The percentage of children in the DHS reporting any fever in the past 2 weeks is high (Table 1). The overall average is 38.5% with a rural-urban differential on the order of 3 percentage points (for a rural/urban ratio of 1.07). There is substantial heterogeneity across countries. Ghana has the lowest reported level of fever (29%) and Benin has the highest (54.5%). While the rural/urban ratio is typically close to 1, it is greater than 1.2 in seven of the countries (Benin, Burkina Faso, Malawi, Mali, Niger, Rwanda and Uganda).
. | Total . | Rural . | Urban . |
---|---|---|---|
Benin 1996 | 54.5 | 57.8 | 47.3 |
Burkina Faso 1999 | 41.0 | 41.8 | 34.1 |
C.A.R. 1994–95 | 35.7 | 36.3 | 34.8 |
Cameroon 1998 | 30.8 | 31.1 | 30.2 |
Chad 1996 | 36.9 | 36.9 | 36.6 |
Comoros 1996 | 50.2 | 50.4 | 49.6 |
Côte d'Ivoire 1994 | 44.2 | 46.3 | 39.8 |
Ghana 1998 | 29.0 | 29.8 | 26.4 |
Kenya 1998 | 42.9 | 42.9 | 43.2 |
Madagascar 1997 | 32.7 | 33.1 | 31.0 |
Malawi 1998 | 47.6 | 48.8 | 37.5 |
Mali 1995–96 | 39.3 | 41.2 | 34.1 |
Mozambique 1997 | 44.7 | 43.2 | 50.4 |
Níger 1997 | 49.4 | 51.1 | 41.2 |
Nigeria 1999 | 31.2 | 32.6 | 27.3 |
Rwanda 1992 | 48.6 | 49.0 | 39.2 |
Senegal 1992–93 | 45.7 | 48.3 | 40.8 |
Tanzania 1996 | 35.8 | 35.7 | 36.1 |
Togo 1998 | 37.7 | 37.7 | 37.8 |
Uganda 1995 | 50.3 | 51.5 | 40.6 |
Zambia 1996–97 | 46.4 | 47.6 | 44.5 |
Zimbabwe 1999 | 31.1 | 32.5 | 28.2 |
All surveys (pooled) | 38.5 | 39.0 | 36.3 |
. | Total . | Rural . | Urban . |
---|---|---|---|
Benin 1996 | 54.5 | 57.8 | 47.3 |
Burkina Faso 1999 | 41.0 | 41.8 | 34.1 |
C.A.R. 1994–95 | 35.7 | 36.3 | 34.8 |
Cameroon 1998 | 30.8 | 31.1 | 30.2 |
Chad 1996 | 36.9 | 36.9 | 36.6 |
Comoros 1996 | 50.2 | 50.4 | 49.6 |
Côte d'Ivoire 1994 | 44.2 | 46.3 | 39.8 |
Ghana 1998 | 29.0 | 29.8 | 26.4 |
Kenya 1998 | 42.9 | 42.9 | 43.2 |
Madagascar 1997 | 32.7 | 33.1 | 31.0 |
Malawi 1998 | 47.6 | 48.8 | 37.5 |
Mali 1995–96 | 39.3 | 41.2 | 34.1 |
Mozambique 1997 | 44.7 | 43.2 | 50.4 |
Níger 1997 | 49.4 | 51.1 | 41.2 |
Nigeria 1999 | 31.2 | 32.6 | 27.3 |
Rwanda 1992 | 48.6 | 49.0 | 39.2 |
Senegal 1992–93 | 45.7 | 48.3 | 40.8 |
Tanzania 1996 | 35.8 | 35.7 | 36.1 |
Togo 1998 | 37.7 | 37.7 | 37.8 |
Uganda 1995 | 50.3 | 51.5 | 40.6 |
Zambia 1996–97 | 46.4 | 47.6 | 44.5 |
Zimbabwe 1999 | 31.1 | 32.5 | 28.2 |
All surveys (pooled) | 38.5 | 39.0 | 36.3 |
Note: Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from weighted DHS data.
. | Total . | Rural . | Urban . |
---|---|---|---|
Benin 1996 | 54.5 | 57.8 | 47.3 |
Burkina Faso 1999 | 41.0 | 41.8 | 34.1 |
C.A.R. 1994–95 | 35.7 | 36.3 | 34.8 |
Cameroon 1998 | 30.8 | 31.1 | 30.2 |
Chad 1996 | 36.9 | 36.9 | 36.6 |
Comoros 1996 | 50.2 | 50.4 | 49.6 |
Côte d'Ivoire 1994 | 44.2 | 46.3 | 39.8 |
Ghana 1998 | 29.0 | 29.8 | 26.4 |
Kenya 1998 | 42.9 | 42.9 | 43.2 |
Madagascar 1997 | 32.7 | 33.1 | 31.0 |
Malawi 1998 | 47.6 | 48.8 | 37.5 |
Mali 1995–96 | 39.3 | 41.2 | 34.1 |
Mozambique 1997 | 44.7 | 43.2 | 50.4 |
Níger 1997 | 49.4 | 51.1 | 41.2 |
Nigeria 1999 | 31.2 | 32.6 | 27.3 |
Rwanda 1992 | 48.6 | 49.0 | 39.2 |
Senegal 1992–93 | 45.7 | 48.3 | 40.8 |
Tanzania 1996 | 35.8 | 35.7 | 36.1 |
Togo 1998 | 37.7 | 37.7 | 37.8 |
Uganda 1995 | 50.3 | 51.5 | 40.6 |
Zambia 1996–97 | 46.4 | 47.6 | 44.5 |
Zimbabwe 1999 | 31.1 | 32.5 | 28.2 |
All surveys (pooled) | 38.5 | 39.0 | 36.3 |
. | Total . | Rural . | Urban . |
---|---|---|---|
Benin 1996 | 54.5 | 57.8 | 47.3 |
Burkina Faso 1999 | 41.0 | 41.8 | 34.1 |
C.A.R. 1994–95 | 35.7 | 36.3 | 34.8 |
Cameroon 1998 | 30.8 | 31.1 | 30.2 |
Chad 1996 | 36.9 | 36.9 | 36.6 |
Comoros 1996 | 50.2 | 50.4 | 49.6 |
Côte d'Ivoire 1994 | 44.2 | 46.3 | 39.8 |
Ghana 1998 | 29.0 | 29.8 | 26.4 |
Kenya 1998 | 42.9 | 42.9 | 43.2 |
Madagascar 1997 | 32.7 | 33.1 | 31.0 |
Malawi 1998 | 47.6 | 48.8 | 37.5 |
Mali 1995–96 | 39.3 | 41.2 | 34.1 |
Mozambique 1997 | 44.7 | 43.2 | 50.4 |
Níger 1997 | 49.4 | 51.1 | 41.2 |
Nigeria 1999 | 31.2 | 32.6 | 27.3 |
Rwanda 1992 | 48.6 | 49.0 | 39.2 |
Senegal 1992–93 | 45.7 | 48.3 | 40.8 |
Tanzania 1996 | 35.8 | 35.7 | 36.1 |
Togo 1998 | 37.7 | 37.7 | 37.8 |
Uganda 1995 | 50.3 | 51.5 | 40.6 |
Zambia 1996–97 | 46.4 | 47.6 | 44.5 |
Zimbabwe 1999 | 31.1 | 32.5 | 28.2 |
All surveys (pooled) | 38.5 | 39.0 | 36.3 |
Note: Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from weighted DHS data.
Aggregate data on the percentage of children under three who have had a fever in the past 2 weeks by wealth quintile is shown in Figure 1. Although the incidence is generally higher in the poorest quintile than in the richest quintile, it remains high in all quintiles. The gap between the richest and poorest quintiles in the incidence of fever is of the order of 10 percentage points (41% in the poorest quintile versus 32% in the richest quintile). The variation by wealth is not uniform across the distribution: incidence remains fairly constant across most quintiles, dropping off only at the upper levels. The gap relative to the poorest quintile is only statistically significant for the richest quintile.
The pattern is similar in rural and urban areas. Reported incidence is somewhat higher in the poorest quintile in urban areas, but the general pattern follows that in rural areas. In both areas, it is only in the richest quintile that incidence is statistically significantly different from that in the poorest quintile.
These bivariate results do not control for potentially confounding characteristics. These are variables that are correlated simultaneously with the incidence of fever as well as household wealth, such as parents’ education or urban/rural residence. Table 2 reports the marginal effects from multivariate Probit models relating fever incidence to wealth quintile and other characteristics (for binary variables this is the change in the outcome resulting from a change in a dummy variable from zero to one, evaluated with other variables set to their means). In addition to parents’ education and urban/rural residence, the model includes age, age squared, gender and a full set of dummy variables for sub-national region, month of interview and the interaction of these in order to account for geographically localized effects as well as for variation in the timing of the surveys.
. | Quintile 2 . | Quintile 3 . | Quintile 4 . | Quintile 5 . | Test for joint significance of quintiles (p-value) . | Test for joint significance of cluster quintiles (p-value) . | Cluster average incidence of fever . | N . |
---|---|---|---|---|---|---|---|---|
Benin 1996 | 0.040 | −0.001 | −0.001 | −0.039 | 0.536 | 0.562 | 0.142 | 2718 |
Burkina Faso 1999 | −0.007 | 0.023 | 0.051 | 0.015 | 0.410 | 0.142 | 0.245** | 3074 |
C.A.R. 1994–95 | 0.014 | 0.052 | 0.038 | 0.071 | 0.442 | 0.041* | 0.086 | 2494 |
Cameroon 1998 | 0.075 | 0.044 | −0.004 | −0.031 | 0.071 | 0.632 | 0.397** | 2063 |
Chad 1996 | 0.013 | 0.040 | 0.027 | −0.031 | 0.212 | 0.730 | 0.418** | 3850 |
Comoros 1996 | −0.055 | 0.024 | 0.022 | 0.061 | 0.386 | 0.593 | 0.157 | 1029 |
Côte d'Ivoire 1994 | −0.027 | −0.035 | 0.001 | −0.095* | 0.031* | 0.563 | 0.059 | 3598 |
Ghana 1998 | −0.077* | −0.046 | −0.078 | −0.072 | 0.258 | 0.014* | 0.070 | 1780 |
Kenya 1998 | 0.078* | 0.034 | 0.043 | 0.027 | 0.248 | 0.312 | 0.165** | 3211 |
Madagascar 1997 | 0.015 | 0.029 | 0.017 | 0.026 | 0.872 | 0.640 | 0.207** | 3269 |
Malawi 1998 | 0.008 | 0.031 | 0.015 | 0.011 | 0.988 | 0.188 | 0.063 | 1292 |
Mali 1995–96 | −0.005 | −0.011 | −0.039 | −0.053 | 0.467 | 0.711 | 0.344** | 5148 |
Mozambique 1997 | −0.043 | −0.040 | −0.016 | −0.053 | 0.844 | 0.157 | −0.049 | 3698 |
Níger 1997 | 0.049 | 0.020 | 0.005 | −0.019 | 0.468 | 0.350 | 0.311** | 4157 |
Nigeria 1999 | −0.004 | 0.015 | −0.032 | −0.044 | 0.502 | 0.653 | 0.284** | 3088 |
Rwanda 1992 | −0.004 | −0.027 | −0.002 | −0.023 | 0.860 | 0.125 | 0.234** | 2986 |
Senegal 1992–93 | 0.044 | 0.079* | 0.136** | 0.092 | 0.083 | 0.134 | 0.201** | 3030 |
Tanzania 1996 | 0.053 | 0.073* | 0.112** | 0.047 | 0.016* | 0.031* | 0.095 | 3708 |
Togo 1998 | 0.006 | −0.013 | −0.015 | −0.017 | 0.945 | 0.463 | 0.370** | 3824 |
Uganda 1995 | 0.042 | 0.022 | 0.036 | 0.078 | 0.617 | 0.057 | 0.451** | 3893 |
Zambia 1996–97 | 0.060* | 0.029 | −0.040 | −0.107 | 0.019* | 0.079 | 0.220** | 3903 |
Zimbabwe 1999 | 0.031 | 0.003 | 0.021 | 0.010 | 0.898 | 0.907 | 0.117 | 1956 |
All surveys | −0.002 | −0.001 | 0.006 | −0.023 | 0.607 | 0.382 | 0.255** | 67 770 |
All surveys: rural only | −0.006 | −0.001 | −0.007 | −0.045 | 0.550 | 0.400 | 0.278** | 49 038 |
. | Quintile 2 . | Quintile 3 . | Quintile 4 . | Quintile 5 . | Test for joint significance of quintiles (p-value) . | Test for joint significance of cluster quintiles (p-value) . | Cluster average incidence of fever . | N . |
---|---|---|---|---|---|---|---|---|
Benin 1996 | 0.040 | −0.001 | −0.001 | −0.039 | 0.536 | 0.562 | 0.142 | 2718 |
Burkina Faso 1999 | −0.007 | 0.023 | 0.051 | 0.015 | 0.410 | 0.142 | 0.245** | 3074 |
C.A.R. 1994–95 | 0.014 | 0.052 | 0.038 | 0.071 | 0.442 | 0.041* | 0.086 | 2494 |
Cameroon 1998 | 0.075 | 0.044 | −0.004 | −0.031 | 0.071 | 0.632 | 0.397** | 2063 |
Chad 1996 | 0.013 | 0.040 | 0.027 | −0.031 | 0.212 | 0.730 | 0.418** | 3850 |
Comoros 1996 | −0.055 | 0.024 | 0.022 | 0.061 | 0.386 | 0.593 | 0.157 | 1029 |
Côte d'Ivoire 1994 | −0.027 | −0.035 | 0.001 | −0.095* | 0.031* | 0.563 | 0.059 | 3598 |
Ghana 1998 | −0.077* | −0.046 | −0.078 | −0.072 | 0.258 | 0.014* | 0.070 | 1780 |
Kenya 1998 | 0.078* | 0.034 | 0.043 | 0.027 | 0.248 | 0.312 | 0.165** | 3211 |
Madagascar 1997 | 0.015 | 0.029 | 0.017 | 0.026 | 0.872 | 0.640 | 0.207** | 3269 |
Malawi 1998 | 0.008 | 0.031 | 0.015 | 0.011 | 0.988 | 0.188 | 0.063 | 1292 |
Mali 1995–96 | −0.005 | −0.011 | −0.039 | −0.053 | 0.467 | 0.711 | 0.344** | 5148 |
Mozambique 1997 | −0.043 | −0.040 | −0.016 | −0.053 | 0.844 | 0.157 | −0.049 | 3698 |
Níger 1997 | 0.049 | 0.020 | 0.005 | −0.019 | 0.468 | 0.350 | 0.311** | 4157 |
Nigeria 1999 | −0.004 | 0.015 | −0.032 | −0.044 | 0.502 | 0.653 | 0.284** | 3088 |
Rwanda 1992 | −0.004 | −0.027 | −0.002 | −0.023 | 0.860 | 0.125 | 0.234** | 2986 |
Senegal 1992–93 | 0.044 | 0.079* | 0.136** | 0.092 | 0.083 | 0.134 | 0.201** | 3030 |
Tanzania 1996 | 0.053 | 0.073* | 0.112** | 0.047 | 0.016* | 0.031* | 0.095 | 3708 |
Togo 1998 | 0.006 | −0.013 | −0.015 | −0.017 | 0.945 | 0.463 | 0.370** | 3824 |
Uganda 1995 | 0.042 | 0.022 | 0.036 | 0.078 | 0.617 | 0.057 | 0.451** | 3893 |
Zambia 1996–97 | 0.060* | 0.029 | −0.040 | −0.107 | 0.019* | 0.079 | 0.220** | 3903 |
Zimbabwe 1999 | 0.031 | 0.003 | 0.021 | 0.010 | 0.898 | 0.907 | 0.117 | 1956 |
All surveys | −0.002 | −0.001 | 0.006 | −0.023 | 0.607 | 0.382 | 0.255** | 67 770 |
All surveys: rural only | −0.006 | −0.001 | −0.007 | −0.045 | 0.550 | 0.400 | 0.278** | 49 038 |
Note: In addition to the reported variables, models include: the percentage of the cluster population in each quintile; child sex, age and age squared, mother's and father's years of schooling; a dummy variable for urban residence; dummy variables for sub-national region, month of interview, and the interaction of these.
Significance tests based on robust standard errors, adjusted for clustering: *significant at 5%; **significant at 1%. Cluster averages are ‘non-self means’, i.e. they exclude the child in question. Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from pooled and weighted DHS.
. | Quintile 2 . | Quintile 3 . | Quintile 4 . | Quintile 5 . | Test for joint significance of quintiles (p-value) . | Test for joint significance of cluster quintiles (p-value) . | Cluster average incidence of fever . | N . |
---|---|---|---|---|---|---|---|---|
Benin 1996 | 0.040 | −0.001 | −0.001 | −0.039 | 0.536 | 0.562 | 0.142 | 2718 |
Burkina Faso 1999 | −0.007 | 0.023 | 0.051 | 0.015 | 0.410 | 0.142 | 0.245** | 3074 |
C.A.R. 1994–95 | 0.014 | 0.052 | 0.038 | 0.071 | 0.442 | 0.041* | 0.086 | 2494 |
Cameroon 1998 | 0.075 | 0.044 | −0.004 | −0.031 | 0.071 | 0.632 | 0.397** | 2063 |
Chad 1996 | 0.013 | 0.040 | 0.027 | −0.031 | 0.212 | 0.730 | 0.418** | 3850 |
Comoros 1996 | −0.055 | 0.024 | 0.022 | 0.061 | 0.386 | 0.593 | 0.157 | 1029 |
Côte d'Ivoire 1994 | −0.027 | −0.035 | 0.001 | −0.095* | 0.031* | 0.563 | 0.059 | 3598 |
Ghana 1998 | −0.077* | −0.046 | −0.078 | −0.072 | 0.258 | 0.014* | 0.070 | 1780 |
Kenya 1998 | 0.078* | 0.034 | 0.043 | 0.027 | 0.248 | 0.312 | 0.165** | 3211 |
Madagascar 1997 | 0.015 | 0.029 | 0.017 | 0.026 | 0.872 | 0.640 | 0.207** | 3269 |
Malawi 1998 | 0.008 | 0.031 | 0.015 | 0.011 | 0.988 | 0.188 | 0.063 | 1292 |
Mali 1995–96 | −0.005 | −0.011 | −0.039 | −0.053 | 0.467 | 0.711 | 0.344** | 5148 |
Mozambique 1997 | −0.043 | −0.040 | −0.016 | −0.053 | 0.844 | 0.157 | −0.049 | 3698 |
Níger 1997 | 0.049 | 0.020 | 0.005 | −0.019 | 0.468 | 0.350 | 0.311** | 4157 |
Nigeria 1999 | −0.004 | 0.015 | −0.032 | −0.044 | 0.502 | 0.653 | 0.284** | 3088 |
Rwanda 1992 | −0.004 | −0.027 | −0.002 | −0.023 | 0.860 | 0.125 | 0.234** | 2986 |
Senegal 1992–93 | 0.044 | 0.079* | 0.136** | 0.092 | 0.083 | 0.134 | 0.201** | 3030 |
Tanzania 1996 | 0.053 | 0.073* | 0.112** | 0.047 | 0.016* | 0.031* | 0.095 | 3708 |
Togo 1998 | 0.006 | −0.013 | −0.015 | −0.017 | 0.945 | 0.463 | 0.370** | 3824 |
Uganda 1995 | 0.042 | 0.022 | 0.036 | 0.078 | 0.617 | 0.057 | 0.451** | 3893 |
Zambia 1996–97 | 0.060* | 0.029 | −0.040 | −0.107 | 0.019* | 0.079 | 0.220** | 3903 |
Zimbabwe 1999 | 0.031 | 0.003 | 0.021 | 0.010 | 0.898 | 0.907 | 0.117 | 1956 |
All surveys | −0.002 | −0.001 | 0.006 | −0.023 | 0.607 | 0.382 | 0.255** | 67 770 |
All surveys: rural only | −0.006 | −0.001 | −0.007 | −0.045 | 0.550 | 0.400 | 0.278** | 49 038 |
. | Quintile 2 . | Quintile 3 . | Quintile 4 . | Quintile 5 . | Test for joint significance of quintiles (p-value) . | Test for joint significance of cluster quintiles (p-value) . | Cluster average incidence of fever . | N . |
---|---|---|---|---|---|---|---|---|
Benin 1996 | 0.040 | −0.001 | −0.001 | −0.039 | 0.536 | 0.562 | 0.142 | 2718 |
Burkina Faso 1999 | −0.007 | 0.023 | 0.051 | 0.015 | 0.410 | 0.142 | 0.245** | 3074 |
C.A.R. 1994–95 | 0.014 | 0.052 | 0.038 | 0.071 | 0.442 | 0.041* | 0.086 | 2494 |
Cameroon 1998 | 0.075 | 0.044 | −0.004 | −0.031 | 0.071 | 0.632 | 0.397** | 2063 |
Chad 1996 | 0.013 | 0.040 | 0.027 | −0.031 | 0.212 | 0.730 | 0.418** | 3850 |
Comoros 1996 | −0.055 | 0.024 | 0.022 | 0.061 | 0.386 | 0.593 | 0.157 | 1029 |
Côte d'Ivoire 1994 | −0.027 | −0.035 | 0.001 | −0.095* | 0.031* | 0.563 | 0.059 | 3598 |
Ghana 1998 | −0.077* | −0.046 | −0.078 | −0.072 | 0.258 | 0.014* | 0.070 | 1780 |
Kenya 1998 | 0.078* | 0.034 | 0.043 | 0.027 | 0.248 | 0.312 | 0.165** | 3211 |
Madagascar 1997 | 0.015 | 0.029 | 0.017 | 0.026 | 0.872 | 0.640 | 0.207** | 3269 |
Malawi 1998 | 0.008 | 0.031 | 0.015 | 0.011 | 0.988 | 0.188 | 0.063 | 1292 |
Mali 1995–96 | −0.005 | −0.011 | −0.039 | −0.053 | 0.467 | 0.711 | 0.344** | 5148 |
Mozambique 1997 | −0.043 | −0.040 | −0.016 | −0.053 | 0.844 | 0.157 | −0.049 | 3698 |
Níger 1997 | 0.049 | 0.020 | 0.005 | −0.019 | 0.468 | 0.350 | 0.311** | 4157 |
Nigeria 1999 | −0.004 | 0.015 | −0.032 | −0.044 | 0.502 | 0.653 | 0.284** | 3088 |
Rwanda 1992 | −0.004 | −0.027 | −0.002 | −0.023 | 0.860 | 0.125 | 0.234** | 2986 |
Senegal 1992–93 | 0.044 | 0.079* | 0.136** | 0.092 | 0.083 | 0.134 | 0.201** | 3030 |
Tanzania 1996 | 0.053 | 0.073* | 0.112** | 0.047 | 0.016* | 0.031* | 0.095 | 3708 |
Togo 1998 | 0.006 | −0.013 | −0.015 | −0.017 | 0.945 | 0.463 | 0.370** | 3824 |
Uganda 1995 | 0.042 | 0.022 | 0.036 | 0.078 | 0.617 | 0.057 | 0.451** | 3893 |
Zambia 1996–97 | 0.060* | 0.029 | −0.040 | −0.107 | 0.019* | 0.079 | 0.220** | 3903 |
Zimbabwe 1999 | 0.031 | 0.003 | 0.021 | 0.010 | 0.898 | 0.907 | 0.117 | 1956 |
All surveys | −0.002 | −0.001 | 0.006 | −0.023 | 0.607 | 0.382 | 0.255** | 67 770 |
All surveys: rural only | −0.006 | −0.001 | −0.007 | −0.045 | 0.550 | 0.400 | 0.278** | 49 038 |
Note: In addition to the reported variables, models include: the percentage of the cluster population in each quintile; child sex, age and age squared, mother's and father's years of schooling; a dummy variable for urban residence; dummy variables for sub-national region, month of interview, and the interaction of these.
Significance tests based on robust standard errors, adjusted for clustering: *significant at 5%; **significant at 1%. Cluster averages are ‘non-self means’, i.e. they exclude the child in question. Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from pooled and weighted DHS.
These multivariate results show a very weak conditional association between incidence and household wealth. In the model that aggregates data across all the surveys, there is about a 2 percentage point gap in the incidence of fever between the richest and poorest quintiles, and this difference is not statistically significantly different from zero.
Models that are limited to urban or rural areas suggest that these results are not idiosyncratic to one or the other of these areas. While the magnitude of the gap in fever incidence between the richest and poorest quintile is somewhat larger in rural areas (4.5 percentage points), it is still not statistically significantly different from zero.
The models include variables capturing the average level of wealth of other households (i.e. the ‘non-self mean’) in the cluster in which the child lives, as well as the average incidence of fever among other children in the cluster (excluding these variables does not affect the coefficients on own wealth).6
6Clusters 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.
The percentage of other children in the cluster who are reported to have a fever is strongly statistically associated with the incidence of fever. A 10% increase in the incidence of fever among other children in the cluster is associated with a 2 to 3 percentage point increase in the likelihood that any given child will report a fever. While this may not appear to be surprising, note that this is conditional on many other individual, household, cluster and regional characteristics, suggesting a strong underlying geographic concentration of episodes of fever. On the other hand, the wealth of other households in the cluster is not statistically significantly associated with the incidence of fever (a finding that withstands removing the cluster average incidence of fever).
Other results (not reported here but available as an Annex table from the author) show that boys have a statistically significant higher reported incidence of fever in rural areas – but the magnitude is small, on the order of two percentage points higher than girls. Age has a statistically significant inverse-U shaped relationship with incidence in both regions, with the highest incidence occurring at about 1 year old. Mother's years of schooling is negatively related to the reported incidence of fever, but the effect is small and only statistically significantly different from zero at the 10% level.
It is possible that there are still too many unobserved confounding factors obscuring a statistically significant relationship. In particular, country-to-country differences in the exposure to malaria, the timing of that exposure, 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 sub-national region, month of survey and the interaction of these). In addition, the pooled approach to estimating wealth might overemphasize differences between countries compared with relatively smaller differences within countries. Table 2 therefore also reports marginal effects of the wealth quintile variables in country-by-country multivariate estimates of the association between fever incidence, wealth and other characteristics. The country-specific models include all the same variables as the aggregated model but allow all coefficients to differ across countries.
Again, the results do not provide evidence for a strong relationship between reported fever and poverty. Wealth – both at the household and cluster levels – is typically not statistically significantly associated with fever. Côte d'Ivoire is the only country in which there is a statistically significantly lower incidence in the richest quintile; about a 10 percentage point lower incidence than in the poorest quintile. There are only four other countries where incidence is more than 5 percentage points lower in the richest than in the poorest quintile (Ghana, Mali, Mozambique and Zambia), although the gap is not statistically significantly different from zero in any of these. Clearly, wealth gaps in the incidence of fever are not large, nor typically statistically significant.
The geographic concentration of incidence occurs in most of the country-level models. In many of the countries, a 10% rise in the cluster level incidence is associated with more than a 2 percentage point rise in the probability that a child has had a fever. In three of the countries, it is associated with as much as a 4 percentage point increase in the probability of fever (Cameroon, Chad and Uganda).
Poverty and treatment-seeking behaviour
An episode of fever typically results in some type of action on the part of the caregiver: 60% of cases resulted in a visit to a modern medical provider (public hospital, public lower-level facility or private medical provider), a commercial source (such as a pharmacy or shop), or a traditional healer (Table 3).
. | No treatment/advice . | Public hospital . | Public, lower level . | Private, medical . | Pharmacy/Shop . | Traditional . | Total . |
---|---|---|---|---|---|---|---|
Burkina Faso 1992/3 | 76.6 | 1.7 | 16.5 | 1.7 | 0.0 | 3.5 | 100 |
Cameroon 1991 | 52.1 | 8.1 | 19.2 | 14.5 | 2.3 | 3.8 | 100 |
Côte d'Ivoire 1994/5 | 15.0 | 26.8 | 32.1 | 17.5 | 5.4 | 3.3 | 100 |
Ghana 1993 | 31.3 | 22.3 | 15.6 | 12.5 | 14.3 | 4.1 | 100 |
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 |
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 |
Rwanda 1992 | 54.6 | 5.4 | 26.3 | 7.5 | 3.0 | 3.2 | 100 |
Senegal 1992/3 | 60.7 | 4.0 | 25.7 | 4.8 | 2.1 | 2.6 | 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 |
All surveys (pooled) | 40.0 | 10.3 | 23.5 | 13.1 | 11.3 | 1.9 | 100 |
All surveys: rural only | 43.4 | 7.0 | 23.9 | 12.4 | 11.3 | 2.0 | 100 |
. | No treatment/advice . | Public hospital . | Public, lower level . | Private, medical . | Pharmacy/Shop . | Traditional . | Total . |
---|---|---|---|---|---|---|---|
Burkina Faso 1992/3 | 76.6 | 1.7 | 16.5 | 1.7 | 0.0 | 3.5 | 100 |
Cameroon 1991 | 52.1 | 8.1 | 19.2 | 14.5 | 2.3 | 3.8 | 100 |
Côte d'Ivoire 1994/5 | 15.0 | 26.8 | 32.1 | 17.5 | 5.4 | 3.3 | 100 |
Ghana 1993 | 31.3 | 22.3 | 15.6 | 12.5 | 14.3 | 4.1 | 100 |
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 |
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 |
Rwanda 1992 | 54.6 | 5.4 | 26.3 | 7.5 | 3.0 | 3.2 | 100 |
Senegal 1992/3 | 60.7 | 4.0 | 25.7 | 4.8 | 2.1 | 2.6 | 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 |
All surveys (pooled) | 40.0 | 10.3 | 23.5 | 13.1 | 11.3 | 1.9 | 100 |
All surveys: rural only | 43.4 | 7.0 | 23.9 | 12.4 | 11.3 | 2.0 | 100 |
Note: ‘Public lower-level’ is government health centre, government health post, mobile clinic, community health worker; ‘private medical’ is private hospital/clinic, private doctor, private mobile clinic. In some countries, additional options have been mapped to this classification. The ‘no treatment/advice’ category includes advice from friends or family. Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from pooled and weighted DHS data.
. | No treatment/advice . | Public hospital . | Public, lower level . | Private, medical . | Pharmacy/Shop . | Traditional . | Total . |
---|---|---|---|---|---|---|---|
Burkina Faso 1992/3 | 76.6 | 1.7 | 16.5 | 1.7 | 0.0 | 3.5 | 100 |
Cameroon 1991 | 52.1 | 8.1 | 19.2 | 14.5 | 2.3 | 3.8 | 100 |
Côte d'Ivoire 1994/5 | 15.0 | 26.8 | 32.1 | 17.5 | 5.4 | 3.3 | 100 |
Ghana 1993 | 31.3 | 22.3 | 15.6 | 12.5 | 14.3 | 4.1 | 100 |
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 |
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 |
Rwanda 1992 | 54.6 | 5.4 | 26.3 | 7.5 | 3.0 | 3.2 | 100 |
Senegal 1992/3 | 60.7 | 4.0 | 25.7 | 4.8 | 2.1 | 2.6 | 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 |
All surveys (pooled) | 40.0 | 10.3 | 23.5 | 13.1 | 11.3 | 1.9 | 100 |
All surveys: rural only | 43.4 | 7.0 | 23.9 | 12.4 | 11.3 | 2.0 | 100 |
. | No treatment/advice . | Public hospital . | Public, lower level . | Private, medical . | Pharmacy/Shop . | Traditional . | Total . |
---|---|---|---|---|---|---|---|
Burkina Faso 1992/3 | 76.6 | 1.7 | 16.5 | 1.7 | 0.0 | 3.5 | 100 |
Cameroon 1991 | 52.1 | 8.1 | 19.2 | 14.5 | 2.3 | 3.8 | 100 |
Côte d'Ivoire 1994/5 | 15.0 | 26.8 | 32.1 | 17.5 | 5.4 | 3.3 | 100 |
Ghana 1993 | 31.3 | 22.3 | 15.6 | 12.5 | 14.3 | 4.1 | 100 |
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 |
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 |
Rwanda 1992 | 54.6 | 5.4 | 26.3 | 7.5 | 3.0 | 3.2 | 100 |
Senegal 1992/3 | 60.7 | 4.0 | 25.7 | 4.8 | 2.1 | 2.6 | 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 |
All surveys (pooled) | 40.0 | 10.3 | 23.5 | 13.1 | 11.3 | 1.9 | 100 |
All surveys: rural only | 43.4 | 7.0 | 23.9 | 12.4 | 11.3 | 2.0 | 100 |
Note: ‘Public lower-level’ is government health centre, government health post, mobile clinic, community health worker; ‘private medical’ is private hospital/clinic, private doctor, private mobile clinic. In some countries, additional options have been mapped to this classification. The ‘no treatment/advice’ category includes advice from friends or family. Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from pooled and weighted DHS data.
There is substantial variability across countries, owing in part to differences in preferences, knowledge and, of course, the availability and quality of health services. Whereas only about 15% of cases of fever among children in Côte d'Ivoire did not result in any medical advice sought, about 75% in Burkina Faso and Niger did not. Similarly, only 10% of cases of fever in children in Niger resulted in a visit to a lower-level public facility, but 41% did so in Zambia. The data are consistent in showing a very small degree of treatment or advice sought from traditional healers: the overall average is 2%, with a range from close to zero in Nigeria and Zimbabwe to 4.1% in Ghana.
Figure 2 shows the association between treatment seeking and wealth quintile for the data aggregated across all surveys. Wealthier households are more likely to seek treatment or advice. The percentage who seek no care is about 60% higher in the poorest than in the richest quintile (42% versus 27%).
Children from poorer households are about one-third as likely to be taken to public hospitals than richer children: while only 6% of cases of fever among children in the poorest quintile resulted in a visit to a public hospital, over 18% of those in the richest quintile did. On the other hand, seeking care from lower-level public facilities falls with household wealth: from 28% in the poorest quintile to about 20% in the richest. Overall, the probability of seeking care from public sources (both hospitals and lower levels) is slightly lower among poorer than richer households (34% versus 38% for the poorest and richest quintiles, respectively).
About 23% of cases of fever result in a visit to a private source, either medical or commercial (pharmacy/shop). The use of these sources is substantially higher in the wealthiest households: about 34% of cases in the richest quintile versus about 20% in the other four quintiles.
As in the incidence analysis, these basic statistics mask heterogeneity across countries, and do not control for potentially confounding factors – variables related to both treatment choice as well as household wealth. Allowing for potentially confounding factors can be implemented by using the Multinomial Logit (MNL) model. The approach allows the investigation, among those who report fever in the past 2 weeks, of the partial association between treatment choices and household wealth conditioning on the same set of variables as in the incidence analysis, i.e. observed individual, household characteristics and cluster variables (as well as controlling for the sub-national region and month of interview).7
7Unlike the incidence analysis, the estimated MNL model does not include the interaction of the sub-national and month of interview dummy variables. This is because the model becomes hard to identify as the number of variables increases. Since the interaction terms were included mostly to adjust for potential seasonality in incidence, and the MNL is conditional on a child having a fever, their omission in this model is not too problematic.
Interpreting the magnitude of MNL coefficients is difficult. As a summary, Table 4 reports the ratio of the predicted probability of seeking each type of care between the richest and poorest quintiles for each country, as well as for the data pooled across all surveys. This prediction sets all characteristics to their actual values except the quintile; predicts the probability for each observation setting the quintile variables to each value in turn; and then averages over all observations at each value of the quintile. The significance of the underlying MNL coefficient on the richest quintile is indicated with asterisks.8 (Detailed significance levels for each treatment choice, as well as all underlying coefficients, are available in Annex tables from the author.)
8In this part of the analysis, care from traditional healers is grouped with no treatment because the MNL model cannot be implemented when only few cases choose one particular choice.
. | Predicted probability for richest quintile divided by predicted probability for poorest quintile . | . | . | . | . | P-values of tests of significance . | . | . | ||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
. | No treatment . | Public, higher level . | Public, lower level . | Private, medical . | Pharmacy/shop . | All own wealth variables . | All cluster wealth variables . | Cluster fever variable . | ||||||
Burkina Faso 1992/3 | 0.93 | n.a.** | 1.31 | 0.48 | 0.29** | 0.000** | 0.000** | 0.000** | ||||||
Côte d'Ivoire 1994/5 | 0.30 | 1.52** | 2.08** | 0.42 | 0.52 | 0.000** | 0.342 | 0.284 | ||||||
Cameroon 1991 | 0.78 | 1.64 | 1.17 | 1.46 | n.a.** | 0.000** | 0.000** | 0.065 | ||||||
Ghana 1993 | 0.69 | 1.55 | 1.09 | 0.91 | 1.25 | 0.621 | 0.265 | 0.069 | ||||||
Kenya 1998 | 0.83 | 0.80 | 1.07 | 1.15 | 1.08 | 0.364 | 0.013* | 0.104 | ||||||
Madagascar 1992 | 0.93 | 1.52 | 0.97 | 1.03 | 0.54 | 0.073 | 0.161 | 0.900 | ||||||
Malawi 1996 | 1.02 | 0.31 | 0.98 | 1.21 | 1.07 | 0.012* | 0.000** | 0.039* | ||||||
Niger 1992 | 0.89 | 0.03** | 2.47** | 1.02 | 1.34 | 0.000** | 0.000** | 0.424 | ||||||
Nigeria 1999 | 0.68 | 1.44 | 1.35 | 1.10 | 1.18 | 0.296 | 0.000** | 0.817 | ||||||
Rwanda 1992 | 0.92 | 1.57 | 1.15 | 0.45 | 3.26* | 0.056 | 0.030* | 0.741 | ||||||
Senegal 1992/3 | 0.99 | 0.95 | 1.21 | 0.59 | 0.54 | 0.440 | 0.579 | 0.305 | ||||||
Tanzania 1991/2 | 1.10 | 1.71 | 0.92 | 0.78 | 0.52 | 0.197 | 0.000** | 0.009** | ||||||
Zambia 1992 | 0.94 | 1.82 | 0.78 | 1.20 | 1.16 | 0.551 | 0.129 | 0.581 | ||||||
Zimbabwe 1999 | 0.36 | 0.18 | 1.27* | 1.45 | 3.21* | 0.000** | 0.000** | 0.008** | ||||||
All surveys (pooled) | 0.78 | 1.65** | 1.07** | 1.17** | 1.07 | 0.001** | 0.006** | 0.644 | ||||||
All surveys: rural only | 0.81 | 1.52** | 1.09* | 1.08 | 1.28* | 0.137 | 0.014* | 0.794 |
. | Predicted probability for richest quintile divided by predicted probability for poorest quintile . | . | . | . | . | P-values of tests of significance . | . | . | ||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
. | No treatment . | Public, higher level . | Public, lower level . | Private, medical . | Pharmacy/shop . | All own wealth variables . | All cluster wealth variables . | Cluster fever variable . | ||||||
Burkina Faso 1992/3 | 0.93 | n.a.** | 1.31 | 0.48 | 0.29** | 0.000** | 0.000** | 0.000** | ||||||
Côte d'Ivoire 1994/5 | 0.30 | 1.52** | 2.08** | 0.42 | 0.52 | 0.000** | 0.342 | 0.284 | ||||||
Cameroon 1991 | 0.78 | 1.64 | 1.17 | 1.46 | n.a.** | 0.000** | 0.000** | 0.065 | ||||||
Ghana 1993 | 0.69 | 1.55 | 1.09 | 0.91 | 1.25 | 0.621 | 0.265 | 0.069 | ||||||
Kenya 1998 | 0.83 | 0.80 | 1.07 | 1.15 | 1.08 | 0.364 | 0.013* | 0.104 | ||||||
Madagascar 1992 | 0.93 | 1.52 | 0.97 | 1.03 | 0.54 | 0.073 | 0.161 | 0.900 | ||||||
Malawi 1996 | 1.02 | 0.31 | 0.98 | 1.21 | 1.07 | 0.012* | 0.000** | 0.039* | ||||||
Niger 1992 | 0.89 | 0.03** | 2.47** | 1.02 | 1.34 | 0.000** | 0.000** | 0.424 | ||||||
Nigeria 1999 | 0.68 | 1.44 | 1.35 | 1.10 | 1.18 | 0.296 | 0.000** | 0.817 | ||||||
Rwanda 1992 | 0.92 | 1.57 | 1.15 | 0.45 | 3.26* | 0.056 | 0.030* | 0.741 | ||||||
Senegal 1992/3 | 0.99 | 0.95 | 1.21 | 0.59 | 0.54 | 0.440 | 0.579 | 0.305 | ||||||
Tanzania 1991/2 | 1.10 | 1.71 | 0.92 | 0.78 | 0.52 | 0.197 | 0.000** | 0.009** | ||||||
Zambia 1992 | 0.94 | 1.82 | 0.78 | 1.20 | 1.16 | 0.551 | 0.129 | 0.581 | ||||||
Zimbabwe 1999 | 0.36 | 0.18 | 1.27* | 1.45 | 3.21* | 0.000** | 0.000** | 0.008** | ||||||
All surveys (pooled) | 0.78 | 1.65** | 1.07** | 1.17** | 1.07 | 0.001** | 0.006** | 0.644 | ||||||
All surveys: rural only | 0.81 | 1.52** | 1.09* | 1.08 | 1.28* | 0.137 | 0.014* | 0.794 |
Note: Based on predicted probability of seeking treatment or advice from each source as derived from Multinomial Logit models that include: the percentage of the cluster population in each quintile; child sex, age and age squared, mother's and father's years of schooling; a dummy variable for urban residence; dummy variables for sub-national region, and month of interview.
**indicates that the coefficient on the richest quintile in the MNL model is significantly different from zero at the 1% level; *at the 5% level. Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from pooled and weighted DHS data.
. | Predicted probability for richest quintile divided by predicted probability for poorest quintile . | . | . | . | . | P-values of tests of significance . | . | . | ||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
. | No treatment . | Public, higher level . | Public, lower level . | Private, medical . | Pharmacy/shop . | All own wealth variables . | All cluster wealth variables . | Cluster fever variable . | ||||||
Burkina Faso 1992/3 | 0.93 | n.a.** | 1.31 | 0.48 | 0.29** | 0.000** | 0.000** | 0.000** | ||||||
Côte d'Ivoire 1994/5 | 0.30 | 1.52** | 2.08** | 0.42 | 0.52 | 0.000** | 0.342 | 0.284 | ||||||
Cameroon 1991 | 0.78 | 1.64 | 1.17 | 1.46 | n.a.** | 0.000** | 0.000** | 0.065 | ||||||
Ghana 1993 | 0.69 | 1.55 | 1.09 | 0.91 | 1.25 | 0.621 | 0.265 | 0.069 | ||||||
Kenya 1998 | 0.83 | 0.80 | 1.07 | 1.15 | 1.08 | 0.364 | 0.013* | 0.104 | ||||||
Madagascar 1992 | 0.93 | 1.52 | 0.97 | 1.03 | 0.54 | 0.073 | 0.161 | 0.900 | ||||||
Malawi 1996 | 1.02 | 0.31 | 0.98 | 1.21 | 1.07 | 0.012* | 0.000** | 0.039* | ||||||
Niger 1992 | 0.89 | 0.03** | 2.47** | 1.02 | 1.34 | 0.000** | 0.000** | 0.424 | ||||||
Nigeria 1999 | 0.68 | 1.44 | 1.35 | 1.10 | 1.18 | 0.296 | 0.000** | 0.817 | ||||||
Rwanda 1992 | 0.92 | 1.57 | 1.15 | 0.45 | 3.26* | 0.056 | 0.030* | 0.741 | ||||||
Senegal 1992/3 | 0.99 | 0.95 | 1.21 | 0.59 | 0.54 | 0.440 | 0.579 | 0.305 | ||||||
Tanzania 1991/2 | 1.10 | 1.71 | 0.92 | 0.78 | 0.52 | 0.197 | 0.000** | 0.009** | ||||||
Zambia 1992 | 0.94 | 1.82 | 0.78 | 1.20 | 1.16 | 0.551 | 0.129 | 0.581 | ||||||
Zimbabwe 1999 | 0.36 | 0.18 | 1.27* | 1.45 | 3.21* | 0.000** | 0.000** | 0.008** | ||||||
All surveys (pooled) | 0.78 | 1.65** | 1.07** | 1.17** | 1.07 | 0.001** | 0.006** | 0.644 | ||||||
All surveys: rural only | 0.81 | 1.52** | 1.09* | 1.08 | 1.28* | 0.137 | 0.014* | 0.794 |
. | Predicted probability for richest quintile divided by predicted probability for poorest quintile . | . | . | . | . | P-values of tests of significance . | . | . | ||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
. | No treatment . | Public, higher level . | Public, lower level . | Private, medical . | Pharmacy/shop . | All own wealth variables . | All cluster wealth variables . | Cluster fever variable . | ||||||
Burkina Faso 1992/3 | 0.93 | n.a.** | 1.31 | 0.48 | 0.29** | 0.000** | 0.000** | 0.000** | ||||||
Côte d'Ivoire 1994/5 | 0.30 | 1.52** | 2.08** | 0.42 | 0.52 | 0.000** | 0.342 | 0.284 | ||||||
Cameroon 1991 | 0.78 | 1.64 | 1.17 | 1.46 | n.a.** | 0.000** | 0.000** | 0.065 | ||||||
Ghana 1993 | 0.69 | 1.55 | 1.09 | 0.91 | 1.25 | 0.621 | 0.265 | 0.069 | ||||||
Kenya 1998 | 0.83 | 0.80 | 1.07 | 1.15 | 1.08 | 0.364 | 0.013* | 0.104 | ||||||
Madagascar 1992 | 0.93 | 1.52 | 0.97 | 1.03 | 0.54 | 0.073 | 0.161 | 0.900 | ||||||
Malawi 1996 | 1.02 | 0.31 | 0.98 | 1.21 | 1.07 | 0.012* | 0.000** | 0.039* | ||||||
Niger 1992 | 0.89 | 0.03** | 2.47** | 1.02 | 1.34 | 0.000** | 0.000** | 0.424 | ||||||
Nigeria 1999 | 0.68 | 1.44 | 1.35 | 1.10 | 1.18 | 0.296 | 0.000** | 0.817 | ||||||
Rwanda 1992 | 0.92 | 1.57 | 1.15 | 0.45 | 3.26* | 0.056 | 0.030* | 0.741 | ||||||
Senegal 1992/3 | 0.99 | 0.95 | 1.21 | 0.59 | 0.54 | 0.440 | 0.579 | 0.305 | ||||||
Tanzania 1991/2 | 1.10 | 1.71 | 0.92 | 0.78 | 0.52 | 0.197 | 0.000** | 0.009** | ||||||
Zambia 1992 | 0.94 | 1.82 | 0.78 | 1.20 | 1.16 | 0.551 | 0.129 | 0.581 | ||||||
Zimbabwe 1999 | 0.36 | 0.18 | 1.27* | 1.45 | 3.21* | 0.000** | 0.000** | 0.008** | ||||||
All surveys (pooled) | 0.78 | 1.65** | 1.07** | 1.17** | 1.07 | 0.001** | 0.006** | 0.644 | ||||||
All surveys: rural only | 0.81 | 1.52** | 1.09* | 1.08 | 1.28* | 0.137 | 0.014* | 0.794 |
Note: Based on predicted probability of seeking treatment or advice from each source as derived from Multinomial Logit models that include: the percentage of the cluster population in each quintile; child sex, age and age squared, mother's and father's years of schooling; a dummy variable for urban residence; dummy variables for sub-national region, and month of interview.
**indicates that the coefficient on the richest quintile in the MNL model is significantly different from zero at the 1% level; *at the 5% level. Pooled data adjusts survey weights to account for differing survey size to population size ratios across countries.
Source: Author's calculations from pooled and weighted DHS data.
The aggregated data show similar general patterns to those identified in the bivariate tables, although after controlling for other factors, the use of public lower-level facilities increases with household wealth, along with the use of public hospitals and private medical facilities. Children from the richest quintile are 65% more likely to be taken to a public hospital, 7% more likely to be taken to a public lower-level facility, and 17% more likely to be taken to a private medical facility than those in the poorest quintile (conditional on the other variables in the model). The results for the rural only sample are similar, except that here the use of private medical facilities is not statistically significantly related to wealth while the use of private commercial sources (pharmacy/shop) is. In rural areas, treatment for children in the richest quintile is about 30% more likely to be sought from a pharmacy or a shop than for children in the poorest quintile.
Wealth in the cluster in which a child lives is also significantly related to treatment-seeking patterns. In particular, wealth in the cluster is statistically significantly related to the use of public hospitals and of private medical facilities. In each case, children in wealthier clusters are more likely to be taken for treatment from these sources. (These results are not reported here but in an Annex table available from the author.)
There are several reasons why aggregating the data might mask important differences. In particular, treatment-seeking behaviour will likely vary substantially with the quantity and quality of health services, which is largely country dependent. Table 4 therefore also reports the rich/poor ratio of the predicted probabilities of treatment seeking from each source for each of the countries.
There is indeed substantial variation across countries. In Burkina Faso, children from the poorest quintile are almost never predicted to be taken to a public hospital, which results in an extremely large ratio of rich to poor. In many countries, children from the richest quintile are around 50% more likely to be taken to a public hospital than those from the poorest quintile, but this is not the case in all countries. In Niger, for example, children in the richest quintile are substantially (and statistically significantly) less likely to be taken to a public hospital. In Malawi and Zimbabwe, they are substantially less likely to be, but the association is not statistically significantly different from zero.
In two countries, Côte d'Ivoire and Niger, children from the richest quintile are more than twice as likely to be taken to a public lower-level facility than those from the poorest quintile. This association is not statistically significantly different from zero in any other country, except Zimbabwe where children from the richest quintile are 27% more likely to be taken to a lower-level public facility than those in the poorest quintile. In three countries (Cameroon, Rwanda and Zimbabwe), the likelihood of seeking care from a commercial source (pharmacy/shop) is over three times higher in the richest than in the poorest quintile.
Discussion and conclusions
A recent review of the social science literature on behavioural issues related to malaria control in sub-Saharan Africa identifies various priorities for new research (Williams and Jones 2004). While responding to some of the calls made in that review is impossible using the DHS data, such as those amenable to ethnographic studies, these data are particularly useful for using rigorous statistical analysis in a consistent way in many countries. In particular, despite the small variations in phrasing and survey administration, the DHS overcome a key difficulty identified by Williams and Jones (2004, p. 507) by being very consistent across countries. No other data source comes close in terms of providing such a large database that can be used to analyze patterns of the incidence and treatment of fever in a systematic way across countries. In addition, the short recall period on the incidence of fever (2 weeks) minimizes recall error, which has been identified as a potential source of error in these types of analyses (McCombie 2002).
There are, of course, issues that warrant caution in interpreting the results. Chief among them is the interpretation of self-reported data on the incidence of fever and its use as a marker for malaria. First, caregivers might not recognize fever in children. Recent investigations of whether fever is recognized have not reached firm conclusions (Dunyo et al. 1997; Einterz and Bates 1997; Kofoed et al. 1998; Verhoef et al. 1998). Reassuringly, the analysis using data that most resembles the DHS has reasonably high sensitivity (78% of fever cases were identified by the caregiver) and rarely (0.8%) found reports of fever when it was not present (Dunyo et al. 1997). Note that the findings are consistent with a bias towards reporting fever, even when it is not 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.
A second potential problem is in interpreting fever as malaria since not all fever is malaria (Brinkmann and Brinkmann 1991; Redd et al. 1992, 1996; Perkins et al. 1997). In the countries under study where there is high malaria prevalence, using fever as a marker for malaria is consistent with Integrated Management of Childhood Illness (IMCI) guidelines for treatment in areas of stable malaria (Gove 1997).9 A related issue is that in some contexts ‘fever’ might refer to illnesses other than those involving high temperature. These data cannot be used to explore this potential problem, although to the extent that this happens in all wealth groups, it should not affect the relative incidence across wealth groups. To the extent that it is geographically specific to national or sub-national regions, this should be accounted for in the multivariate analysis.
9The countries of sub-Saharan Africa under study here are virtually all in areas which are suited to stable malaria (MARA 1998). Stable malaria describes areas with year-round transmission, which may be low or high intensity. Northern regions of Chad, Mali and Niger are not suited to stable malaria but Northern Mali is excluded by virtue of DHS sample design, and dummy variables for sub-national region will account for differences in northern Chad and Niger.
A third potential problem in interpreting these data is self-selective reporting, that is, who recognizes fever might differ by socioeconomic group. Past studies have typically concluded that more easily observed symptoms are less likely to suffer from selective reporting.10 The DHS data cannot be used to test whether differential recognition of fever biases the results. However, the multivariate analysis controls for mother's education, which would likely capture a large part of the self-selective nature of reporting.
10For examples of studies addressing this issue see Butler et al. (1987) for an example from the United States, and Sindelar and Thomas (1991), Strauss and Thomas (1996) and Deolalikar (1998) for discussions relating to poor countries.
Another set of problems in interpreting the results, especially comparing results across countries or aggregating the various datasets, is related to the fact that DHS surveys were not fielded in such a way as to consistently be in a low or high period of malaria transmission, and consequently fever incidence. Even though all the countries studied are in areas of stable malaria, there may be a seasonal element to transmission in all or some parts of the countries. Note that such a sensitivity would need to affect not just the levels of incidence and use of facilities, but also their relative magnitudes across socioeconomic groups in order to alter the main conclusions of this analysis. Including dummy variables for the month in which the survey was implemented, and carrying out analysis country-by-country will reduce the potential biases that this might imply. While this approach will mitigate the problem, it will not necessarily remove the possibility that the results might be sensitive to precisely when in the malaria season a survey was carried out. This caveat should be borne in mind; the validity of the comparative and aggregate results clearly rest on this problem not being too severe.
At the global level, there is no doubt that malaria and poverty move together. Based on country-level data, Gwatkin and Guillot (2000) estimated that almost 60% of deaths due to malaria occurred among the poorest 20% of the world's population in 1990. McCarthy et al. (2000) and Gallup and Sachs (2000) have tried to show the extent to which this national-level association is causal and have concluded that malaria causes a substantial percentage reduction in GDP per capita growth. More generally, countries where malaria is prevalent are also poor countries and while the relationship is not perfect, a focus on malaria is a focus on the world's poor. But there has been little work on whether malaria and poverty are linked at the level of households and individuals.
The DHS data suggest only a weak association between household-level poverty and the incidence of fever among children under 3 years. This holds both when aggregating over all the datasets available for this analysis, as well as at the country level, with incidence about 10 percentage points lower in the richest than the poorest quintile in the aggregated data. This gap becomes smaller, and typically statistically insignificant, after controlling for other variables.
While malaria may be more prevalent in poorer parts of the world (and may even contribute to national poverty if the macro-level analyses are correct), the results here suggest that there is not a strong association at the household level, at least to the extent that fever is a valid marker for malaria. This is consistent with the geographic specificity of malaria. Indeed, the DHS results suggest that the local average incidence of fever is consistently related to the chance that an individual child falls ill. There are two main implications of these findings. First, despite the fact that wealthier households may be able to afford preventive anti-malaria products such as bednets, they do not typically use them effectively to prevent fever. Secondly, rich and poor households alike are affected by the overall level of fever.
The focus of this paper on fever (as opposed to malaria) is driven by its availability in the DHS data. Nevertheless, episodes of fever typically are what prompt caregivers to seek treatment, and patients are most often treated presumptively (both by parents and by medical personnel) for malaria. Past studies of this treatment-seeking behaviour have documented large cross-country variability (McCombie 1996, 2002; Williams and Jones 2004). Nevertheless, a typical finding is that approximately 20 to 25% of cases result in a visit to formal health services.11 Very few cases rely exclusively on traditional healers or medicines (none at all for uncomplicated malaria). One study from Malawi found a higher share of episodes, 52%, resulting in a visit to a clinic (Slutsker et al. 1994) and is notable in that higher socioeconomic status was found to be positively correlated with clinic attendance. McCombie (2002) summarizes various studies suggesting that treatment at a clinic (or hospital) is positively associated with educational levels, as well as with prior experience with the disease.
11Brinkmann and Brinkmann (1991) concluded that between 8 and 25% of persons with malaria visit health services, with self-treatment being more common in urban than in rural areas. McCombie (1996) found a substantial variation across countries. On average, close to 50% of cases relied exclusively on self-treatment, usually with antimalarials. Most episodes involved self-treatment using purchased drugs. More recent studies, with relatively consistent results, are Mwenesi et al. (1995), Ruebush et al. (1995) and (Ahorlu et al. 1997). McCombie (2002) summarizes that the rate of self-treatment is high in a number of studies, with perceived severity as a major determinant of self-treatment.
The DHS data generally imply a higher percentage of visits as a result of fever: around 30% result in a visit to a public facility, and 20% result in a visit to a private or commercial facility. But the main new finding is the gradient across the wealth distribution. Children from the richest quintile are substantially and statistically significantly more likely to be taken to a public hospital or a private medical facility than those in the poorest quintile and, conditional on other factors, to a public lower-level facility (such as a clinic) as well.
The DHS data suggest that while these general patterns hold for many of the countries, there is sufficient variation across countries that any policy seeking to reform the health sector in order to better cater to the poor needs to be informed by country-specific work.
This analysis of DHS data provides only limited support for (McCombie's 1996, 2002) finding that the community prevalence of malaria reduces the probability of seeking care from a doctor. The association between cluster incidence and treatment choice is only statistically significant in four of the countries, and the magnitude of the effect is typically small.
Taken as a whole, the results suggest that the incidence of fever and its treatment are related to poverty in sub-Saharan Africa. Incidence is typically lower at the very top of the wealth distribution. The relationship, however, is not strong, especially after controlling for potentially confounding factors. Treatment patterns are strongly related to poverty; wealthier households are more likely to seek care or advice. While it is perhaps unsurprising that treatment from private sources increases with household wealth, government services – despite their public nature – are also typically used more by wealthier households. If a goal of public policy is to enable poor households to use public facilities, then the typical pattern of behaviour does not suggest that this goal is being met.
Biography
Deon Filmer is a Senior Economist in the Development Research Group (Public Services Team) of the World Bank. He received his Ph.D. in Economics from Brown University in 1995 after which he joined the research group at the World Bank. He was a core team member of the World Development Report 2004: Making Services Work for Poor People. His research focuses on how the behaviour of individuals, households and providers interact with public policy in the determination of health and education outcomes.
The author thanks Davidson Gwatkin, Menno Pradhan, Adam Wagstaff, two anonymous referees, 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. The findings, interpretations and conclusions expressed in this paper are entirely those of the author(s) and do not necessarily represent the views of the World Bank, its Executive Directors, or the countries they represent.
References
Ahorlu CK, Dunyo SK, Afari EA, Koram KA, Nkrumah FK.
Bonilla-Chacin M, Hammer JS.
Brinkmann U, Brinkmann A.
Butler JS, Burkhauser RV, Mitchell JM, Pincus TP.
Chima RI, Goodman CA, Mills A.
Deaton A.
Deolalikar A.
Dunyo SK, Koram KA, Nkrumah FK.
Einterz EM, Bates ME.
Filmer D, Pritchett L.
Filmer D, Pritchett L.
Gallup JL, Sachs JD.
Gove S.
Gwatkin DR.
Gwatkin DR, Guillot M.
Gwatkin D, Rutstein S, Johnson K, Pande R, Wagstaff A.
Kofoed PE, Dias F, Lopes F, Rombo L.
MARA.
McCarthy FD, Wolf H, Wu Y.
McCombie SC.
McCombie SC.
Mwenesi H, Harpham T, Snow RW.
Perkins BA, Zucker JR, Otieno J et al.
Redd SC, Bloland PB, Kazembe PN et al.
Redd SC, Kazembe PN, Luby SP et al.
Ruebush TK, Kern MK, Campbell CC, Oloo AJ.
Sahn DE, Stifel DC.
Sindelar J, Thomas D.
Slutsker L, Chitsulo L, Macheso A, Steketee RW.
Stecklov G, Bommier A, Boerma T.
Strauss J, Thomas D.
Verhoef H, Hodgins E, West CE, Carter JY, Kok FJ.
Wagstaff A, Watanbe N.
Williams HA, Jones COH.
World Bank.
World Health Organization.