Use of proxy measures in estimating socioeconomic inequalities in malaria prevalence


Corresponding Author Masha F. Somi, Australian Centre for Economic Research on Health, College of Medicine and Health Sciences, Cnr Mills and Eggleston Roads, Canberra, ACT 0200, Australia. Tel.: +61 2 61253688; Fax: +61 2 61259123; E-mail:


Objectives  To present and compare socioeconomic status (SES) rankings of households using consumption and an asset-based index as two alternative measures of SES; and to compare and evaluate the performance of these two measures in multivariate analyses of the socioeconomic gradient in malaria prevalence.

Methods  Data for the study come from a survey of 557 households in 25 study villages in Tanzania in 2004. Household SES was determined using consumption and an asset-based index calculated using Principal Components Analysis on a set of household variables. In multivariate analyses of malaria prevalence, we also used two other measures of disease prevalence: parasitaemia and self-report of malaria or fever in the 2 weeks before interview.

Results  Household rankings based on the two measures of SES differ substantially. In multivariate analyses, there was a statistically significant negative association between both measures of SES and parasitaemia but not between either measure of SES and self-reported malaria. Age of individual, use of a mosquito net, and wall construction were negatively and significantly associated with parasitaemia, whilst roof construction was positively associated with parasitaemia. Only age remained significant when malaria self-report was used as the measure of disease prevalence.

Conclusions  An asset index is an effective alternative to consumption in measuring the socioeconomic gradient in malaria parasitaemia, but self-report may be an unreliable measure of malaria prevalence for this purpose.


Objectifs  Présenter et comparer les niveaux de statut socioéconomique (SSE) des ménages basés sur la consommation et l’indice des actifs comme deux mesures alternatives du SSE et comparer et évaluer la performance de ces deux mesures dans des analyses multivariées du gradient socioéconomique dans la prévalence de la malaria.

Méthodes  Les données de l’étude proviennent d’une enquête auprès de 557 ménages dans 25 villages en Tanzanie en 2004. Le SSE des ménages a été déterminé sur base de la consommation et d’un indice des actifs calculés en utilisant l’Analyse de Composants Principaux d’un ensemble de variables du ménage. Dans des analyses multivariées de la prévalence de la malaria, deux mesures alternatives de la prévalence de la maladie ont également été utilisées, la parasitémie et la malaria ou la fièvre auto rapportée au cours des deux semaines avant l’interview.

Résultats  Le classification des ménages sur la base des deux mesures du SSE varie de façon importante dans les analyses multivariées, une association négative statistiquement significative a été trouvée entre les deux mesures de SSE et la parasitémie mais pas entre chacune de ces mesures du SSE et la malaria auto rapportée. L’âge de l’individu, l’utilisation d’une moustiquaire et la construction de mur étaient négativement et significativement associés à la parasitémie, alors que la construction d’un toit était positivement associée à la parasitémie. Seul l’âge est demeuré important lorsque la malaria auto rapportée a été utilisée comme mesure de la prévalence de la maladie.

Conclusions  L’indice des actifs est une alternative efficace de la consommation pour mesurer le gradient socioéconomique dans la parasitémie de la malaria, mais la maladie auto rapportée peut être une mesure non fiable de la prévalence de la malaria dans ce but.


Objetivos  Presentar y comparar los rangos del estatus socioeconómico (ESE) de los hogares, utilizando el consumo y un índice basado en activos como dos medidas alternativas de ESE; y comparar y evaluar el desempeño de estas dos medidas mediante análisis multivariado del gradiente socioeconómico en la prevalencia de malaria.

Métodos  Los datos del estudio vienen de una encuesta a 557 hogares en 25 poblados de estudio en Tanzania en el 2004. El SES de los hogares se determinó utilizando el consumo y el índice basado en activos calculado utilizando el Análisis de Componentes Principales en un grupo de variables del hogar. En análisis multivariados de prevalencia de malaria, también se utilizaron dos medidas alternativas de prevalencia de enfermedad – parasitemia y auto reporte de malaria o fiebre en las dos semanas anteriores a la entrevista.

Resultados  Los rangos de hogares basados en las dos medidas de SES difieren sustancialmente. En análisis multivariados, se encontró una asociación negativa significativa entre ambas medidas de SES y la parasitemia pero no entre las medidas de SES y el auto-reporte de malaria. La edad del individuo, el uso de una red mosquitera, y la construcción de las paredes estaban asociadas de forma negativa y significativa con la parasitemia, mientras que la construcción del techo estaba positivamente asociada con la parasitemia. Solo la edad seguía siendo significativa cuando el auto reporte de malaria se utilizaba como medida de la prevalencia de la enfermedad.

Conclusiones  Un índice de activos es una alternativa efectiva al consumo, en medir el gradiente socioeconómico de la parasitemia de malaria, pero el auto reporte puede no ser una medida fiable de prevalencia de malaria con este propósito.


The measurement of socioeconomic status is difficult, particularly in impoverished, rural and developing country settings. Consumption is generally accepted as the most accurate and direct measure of household socioeconomic status (Deaton & Grosh 1998; Castro-Leal et al. 2000; Makinen et al. 2000), but because of the high cost of implementing consumption surveys, alternative measures of welfare, such as asset-based indices, have been developed. The use of asset-based indices has become increasingly wide-spread in recent years, especially after Filmer and Pritchett (2001) found a high predictive value with an asset-based index compared to a consumption survey in estimating the relationship between educational enrolments and socioeconomic status. Houweling et al. (2003) found, however, that the items used in developing asset indices made differences to the numbers of households in each socioeconomic status quintile, and to the allocation of households into quintiles. Further, they argue that items which may have a direct influence on the outcome measure being investigated should not be included in the calculation of the asset index, as this impedes the ability to identify the direct and indirect influences of socioeconomic status on the variable of interest.

Our aim was to assess the effectiveness of an asset-based index in measuring household socioeconomic status, compared to the gold standard consumption survey. First, household rankings using the two methods are compared. An additional comparison is made between two slightly different asset indices, in order to determine the sensitivity of the index used in this study to its composite items. Asset index 1 includes the entire range of assets for which data are available, while asset index 2 excludes wall and roof construction and mosquito net ownership because these factors are likely to be directly associated with the prevalence of malaria. Second, the effectiveness of the two socioeconomic status measures in predicting inequalities in malaria prevalence is assessed. Consistent with Houweling et al. (2003) rationale, asset index 2 is used for this component of the analysis.

Testing for malaria parasites using blood slides is the gold standard measure of malaria prevalence. Self-report of fever or malaria is frequently used as a proxy for malaria infection because parasitaemia testing is invasive, costly and difficult to implement during survey data collection in developing countries, particularly in rural and remote areas. As well as being the first evaluation of asset-based indices in predicting inequalities in malaria prevalence, this is the first evaluation of the malaria prevalence measure ‘self-report of malaria or fever in the previous 2 weeks’ compared to a biological measure (malaria parasitaemia at the time of survey) for inequalities assessments.

Data and methods

Study area

The study took place in 25 villages in the Kilombero and Ulanga Districts, Morogoro Region, south-eastern Tanzania, under the umbrella of the Ifakara Demographic Surveillance Site. The area is described in detail in an INDEPTH monograph (INDEPTH 2002). The most common occupations are subsistence farming, fishing and small scale trading. Rice and maize are the predominant food crops. The site has no paved roads, and some villages are cut off for parts of the year as a result of flooding (usually during the rainy season). Most families have a second house at a farm, where they stay during the planting and harvesting seasons. Malaria transmission at the site is intense and year round, with a mean entomological inoculation rate of 395 infective bites per person per year (G. Killeen, personal communication).

Data collection and analysis

Informed consent was obtained from all participating adults, and from parents or guardians of any children under the age of 11 years. The institutional review boards of the Australian National University, the Ifakara Health Research and Development Centre, the Tanzanian Medical Research Coordinating Committee and the Centers for Disease Control, Atlanta approved the study. Data on household assets and malaria prevalence outcomes were collected as part of the Interdisciplinary Monitoring Project for Antimalarial Combination Therapy (IMPACT-Tz) 1 between May and August 2004. For the requirements of the IMPACT-Tz project, 1591 households from the Ifakara Demographic Surveillance Site were randomly selected to participate in the study; 1295 of these participated. In the first household survey, we elicited information on asset ownership, amount spent on malaria prevention, malarial/fever illness in the previous 2 weeks and took a blood smear to test for malaria parasitaemia. The first interviews coincided with the end of the rainy season, when malaria infection is more common. Of the 1295 households which participated in the IMPACT-Tz study, 600 were randomly selected to participate in a second survey. Of these, 557 participated fully in a survey between September and December 2004 (42 either refused or could not be found and one did not provide sufficient information to be included in the analysis). In the second interview, households provided information on consumption, using three recall periods: 1 year (for expensive and infrequently purchased items such as furniture and travel); 1 month (for items such as clothes and eating out); and 1 week (for frequently purchased items such as food). During the second interview households were also asked about ownership of houses and durable items, and about consumption of items produced within the home.

Data analysis for the consumption survey followed the method described by Deaton and Zaidi (1999), which is a generally accepted method for analysing consumption surveys. Total household consumption comprised imputed rental from all houses the household owned, user fees from all durable items owned and reported household consumption. Total consumption was deflated by a price index (calculated for each household) and household size. Health related spending and tax costs were not included in the total, as neither increases welfare. The value of consumption of items produced within the home was included in the aggregate. Data analysis for the asset index used principal components analysis (PCA) to discriminate between households in asset ownership, a method advocated by Filmer and Pritchett (2001). Because samples varied little for most items in the index, household responses were translated to dichotomous variables (e.g., 0 or 1 for manufactured roof). Two asset indices were developed using this method. Asset index 1 was calculated using wall, roof and floor construction; cooking fuel used; water and light source; access to toilet; and ownership of a bed, mosquito net, watch/clock, mattress, iron, radio, clothing cupboard, bicycle, livestock, sofa, motorbike or car. These are the items that have been used in previous studies at the study site to calculate asset indices. To assist with comparability, this study used the same items. For asset index 2, following the rationale of Houweling et al. (2003), items that may have a direct influence on malaria, for example, wall and roof construction and ownership of a mosquito net, were excluded.

For all three socioeconomic status calculations, households were ranked into quintiles titled ‘Most Poor’, ‘More Poor’, ‘Poor’, ‘Less Poor’ and ‘Least Poor’. The significance of socioeconomic differences in the variables that may influence malaria prevalence was tested using Pearson’s chi-squared for proportions and with anova for continuous variables; differences between two groups were tested with Satterthwaite’s t-test (Casella & Berger 1990). Cuzick’s (1985) test of trend was used to determine if there were trends across socioeconomic status quintiles in the variables that may influence malaria prevalence. These tests are designed to measure the association between socioeconomic status and malaria. The direction of causation between them could be estimated using multivariate analysis, in which one or another of these variables is included as the dependent (left-hand side) variable. But such single equation models are conceptually unsound because the causal relationship between socioeconomic status and malaria is bi-directional. There is much evidence to suggest that there is two-way causation between socioeconomic status and health variables, including qualitative evidence such as the World Bank’s Voices of the Poor Report (Nayaran et al. 2000). In a single equation model, it is assumed that the direction of causation is from the explanatory (right-hand side) variables to the dependant variable with the explanatory variables themselves not being influenced by the dependant variable (i.e. the explanatory variables are exogenous). In the presence of bi-directional causation, the parameter estimates for the explanatory variables will be biased to the extent that the measures of the explanatory variables reflect the influence of that reverse causation. A different estimation technique is needed to remove the bias arising from reverse causation in single equation models.

In this paper, we isolate the influence of socioeconomic status on malaria prevalence using instrumental variable (or two-stage) probit regressions. Two-stage regressions were chosen to overcome the bias arising from the bi-directional relationship between socioeconomic status and malaria discussed above. This estimation method assumes a linear relationship between socioeconomic status (SES) and explanatory variables x1xk, i.e.:


The explanatory variables x1xk used in this relationship must satisfy two conditions. First, they should be plausible determinants of, and hence be correlated with, socioeconomic status, which is the endogenous variable they are predicting. Second, they must be uncorrelated with the structural error of the equation relating malaria to socioeconomic status. If malaria prevalence (M) was a continuous variable, one could easily use the predicted values of SES from the above regression as an instrument (proxy) for SES in a second stage regression to obtain an estimate of the effect of SES on M that would not be contaminated by endogeneity (reverse causality) bias. But, since M is binary, a probit model is appropriate for it, and since probit models are nonlinear, the estimation of the effect of SES on M is not as simple as using the predicted value of SES from a first stage regression as a proxy for SES. Instead, the dependence of the binary variable M on SES is modelled via an underlying continuous latent variable M* in this way,


where y1,…, yn are explanatory variables that are a subset of x1,…,xk with n < k. The instrumental variable probit estimation method assumes that the error terms in the SES and the M* equations are jointly normally distributed, and uses this system of equations to obtain consistent estimates of the effect of SES on M that are not biased. Wooldridge (2002) (Section 15.7.2) provides more details about this estimation procedure.

This instrumental variable method allows us to use the component of socioeconomic status that is uncorrelated with the error term of the structural equation for consistent estimation of the effect of socioeconomic status on malaria. All models were estimated to correct for clustering of individuals at the household level. The reported results are the marginal effects for each variable, calculated for the individual with average characteristics.


Ranking households into quintiles

There was sufficient consumption information to group all the households in the survey into SES quintiles (n = 557). Table 1 shows, for each quintile, the number of households, and the range and average of annual consumption per equivalent adult in Tanzanian shillings. Weekly spending items, including food and cooking requirements, are the largest share of expenses for households in the sample, and on average comprise 76% of total household consumption. Spending on items such as clothes and shoes, entertainment and cleaning products comprise an average 17% of household consumption. Annual user costs from durables average only 0.5%, whilst imputed rents from housing average 2.5% of household consumption. Large expense items, such as saucepans, cutlery, farm and household help and school fees comprise an average 4.0% of household consumption.

Table 1.   Number of households and average annual consumption by SES group
SES quintileNumber of householdsRange (average) of annual household consumption (Tsh)
  1. $US1 = 1030 Tanzanian shillings (Tsh).

Most Poor11245 520–156 054 (114 360)
More Poor111156 242–213 256 (188 053)
Poor112213 654–274 388 (244 677)
Less Poor111275 694–376 378 (325 093)
Least Poor111379 100–2 852 107 (574 410)
Total55745 520–2 852 107 (288 924)

The Most Poor households consume significantly less than the Least Poor households in a year (t-test, P < 0.001). The smallest difference in consumption is in food and cooking requirements, where the Most Poor households consume 39% of the amount that the Least Poor households consume, whilst the largest difference is in large expense items where Most Poor households consume only 16% of the amount the Least Poor consume.

There was sufficient asset information available to group 556 households into socioeconomic status quintiles. Household PCA scores range between −3.13 and 8.21 in asset index 1, and −2.77 and 7.28 in asset index 2. The first principle component explains 22.9% of the variation in asset index 1 and 22.8% of the variation in asset index 2. Table 2 outlines: the items that were used in calculating the indices (note that ownership is at the household level); the proportion of households owning each item; and each item’s PCA weight. The ‘impact on PCA score’ column indicates that a household with a manufactured floor, all other things being equal, will have a PCA score that is higher by 1.11 units in asset index 1 and 1.14 units in asset index 2 than one without a manufactured floor. Similarly, in asset index 1, a household that keeps livestock will, all other things being equal, have a PCA score that is 0.14 units lower than one that does not. The negative impact of three of the variables (having access to piped water or a private well or pump, owning livestock and owning a motorbike in asset index 1) was surprising, though negative signs for items that would be expected to be positively correlated with socioeconomic status have been reported elsewhere. Interestingly, in their asset indices, Njau et al. (2006) and Filmer and Pritchett (2001) also found that piped water had a negative sign. Njau et al. (2006), in the same site as the present study, hypothesized that the negative sign may reflect that households in areas of higher population density (where relatively better off households tend to reside) are more likely to share water supplies. The negative sign associated with livestock is likely to reflect the groups of pastoralists in the study site, who own livestock but few other assets. The behaviour of the sign on the weights for motorcycles is inconsistent between the two indices, but the magnitude of the weights on this asset are very small.

Table 2.   Items used in asset indices, and their weights
Item†Proportion of households owning the itemAsset index 1Asset index 2
WeightImpact on PCA score‡WeightImpact on PCA score
  1. †Households were allocated 1 if at least some of the construction materials or light/cooking fuels fell in these categories.

  2. ‡Impact of a change from 0 to 1 for each variable (weight/standard deviation).

Manufactured floor (tiles/cement)0.090.311.110.321.14
Manufactured walls (fired bricks/cement)0.370.290.60Not used
Manufactured roof (tiles, cement, corrugated iron)0.250.340.78Not used
Cooking fuel is electricity, gas, kerosene or charcoal0.070.250.950.291.10
Water source is piped water inside house or private well or pump0.23−0.06−0.14−0.06−0.13
Household has use of toilet0.970.070.400.080.44
Light source is electricity or hurricane lamp0.230.230.740.360.86
Mosquito net0.900.150.51Not used
Clothing cupboard0.

Table 3 shows the number of households per socioeconomic status quintile, and the range and average of PCA scores per quintile for the two asset indices. There are different numbers of households in each quintile because households with the same PCA score are automatically grouped into the same quintile. When socioeconomic status groupings using the two different asset indices are compared, the number of households in each quintile varies, most notably in the Most Poor and More Poor quintiles. Further, 159 households (29%) were grouped into different quintiles by the two asset indices; of these, 84 moved down one or more quintiles. Of the 16 households that moved more than one socioeconomic status quintile, for example from ‘Poor’ to ‘Most Poor’, 14 moved downwards.

Table 3.   Number of households by SES quintile (asset index)
SES quintileAsset index 1Asset index 2
Number of householdsRange (average) of PCA scoresNumber of householdsRange (average) of PCA scores
Most Poor122−3.13, −1.73 (−2.04)147−2.77, −1.53 (−1.83)
More Poor102−1.7, −1.09 (−1.28)80−1.45, −0.87 (−1.15)
Poor114−1.08, 0.01 (−0.51)107−0.83, −0.43 (−0.56)
Less Poor1090.04, 1.39 (0.68)114 0, 1.45 (0.63)
Least Poor1091.42, 8.81 (3.34)1081.48, 7.28 (3.24)
Total556−3.13, 8.21 (0.1)556−2.77, 7.28 (0.1)

Comparing household rankings

Of the 556 households for which both asset and consumption information was available, the two methods yielded different socioeconomic status groupings in 401 households (72%). Table 4 shows a comparison of household socioeconomic status groupings using the consumption survey and asset index 2. The data in bold are the households that received the same rank from both methods. Of the 401 households which changed ranking, 188 (34%) moved one socioeconomic status group, 118 (29%) moved two groups, 69 (17%) moved three groups and 26 (5%) moved four groups. Figure 1 provides a visual depiction of how household rankings vary when the two socioeconomic status measures are used. Each observation (+) represents the two socioeconomic status rankings for each household. The 45-degree line indicates the position that the observations would have taken if the two socioeconomic status measures produced the same household rankings. The flatter line is the line of best fit, and represents the predicted household consumption ranking based on a linear regression of the consumption ranking on a household’s PCA ranking. The figure further highlights the fact that the two socioeconomic status measures produce very different rankings of households.

Table 4.   Comparison of household SES groupings using consumption and asset index 2
ConsumptionAsset index 2
Most PoorMore PoorPoorLess PoorLeast PoorTotal
Most Poor421924207112
More Poor3722161816109
Less Poor2212212730112
Least Poor1911212239112
Figure 1.

 Comparison of household socioeconomic status rankings using consumption and asset index 2.

Regressing household consumption by the household’s PCA score for asset index 2 produces an R2 of only 0.07, albeit at a very high significance level (P < 0.001). The Spearman rank correlation between household rankings based on consumption and asset index 2 is 0.27, indicating a weak relationship. The use of dichotomous variables in the development of asset index 2 would have contributed to the low correlations with consumption, as they resulted in several households having the same PCA score. As a result, the rankings based on PCA scores are less discriminating than those based on consumption (a continuous measure), with many more tied ranks being observed.

Household ownership of/access to five of the 16 items used in asset index 2 was very significantly associated with household consumption: bicycles; mattress; couch; cooking with manufactured facilities; and use of modern toilet. Regressing household consumption on the 15 items used in asset index 2 (see Table 1 for a list of the items used in asset index 2) produces an R2 of only 0.13, but at very high significance (P < 0.001).

The prevalence of malaria

Of the 2034 individuals in the sample, 1577 provided blood slides for parasitaemia testing. Of these, 402 (25%) were parasite positive on the day of testing. There has been a gradual decline in parasite prevalence in the study site since 2000, when prevalence was 35% (IMPACT-Tz, unpublished data). Table 5 presents and compares malaria prevalence information using the two socioeconomic status measures. There is large variation between the socioeconomic status groups in parasitaemia rates, with lower socioeconomic status individuals significantly more likely to be carrying parasites than higher socioeconomic status individuals (Pearson’s chi2, P < 0.001). The negative trends across quintiles are significant for both measures of socioeconomic status (Cuzick’s test of trend, P = 0.001). Parasite prevalence also varies significantly by age (t-test, P < 0.001).

Table 5.   Differences in individual characteristics by SES quintile using consumption and asset index 2
CharacteristicSocioeconomic statusTotal
Most PoorMore PoorPoorLess PoorLeast Poor
Average age (years)2626222121211921222122
Age under 5 (%)1519202117202318181719
Location = farm (%)4121231819318279316
Mosquito net (%)6464747583809286889079
Treated (%)2013242033312530333727
Eaves (%)8336146148102310
Manufactured roof (%)85191123173427506225
Manufactured walls (%)2122322331284742686838
Ave number of individuals in the house53.
Amount spent on malaria
prevention (Tsh)
Knowledge that malaria is
transmitted by mosquitoes
Parasitaemia (%)3230252827262226211825
Under five parasitaemia (%)4842674238372532251834
Self-report of malaria or fever (%)9131291111101010810
Under five self-report (%)1624201221241514231820
Parasite positive and self-report (%)81310912151511201012

All 2034 individuals in the sample reported whether they had experienced malaria or fever in the previous 2 weeks, and 10% responded that they had. Again there are significant differences by age in reporting rates (t-test, P < 0.001). Variations in reporting rates by individuals in different socioeconomic status groups exist, but are not significant. Trends across quintiles in self-report rates are positive but not significant (Cuzick’s test of trend, P > 0.05 for consumption, and P = 0.054 for asset index 2).

Estimating inequalities in malaria prevalence

Table 5 outlines how characteristics of individuals and households that relate to malaria vary by socioeconomic status quintile, using the two socioeconomic status measures. There are clear and significant trends across socioeconomic status quintiles for all of the variables except the proportion of individuals under the age of five when the asset-based index is used, and the proportion of people reporting malaria or fever in the 2 weeks before interview when both socioeconomic status measures are used (Pearson’s chi2 for proportions and anova for continuous variables, P < 0.05).

In total, four instrumental variable (or two stage) probit regressions were run using a mix of the two socioeconomic status measures (consumption and asset index 2) and two malaria prevalence measures (parasitaemia and self-report of malaria or fever in the 2 weeks before interview). For the first stage regression, the variables used to explain socioeconomic status were the length of time the household head had lived in the area, earning income from a non farming source, and gender and education level of the household head. Education level of the household head was not associated with malaria, satisfying the requirement for it to be used as an instrument. Correlations between socioeconomic status, measured either by consumption or an asset-based index, and the instruments ranged between 0.03 and 0.21. Regressing household consumption and PCA score by the instruments produces R2-values of 0.04 and 0.08, respectively, but at very high significance levels (P < 0.01) for all variables except gender of the household head (P < 0.05). The F-statistics were 21.01 (P < 0.001) and 45.23 (P < 0.001) for consumption and the asset-based index, respectively. Whilst the low R2-values suggest the instruments are weak, the F-statistics indicate that they are not weak enough to distort inference on the structural parameters in the second stage model (Stock & Yogo 2002). Post-estimation tests confirmed that these variables were not associated with the error term from the probit regressions (P > 0.05). These findings indicate that these four variables satisfy the conditions to be instruments for socioeconomic status.

Table 6 outlines the results from the four regressions undertaken. The results are the marginal effects for the individual with average characteristics: the marginal effects show the effect of a change in an explanatory variable on the probability of malaria, holding the values of all other explanatory variables constant. When parasitaemia is used as the measure of malaria prevalence, the two socioeconomic status measures, consumption and asset index 2, are both significant. The two equations containing these measures produce similar coefficients and significance levels for many of the variables used in the regressions, with location of interview and number of people living in the house being the notable exceptions. Whilst the two socioeconomic status measures had low correlation with each other, their association with malaria parasitaemia is negative and significant. Other variables that are significantly associated with parasitaemia include age, sleeping under a mosquito net on the night before interview, and the household’s wall and roof construction.

Table 6.   Marginal effects from IV probit regressions of malaria prevalence based on parasitologic data or self-report
Explanatory variableMalaria prevalence measure (dependent variable)
  1. Significantly different from zero at 95 (*) and 99 (**)% confidence.

SES consumption−0.003* 0.0009 
SES asset index 2 −0.074* 0.018
Age (years)−0.006**−0.006**−0.0007−0.001*
Location (0 = village, 1 = farm)0.0580.111**0.0140.001
Mosquito net−0.082*−0.088*−0.004−0.004
Treated net0.0220.019−0.026−0.023
Number of people living in the house−0.0020.030**0.002−0.007
Manufactured roof0.140*0.179*−0.019−0.021
Manufactured walls−0.100*−0.109**−0.009−0.006
Amount spent on malaria prevention<0.001<0.001<0.001<0.001
Know that malaria caused by mosquitos−0.074−0.0610.0290.034
Number of individuals1566156620152015
Wald χ2185**182**109

The positive association between manufactured roofs and parasitaemia is counterintuitive, and was thus investigated further. Given that the pair-wise correlation between manufactured walls and roofs is the highest amongst all of the explanatory variables in the model (0.69), we defined mutually exclusive dummy variables (manufactured walls but not roof, manufactured roof but not walls, manufactured roof and walls), and estimated the model with these dummy variables instead of manufactured roof and manufactured walls. These results show that having only manufactured walls has a significant effect on malaria (P = 0.004), but having manufactured roof does not (P = 0.928). Additionally, having both manufactured roof and walls has an insignificant effect on malaria (P = 0.099).

The explanation for this counterintuitive result lies in the first stage regression of socioeconomic status on the exogenous variables. In these regressions the manufactured roof and walls dummy has the largest partial correlation with socioeconomic status and is the most significant explanatory variable (its t-statistics were 16.6 in the PCA regression and 8.99 in the consumption regression). This means that, even though manufactured roof and walls are not used in the construction of PCA scores, living in a residence with both manufactured walls and roof is highly correlated with socioeconomic status even after controlling for all other explanatory variables. Therefore, with the socioeconomic status variable in the model, the data cannot separately identify the effect of living in a house with manufactured roof and walls on malaria. A confirmation of the validity of this conclusion is apparent in the reduced form probit model where malaria is regressed on all exogenous variables (that is, when the socioeconomic status variable is excluded) - we obtain negative effects for the manufactured walls (P < 0.001) and roof dummy (P = 0.095) variables. It is important to note that the estimate of the effect of socioeconomic status on malaria parasitaemia prevalence, which is the key parameter of interest, is non-fragile and remains significant and negative with any combination of manufactured roof and manufactured walls dummy variables in the model.

When self-report of malaria or fever in the previous 2 weeks is used as the measure of malaria illness, again the associations between malaria and the two socioeconomic status measures are similar. In these regressions, however, age is the only variable that is significantly associated with reported malaria infection, and this was true only when socioeconomic status is measured by asset index 2.


The first objective of this paper was to compare two socioeconomic status measures: consumption and an asset-based index. The study results indicate that the two measures have low correlation and produce very different household rankings and quintile numbers. The second study objective was to evaluate the performance of the two socioeconomic status measures in a multivariate analysis of socioeconomic inequalities in malaria prevalence. Despite the low agreement between the two measures in ranking households, their association with malaria is very similar. The finding that the two measures of socioeconomic status produce very similar results in multivariate analyses of the household and individual factors that influence malaria prevalence indicates that an asset-based index is an effective alternative to consumption in determining the socioeconomic gradient in malaria prevalence. Thus the cheaper and easier to collect measure of socioeconomic status (the asset index) can and should be used more frequently for studies interested in calculating socioeconomic status in rural subsistence communities.

The regression results provide evidence that consumption and the asset-based index are both acting as proxies for an unobserved underlying cause of difference between the households, that is, household welfare. It is likely that the proxies encapsulate different aspects of household welfare: consumption measures absolute access to resources that maximise welfare in the short run (accepting that households smooth consumption over the long run); whilst an asset index measures a household’s ability (relative to others) to purchase items that are expected to increase welfare. Figure 2 outlines the hypothesized interaction between household welfare, consumption and PCA scores, with an underlying assumption that it is possible for the two proxies to have low correlation yet still accurately reflect the underlying variable they represent. This is an area of possible future research.

Figure 2.

 Relationships between household socioeconomic status, consumption and PCA scores.

The strength and magnitude of the impact of socioeconomic status on malaria prevalence (particularly when malaria is measured by parasitaemia) is a very important finding from this study, and the results of further work on this topic are reported by Somi et al. (2007). The results of regressions using self-report of malaria or fever as the measure of malaria infection show a surprising lack of association with socioeconomic status and the base variables. Given the clear associations found between these variables when parasitaemia is used as the measure of malaria, the lack of association suggests that there might be a problem with the measure itself. It is important to note, however, that the two measures are in effect not measuring precisely the same dimension of malaria, that is, self-identified clinical manifestation of illness versus microscopic evidence of malaria infection. The results of the regression analysis indicate that there are factors mediating the association between actual parasite prevalence and self-reported malaria. The factors causing the difference could not be determined, but may reflect ‘noise in the signal’ between the two variables. This noise could be caused, in part, by the variety of diseases present in the study site and the difficulties associated with distinguishing them using symptoms alone. The difference may also represent the different temporal locations of the two measures: self-report of malaria is a retrospective measure, whilst parasitaemia represents recent past risk of exposure to malaria and a higher risk of future malarial illness. It may also be the case that a person who has undertaken effective treatment for malaria may be accurately self-reporting malaria in the previous 2 weeks and be free of parasites on the day of interview.

Wagstaff (2002) notes that the use of subjective indicators for health inequalities assessments, for example, self-report of illness in the previous 2 weeks, is likely to cause bias so that better-off individuals report worse health than the poor. In their review of Livings Standards Measurement Surveys from five developing countries, Baker and van der Gaag (1993) found that self-report of illness or injury within the previous 4 weeks did not follow distinctive patterns across socioeconomic status quintiles, and that in four of the five countries the prevalence of illness or injury was highest in the Least and Less Poor quintiles. Baker and van der Gaag also found that cross-country comparisons of life expectancy and infant mortality on the one hand and self-reported illness or injury on the other, did not follow the expected positive association between Gross National Product (GNP) and health indicators. Peru, with the highest GNP and average life expectancy and lowest infant mortality of the five countries, had the highest reported illness and injury, whilst Bolivia, with low GNP and very poor health indicators, reported low frequencies of illness and injuries.

If better-off individuals were to report ill health more frequently than poorer individuals, any associations between socioeconomic status and true malaria prevalence would be impossible to determine based on self-report alone, as higher (lower) socioeconomic status individuals would over (under) report infection. It is possible that a larger sample size would allow for associations using self-report of malaria to be detected, but this hypothesis cannot be tested using the current data set. Further studies need to be undertaken on the relationship between parasitaemia and self-report of malaria or fever, in order to further assess the effectiveness of this measure of malaria, particularly as it is often used as a proxy for parasitaemia. This is particularly pertinent given the recent finding in Uganda that a high proportion of individuals with asymptomatic parasitaemia at one point in time developed symptomatic malaria within 30 days (Njama-Meya et al. 2004), suggesting that inequalities in malaria parasitaemia result in inequalities in health outcomes over time.

Whilst the study results imply that self-reported malaria or fever should not be used for studies interested in the socioeconomic gradient in malaria prevalence, the measure is still appropriate for use in studies interested in household responses to malaria illness, for example, the economic burden of malaria; this is because it is the perception of illness (and its associated severity) that determines behaviour such as treatment seeking (Mugisha et al. 2002).

Several of the study findings were unexpected, notably the lack of significant negative associations between malaria parasitaemia and insecticide treatment of mosquito nets, roof construction, eaves, knowledge that malaria is transmitted by mosquitoes and the amount spent on malaria prevention in the month before interview. But, in settings of low or exclusively seasonal malaria transmission, earlier work has established that untreated mosquito nets are sufficient to reduce malaria (Clarke et al. 2001). Regular insecticide treatment is considered more important in areas of intense transmission (such as the site of the study), and regular treatment of mosquito nets with insecticide substantially improves the public health benefits of using a net (Phillips-Howard et al. 2003). It is possible that reporting of insecticide treatment of mosquito nets in our study was inaccurate (i.e. nets reported as treated were in fact not treated, or had been retreated more than 6 months before the survey). In the study site, Erlanger et al. (2004) found that reported insecticide use correlated poorly with measurable insecticide on nets collected from households.

Several African studies have demonstrated the importance of housing construction in reducing malaria transmission (Lindsay & Snow 1988; Adiamah et al. 1993; Ghebreyesus et al. 2000; Charlwood et al. 2003; Lindsay et al. 2003). Eaves and manufactured walls and roofs reduce malaria transmission by preventing mosquitoes from entering the house through cracks and resting in them after blood meals. A review of the methods used in the other African studies (as identified above) indicates that none were able to control for the presence of walls when undertaking their analyses. Only one study (Ghebreyesus et al. 2000) used a multivariate model and did not include walls in the analysis. Ours is the first African study to use multivariate analysis to isolate the contribution that each of the housing construction features makes to malaria prevalence.

Neither household knowledge that malaria is transmitted by mosquitoes nor the amount that a household spent on malaria prevention in the previous month was associated with malaria parasitaemia. These findings may be explained by two reasons. First, households may not be able to undertake activities to reduce contact with mosquitoes, for example, use mosquito nets, despite understanding the need to do so. Second, spending on malaria preventive activities in the research site is not entirely effective in reducing malaria parasitaemia, once mosquito nets have been controlled for. Note that most malaria preventive spending at the study site is on mosquito nets, which also explains the counter-intuitive finding that the poorest households spent the most on malaria prevention (i.e. relatively wealthy households have already purchased mosquito nets and the poorer households are in a catch-up phase). Further research needs to be undertaken in order to find ways to increase the effectiveness of short run malaria preventive spending in these communities, so that maximum benefit is derived from the household’s scarce resources.

Whilst parasitaemia has proven to be an effective measure for predicting inequalities in malaria prevalence, it is difficult to implement parasite testing in developing countries, especially in the isolated rural communities where most malaria transmission, illness and mortality occurs. Facilities are limited, and there is a dearth of adequately trained staff available for both blood slide collection and reading. Our results suggest that using self-report as a proxy will not allow for the detection of existing inequalities in malaria prevalence.

Whilst the findings from the study are clear, several limitations must be borne in mind when evaluating their significance. First, data for the study were collected across two seasons: parasitaemia and self-report data were collected during the rainy season; and consumption data were collected in the dry season. There is evidence that seasonality affects both malaria and consumption; however, financial and time constraints meant that data collection had to be split across two separate interview rounds. Of particular concern to the study team was the amount of time households could be expected to dedicate to each interview. Second, the two sites in the study are relatively homogeneous: they are both relatively poor rural sites. Before definite conclusions can be drawn from these results, they must be replicated in other settings, for example, relatively wealthy rural areas and urban areas. Third, in looking at some individual and household features that may influence malaria parasitaemia prevalence, many individual and environmental features fell beyond the scope of the study. Features such as host immunity and transmission rates were not available for the study sample. Whilst this information would have been useful in further understanding the aetiology of disease, it would not have added to our understanding of the relationship between socioeconomic status and malaria infection. The variables included in the analysis covered many potential mechanisms through which socioeconomic status might affect malaria, and this may have tended to mediate any direct relationship between socioeconomic status and malaria. Nevertheless, the analysis still found a significant direct influence of socioeconomic status on malaria.


The results from this analysis have yielded three key findings. First, asset-based indices are an effective alternative to consumption in measuring socioeconomic status for the purpose of measuring socioeconomic inequalities in malaria, but only if used in conjunction with an objective measure of malaria prevalence. These findings may be applicable to other health outcomes. Further research needs to elucidate the mechanisms of these associations, given that consumption and the asset-based index had low correlations with each other but produced similar results in the multivariate analyses. Second, there is a significant negative relationship between socioeconomic status and malaria prevalence, which can be clearly seen when malaria is measured through parasitaemia. When self-report of illness from malaria or fever is used, the relationship between these two variables is not discernible. Third, self-report of malaria is not a good measure for assessing the relationship between socioeconomic status and malaria. Our data indicate that using self-reported malaria or fever instead of parasitaemia does not allow for socioeconomic inequalities to be detected. This may be true for other health outcomes where the link between the disease and its symptoms is neither strong nor specific, but further work needs to be done on this topic. Self-reported malaria or febrile illness is still a worthwhile measure for tracking trends in care seeking for potential malaria episodes.


  • 1

    IMPACT-Tz is a multiyear implementation research evaluation project that rests on a collaborative platform compris-ing the US Centers for Disease Control and Prevention (CDC), Ifakara Health Research and Development Centre, the National Institute for Medical Research, London School of Hygiene and Tropical Medicine (UK), and the Tanzanian Min-istry of Health, including its National Malaria Control Programme, the Tanzania Essential Health Interventions Project, and the Council Health Management Teams of Rufiji, Morogoro, Mvomeru, Kilombero and Ulanga Districts. IMPACT-Tz is primarily supported by funding from CDC, the United States Agency for International Development, and Wellcome Trust.


We thank the Ifakara Health Research and Development Centre for their support and assistance throughout the data collection process. The IMPACT-Tz project was also very supportive during data collection and shared information with this project. The field work would not have been possible without the help of Rashid Khatib, Jensen Charles and Chrisostom Mahutanga and Berty F. Elling. The findings and conclusions presented in this paper are those of the authors and do not necessarily represent those of the United States Public Health Service or the Centers for Disease Control and Prevention.