Explaining household socio-economic related child health inequalities using multiple methods in three diverse settings in South Africa

Background: Despite free healthcare to pregnant women and children under the age of six, access to healthcare has failed to secure better child health outcomes amongst all children of the country. There is growing evidence of socioeconomic gradient on child health outcomes. Methods. The objectives of this study were to measure inequalities in child mortality, HIV transmission and vaccination coverage within a cohort of infants in South Africa. We also used the decomposition technique to identify the factors that contribute to the inequalities in these three child health outcomes. We used data from a prospective cohort study of mother-child pairs in three sites in South African. A relative index of household socio-economic status was developed using principal component analysis. This paper uses the concentration index to summarise inequalities in child mortality, HIV transmission and vaccination coverage. Results: We observed disparities in the availability of infrastructure between least poor and most poor families, and inequalities in all measured child health outcomes. Overall, 75 (8.5%) infants died between birth and 36 weeks. Infant mortality and HIV transmission was higher among the poorest families within the sample. Immunisation coverage was higher among the least poor. The inequalities were mainly due to the area of residence and socio-economic position. Conclusion: This study provides evidence that socio-economic inequalities are highly prevalent within the relatively poor black population. Poor socio-economic position exposes infants to ill health. In addition, the use of immunisation services was lower in the poor households. These inequalities need to be explicitly addressed in future programme planning to improve child health for all South Africans. © 2011 Nkonki et al; licensee BioMed Central Ltd.

Three child health outcomes were chosen: infant mortality (<9 months), HIV transmission and immunisation coverage. Infant mortality was described as any death of an infant between birth and nine months of age (<9 months). HIV transmission amongst infants was described as the number of HIV infected infants at 3, 24 or 36 weeks of age. Immunisation coverage was described as the number of infants who received complete immunisation (i.e. BCG, OPV3 and DTP3) at 24 weeks. The latter indicator is not only a child health indicator but also a health service use indicator. All three outcomes were binary variables. Krieger, Williams and Moss [14] use the term socio-economic position to refer to various components of economic and social well-being, as related to class position. They argue that the commonly used term 'socio-economic status' blurs the distinction between two different aspects of 'socio-economic position', namely actual resources and status, meaning prestige-or rank-related characteristics. In light of this argument we use the term 'socio-economic position' instead of 'socio-economic status'. There is no best practice on how to select variables to proxy living standards. Researchers have used variables, such as access to utilities and infrastructure (for example sanitation facilities, electricity and sources of water), durable goods [15,16], ownership of live stock[17], ownership of land[12], and the ratio of the number of people to the number of rooms in a household [18]. The number of variables used in studies has ranged from 10[12] to 30[19]. However, some researchers have used parsimonious sets of variables [20-26]. These sets include only 3-4 variables, for instance radio, television, type of floor [20-26]or electricity in a home, piped water and high status occupation[22]. For this study, we collected information on durable asset ownership, access to utilities and infrastructure, food availability at 24 weeks postpartum, maternal education, and household income. We avoided combining socio-economic indicators that are hypothesised to interact with health outcomes in different ways. In deciding on which variables to include in the index, we chose variables that would not increase households' ability to purchase health services or increase their understanding of health messages. Therefore, we did not include maternal education, food availability and household income in the index. The variables we considered including in the asset index were: fridge, radio, TV, stove, telephone/mobile phone, car, and infrastructural variables, such as sources of water supply, type of toilet and main fuel used for cooking. Including of infrastructural variables in an asset index aimed at measuring household wealth can be viewed as inappropriate since these variables are publicly provided and therefore depend on availability of infrastructure at the community level. The argument of whether the asset index is reflects household's wealth or community wealth depends on how the assets were financed. In this study, the source of finance for the assets is less relevant because we are interested in the relative advantage that having these items presents to a household, and not who paid for them. For instance a household of high income earners within an area that does not have running water, electricity and adequate sanitation would be subject to the same constraints as other households within the same community. In addition, they would be at a disadvantage compared to their counterparts within households in areas that do not have these publicly provided indicators. Therefore, a comprehensive asset index should acknowledge that infrastructural variables play a role in living standards. Another disclaimer to the including infrastructural variables in the asset index is related to the health outcome of choice. Infrastructural variables have a direct relationship with infant mortality over and above their indirect relationship as measures of socio-economic position. Excluding these variables is expected to yield smaller inequalities in outcomes such as infant mortality [27]. In light of infrastructural community level variables having a direct relationship with infant mortality, we constructed three indices; the first index only included consumer durables. This index included a fridge, radio, TV, stove, telephone/mobile phone and car. We did not base this choice of items on any economic value of the items themselves; it was based on the availability of items, which were only indicators of socio economic position on the original data set. The second index comprised infrastructural variables. The third index was a combination of both Index 1 and Index 2. We assigned variable weights for all three indices, using the method of principal components analysis (PCA) [15]. PCA can be used as a data reduction or classification tool. In this paper, we have it as a tool for summarising variability within a set of variables. This method describes the variation of a set of multivariate data in terms of a set of uncorrelated linear indices or components of the original variables. Each consecutive linear combination is derived so that it explains as much as possible the variation in the original data, while being uncorrelated with other linear combinations [17]. The components are ordered so that the first component explains the largest possible variation in the original data. The subsequent components are uncorrelated with the preceding component and explain additional but less variation than the first component. McKenzie[19]demonstrated that only the first principal component was necessary for measuring wealth. Therefore, we extracted only the first principal component in this study. The first linear combination of variables (the first principal component [c1]) contains the most information on the variation in the underlying set of variables. The xij terms refer to variable i for household j, and the yhi terms refer to the factor loadings (linear coefficients) for component h and variable i. The first linear combination is: Using the first principal component, the percentage of variance were 44%, 69% and 42% for the consumer durables index (Index1), the index with infrastructural variables (Index2) and the index that combined the first two indices respectively. Index 1 and 2 were short and consisted of more homogeneous sets of items. As a consequence, households were grouped together in a small number of distinct clusters (clumping). Hence, it was not possible to categorise households into to five wealth groups. This problem is not unique to this study, McKenzie [19] identified problems of clumping and truncation as a major challenge for PCA-based asset indices. Therefore, we chose Index 3 as the measure of socio-economic position, since it included a range of asset variables that were broad enough to avoid clumping problems. We assessed the reliability of the asset index in two dimensions. Firstly we assessed whether the asset index produced clean separations across the least poor to the most poor for assets that are indicative of least poor and assets that are indicative of most poor. Secondly, we assessed the relationship between education, income and the food inventory index, and the asset index using cross-tabulations. We were interested in observing whether the asset index agreed with other measures of socio-economic position. We measured maternal education as the last standard passed at school. We measured the standard in terms of three categorical variables: no education and primary education, higher primary education, and successfully completed matric. We defined household income as the total household monthly income, including all sources of income. We did not collect information on household size. We based the food inventory index on food items directly observed (by field researchers) in the households at 24 weeks after delivery. Similarly to the asset index variables, we assigned weights using PCA. We did not measure the quantities of these food items, and despite this limitation, observing that these items were present avoids the recall bias inherent in recall methods. For instance, it has been observed that households that have experienced an adverse child health outcome (such as death) have a heightened awareness of all the events prior to the death of the child compared to households who have not experienced that outcome. Similar observations have been made in households that have experienced severe hunger compared to those who have not. Nevertheless, this is a crude approximation of consumption and should be interpreted with caution. We used two different measures to explore the presence of inequality. Firstly, we divided the first principal component into quintiles so that each household was classified as most poor, poor, less poor or least poor, in terms of socio-economic position, with mean scores (i.e. first principal component) of -2.55, -1.50, -0.05, 1.37 and 2.73 respectively. Secondly, we use the concentration index to further explore drivers of inequality. Presenting findings using quintiles is common practice, especially in public health journals. This method is easier to understand since the study population is categorised into five equal groups representing the least poor to the most poor. The concentration index quantifies the degree of socioeconomic related inequality in a health variable [28]. Where: yi is the health variable of interest for the ith person; μ is the mean of y; Ri is the ith-ranked individual in the socioeconomic distribution from the most disadvantaged (i.e. poorest) to the least disadvantaged (i.e., richest); n is the number of persons Unlike quintiles, the concentration index reflects the experiences of the entire population. Another advantage of the concentration index is that it is sensitive to the changes in size of the various groups, even if their health outcome mean has not changed [29]. The bounds of the concentration index are -1 to 1, where the sign indicates the direction of inequality, and the magnitude reflects both the strength of the relationship and the degree of variability in the health variable. The concentration index has been used in other studies to analyse inequalities in child health [30-32]. Using the asset index to measure socio-economic position presents a methodological challenge. The PCA has a mean of zero and takes negative values for about half of the households. Many measures of inequalities are divided by the mean and so do not apply to data that take negative values [33]. In order to address these challenges, we used the properties of the concentration index to rescale the index for the principal component analysis. This is explained in further detail below. The concentration index depends on the relationship between the health variable and the rank of the living standard, and not on the variation in the living standard itself. The change in the living standard should therefore not affect the concentration index [34]. We rescaled the asset index by adding a constant of 3.0, which was the minimum whole number required to eliminate negative values. The range of the asset index prior to rescaling was -2.75 to 3.48. After rescaling the range was 0.25 to 6.48. This rescaling does not affect the contribution of each variable to the concentration index, since the rank ordering is unchanged. However, the relative magnitude of the elasticity and concentration index in the decomposition does change. We calculated the concentration index using covariance and regression methods [34], and both yielded the same result. Inequalities are sometimes unavoidable, for example, there may be an unequal distribution due to biological factors or age. Alternatively, inequality adds a value judgement on the observed disparities; it often includes assessing whether the disparities are unjust, unfair and remediable. Inequalities in child health outcomes are caused by inequalities in the factors that affect the variable of interest. Hence, an important policy question is what the relative contribution of each of these various inequalities in explaining child health outcomes inequality is. In order to address this question, we decomposed the concentration index. Wagstaff, van Doorslaer and Watanabe [32]} demonstrated that the concentration index of health can be expressed as the sum of contributions of various factors, together with an unexplained residual component. Together, the linear additive relationship between the health outcome variable yi, the intercept α, the relative contributions of xkdeterminants and the residual error εi give the formula: Equation 4 demonstrates that the overall inequality in health outcome has two components: the explained component and the unexplained component. In the explained component, βkis the coefficient from a regression of health outcome on determinant k, is the mean of determinant k, μ is the mean health outcome and Ck is the concentration index for determinant k. In the unexplained component, GCεis the generalised concentration index for the error term: The decomposition framework focuses on two main elements: the impact each determinant has on health outcomes and the degree of unequal distribution of each determinant across income groups: The decomposition method was first introduced to be used with a linear, additively separable model. The child health outcomes in this study are non-linear. The two common choices common choices that yield probabilities in the range (0, 1) are the logit model and the probit model, both of which are fitted by maximum likelihood. One possibility when dealing with a discrete change from 0 to 1 is to use marginal or partial effects (dy/dx), which give the change in predicted probability associated with unit change in an explanatory variable. Therefore, an approximate of the non-linear relationship using marginal effect approximately restores the mechanism of the decomposition framework in Equations 3-5. The linear approximation of the non-linear estimations is given by Equation 5, where μ is the error generated by the linear approximation used to obtain the marginal effects. Data for this study were obtained from a prospective cohort study (henceforth referred to as 'the Good Start study') from three diverse sites in South Africa. The background of this study and the selection of sites are described below. The aim of the Good Start study was to determine the impact of a PMTCT programme on vertical transmission of HIV. The Good Start study had three sites that were purposively selected. These sites represented a variety of settings that exist in South Africa in three respects: area of residence, antenatal HIV prevalence and health systems functioning [35]. Site A is a peri-urban farm area that had an antenatal HIV prevalence of 9% at the start of the study. Site B is a rural area in one of the poorest regions of South Africa, with a poorly resourced health system and which had an antenatal HIV prevalence of 28% at the start of the study. Site C is a peri-urban township area with a moderately well-resourced health system compared to the other two sites and which had an antenatal HIV prevalence of 47%. We followed the enrolled mother-child pairs from delivery to 9 months of age. We measured the HIV status of the infant at 3-4 weeks, 24 weeks and 36 weeks. Of the 891 women initially enrolled in the study, we obtained complete follow-up data for 701 mother-child pairs (78.7%), including 75 (8.5%) child mortality (<9 months) (Figure ​(Figure1).1). Doherty [36] has published full details of the data-collection protocol. Study profile. The left hand side of the figure above indicates all completed interviews at different interviewing periods. The positive and negative sign (+/-) indicates HIV positive and negative women respectively. The right hand side of the figure indicates all missing data which was a result of either mother moving or withdrawal from the study and child mortality. The socio-economic position data was collected at 3 weeks and was not done during the first interview. Therefore, we do not have data on socio-economic position for 97 of mother-child pairs at this stage. For the remaining outcomes of interest (immunisation coverage and HIV transmission), we have data on the socio-economic position for 133 and 113 respondents respectively. Table ​Table11 contains the profile of missing data on the above-mentioned outcomes by socio-economic position. Mother and child pairs, missing data on immunisation coverage and HIV transmission by socioeconomic position. We did not restrict the analysis to children that had complete data for all variables. Instead, it was done separately for each variable. Consequently, the number of children with missing data varies through the results. We carried out the analysis using Stata version 9 [37] and exploratory data analysis using frequency tables. We summarised numerical and categorical data using the appropriate descriptive statistics. We used the Shapiro-Wilk test and histograms to detect departures from normality, and correlation analysis to assess linear associations between numerical variables. We calculated Pearson's product moment correlation coefficient for normally distributed data and Spearman's rank correlation for skewed data. Statistical significance was determined at 5% level. We included the following explanatory variables: the asset index (our chosen measure of socio-economic position), marital status, site, mother's education, income, mother's age and mother's viral load (only in the case of the HIV transmission outcome). Jackson et al. [36] identified maternal viral load as the single most important factor associated with HIV transmission or death. Therefore, we controlled for maternal viral load at 3 and 36 weeks for the HIV outcome.

N/A

• Dima Health