Growing inequities in maternal health in South Africa: A comparison of serial national household surveys

Background: Rates of maternal mortality and morbidity vary markedly, both between and within countries. Documenting these variations, in a very unequal society like South Africa, provides useful information to direct initiatives to improve services. The study describes inequalities over time in access to maternal health services in South Africa, and identifies differences in maternal health outcomes between population groups and across geographical areas. Methods: Data were analysed from serial population-level household surveys that applied multistage-stratified sampling. Access to maternal health services and health outcomes in 2008 (n = 1121) were compared with those in 2012 (n = 1648). Differences between socio-economic quartiles were quantified using the relative (RII) and slope (SII) index of inequality, based on survey weights. Results: High levels of inequalities were noted in most measures of service access in both 2008 and 2012. Inequalities between socio-economic quartiles worsened over time in antenatal clinic attendance, with overall coverage falling from 97.0 to 90.2 %. Nationally, skilled birth attendance remained about 95 %, with persistent high inequalities (SII = 0.11, RII = 1.12 in 2012). In 2012, having a doctor present at childbirth was higher than in 2008 (34.4 % versus 27.8 %), but inequalities worsened. Countrywide, levels of planned pregnancy declined from 44.6 % in 2008 to 34.7 % in 2012. The RII and SII rose over this period and in 2012, only 22.4 % of the poorest quartile had a planned pregnancy. HIV testing increased substantially by 2012, though remains low in groups with a high HIV prevalence, such as women in rural formal areas, and from Gauteng and Mpumalanga provinces. Marked deficiencies in service access were noted in the Eastern Cape ad North West provinces. Conclusions: Though some population-level improvements occurred in access to services, inequalities generally worsened. Low levels of planned pregnancy, antenatal clinic access and having a doctor present at childbirth among poor women are of most concern. Policy makers should carefully balance efforts to increase service access nationally, against the need for programs targeting underserved populations. This paper is a sub-analysis of the third (2008) and fourth (2012) South African National HIV Prevalence, Incidence, Behaviour and Communication Surveys [9, 10]. The two surveys employed multistage stratified sampling, taking into account province; locality (urban formal, urban informal, rural formal including commercial farms, and rural informal or tribal areas); and race groups. Full details of the survey methods, response rates and ethical procedures are detailed elsewhere [9, 11]. In brief, the sampling frames were based on enumerator areas (EAs) used in the South Africa national census. The primary sampling units consisted of 1000 EAs, which were selected from a database of 86,000 EAs. Fifteen households within each selected EA constituted the secondary sampling units (15,000 households). The same EAs were used in both surveys, but different households were selected. The final sampling unit was made up of eligible individuals within households. Anyone who slept in the household on the night preceding the survey (including visitors) was considered a household member. In 2008, only four persons were eligible to participate from each household; one in each age group (0–1, 2–11, 12–14, and above 15 years) [8]. In the 2012 survey, all persons in the selected households were eligible. Consenting participants responded to individual questionnaires. Dried blood spot specimens were collected from consenting participants, tested for HIV antibodies and linked anonymously with the questionnaires administered to study participants [12]. The analysis includes data collected from two groups of women aged 15–55 years: those who had been pregnant in the preceding 2 years and those interviewed as the parent or guardian of a child below two years. The study variables are described in the publication of the 2008 survey findings [8], and are only overviewed here. Socio-economic quartiles (SEQ) were derived from an asset score based on measures of household-living standards, and were generated using multiple correspondence analysis [13, 14]. Quartiles were preferred over quintiles as the socio-economic differentials are very narrow in many areas of the country, given that many women perform the same income-generation activities and thus have similar incomes and asset levels [8, 15]. Quintiles would thus have been unable to differentiate between women in Q1 and Q2, who have essentially the same living standards. Access to maternal health services was measured by: utilisation of antenatal clinics; HIV testing coverage; and the presence of a skilled birth attendant (SBA) or doctor at birth [8]. Maternal health status was not assessed in detail within the survey, thus proxy indicators were used. Women who said they had a fair or poor health status were categorised as having a lower self-assessed health status, and compared with those reporting good or excellent health. Planned pregnancy, multiparity (five or more children), and prevalence of HIV infection were used as indicators of maternal health status, given their links with pregnancy outcomes for women and children [16]. Though the study focused on the distribution of access and outcomes by SEQ, variation was also assessed across other categories of social differentiation. The applicable categories of the PROGRESS-Plus acronym were used, namely: Place of Residence (province; locality as urban formal and informal, and rural formal and informal), Race, Occupation, Education, Socio-economic Status (SEQ and employment of the mother), and Age and HIV status representing the Plus category [17]. We also examined whether there were systematic differences in access to services between those with and without HIV infection. Data were analysed using Stata version 13.0 (College Station, Texas, United States [18]), taking into account the complex multilevel sampling design (by age, race group and province) and participant non-response. Summary indices for descriptive analysis are weighted percentages, while unweighted counts are provided. Clustering was not accounted for given that the large number of primary sampling units (1000) in the study is comparable to respondent number, diminishing such effects [8]. Socio-economic inequalities in maternal health were calculated using three inequality measures: the Slope Index of Inequality (SII) for quantifying absolute inequalities, and the Relative Index of Inequality (RII) and Concentration Index (CI) for assessing the magnitude of relative inequalities [19–21]. ArcGIS Desktop Version 10.0 was used to show the geographical variation in access to antenatal services and a skilled birth attendant, planned pregnancy and health status. The SII and RII indices are regression based and take the whole wealth distribution into account, rather than only comparing the two most extreme groups (e.g., the wealthiest and poorest quartiles), such as done with a rate difference and rate ratio [20]. The RII and SII are “recommended when making comparisons over time or across populations” [22, 23]. While most trend studies focus on relative, as opposed to absolute inequalities [24], the use of both provides a more complete assessment of patterns of inequalities and changes over time [25, 26]. To derive the SII and RII, each woman in the study population was assigned a notional socio-economic rank score, scaled to take values between 0 (bottom of hierarchy (Q1)) and 1 (top of hierarchy (Q4)) [8]. The rank score equals the midpoint of the range in the cumulative distribution of the population of participants in a given SEQ [24]. For example, if the Q1 women comprise 34.5 % of the population, the women in this category are assigned a rank score of 0.17 (0.345/2), and if the Q2 women comprises 32.5 % of the population, the corresponding rank score is 0.51 (0.345 + [0.325/2]) and so forth. The generated individual data is self-weighted and the only weight applied in the analysis is the survey weight to correct for survey sample design. We then used generalised linear models (GLM) to fit binomial models (Eq. 1) to generate inequality measures, as has been suggested by several authors [24, 27–30]. where Y = 1 if outcome is present and Y = 0 if absent, g(Y) = Y is the identity link function (i.e. binomial regression) generating SII together with the 95 % confidence and, g(Y) = log(Y) is log link function (i.e. log-binomial regression) generating the RII and the 95 % confidence interval, rscore is the notional socio-economic rank score for each woman, β1 and β2 are the regression coefficients. Survey equals 1 for the 2008 survey and 2 for the 2012 survey, and ε is the error term with a binomial distribution. The SII, the β1 under binomial regression, represent the estimated difference in predicted value of the outcome between those at the top (wealthiest) and those at the bottom (poorest) of the social hierarchy. SII is the absolute effect on health outcome of moving from the poorest to the wealthiest group [20, 31]. A positive SII represent inequality in favour of the wealthy, while a negative SII is inequality in favour of poor. The RII, the exponential of the slope, exp(β1) under log-binomial regression, represents the proportionate difference in outcome across the distribution of socio-economic position; or the likelihood of having an outcome, relative to one’s SES level. The RII increases from zero, with higher values indicating higher inequality. An RII above one indicates that the outcome is more prevalent among wealthy women, compared to their poor counterparts. To deal with lack of model convergence, we fitted Poisson regression with robust variance [32]. Poisson regression is suitable when outcomes are not rare, as in this study where most had prevalence greater than 10 %[28]. A decline in both SII and RII is the best evidence of progress in closing the inequality gap [24]. For each outcome, the linear trends of the RII and SII over the five year period 2008–2012 were tested by estimating the p-value for an interaction term between rank score and years since baseline, i.e. 2008 survey coded 1, 2012 coded 2, to account for the different time intervals between surveys (Eq. 2). Positive and significant coefficients, β3 greater than one, for the interaction term indicate widening RII (SII) inequalities over time. The concentration index is defined with reference to the concentration curve, which plots the cumulative proportion of health outcome against the cumulative proportion of the population, ranked by SEQ beginning with the poorest [8, 33]. If health access is equally distributed across SEQ, concentration curves coincide with the diagonal line of equality. The Concentration Index is given by twice the area between the concentration curve and line of equality and ranges from −1 to 1. Zero represents perfect equality, while positive values indicate richer individuals have greater coverage (or good health outcomes) than poorer individuals [33].