Menu
Menu
Menu
Menu
Background Understanding the effects of socioeconomic disparities in health outcomes is important to implement specific preventive actions. We assessed socioeconomic disparities in mortality indicators in a rural South African population over the period 2001–13. Methods We used data from 21 villages of the Agincourt Health and socio-Demographic Surveillance System (HDSS). We calculated the probabilities of death from birth to age 5 years and from age 15 to 60 years, life expectancy at birth, and cause-specific and age-specific mortality by sex (not in children <5 years), time period, and socioeconomic status (household wealth) quintile for HIV/AIDS and tuberculosis, other communicable diseases (excluding HIV/AIDS and tuberculosis) and maternal, perinatal, and nutritional causes, non-communicable diseases, and injury. We also quantified differences with relative risk ratios and relative and slope indices of inequality. Findings Between 2001 and 2013, 10 414 deaths were registered over 1 058 538 person-years of follow-up, meaning the overall crude mortality was 9·8 deaths per 1000 person-years. We found significant socioecomonic status gradients for mortality and life expectancy at birth, with outcomes improving with increasing socioeconomic status. An inverse relation was seen for HIV/AIDS and tuberculosis mortality and socioeconomic status that persisted from 2001 to 2013. Deaths from non-communicable diseases increased over time in both sexes, and injury was an important cause of death in men and boys. Neither of these causes of death, however, showed consistent significant associations with household socioeconomic status. Interpretation The poorest people in the population continue to bear a high burden of HIV/AIDS and tuberculosis mortality, despite free antiretroviral therapy being made available from public health facilities. Associations between socioeconomic status and increasing burden of mortality from non-communicable diseases is likely to become prominent. Integrated strategies are needed to improve access to and uptake of HIV testing, care, and treatment, and management of non-communicable diseases in the poorest populations. Funding Wellcome Trust, South African Medical Research Council, and University of the Witwatersrand, South Africa.
We used data from the ongoing Agincourt Health and socio-Demographic Surveillance System (HDSS), which was established in 1992.28, 29 Agincourt is located in a resource-poor rural setting in Bushbuckridge Municipality in northeast South Africa, close to the Mozambique border. The Agincourt HDSS has generated detailed longitudinal data on births, deaths, and migration and complementary data covering health and socioeconomic indicators. The study area included 21 villages spread over 402 km2 until 2006,22 and was extended to 26 villages in 2007 and to 31 villages from 2010 to 2012.17 Most people speak Shangaan. About a third of the population is made up of immigrants from Mozambique, who arrived in the area in the early to mid-1980s, and their descendants. Data have been collected annually since 1999. Detailed documentation describing the Agincourt HDSS data and an anonymised database containing data from 10% of the surveillance households are available for public access. The Agincourt HDSS core demographic data are also routinely deposited for public access in the INDEPTH Network Data Repository. In this study we have used only data from the original 21 villages to maximise the duration of follow-up at the village level. These customised data are available on request to interested researchers. Ethics approval was obtained from the Human Research Ethics Committee (Medical) of the University of the Witwatersrand, Johannesburg, South Africa, for surveillance activities in the Agincourt HDSS (protocols M960720 and M110138) and for the analyses reported in this study (protocol M120488). Informed verbal consent was obtained at every surveillance visit from the head of the household or another eligible adult in the household. The person giving consent was noted in the household roster, and the details and date of the process were recorded by the responsible fieldworker. For every death recorded from 2001 to 2013, we used the InterVA-4 probabilistic model (version 4.03) to assign the most probable cause, rather than the more traditional, clinically oriented underlying cause. This model enables a standardised, automated assignment of cause of death that is much quicker and more consistent than physician assessment, and is particularly useful for assessing changes over time and across settings. It assigns each death to a maximum of three likely causes, with associated likelihoods based on information about signs and symptoms of illness or injury collected through verbal autopsy interviews.30 In the annual surveillance updates of the Agincourt HDSS, caregivers of individuals who had died since the previous visit were interviewed with a questionnaire in Shangaan that had been locally validated.29, 31 Thus, timing of the interviews ranged from 1 to 11 months after death. The cause of death was categorised as indeterminate when inadequate information was obtained for the model to assign a cause of death. The causes of death generated by the InterVA-4 model are based on the WHO 2012 verbal autopsy standards and correspond to the International Classification of Diseases, tenth edition.30 We categorised the most probable causes of death into five broad groups: HIV/AIDS and tuberculosis; other communicable diseases (excluding HIV/AIDS and tuberculosis) and maternal, perinatal, and nutritional causes; non-communicable diseases; injury; and indeterminate. The first four categories are consistent with the burden of disease classification system used in South Africa.27 We combined HIV/AIDS and tuberculosis because HIV is an underlying cause in most tuberculosis deaths and distinguishing those that are HIV related from those that are not is difficult with the verbal autopsy method.23 We measured socioeconomic status with an absolute household wealth index computed from a list of household asset indicators that were grouped in the following categories: construction materials in the main dwelling; type of toilet facilities and sources of water; sources of energy; ownership of modern assets; and livestock.17, 23, 32 For each household, after categorisation, asset indicators were assigned weights, with higher values corresponding to higher socioeconomic status. The value assigned to each item was divided by the highest value for all households to obtain normalised values that fell in the range of 0–1. The normalised values within each category were summed to obtain category-specific values, normalised by the same method, then summed to produce an overall household wealth index value that fell in the range 0–5. Once constructed, the wealth index was divided into household wealth quintiles, in which the first quintile represented the poorest households and the fifth the richest households. Data on household asset indicators were collected in 2001, 2003, 2005, 2007, 2009, 2011, and 2013. For each individual we organised data into a person-year file that contained one record for each full year lived, similar to the methods of Houle and colleagues23, 33, 34 and Kabudula and colleagues.26 We included only records for completely observed person-years plus the year in which the individual died irrespective of whether the person-year was complete. Covariates recorded were sex, age (<5, 5–14, 15–49, 50–64, or ≥65 years), time period (2001–03, 2004–07, 2008–10, and 2011–13), date of death, likely cause of death, and household wealth quintile. For covariates that change over time, such as age and household wealth quintile, we used the value at the beginning of the relevant person-year. For completed person-years the death indicator was set to 0, and it was set to 1 in records where there was a death during the year. Time periods were split across years to contextualise the dynamics of the HIV/AIDS epidemic and the roll-out of services for prevention of mother-to-child transmission and ART. We used the person-year file to calculate the probabilities of death from birth to age 5 years and from age 15 to 60 years, life expectancy at birth, and age-specific and cause-specific mortality by sex (excluding children <5 years), time period, and household wealth quintile. Thereafter, we estimated relative and absolute socioeconomic differences in the mortality indicators with the relative index of inequality (RII) and the slope index of inequality (SII), respectively (appendix).35 These measures take into account the whole socioeconomic distribution and the effects on mortality indicators of a person moving from the lowest to the highest quintile.35, 36 RII=1 and SII=0 imply no difference between the lower and higher ends of the socioeconomic continuum. RII values greater than 1 and positive SII values imply greater mortality at the lower end of the continuum, and RII values less than 1 and negative SII values imply greater mortality at the higher end. We fitted separate models for each time period and sex (except for children <5 years) to calculate RIIs and SIIs for mortality in children and adults and life expectancy at birth, with the modified ridit score (appendix)37, 38 as the independent variable. To calculate RIIs and SIIs for cause-specific mortality, we fitted separate models for each cause-of-death category, time period, and sex, with the modified ridit score and age group as independent variables. We also fitted models with two-way interaction terms between the modified ridit score and time period to assess trends in socioeconomic differences in the mortality indicators over time. We also calculated relative risk ratios and 95% CIs to investigate associations between relative inequalities and household wealth quintile, which we obtained from multinomial logistic regression models,39, 40, 41, 42, 43 with cause of death as an indicator of mortality used as the dependent variable and household wealth quintiles, sex, age group, and time period as independent variables. Although socioeconomic status can be measured at the individual level with factors such as education and occupation,44 samples are necessarily restricted to people who have reached a certain age to make the indicators meaningful (eg, age beyond which individuals are unlikely to advance their eduction further or working age). Instead, we used unadjusted household socioeconomic status to maximise the sample size because these data are collected more frequently than individual data and because all individuals in the household are affected by the household environment. Household socioeconomic status provides a good cumulative indicator of material living standards,44, 45 which strongly affect individual household members. Data on household asset indicators used for calculating the household wealth index were collected in alternate years from 2001 onwards and, therefore, we used multiple imputation to minimise the loss of data due to missing values. We used partial mean matching (based on the nearest two neighbours) to generate five imputed datasets and derive parameter estimates and SEs by averaging across the imputations and adjusting for variance. As done by Houle and colleagues,33 the imputations are generated from a household-year data set that includes counts of men, boys, women, and girls, Mozambicans and South Africans, individuals aged younger than 20 years, 20–59 years, and 60 years and older, and 1–2-year lags of household wealth index. We did all analyses with Stata version 14.1. Estimates with p values less than 0·05 were taken to be significant. The funders had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.
