The unfolding counter-transition in rural South Africa: Mortality and cause of death, 1994-2009

The HIV pandemic has led to dramatic increases and inequalities in adult mortality, and the diffusion of antiretroviral treatment, together with demographic and socioeconomic shifts in sub-Saharan Africa, has further changed mortality patterns. We describe all-cause and cause-specific mortality patterns in rural South Africa, analyzing data from the Agincourt health and socio-demographic surveillance system from 1994 to 2009 for those aged 5 years and older. Mortality increased during that period, particularly after 2002 for ages 30-69. HIV/AIDS and TB deaths increased and recently plateaued at high levels in people under age 60. Noncommunicable disease deaths increased among those under 60, and recently also increased among those over 60. There was an inverse gradient between mortality and household SES, particularly for deaths due to HIV/AIDS and TB and noncommunicable diseases. A smaller and less consistent gradient emerged for deaths due to other communicable diseases. Deaths due to injuries remained an important mortality risk for males but did not vary by SES. Rural South Africa continues to have a high burden of HIV/AIDS and TB mortality while deaths from noncommunicable diseases have increased, and both of these cause-categories show social inequalities in mortality. © 2014 Houle et al. The Agincourt health and socio-demographic surveillance system (HDSS) was reviewed and approved by the Committee for Research on Human Subjects (Medical) of the University of the Witwatersrand (protocol M960720 and M081145). Community consent from civic and traditional leadership was secured at the start of surveillance and is reaffirmed from time to time, while informed verbal consent is obtained at individual and household level at each annual follow-up visit and approved by the ethics committees. A record of participant consent is kept of the household respondent who consented to interview as well as the responsible fieldworker. This information is captured on each household roster (populated census form). We use data from the HDSS to describe the population living in the Agincourt subdistrict of Bushbuckridge, Mpumalanga Province, South Africa [35], [36]. The study area can be considered a border region of rural Southern Africa as well as South Africa: over 30% of the population are Mozambican immigrants, formerly refugees who entered the area following the Mozambican war and share family and kinship ties with the host South African population [35]. The data contain records of some 82,000 people (the exact number changes over time) living in 21 rural villages. Trained fieldworkers collect data by interviewing the most knowledgeable person in each household, including vital events (deaths, births), spatial movements, nuptial events and an array of other information. We use data from 1994 through 2009. Household socioeconomic status has been collected every two years since 2001 using a 34-item asset status survey [37]. The survey covers the living conditions and assets of the household, including factors such as access to water and electricity, and ownership of appliances, transport, and livestock. Each variable is normalized to have range zero to one. An index is created by categorizing assets into five groups: modern assets, livestock assets, power supply, water and sanitation, and dwelling structure. The scaled asset values are summed within each group and then scaled again on [0,1]. Finally, these five asset groups are summed to yield an overall asset score for each household in the range of zero to five. We operationalize household socioeconomic status by taking quintiles of this asset score over all the years in which it was collected. Our cause-specific models use cause-of-death information from verbal autopsy (VA) interviews. For every death recorded in an annual census round, trained lay fieldworkers interview the closest relative of the deceased person, using a standardized, validated instrument and approach [36], [38]. To identify the most likely cause of death, we analyze this information using InterVA-4 (http://www.interva.net), a Bayesian probability model for interpreting VA data [39], [40]. We specify the model to suit a local setting with high prevalence of HIV and very low prevalence of malaria. We have used InterVA-4 rather than a physician-based VA approach because InterVA-4 ensures consistent coding and hence comparability across time, whereas physician coders change periodically. A complementary analysis using physician-based assessments produced similar findings. Indeterminate causes of death were not included in the cause-specific analysis – these individuals were considered censored at the time of their death. An alternate model including these indeterminate causes as a separate cause of death category yielded similar findings. We model mortality using discrete time event history analysis [41] for non-repeating events [42]. Data are organized as person-years with one record for each fully-observed year lived by each person aged 5+ years, including the year they died if there was a death. Values for covariates are set at the beginning of each person year, and the death indicator is set to ‘1’ if there was a death during the year. We use logistic regression to estimate the yearly probability of dying from all possible causes and the relative importance of sex, age, time period, and household SES as predictors of death during a year. We use multinomial logistic regression to estimate the yearly probability of dying by specific cause-of-death categories and include covariates sex, age, time period, and household SES. Interactions between predictor variables are tested using likelihood ratio tests for nested models. Time periods are selected to simultaneously maximize our ability to detect change along the various dimensions included in the models, while acknowledging the fundamental dynamics of the HIV epidemic in the study population. HIV began affecting mortality in the period 1997–8, and consequently we include a break-point in the time periods between 1997 and 1998. For most models we use relatively wide four-year intervals to capture temporal change because this ensures sufficient deaths in each cell. For the SES model we use two-year time intervals because SES is updated on a two-year cycle. We categorize causes of death into four groups based on the Global Burden of Disease Study [43]: (1) HIV/AIDS and TB; (2) other communicable, maternal, perinatal, and nutritional diseases (excluding HIV/AIDS and TB); (3) noncommunicable diseases; and (4) injuries. Group 1 includes HIV and TB because HIV is an underlying cause in most TB deaths and the VA method does not easily distinguish HIV-related from non-HIV-related TB deaths. Group 2 includes among others diarrhea, malaria, and respiratory diseases. Group 3 includes cancer, cardiovascular disease, congenital diseases, diabetes, epilepsy, kidney disease, liver disease, and upper gastrointestinal bleeds. Group 4 involves injuries from accidents (including transport accidents), assault, suicide, and other external causes. SES models are restricted to years 2001–2009. We estimate all models using Stata [44]. We summarize these models using predictive probabilities for discrete sex-age-period groups. Differences over time for each sex-age group are compared relative to the referent (earliest) time period. Full documentation associated with the Agincourt HDSS, as well as an anonymized 10% sample of the full database are available at the Agincourt data website (http://www.agincourt.co.za/). Customized data extraction can be requested from Dr. F. Xavier Gómez-Olivé (az.oc.truocniga@reivaX). Full details of data sharing and collaborations are detailed elsewhere [35].