Progression of the epidemiological transition in a rural South African setting: findings from population surveillance in Agincourt, 1993-2013

Background: Virtually all low- and middle-income countries are undergoing an epidemiological transition whose progression is more varied than experienced in high-income countries. Observed changes in mortality and disease patterns reveal that the transition in most low- and middle-income countries is characterized by reversals, partial changes and the simultaneous occurrence of different types of diseases of varying magnitude. Localized characterization of this shifting burden, frequently lacking, is essential to guide decentralised health and social systems on the effective targeting of limited resources. Based on a rigorous compilation of mortality data over two decades, this paper provides a comprehensive assessment of the epidemiological transition in a rural South African population. Methods: We estimate overall and cause-specific hazards of death as functions of sex, age and time period from mortality data from the Agincourt Health and socio-Demographic Surveillance System and conduct statistical tests of changes and differentials to assess the progression of the epidemiological transition over the period 1993-2013. Results: From the early 1990s until 2007 the population experienced a reversal in its epidemiological transition, driven mostly by increased HIV/AIDS and TB related mortality. In recent years, the transition is following a positive trajectory as a result of declining HIV/AIDS and TB related mortality. However, in most age groups the cause of death distribution is yet to reach the levels it occupied in the early 1990s. The transition is also characterized by persistent gender differences with more rapid positive progression in females than males. Conclusions: This typical rural South African population is experiencing a protracted epidemiological transition. The intersection and interaction of HIV/AIDS and antiretroviral treatment, non-communicable disease risk factors and complex social and behavioral changes will impact on continued progress in reducing preventable mortality and improving health across the life course. Integrated healthcare planning and program delivery is required to improve access and adherence for HIV and non-communicable disease treatment. These findings from a local, rural setting over an extended period contribute to the evidence needed to inform further refinement and advancement of epidemiological transition theory. We use mortality and cause of death data collected from 1993 to 2013 as part of annual updates of vital events conducted using the Agincourt HDSS in a population occupying 27 villages in rural northeast South Africa [32, 33]. The population is largely Shangaan (Tsonga)-speaking. Former Mozambican refugees, who arrived in the study area in the early to mid-1980s in the course and aftermath of civil war, and their descendants, make up about 30% of the population. The population has been under epidemiological and demographic surveillance since 1992 and vital events were updated at approximately 15- to 18-month intervals between 1993 and 1999, and annually since 1999. Although the population has limited access to infrastructure and public sector services, it has experienced substantial socioeconomic changes over the years. As documented in our earlier study [34], the proportion of households that own assets associated with greater modern wealth has increased substantially over time. For example, the proportion of households with dwellings constructed with either brick or cement walls increased from 76% in 2001 to 98% in 2013; and the prevalence of tiles as roof and floor materials of dwellings increased respectively from 3% and 0.5% in 2001 to 15% and 14% in 2013. In addition, the use of electricity for lighting and cooking respectively increased from 69% and 4% of households in 2001 to 96% and 50% of households in 2013. Other notable increases are in the proportion of households owning stove, fridge, cellphone and car respectively from 41%, 40%, 37% and 14% in 2001 to 85%, 86%, 98% and 20% in 2013. For individuals identified as having died between the annual surveillance update rounds, verbal autopsy (VA) interviews were conducted with their caregivers to elicit signs and symptoms of the illness or injury prior to their death. The interviews were conducted one to 11 months after death using a locally validated, local-language VA instrument [33, 35]. Given the rigorous processes involved in the collection, quality assurance and processing of HDSS data [14, 36], the data from the Agincourt HDSS population is one of the rare high-quality and methodologically consistent longitudinal health and demographic dataset for populations in resource-poor low- and middle-income settings. The available mortality and cause of death information by age and sex over an extended period provides a unique opportunity for assessing how populations in low- and middle-income settings, including those in rural sub-Saharan Africa are currently experiencing the epidemiological transition. We use the InterVA-4 probabilistic model (version 4.03) to assign probable causes of death to every death with a complete VA interview. For each death, the InterVA-4 model assigns up to three likely causes of death with associated likelihoods [37]. An indeterminate cause of death is assigned when the VA information is inadequate for the model to arrive at any cause of death. We opted for InterVA-4 as opposed to physician-coded causes of death because the InterVA-4 model assigns causes of death in a standardized, automated manner that is much quicker and more consistent than the former (particularly for assessing changes over time and across settings). Additionally, causes of death derived from InterVA-4 have been found to not substantially differ from those generated by physician coding [38]. Similar to some earlier studies [28, 39], we use discrete-time event history analysis (DTEH) [40] to estimate overall and cause-specific annual hazards of death as functions of sex, age and time period. The annual hazard of dying is the probability of dying during a one-year interval starting on a particular date experienced by living individuals, conditional on their state at the beginning of the interval. An individual’s continuously evolving state is described by the combination of values taken by both constant and time-varying variables, for this study, sex, age and time period. One of the basic requirements of DTEH is the splitting of each individual’s survival history into a set of discrete person years [40]. We create a person-year file that contains one record for each full year lived by each individual in the study population. For example, individuals who died after one year of surveillance contribute one person-year each while those who died after five years of surveillance contribute five person-years. Only completely observed person-years are included in the data set except when an individual dies before completing a person-year time unit. Survival histories are truncated for individuals who were alive at the beginning or end of the study and for those who migrated in/out during the study. After constructing the person-year file we estimate the annual hazards of dying using logistic regression models [40–44]. Binary logistic regression models are used for estimates of the risk of dying from all possible causes, and multinomial logistic regression models are used to obtain estimates of the risk of dying from causes in broad cause of death categories. Using the estimated annual hazards of death, we construct standard life tables to derive life expectancies at birth and adult mortality rates (the probability of dying between ages 15 and 60 for those who survive to age 15 if subjected to age-specific mortality rates between those ages for the specified calendar year). In order to contextualize the dynamics of the HIV epidemic and the availability of antiretroviral treatment over time, the years of the study are divided into the following time periods: 1993–1997, 1998–2000, 2001–2003, 2004–2007, 2008–2010 and 2011–2013. We also categorize age into the following commonly used age groups: 0–4, 5–14, 15–49, 50–64 and 65+. For the cause-specific analyses, the most likely causes of death generated by the InterVA-4 model except indeterminate are categorized into four broad groups: (1) HIV/AIDS and TB; (2) other communicable, maternal, perinatal, and nutritional diseases (excluding HIV/AIDS and TB); (3) non-communicable diseases; and (4) injuries, consistent with the burden of disease classification system in South Africa [23]. Following a common, standard approach to analyzing changes in mortality and cause of death patterns, we divide the most likely causes of death generated by the InterVA-4 model into three broad cause groups that can be compared with existing publications: Group I (communicable diseases, maternal, and perinatal conditions and nutritional deficiencies), Group II (non- communicable diseases), and Group III (accidents and injuries) [45, 46]. The proportion of deaths attributed to each cause group ranges from 0 to 1 and the set of proportions for all of the cause groups sums to 1 after excluding indeterminate causes. We follow Salomon and Murray [46] to relate the distribution of deaths among cause groups to the overall level of mortality. We fit estimates of age and cause-specific mortality fractions to a set of regression equations of the following form where i indexes age; Y i1 and Y i2 are the log ratios of the cause-specific fractions for Group II causes (P2) and Group III causes (P3) relative to the cause-specific fraction for Group I causes (P1): Yi1=lnP2P1 and Yi2=lnP3P1; M i is the all-cause mortality rate; β 0 and γ 0 are constant terms and ε i1 and ε i2 are error terms. The coefficients are estimated using seemingly unrelated regression models, separately for each sex and age group. These models provide efficient means of jointly obtaining estimates from a set of equations each with its own error term that may be correlated with the error terms of other equations. As in Salomon and Murray [46] we compute predicted values for Y1 and Y2 for each observation in the dataset. Those predicted values are transformed into predicted proportions for each cause group using the multivariate logistic transformation: where J = 3 and P 3 is 1 − P 1−P 2. All analyses have been conducted using Stata version 14.1 (Stata Corp., College Station, USA).