The socioeconomic profile of alcohol-attributable mortality in South Africa: a modelling study

A comparative risk assessment was performed for the adult (15+ years) general population of South Africa in the year 2015, using individual data, aggregate data, and risks relations reported in the literature [16, 17]. A schematic overview over the data sources used in the three main steps of calculation is shown in Fig. 1.

All calculation steps were performed for sociodemographic subgroups defined by SES, age, and sex. SES was measured through an asset score [18‐20], incorporating information on ownership of household assets such as a refrigerator, TV, or car as well as access to water, sanitation, and electricity. Asset scores were calculated using Multiple Correspondence Analysis [21, 22], resulting in three approximately equally sized groups representing high, middle, and low SES. Details on the calculation of asset scores are presented in the Additional file 1. Age was measured in 5-year age groups for all metrics of mortality and 20-year age groups (15–34, 35–54, and > 55 years) for exposure to alcohol and the respective AAFs.

Exposure to alcohol was categorized into lifetime abstinence (never used alcohol), former drinking (no alcohol consumption in the past 12 months), current drinking, and binge drinking (5+ standard drinks per usual occasion). The level of exposure was quantified in average grams of pure alcohol per day among current drinkers. Risk relationships between exposure to alcohol and cause-specific mortality were taken from Rehm et al. [23] (see also [24]). Based on the results of a large representative study on South Africa [25], an interaction effect was modeled for testing HIV positive in a way that, in addition to the main effects of current alcohol use and SES, an interaction variable for alcohol use * SES was added to the equation. This interaction effect was accounted for in the current comparative risk assessment, leading to an increased relative risk for current drinkers of low SES when calculating the AAFs. For more details on the implementation see Probst et al. [26].

In total, 19 cause of death categories were investigated, including communicable diseases, namely HIV/AIDS, lower respiratory infections, and tuberculosis; non-communicable diseases such as ischemic heart disease, hemorrhagic stroke, or diabetes; and all injuries (for details see Additional file 1: Table S1). Causes of death were selected to include all major causes of death for which alcohol use has been shown to be a causal risk factor [23].

Secondary, individual-level data on alcohol consumption in each sociodemographic subgroup came from the National Income Dynamics Study (NIDS) [27]. The survey was conducted from October 2014 to August 2015, with an overall response rate of 86.5%. After removing 4078 observations with missing data on age, sex, SES, alcohol use, or survey weights, the sample comprised 22,741 adults (15+ years) with complete data. Although the survey was designed to be nationally representative, given the household sampling frame, some populations, such as homeless people, were excluded (for details see Additional file 1).

Alcohol use was assessed as usual frequency (ranging from never to daily) and quantity (number of standard drinks per occasion ranging from 1 or 2 to 13 or more). Asset variables and the generation of the asset score using NIDS data are described in Additional file 1.

The aggregated total adult per capita consumption in liters of pure alcohol in 2015 (9.5 L) was obtained from the World Health Organization (WHO) Global Information System on Alcohol and Health [28]. In order to account for possible underreporting in medical epidemiological studies and for alcohol that was produced and sold but not consumed, 80% of the total adult per capita consumption was used [29].

Cohort data containing information on cause of death, age, sex, and SES were obtained from the Demographic Surveillance Area (DSA) of the Africa Centre for Health and Population Studies for the years 2000 to 2014 [30, 31]. The open cohort covering approximately 11,000 households was based on a full assessment of a predominantly rural area in the Umkhanyakude district of KwaZulu-Natal, comprising also periurban informal settlements and a township. After removing 923 participants for which at least one of the relevant variables were missing, the sample comprised 87,029 adults, 757,404 person-years and 23,002 deaths. Each household was visited at least twice a year. The cause of death was established using verbal autopsies [32] and information on household assets and basic services was obtained in face-to-face interviews. The earliest observation with information on household assets was used as baseline asset score. Details on generating the asset score are described in Additional file 1. In order to correct for the predominantly rural population of the study area of the DSA data, the asset score was ‘projected’ onto nationally representative data. This was done by sub-setting the national survey data to rural and periurban areas of Kwazulu-Natal and using the percentiles that corresponded to the tertiles of the national distribution as cut-offs to split the asset score of the DSA data [26].

Aggregated death counts on the national level in 2015 by cause of death, 5-year age group, and sex were obtained from the Institute for Health Metrics and Evaluation [33]. Population counts were based on estimates from the Community Survey 2016 [34].

Step 1: AAFs represent the proportion of deaths that would not have occurred in the hypothetical counterfactual scenario of everyone being a lifetime abstainer. AAFs were calculated by combining the level of exposure to alcohol with the cause-specific mortality risk related to the exposure [29].

Measures of exposure to alcohol by sociodemographic subgroup were calculated using NIDS data. The average grams of pure alcohol per day among current drinkers were calculated using the standard quantity-frequency approach [35]. In order to account for the underestimation of the grams of pure alcohol per day based on survey data [36], information on the relative quantity consumed by each sociodemographic subgroup (based on NIDS data) were combined with the aggregated total adult per capita consumption in liters of pure alcohol (taken from the WHO Global Information System on Alcohol and Health). The standard triangulation technique of the Global Burden of Disease study and WHO Global Status Reports was used [29, 37].

Step 2: AAFs were applied to cause-specific death counts in each subgroup. As the aggregate Institute for Health Metrics and Evaluation death counts were only available by age and sex, hazard ratios (HRs) were used to split death counts in each subgroup by SES. The HRs were derived from DSA data using Cox proportional hazards survival analyses for specific causes of death as well as broader cause categories, adjusting for age at baseline and sex. A sensitivity analysis was performed to examine the stability of the results when using a simple tertile-split asset score for calculating HRs (i.e., without projection onto the national distribution). Details on the HRs are shown in Additional file 1: Table S2.

Step 3: In order to calculate alcohol-attributable mortality rates per 100,000 adult population, the death counts in each subgroup were divided by population counts. Age- and sex-specific population counts were split by SES using the proportions observed in the nationally representative NIDS data. Death rates were age-standardized using the WHO reference population [38].

Uncertainty of AAFs and mortality rates was estimated using Monte Carlo simulations, sampling from the distribution of each lowest level parameter 100,000 times [39]. The 95% uncertainty interval (UI) was determined using the 2.5th and 97.5th percentile of the resulting distribution of the metric. Taking the uncertainty as expressed in the standard error of all lowest level metrics that enter the calculation of the final estimate into account, the true point estimate would fall into this range 95% of the time. This, of course, relies on the assumption that the underlying lowest level metrics follow the assumed (usually normal or binomial) distribution.

The prevalence of current drinking decreased from high to low SES, and persons of lower SES were more likely to be lifetime abstainers (Table 1). However, the grams of pure alcohol consumed among current drinkers decreased from low to high SES and the prevalence of binge drinking was highest in the middle SES group, followed by the lowest SES group, particularly among men.

AAFs for major causes of death are shown in Fig. 2, for details see Additional file 1: Tables S3 and S4. Among men, socioeconomic differences in AAFs were pronounced. Particularly among the middle and older age group (35+), men of middle and low SES had higher AAFs for outcomes such as HIV/AIDS, tuberculosis, injuries, liver cirrhosis, or pancreatitis. For other outcomes, such as cardiovascular diseases, the AAFs were very similar between SES groups. Across SES groups, middle-aged men (35–54) had the highest AAFs. Among women, the socioeconomic pattern was more complex and differences were more pronounced in the youngest age group (15–34). In the latter, women of high SES showed elevated AAFs for some outcomes such as liver cirrhosis, epilepsy, hemorrhagic stroke, and some cancers. However, middle-aged women (35–54) of low SES showed higher AAFs for HIV/AIDS, ischemic stroke, tuberculosis, and pancreatitis.

Overall, approximately 62,300 (95% UI 27,000–103,000) adults died from alcohol-attributable causes of death in South Africa in 2015 (Table 1). With a total of approximately 529,400 deaths from all causes in 2015 [40], roughly one in ten deaths was attributable to alcohol use. Roughly 60% of all alcohol-attributable deaths (~39,300 deaths, 95% UI 13,800–71,200) occurred in the low SES group. About a quarter (~17,000 deaths, 95% UI 6600–31,900) occurred in the middle SES group, and roughly 15% (~9000 deaths, 95% UI 3100–17,000) in the high SES group (Table 1). In the high SES group, deaths due to chronic diseases constituted the largest subcategory of alcohol-attributable deaths (~4200 deaths, 95% UI 900–8800). In the middle SES group, chronic (~6200 deaths, 95% UI 1600–13,100) and infectious diseases (~5900 deaths, 95% UI 1900–11,700) contributed approximately the same number of deaths. In the low SES group, the majority of the alcohol-attributable deaths (~24,800 deaths, 95% UI 7000–46,700) occurred due to infectious diseases. The age distribution of all alcohol-attributable deaths by SES and sex is shown in Fig. 3.

Cause-specific, age-standardized rates of alcohol-attributable mortality (deaths per 100,000 adults) are shown in Fig. 4 (for complete results see Additional file 1: Table S5). Overall, the age-standardized mortality rate from all alcohol-attributable deaths was 727 (95% UI 354–1208) deaths per 100,000 adults in the low, 377 (95% UI 165–687) deaths per 100,000 adults in the middle, and 163 (95% UI 71–289) deaths per 100,000 adults in the high SES group. Thus, persons of low SES had an approximately 4.5-fold alcohol-attributable mortality rate compared to persons of high SES. For persons of middle SES, the alcohol-attributable mortality rate was 1.9 times higher than for persons of high SES. This compares to HRs of all-cause mortality of 2.73 (95% CI 2.30–3.24) for low compared to high SES and 2.07 (95% CI 1.74–2.46) for middle compared to high SES (Additional file 1: Table S2).

The complete analysis was repeated using HRs from the DSA data that were based on an asset score that was not projected onto the national distribution. Overall, the sensitivity analysis confirmed the findings of the main analysis. However, as could be expected, the socioeconomic differences were less pronounced, with a total of approximately 12,500 deaths (95% UI 4400–20,900) in the high, 17,200 deaths (95% UI 8500–27,200) in the middle, and 32,600 deaths (95% UI 13100–55,000) in the low SES group. This corresponded to approximately 20%, 28%, and 52% of all alcohol-attributable deaths occurring in the high, middle, and low SES groups, respectively. Detailed results of the sensitivity analyses are shown in Additional file 1: Table S6 and Figures S2 and S3.

This study applied a rigorous methodology using the most up to date data available to estimate nationally representative alcohol-attributable mortality rates for all major causes of death known to be causally attributable to alcohol use by age, sex, and SES.

A main limitation of the study relates to the assessment of alcohol exposure in each sociodemographic subgroup. Even though binge drinking and heavy alcohol use are known to be highly prevalent in South Africa [51‐53], estimates of over 60 g of pure alcohol per day, on average, are very high. The high levels of consumption among current drinkers could have resulted from a low coverage of alcohol use that has been observed for all major, nationally representative surveys in South Africa in recent years [36]. The triangulation technique, used to estimate the ‘true’ exposure to alcohol, relied on nationally representative estimates of the prevalence of alcohol use in each subgroup and relative levels of alcohol use between subgroups [37]. Consequently, a underestimation of the prevalence of current alcohol use based on the survey data could have led to an overestimation of the levels of alcohol use among current drinkers.

Underreporting and denial of alcohol use due to stigma or memory bias, a high prevalence of heavy and irregular drinking patterns, systematic non-observation of heavy alcohol use due to the sampling frame, and selective non-response were identified as potential causes of the low coverage [54‐58]. Recent research estimated that 93% of the alcohol used in Pretoria, South Africa, was consumed in heavy drinking occasions, a drinking pattern that likely contributed to the low coverage [59]. The latter study also used a much more elaborate assessment of alcohol use, which led to considerably higher levels of consumption among current drinkers compared to the assessment in large nationally representative surveys [53].

As a consequence of the low coverage, AAFs could have been overestimated for some causes of death such as cardiovascular diseases [60], liver cirrhosis [61], or pancreatitis [62], with risk functions that are sensitive to high levels of consumption. At the same time, the potential underestimation of the prevalence of current alcohol use could have led to an underestimation of AAFs for causes of death with flatter risk curves and an elevated risk at low levels of consumption such as cancers [63] or lower respiratory infections [64].

As nationally representative cause of death data including reliable information on the SES of the deceased are not available, HRs were used to split the deaths into the three SES groups. Using the projected asset score relied on the assumption that the population of the DSA was representative of the rural and periurban population in Kwazulu-Natal as assessed in the nationally representative survey.

The HRs used for splitting deaths by SES were not age and sex specific, but rather adjusted for age and sex. However, previous studies did not find systematic sex differences in the socioeconomic gradients of cause-specific mortality that could have led to an over- or underestimation of deaths in a specific sociodemographic group [65].

Uncertainty of all estimates was estimated using a Monte Carlo approach. While this is in line with current standards, the intervals depend on the assumed distributions. An alternative approach for future research could be a systematic analysis of the variation in the resulting point estimates when using plausible extreme values of the lowest level parameters similar to a Latin Hypercube approach [66].

There are a series of effective policy measures to prevent alcohol-attributable harm such as limitations of availability and affordability of alcohol, restrictions of alcohol-marketing, and improvement of the healthcare system [67‐69]. However, such broad brushed measures might fail to address the causes of death found to be most relevant as well as the high risk groups identified in the current study.

When considering alcohol policies in South Africa, it should be taken into account that approximately 23% of the total alcohol consumed in South Africa in 2015 was unrecorded [70]. It is likely that the lion’s share of unrecorded alcohol is consumed by people of lower SES [71]. This means that policies and interventions targeting the consumption of unrecorded alcohol might be more suitable for reducing alcohol-attributable harm in people of low SES than the national alcohol policies listed above. Furthermore, policies could address drinking venues frequented by people of low SES, such as unlicensed alcohol outlets, also called shebeens [72]. Even when selling recorded alcohol, shebeens often operate outside the legal market, and policy measures to restrict availability and hours of sales do not reach unlicensed shebeens or their customers.

There have been attempts to integrate shebeens into the legal market; however, the owners are often not able to afford license fees and related taxes or apply the required changes to adhere to the guidelines or they are situated in areas not zoned for business use [73, 74]. The current political strategy seems to focus on police raids, confiscations of liquor, and closing down of unlicensed outlets [73, 74]. However, this practice fails to acknowledge the economic and social importance shebeens have for owners and customers [72]. Drivdal and Lawhon [74] proposed a plural regulation of shebeens based on a concerted effort of community leaders, shebeen owners, and residents, which could be the first step towards enforcement of closing hours, prevention of sales to minors and intoxicated people, and reduction of violence in and around shebeens [75].

Targeted interventions on HIV transmissions under the influence of alcohol are another approach to prevent alcohol-attributable mortality and related socioeconomic differences. Alcohol-related HIV risk-reduction interventions, targeted at drinkers in under-resourced areas, have been shown to be effective in reducing unprotected intercourse under the influence of alcohol [76]. Alternatively, HIV/AIDS risk reduction counseling and brief interventions could be targeted at drinkers in sexually transmitted infections clinics and more broadly in primary healthcare clinics and trauma units [77].

Apart from interventions that address alcohol use and drinking environments of people of low SES, socioeconomic differences can be addressed on a structural level [4]. Healthy communities require living environments with affordable housing, clean water, sanitation and electricity, infrastructure and public transit, access to education and healthcare, and safe opportunities to spend leisure time [78, 79]. Marmot et al. suggested that “health and health equity might not be the aim of all social and economic policies, but they will be a fundamental result” ([79], p. 1661). Therefore, all policies should be evaluated with respect to their potential effects on health and its respective inequalities.