Effects of minimum unit pricing for alcohol in South Africa across different drinker groups and wealth quintiles: a modelling study

Objectives To quantify the potential impact of minimum unit pricing (MUP) for alcohol on alcohol consumption, spending and health in South Africa. We provide these estimates disaggregated by different drinker groups and wealth quintiles. Design We developed an epidemiological policy appraisal model to estimate the effects of MUP across sex, drinker groups (moderate, occasional binge, heavy) and wealth quintiles. Stakeholder interviews and workshops informed model development and ensured policy relevance. Setting South African drinking population aged 15+. Participants The population (aged 15+) of South Africa in 2018 stratified by drinking group and wealth quintiles, with a model time horizon of 20 years. Main outcome measures Change in standard drinks (SDs) (12 g of ethanol) consumed, weekly spend on alcohol, annual number of cases and deaths for five alcohol-related health conditions (HIV, intentional injury, road injury, liver cirrhosis and breast cancer), reported by drinker groups and wealth quintile. Results We estimate an MUP of R10 per SD would lead to an immediate reduction in consumption of 4.40% (−0.93 SD/week) and an increase in spend of 18.09%. The absolute reduction is greatest for heavy drinkers (−1.48 SD/week), followed by occasional binge drinkers (−0.41 SD/week) and moderate drinkers (−0.40 SD/week). Over 20 years, we estimate 20 585 fewer deaths and 9 00 332 cases averted across the five health-modelled harms. Poorer drinkers would see greater impacts from the policy (consumption: −7.75% in the poorest quintile, −3.19% in richest quintile). Among the heavy drinkers, 85% of the cases averted and 86% of the lives saved accrue to the bottom three wealth quintiles. Conclusions We estimate that MUP would reduce alcohol consumption in South Africa, improving health outcomes while raising retail and tax revenue. Consumption and harm reductions would be greater in poorer groups.


Supplementary Material
Price to consumption Our model starts by estimating mean and peak alcohol consumption at current alcohol prices at the individual level. The proportion of alcohol consumption which is homebrew is also estimated. This process utilised both alcohol frequency questions and seven day recall questions asked in the same survey. As survey data significantly underreports consumption we calibrate these estimates to market research data using statistical methods established in the literature [1][2][3] . Following the shift of mean consumption, peak consumption is re-estimated using a simple regression model created at baseline. We categorise drinkers into three exhaustive and mutually exclusive groups; moderate (less than 15 standard drinks per week); occasional binge (less than 15 drinks per week but more than 5 on one occasion); and heavy (15 or more drinks per week). A standard drink in South Africa is currently 15ml or 12 grams of pure ethanol. We compute a regression model for wealth quintiles using the South African Demographic and Health Survey (SADHS) data and use it to predict wealth quintiles in the International Alcohol Control (IAC) dataset to generate price distributions for wealth and drinker groups. Alcohol is treated as one commodity due to data constraints.

Estimating baseline consumption using South African Demographic and Health Survey (SADHS)
The SADHS survey asked the following questions:

SADHS 2016
Have you ever consumed a drink that contains alcohol such as beer, wine, ciders, spirits, or sorghum beer? Probe: Even one drink? [yes, no] Was this within the last 12 months? [yes, no] In the last 12 months, how frequently have you had at least one drink? [5 or more days a week, 1-4 days per week, 1-3 days a month, less often than once a month] During each of the last 7 days, how many standard drinks did you have? [use showcard, record total number of drinks consumed each day starting with the day before the day of the interview and proceeding backwards] During the last 7 days, how many standard home-made beers or other homemade alcohol did you have? [use showcard, record number] In the past 30 days, have you consumed five or more standard drinks on at least one occasion? [yes, no] BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s)

Process of adjusting the SADHS estimates
Drinkers were categorised by their drinking frequency and by whether or not they had reported any drinking in the last seven days. Readjusting those with a seven day drinking pattern (pink numbers) The pink numbers are respondents who say they only drink 1 -3 days per month or less often than once a month but have drunk in the last 7 days. If this were multiplied by 52 it would be an overestimate. Therefore, we assumed for those that drink 1-3 days per month we have captured their one drinking week in the month and multiply by 12 to get their annual consumption. There are 799 people in this category. We assumed for those who drink less often than once a month but who did drink in the last week we have caught their one drinking week that occurs every two months. We multiplied by six, to get the annual figure. There are 404 people in this category. The yellow numbers do not require adjustment as respondents report drinking every week and have a seven day drinking pattern.
Readjusting those without a seven day drinking pattern but who say they drink (blue and red numbers) For those with a drink frequency of five or more days per week we used the mean standard drinks for drinkers who reported the same frequency but who do have a seven day pattern, there are 27 people that this applies to (blue).
For those with a drink frequency of 1 -4 days per week we used the mean standard drinks for drinkers who report the same frequency but who do have a seven day pattern, there are 103 people in this group (blue).
For those with a drink frequency of 1 -3 days per month we used the mean adjusted annual drinks (adjusted in 2.2.1.3.1) of the equivalent frequency group who did report a drinking pattern. There are 364 drinkers in this group (red).
All of the above estimates were computed for sex and binge drinking subgroups.
For those with a drink frequency of less than once per month we used the mean adjusted annual drinks (adjusted in 2.2.1.3.1) of the equivalent frequency group who did report a drinking pattern. This is computed for subgroups based on sex and binge drinking. There are 783 people in this group (red)

Process of adjusting peak drinks
Using the same process as above we applied a peak drink to those observations without one. As an additional check we validated that all those reporting binge drinking had a peak drink at minimum of 5.
Comparing the adjusted SADHS data with the estimates using only 7 day recall as expected prevalence of drinking increases and per capita estimates reduce ( Table 6). The prevalence estimates are now broadly similar to the NiDs and GISAH estimates (Table 3).   Incorporating the frequency data into the seven day recall moves the distribution towards the left (Figures 1 and  2). This is logical as the sample will now include those drinkers who stated that they drink but did not record any for the last seven days, it also adjusted down those who claim to drink less than weekly but who did recall drinks for the last seven days. This pattern gives some confidence in the dataset and utilises the strengths of capturing heavy drinking well and including occasional drinkers.

Uplifting consumption
Surveys provide important data about drinking patterns within the population but total consumption estimates are far smaller than that indicated by administrative sources 4 . As this is a global phenomenon there are established statistical calibration methods in the academic literature. The steps are broadly as follows:  compute the ratio between survey and sales per capita consumption (known as coverage)  use this ratio to adjust the mean for each subpopulation of interest  use the new mean to estimate an associated standard deviation based on a published relationship, estimated using regression on a large global dataset 2 = 1.174 × ̂ + 1.003 ×  use the new mean and standard deviation to generate the shape and rate parameters and fit a gamma distribution This method relies on three assumptions. Firstly that the sales data accurately reflects per capita consumption. Secondly, that the true proportion of abstainers has been captured by the survey, and finally that under-estimation of consumption is the same across all population groups.
Two additional key limitations have been identified with regards to this method. Firstly, there is no empirical evidence that under-coverage is distributed as implied by the shifts needed to fit the adjusted consumption to the gamma. Secondly, that shifting consumption to a gamma can artificially reduce the long tail of heavy drinkers 3 . To address the second point a proposed method is to fit a gamma distribution to the survey and for each percentile of the distribution calculate the percentage consumption increase and apply these percentage shifts to the corresponding percentile of the survey data.
The following steps outline, in detail, how we calibrated the SADHS dataset to Euromonitor figures:  First a cap was applied to all drinkers of 68 litres of alcohol per year or 150 grams of alcohol per day. As the model includes long term effects (20 years) the cap is needed as a higher level of alcohol cannot be sustained in the long term 5 . This cap impacted one woman and ten men. Of this small group only two men drunk both homebrew and recorded alcohol and so their total consumption was reduced to 68 litres and then split into recorded and homebrew using their previous percentage split.
 Survey coverage level was calculated as the difference between total per capita consumption recorded in the SADHS survey and per capita consumption using Euromonitor recorded sales data for 2018. 80% of the sales data is used to account for spillage, stockpiling and tourist consumption. This sales figure was then increased to take account of the 4.15% of total alcohol consumed in the SADHS survey reported as homebrew (representing unrecorded alcohol in the model). The comparison of total consumption according to the survey and the adjusted official sales data was used to calculate a coverage of 27%.
 For female and male subgroups the mean litres of alcohol was adjusted by the multiplication factor. This adjusted mean was used to estimate an associated standard deviation based on a previously established relationship between the two. These were then used to fit a "shifted" gamma distribution (maintaining the cap of 68 litres), calculated for male and females separately.
 A gamma distribution was fitted to the original sample of drinkers, by sex, and percentiles were taken across this and the shifted distribution. Percentage differences in consumption were calculated. These increases were then applied to the percentiles of the original survey sample.
 Each individual's total consumption was split into homebrew and recorded alcohol using the original percentage split (this assumes underreporting is equal across homebrew and recorded alcohol).
 Results were compared visually and via a table (Table 7 and Figures 8 and 9). There is a small difference between the two methods, more visible for males than females. It appears adjusting by percentiles only makes a difference at the extremes, lowering the left hand peak slightly but also falling below the Gamma shifted distribution after 60 litres of alcohol per year for men meaning there is a smaller number of the very high drinkers.
The percentile adjusted distribution was used for the main model base on expert opinion.

Uplifting peak consumption
Peak drinking measures the highest number of drinks consumed on a single drinking occasion and therefore relates to intoxication which is associated with harms such as road injury, interpersonal violence and self-harm. Following the method used in the Sheffield Alcohol Policy Model 6 , the following linear regression model was fitted, for all drinkers, to the non-shifted SADHS data, relating peak drinking to mean consumption, age and sex.
The model was used to compute fitted values for the non-shifted data. The model assumes there is a linear relationship between peak and mean consumption, the magnitude of which is allowed to vary by age and sex.
After the mean consumption was shifted as above the corresponding new peak consumption was computed using the following formula: The linear relationship between mean and peak estimated from the SADHS survey is maintained for the shifted mean and peak consumption, this assumes individuals under reported peak and mean consumption by the same magnitude. The method also assumes the prediction error for the model is of the same magnitude for all levels of consumption.
The predictions were checked to ensure that peak estimates were not below mean daily drinking. There were 88 people (out of the 3311 drinkers) for whom this was true. These people had their peak drinking increased to match their mean daily drinking.

Wealth quintiles
In order to match wealth groups between the two datasets an ordered choice model was created using SADHS data with wealth quintile (1 -5) as the dependent variable, using the MASS package in R 7 . Wealth groups were chosen as the best available measure to capture socioeconomic status that allowed us to match between the SADHS and IAC dataset. Although income was asked in the IAC dataset many of the respondents refused to answer resulting in a very small sample.
All the variables that were common across the two datasets were included in the initial model, these were not just asset ownership but also age, sex, educational level and population group (race). Stepwise regression was performed using the step.AIC function. This chooses the best variables to include by running the regression with all variables in and then taking one out and computing a goodness of fit measure (the AIC). If the goodness of fit measure is improved then that model is preferred, it runs this for many models until it finds the model with the highest AIC. This method resulted in the selection of the following variables: age, sex, population group, education level, car, landline, electricity, fridge, computer, radio, tv. The only variable it removed was mobile phone which fitted anecdotally with conversations we had with stakeholders in South Africa regarding how much poorer people prioritise mobile phones.
The goodness of fit matrix evaluates the success of the model, comparing the closeness of the predicted and observed outcome ( Table 5). The model never predicts the poorest as the richest or the richest as the poorest.

International Alcohol Control Study 2014 for prices
The IAC dataset provides prices by drinking location by beverage, by container size and also asks whether the individual binge drinks, demographic data is also collected. The survey asked for the price in Rands by location, for example they ask for the price of a beer paid at a pub for each container size. There are 17 drinking locations (12 on trade and 5 off trade) and 12 drink types. On-trade is where the alcohol is consumed on the premises it is purchased (e.g. hotels, restaurants, pubs), off-trade is where the alcohol is consumed off the premises it was purchased at (e.g. supermarket or bottle store).
Prices were disaggregated by population subgroups rather than by drink type (wine/beer/spirits etc). This was consistent with the South Africa specific price elasticities which were calculated for drinker groups whilst treating alcohol as a single commodity. The IAC respondents were categorised into drinker groups using the definitions above. Each price was weighted by the number of units (e.g. bottles, glasses, cans) sold, the container size of those units and the number of drinking occasions in 6 months ( Figure 5). Every price observation was validated using data from the South African Consumer Price Index. Prices were increased to 2018 to account for inflation. On-trade is where the alcohol is consumed on the premises it is purchased (e.g. hotels, restaurants, pubs), off-trade is where the alcohol is consumed off the premises it was purchased at (e.g. supermarket or bottle store). The off-trade wine prices were adjusted using data from the South Africa Wine Industry Statistics 8 who report the proportions of still wine sold (which makes up 93% of total volume of wine sold) in the off-trade in 2018 that falls within different price bands, this data was used to adjust downwards the off-trade wine ( Table 6). The price observations were sorted in ascending order and a cumulative volume variable created. The price closest to the 49th percentile was then adjusted down to R3.74 and all prices below adjusted using the same proportion. The prices at the very bottom were adjusted so they could not go below R2.50. The same adjustment process was applied to each of the four groups. As the Tschwane prices were collected in one locality, they were validated against national data sources. Beer is by far the most popular drink, accounting for over 50% of the alcohol sold so beer prices are critical. We accessed data from the South Africa Consumer Price Index for January 2020 to compare the Gauteng province (where Tshwane is located) with other provinces. Beer, which accounts for over 50% of alcohol sold in South Africa, Guateng is at R13.76 for a 330ml can. The average across the eight prices listed above is R13.66 which is very close to Guateng's price, therefore we assume the same price distributions across the whole of South Africa.
Finally, prices were validated with all stakeholders including individuals resident in townships who could provide anecdotal evidence relating to cheap alcohol available at shebeens.

Base prices by subgroup
All IAC drinkers were now categorised by drinker type and by wealth quintile (Table 7). Wealth quintile was predicted using the ordered choice model created using the SADHS data. Drinkers in the lowest wealth quintile appear the least likely to drink in moderation leaving a very small sample size (this is not weighted by number of drinks). It is therefore not possible to create price distributions for all 15 categories. The mean price for each of these drinker categories demonstrates there is wealth gradient (Table 8). In order to ensure adequate sample size the poorest/poorer/middle and richer/richest categories were aggregated for moderate drinkers (Table 9). This represents the final group of prices used in the model.

Adjusting the elasticities
The starting point for elasticities -0.4, -0.22 and -0.18 for moderate, occasional binge and heavy drinkers respectively 9 . We adjusted these elasticities to incorporate an income gradient using -0.86 and -0.5 elasticity for low and high socioeconomic status 10 . To remain on the conservative side we will count the bottom two quintiles as low SES and the top three as high.

Individual spend, tax and retail revenue
Alcohol consumption expenditure The total retail spend at baseline, and each scenario, was computed by adding up all the individual spends multiplied by their population weights. When the SADHS consumption estimates were shifted to calibrate to market research data only 80% of the consumption figure was used to take account of spillage, stockpiling and tourism, but the 20% of alcohol remains in the headline sales revenue. Therefore to make it comparable we estimate the total sales revenue by increasing the modelled alcohol consumption revenue by 1.25 (100/80).
Government revenue, VAT, excise tax and retail revenue The following steps outline how we computed government and retail revenue: 1. Calculate VAT by assuming 15% of the base retail spend is VAT 2. Import 2018 base excise tax from Treasury Budget Report 11 3. Calculate total volume consumed of alcohol at all four scenarios (baseline/R5/R10/R15) 4. Calculate the percentage change in volume from baseline for each of the three policies 5. Apply the percentage change in volume to base excise tax (we assume a fixed ratio between volume and excise tax) 6. Calculate retail revenue by: spend -vat -excise tax It is likely this is a conservative approach to modelling excise tax revenue as generally the cheaper alcohol, which this policy targets, generates a lower proportion of excise tax than the more expensive, so we can consider this a lower band on the excise tax revenue.

. Relative risks
Relative risks were calculated for each of the health outcomes of interest at baseline, and each policy scenario using published relative risk equations 12,13 . The same relative risk equations are used for morbidity (or prevalence) and mortality. HIV risk is derived from a stepped function for mean drinking differing by socioeconomic status, intentional injuries and road injury from a continuous function of mean drinking differing by whether the individual binge drinks, liver cirrhosis and breast cancer from a continuous function of mean drinking, for breast cancer this is only for females (Table 10).

Potential impact fractions
Potential impact fractions (PIFs) were calculated by dividing relative risk under each policy by relative risk at baseline. These incorporated population weights and were computed by sex (i), wealth group (j) and drinker group (k).

Socioeconomic gradients of ill health
Health outcomes in South Africa are not evenly distributed throughout the population, with the poor often bearing a higher burden of disease, depending on the illness. Data analysis was carried out using General Household Survey (GHS) data for 2018. The ordered choice regression model computed previously, using SADHS data, was applied to the GHS data to split the survey population into wealth quintiles compatible with the foundational dataset (SADHS). Percentage within each wealth quintile with the disease was computed (Table 11). Liver cirrhosis was not one of the health conditions included in the survey and breast cancer was not specifically included although the broader category of cancer was. Sensitivity analysis was carried out using alternative gradients.

Distributing baseline deaths and cases and calculating probabilities
The deaths/cases (which come disaggregated by sex) at baseline is split between the five wealth quintiles using the GHS data to account for the socioeconomic gradient, as explained above. However, a preparatory step was necessary as the proportions of the population (using the SADHS proportions) in each quintile were not perfectly equal, for example for Q1, Q2, Q3, Q4, Q5 corresponded to 0.19, 0.19, 0.20, 0.21, 0.21 for females and 0.19, 0.20, 0.21, 0.20, 0.21 for males. The probability of death was calculated for each quintile first by assuming the population was split into quintiles of equal size. The total deaths/cases for each quintile using the SADHS proportions was then calculated by applying the relevant probability of death/cases for that part of the quintile which overlapped with the underlying equally sized quintile. This concept can be best illustrated on a graph.
The existence of relative risk equations implies that the baseline mortality/morbidity will also not be distributed equally between drinker groups, one would expect a higher proportion of the baseline cases to exist amongst heavy drinkers, followed by occasional binge, moderate then abstainers. In order for the baseline mortality/morbidity to vary by drinker group the total risk, for each disease, is calculated for each drinker group group, by sex and wealth quintile. The proportional share of risk between drinker groups is then calculated and used to distribute the mortality/morbidity, which has already been assigned to each quintile, between each drinker group within that quintile.
The model uses iHME data for deaths and cases of disease and population statistics (Statistics South Africa) from 2018. Life tables to get the probability of death by single year of age were only available for 2017 from iHME so these were used. The 2018 population is split proportionally into the sex/wealth/drinker groups using the SADHS proportions. The probability of death for each disease is calculated for the baseline scenario and taken away from overall probability of death for each single year of age given in the life table to give a probability of death from non-modelled causes. This probability of death from non-modelled causes remains constant at every policy scenario. The probability of death from the five diseases of interest then vary according to the policy level and the corresponding potential impact fraction.
We model counterfactual population structure (i.e. in the absence of the policy) over 20 years, starting from 2018 using current population estimates from Statistics South Africa, plus birth projections for 2020 to 2023 and assume current age-, sex-and wealth-specific mortality rates remain constant 14 . Birth cohorts for years beyond 2023 are not modelled as they would not have reached the age at which we model alcohol consumption (15+) within the time horizon.
We create multistate life tables in which the population faces a probability of mortality for each of the five disease/injury conditions and for other cause mortality each year. This approach allows us to simulate prevalence of and mortality from multiple diseases simultaneously, assuming diseases are independent of one another. The model generates alternative population impact fractions (as above) for baseline and for each policy scenario. Using the relevant population impact fraction and rerunning the multistate life table enables a calculation of the difference between baseline and the policy.
13. Baseline health and lagged health impact HIV, road injuries and intentional injuries realise the full impact of the reduction in drinking immediately whereas the health impact on liver cirrhosis and breast cancer are subject to lags in the effect, meaning the reduced drinking does not translate to a reduced health risk immediately 15 . Breast cancer starts to see an impact at year 11 and it is 20 years until full effect, liver cirrhosis sees some impact from year one but does not realise the full effect until year 20 (Appendix part 9).
The life tables for the 20 year time horizon are saved for each of the policy scenarios. They are then used in combination with the probability of having the disease and the potential impact fraction under each policy, to estimate the number of cases.
HIV, road injuries and intentional injuries realise the full impact of the reduction in drinking from the first year of the drinking reduction whereas liver cirrhosis and breast cancer are subject to lags in the effect. Breast cancer only starts to see an impact at year 11 and it is 20 years until full effect, liver cirrhosis sees some impact from year one but does not realise the full effect until year 20 (Table 12).

Hospital multipliers and costs
The prevalence of disease/injury at each policy scenario for each year of the 20 year time horizon was multiplied by the proportion who would then go on to receive hospital treatment (Table 13) and the relevant hospital cost applied (Table 14). The costs taken from the literature were increased by inflation where necessary to reach the baseline year of 2018. Future costs were discounted at 5% as recommended by the Department of Health in the guidelines for pharmacoeconomic submissions 16 . All sources were sense checked with a South African stakeholder with health economics expertise.  The assumption that drinkers will make up 30% of the reduction in drinking recorded alcohol with homebrew comes from consultation with the stakeholders at workshop two.
To test the importance of this assumption on the results a null impact and a 100% impact are introduced. 100% would mean that any homebrew drinkers will not receive any positive health impacts from the policy as all of their reduction in recorded alcohol will be replaced with homebrew alcohol.