Long-Run Impacts of Increasing Tobacco Taxes: Evidence from South Africa

Tobacco taxes are considered an effective policy tool to reduce tobacco consumption and produce long-run benefits that outweigh the costs associated with a price increase. Through this policy, some of the most adverse effects and economic costs of smoking can be reduced, including shorter life expectancy, higher medical expenses, added years of disability among smokers, and the effects of secondhand smoke. Nonetheless, tobacco taxes are often considered regressive because low-income households tend to allocate a larger share of their budgets to purchasing tobacco products. This paper uses an extended cost-benefit analysis to estimate the distributional effect of tobacco taxes on household welfare in South Africa. The analysis considers the effect on household income through an increase in tobacco prices, changes in medical expenses, and the prolongation of working years. The results indicate that a rise in tobacco prices initially generates negative income variations across all groups in the population. If benefits through lower medical expenses and an expansion in working years are considered, the negative effect is reduced, particularly in medium- and upper-bound elasticities. Consequently, the aggregate net effect is progressive and benefits the bottom deciles more than the richer ones. Overall, tobacco tax increases exert a small, but positive effect in the presence of low conditional tobacco price elasticity. If the population is more responsive to tobacco price changes (or participation elasticity estimates are included), then they would experience even more gains from the health and work benefits. More research is needed to clarify the distributional effects of tobacco taxation in South Africa.

Tobacco taxes are considered an effective policy tool to reduce tobacco consumption and produce long-run benefits that outweigh the costs associated with a price increase. Through this policy, some of the most adverse effects and economic costs of smoking can be reduced, including shorter life expectancy, higher medical expenses, added years of disability among smokers, and the effects of secondhand smoke. Nonetheless, tobacco taxes are often considered regressive because low-income households tend to allocate a larger share of their budgets to purchasing tobacco products. This paper uses an extended cost-benefit analysis to estimate the distributional effect of tobacco taxes on household welfare in South Africa. The analysis considers the effect on household income through an increase in tobacco prices, changes in medical expenses, and the prolongation of working years. The results indicate that a rise in tobacco prices initially generates negative income variations across all groups in the population. If benefits through lower medical expenses and an expansion in working years are considered, the negative effect is reduced, particularly in medium-and upper-bound elasticities. Consequently, the aggregate net effect is progressive and benefits the bottom deciles more than the richer ones. Overall, tobacco tax increases exert a small, but positive effect in the presence of low conditional tobacco price elasticity. If the population is more responsive to tobacco price changes (or participation elasticity estimates are included), then they would experience even more gains from the health and work benefits. More research is needed to clarify the distributional effects of tobacco taxation in South Africa.

Introduction
Tobacco is the second leading cause of death and disability worldwide, accounting for 6.3 percent of the total burden (Ng et al. 2014). Moreover, smoking is among the major preventable causes of disease and premature death globally (Doll and Hill 1956;Wynder and Graham 1950). Diseases associated with tobacco use include lung cancer, stroke, ischemic heart disease, and respiratory diseases (DHHS 2004). The World Health Organization (WHO 2017) estimates that tobacco kills more than 7 million people worldwide each year. Lowand middle-income countries, including South Africa, harbor nearly 80 percent of the world's smokers and are less likely to be informed about the adverse health effects of tobacco use relative to individuals in high-income countries. In South Africa, over 170,000 deaths annually, equivalent to 36 percent of all adult deaths in 2015, were attributed to tobacco use (Stats SA 2017; see table 3).
The proportion of regular smokers among adults in South Africa has shown a marked decline over the past two decades (31.0 percent in 1994 and 18.2 percent in 2012). This is largely attributed to the country's aggressive tobacco tax policy, one that has made South Africa a global leader in tobacco control. In 1994, the government announced it would increase the tax on tobacco, the excise tax, and the value added tax combined, from 32 percent of the retail price to 50 percent. 2 By 1997, the target had been achieved, and the excise tax was adjusted annually until 2005 to maintain the 50 percent threshold. In 2006, the total tax target was increased to 52 percent and has remained unchanged since. Overall, cigarette sales declined by a third, and government revenue from tobacco taxes rose from R 1 billion in 1993 to R 9 billion in 2009 (ACS 2012). Furthermore, in 1993, with the Tobacco Products Control Act, health warnings were introduced on cigarette packs, and advertising material and smoking was banned in public transport. In 1999, the original legislation was further strengthened: tobacco advertising, smoking in all indoor public areas, and the sale of tobacco to minors were all prohibited.
Even though increasing taxes on tobacco seem to be one of the most efficient measures for reducing tobacco consumption and increasing government revenue, its effectiveness largely depends on how the tax increase impacts the final price paid by consumers (IARC 2011;World Bank 1999). In South Africa, the tobacco industry enhanced the impact of the excise tax and the industry's profitability-by raising the real retail price by more than the increase in the real excise tax. British American Tobacco has been the dominant cigarette producer and distributor in the country and, prior to 2010, the undisputed price setter. Between 2001 and 2010, the real price of cigarettes rose by 64 percent. Post-2010, the high profits earned by multinationals attracted small cigarette producers that sold at prices significantly lower than the economy brands of their competitors. 3 This changed substantially South Africa's cigarette market: the real price of cigarettes remained relatively constant after 2010. These changes have made passing excise tax increases onto consumers more difficult, rendering cigarettes more affordable and ultimately resulting in a less effective tool to reduce tobacco consumption (Linegar and van Walbeek 2017). 4 Over 340,000 children and more than 5 million adults consume tobacco in South Africa every day (ACS and WLF 2010). 5 Despite the progress of the past two decades, smoking rates are still high among men. Fewer than 8 percent of South African women ages 15 or older smoke relative to 35 percent of men. Similarly, colored adults exhibit a higher smoking prevalence (45 percent) than white adults (25 percent), Indians (20 percent), or black Africans (17 percent). 6 Yet, daily smoking rates in South Africa remain comparable with those in countries of the Organisation for Economic 19.0 percent and 18.2 percent, respectively (OECD 2017).
Tobacco taxes are often considered regressive because the share of household budgets allocated to tobacco products is larger among low-income households than among high-income households. This paper argues that if indirect effects, especially on health, are taken into account, this is no longer valid. The long-run benefits of not smoking offset the costs associated with tobacco taxes among low-income groups and the overall population. Potential benefits include a reduction in medical expenditures and an increase in healthy life years, factors that translate into economic benefits that outweigh the losses created by tax increases if consumers stop or never start smoking. This paper describes and quantifies the effects of tobacco tax increases on aggregate household welfare through three channels. The first implies that higher tobacco prices because of higher taxes induce a behavioral response involving a reduction in tobacco consumption that is manifest particularly among people who discontinue smoking and younger individuals who do not start smoking. The second channel is associated with a reduction in medical expenses, and the third is a rise in incomes because of gains in years of employment. To assess the impact of these effects, this paper estimates the price elasticity of tobacco, simulates upper-and lower-bound scenarios, and calculates the welfare gains among various income groups.
The study is structured as follows. Section 2 briefly reviews the literature on the health effects of tobacco, the economic costs associated with tobacco-related diseases, tobacco tax policies, and price elasticities. Section 3 describes the methodology. Section 4 presents an overview of the data used to forecast the impact of the tobacco tax. Section 5 examines the estimation results. The final section concludes with a discussion on policy implications. Blecher and van Walbeek (2008) estimate cigarettes became more affordable in developing countries between 1997 and 2006. This means these countries are more likely to bear the major health impacts of tobacco consumption. Furthermore, low-and middle-income countries are experiencing a rise in non-communicable diseases. This has negative consequences on human capital development and imposes an increasing economic burden because smoking decreases earnings potential and labor productivity (WHO 2015a). Atun (2014) estimates that the incidence of premature disability and mortality could be reduced by a fifth in South Africa if the risk factors associated with non-communicable diseases are addressed. Approximately 40 percent of deaths in South Africa are related to non-communicable diseases, and a high prevalence is attributed to avoidable risk factors such as tobacco use (WHO 2013). In 2015, over 12,000 deaths of individuals ages 15 or older who smoked were attributed to cardiovascular diseases, and over 16,000 to respiratory diseases (OECD 2017).

Literature review
Several studies have quantified the economic cost of smoking, though most have been carried out in highincome countries. Annual tobacco-related health costs are estimated at US$81 billion in the United States, nearly US$7 billion in Germany, and US$1 billion in Australia (Guindon et al. 2007). Goodchild, Nargis, and Tursan d'Espaignet (2018) find that tobacco-related diseases accounted for 5.7 percent of global health expenditure in 2012 and that the total economic costs of smoking, including health expenditure and productivity losses, were equivalent to 1.8 percent of the world's gross domestic product (GDP) (US$1.85 trillion in purchasing power parity U.S. dollars). The highest share, according to these authors, was in high-income countries (US$1.12 trillion in purchasing power parity dollars), where the tobacco epidemic is the most advanced. 7 Nearly 40 percent of these costs are concentrated in developing countries, reflecting the substantial burden experienced by this group of countries. Earlier estimates of Lightwood et al. (2000) indicate that the gross health cost of tobacco in high-income countries is between 0.1 percent and 1.0 percent of GDP. Likewise, Pichón-Riviere et al. (2014) estimate the annual direct cost of tobacco-related disease in the Chilean health system at approximately 0.6 percent of GDP.
Tobacco price increases are also associated with expansion in productive life years. Verguet et al. (2015) analyze the health effects of a price increases in China and conclude that a 50 percent rise in prices would result in 231 million life years gained over 50 years and would have a significant impact among the poor. In contrast, Pichón-Riviere et al. (2014) estimate that tobacco use in Chile would reduce life expectancy for smokers by nearly 4.0 years among women and 4.3 years among men.
Similarly, exposure to secondhand smoke has a strong relationship with many respiratory diseases among children and adults (DHHS 2004(DHHS , 2014Mason, Wheeler, and Brown 2015;Öberg et al. 2011). According to the World Health Organization, secondhand smoke is responsible for over 890,000 premature deaths per year (WHO 2017). In the United States, exposure to secondhand smoke costs an estimated US$5 billion annually in direct medical costs and over US$5 billion more in indirect medical costs, that is, disability and lost wages (Behan, Eriksen, and Lin 2005). In the state of Indiana, the health-related costs of secondhand smoking have been estimated at more than US$1.3 billion annually (Mason, Wheeler, and Brown 2015). In contrast, McGhee et al. (2006) estimate the cost of direct medical care, long-term care, and productivity losses because of secondhand smoke exposure in Hong Kong SAR, China, at approximately US$156 million annually. Results from the 2011 Global Youth Tobacco Survey in South Africa indicate that 1 in 3 students ages 11-18 in 10 live in homes where someone smokes, and 4 in 10 are around others who smoke in places outside the home.
Tobacco taxation has been recognized as one of the most effective strategies to decrease smoking. In highincome countries, a 10 percent increase in the price of cigarettes is associated with a decrease in the demand for cigarettes of approximately 4 percent (World Bank 1999). In low-and middle-income countries, an equivalent increase is associated with an average 6 percent reduction in cigarette consumption (IARC 2011). Higher taxes have the additional benefits of reducing exposure to secondhand smoke and increasing government revenues.
In South Africa, the 1994 excise tax-induced increase in cigarette prices led to significant reductions in tobacco use (Chaloupka et al. 2000;van Walbeek 2002a). Chelwa, van Walbeek, and Blecher (2017) estimate that, by 2004, per capita cigarette consumption was 36 percent lower than it would have been without South Africa's strong tobacco tax policy. Similarly, van Walbeek (2005) finds that price increases in 1990-2000 reduced the regressivity of the cigarette excise tax in South Africa. Stacey et al. (2018) argue that more aggressive excise tax policies on tobacco in South Africa could lead to improvements in health and revenue. They estimate that an excise rate of 60 percent on tobacco would result in a gain of 858,923 life years.
Other tobacco control interventions are also relevant in decreasing demand but have a smaller impact on tobacco consumption. Several studies have found that health publicity has contributed to a reduction in cigarette use, but the impact has generally been small and, in some cases, temporary. 8 Levy et al. (2012) estimate that the vast reduction in tobacco use in Brazil was mostly caused by higher tobacco prices (46 percent of the impact) and, to a lesser extent, by smoke-free policies (14 percent). The World Health Organization argues that smoke-free environments are the only way to mitigate the harmful impacts of secondhand smoking (WHO 2015b). As part of South Africa's Tobacco Products Control Act, tobacco advertising, smoking in all indoor public areas, and the sale of tobacco to minors were all prohibited in 1999. The impact of the tobacco control interventions can be seen in the decrease in mortality rates associated with tobacco-related diseases in South Africa, such as ischemic heart disease, lung cancer, chronic obstructive pulmonary disease, and asthma Peer et al. 2009).
Price elasticities are crucial in the design of effective tobacco taxation systems. With these, policy makers can determine the sensitivity of demand to a change in tobacco prices. Tax increases tend to generate more impact on tobacco consumption in low-and middle-income countries relative to high-income ones (WHO 2015b). There is an extensive literature estimating the relationship between tobacco prices and consumption. Chaloupka and Grossman (1996) and Lewit and Coate (1982) estimate the elasticity among the under-18 population in the United States at between −1.44 and −1.31 and, among adults ages 18 years or older, at between −0.27 and −0.42. Gallus et al. (2006) estimate a price elasticity of −0.46 for 52 countries in Europe. Cigarette price elasticities across income groups in India range from −0.83 for the lowest income group and −0.26 for the highest (Selvaraj, Srivastava, and Karan 2015). In the United Kingdom, price elasticity is estimated at −0.5 and, in Hungary, between −0.44 and −0.37 (Szilágyi 2007;Townsend, Roderick, and Cooper 1994). Denisova and Kuznetsova (2014) estimate price elasticities in Ukraine by income deciles, ranging from −0.44 for the lowest income group to −0.11 for the highest. Fuchs and Meneses (2017a) also estimate decile-level price elasticities in Ukraine and find a higher average price elasticity (−0.45), ranging from −0.33 for the richest income group and −0.59 for the poorest. Similarly, Krasovsky et al. (2002) estimate an average price elasticity of −0.24 for Ukraine, with variations by income group and age.
The average price elasticity of cigarettes in South Africa ranges from −0.5 to −0.87 (Boshoff 2008;Reekie 1994;van der Merwe and Annett 1998;van Walbeek 2000). Van Walbeek (1996) finds evidence of long-run price elasticities ranging from −0.53 to −1.52 based on data from 1970-90. Van Walbeek (2002b) estimates prices elasticities across income quartiles, controlling for income changes, and finds elasticities of −1.39 and −0.81 for the poorest and richest income quartile, respectively. These estimates are a bit above of the expected price elasticity for developing countries (−0.4 and −0.8) (Chaloupka et al. 2000). None of these have relied entirely on household data to estimate price elasticities; instead, they use annual or quarterly data on aggregate tobacco prices and consumption. 9 Moreover, these estimates likely do not reflect the reality in South Africa because smoking prevalence has decreased significantly in the past two decades.
Age and income are two key factors in determining tobacco price elasticities. Individuals in low-income groups and young adults are more responsive to price changes relative to their peers. This makes them particularly susceptible to tobacco tax increases because they tend to be less dependent on nicotine, more affected by peer effects, and possess less disposable income (Jha and Peto 2014). Several studies in the United States have consistently shown that younger groups show higher elasticities relative to older ones (CDC 1998;Chaloupka and Grossman 1996;Lewit and Coate 1982). 8 Atkinson and Skegg (1973); Bardsley and Olekalns (1999); Stavrinos (1987); Sumner (1971); Townsend (1987); Townsend, Roderick, and Cooper (1994); Witt and Pass (1981). 9 Even though van Walbeek (2002b) uses the South African Income and Expenditure Survey to estimate tobacco price elasticities by income group, the average retail price of cigarettes is applied to all households. Thus, no distinction is made in cigarette quality or price variations by brand.

Model
The impact of rising tobacco taxes in South Africa is estimated using an extended cost-benefit analysis as in other studies (Pichón-Riviere et al. 2014;Verguet et al. 2015). The paper analyzes three factors to estimate how tobacco taxes could affect household income. First, assuming tobacco consumption does not change, tobacco taxes directly affect household income as the share of household budgets allocated to tobacco purchases increases with the rise in taxes. Second, household medical expenses could decrease as a result of reduced tobacco consumption. Finally, households could also experience a positive income change because of additional years of labor recovered through the extension of life expectancy. The aggregate effect of a tax policy is estimated as follows: Income effect = change in tobacco expenditure (A) + lower medical expenses (B) + rise in income (C) (1) A partial equilibrium model is used to assess the distributional effects of a tobacco tax. This approach is used to evaluate the change in prices and relies mainly on household expenditure patterns. This decision implies that only first-order effects are assessed; furthermore, behavioral changes of economic agents such as increases in the consumption of other goods are excluded from the analysis. The model therefore estimates the effects of the short-term response. Moreover, productivity gains from improvements in health deriving from the reduced use of tobacco products are not incorporated in the model primarily because data on the number of days lost or on the depreciation of human capital as a result of tobacco diseases are not readily available. Thus, the estimated income effect should be considered a lower-bound estimate.
The model uses the share of tobacco consumption in household budgets relative to price increases. The loss of real income arising from price increases in products i = 1, …, n is obtained by: where is the share of product i in total household expenditure; ∆ is the percent price increase, and ∆ is the change in consumption of the taxed good that depends on the price elasticity of the product. 10 For example, if 10 percent of the total budget is destined for cigarettes, and the price of cigarettes increases by 10 percent, the real loss in income amounts to 1 percent.

Change in tobacco expenditure
To estimate the variation in tobacco consumption after the tax increase, the model considers the change in prices (∆ ), tobacco price elasticity ( ) for decile i, and tobacco expenditure as a proportion of total household expenditure of decile i in period 0 .
The change in tobacco expenditure is divided by the total expenditure for each decile group i, thereby obtaining a comparable per household measure of the change in tobacco expenditure relative to the total expenditure of each decile group. ∆ .

Medical expenses
The change in medical expenses from tobacco-related diseases is estimated using equation (5), where the cost of treatment of tobacco-related diseases for income decile i is obtained from administrative data and adjusted according to the expenditure survey. The cost of tobacco-related medical expenses is distributed across income decile i according to the share of households that consume tobacco in decile i. Equation (5) shows the income 7 gains associated with the reduction of medical expenses because of reduced tobacco consumption in the long term.

(5)
A reduction in tobacco consumption in the long run would be strongly related to a reduction in tobacco-related diseases. The model assumes that the health effects of tobacco-related diseases will immediately diminish with the reduction in tobacco consumption. 12 Even though this assumption is implausible in the short term because changes in the effects of tobacco-related diseases take some time to materialize, it provides an upper-bound estimate of the effects of tax increases.

Increase in working life years
Finally, the model estimates the impact on income arising from the increase in working years (equation 7). To estimate the increase in working years, the years of life lost (YLL) from tobacco-related diseases are distributed across deciles (i) proportionately to the number of households that consume tobacco (equation 6). 13 Subsequently, the income lost is estimated as the average income per household in decile i. Overall, the model anticipates that incomes will increase as the number of years lost because of premature deaths from tobacco consumption decline.
Lastly, the total income gains in each income group are estimated by adding the results of the increase in tobacco expenditures, the reduction in medical treatments, and the gains in working years (equation 1).

Tobacco prices
Data on household consumption of and expenditure on tobacco products in South Africa come from the National Income Dynamic Study (NIDS), a survey that has been collected in four waves since 2008. 14 The survey asks households how much on "average [was] spent in the last 30 days on cigarettes and tobacco" and, on "average [how many] cigarettes [were] smoked per day." These questions allow us to estimate the average price paid by households for tobacco products at four points in time and to estimate individual-level price elasticities by income group. 15 Moreover, the NIDS also has the advantage of being a panel study that follows the same individuals over time and thus allows us to control for individual unobserved heterogeneity to estimate tobacco price elasticities. 16 12 Other studies have estimated the pass-through between the decline in tobacco consumption and the effect on medical expenditures. These estimates may also differentiate the effect associated with people who stop consuming tobacco versus people who do not start at all because of the tax policies (Verguet et al. 2015). Because of data restrictions, these assumptions are not included in the analysis. 13 Life expectancy of 65.6 for women and 58.5 for men are used to estimate years of life lost in South Africa (World Development Indicators). 14 The South African Income and Expenditure Survey was considered; however, the survey does not include quantity information on tobacco products, limiting the analysis that could be derived from the survey. 15 The quantity of cigarettes is converted to monthly terms to estimate monthly prices (expenditure on cigarettes/quantity). 16 Income deciles were created using the latest available data set, wave 4 (2014/15). Table 1 shows significant variation across deciles in cigarette prices. For instance, the poorest decile in 2015 paid an average R 6.65 for 20 cigarettes, whereas the richest decile spent R 20.14. Even though the cigarette prices obtained from the survey are lower than those shown by Statistics South Africa, they follow the same market trend for the available years (figure 1). Moreover, these differences are likely caused by the phrasing of the relevant questions in the NIDS, household recall error (Biedman 2010) and that using household expenditure on tobacco accounts for the illicit market. In addition, there is significant price variation by cigarette brand; the most expensive sold at an average price of R 39, and the cheapest at R 18.7 in 2015. 17 It is crucial to account for these variations to estimate tobacco price elasticities; thus, applying the same price for all households would not reflect the differences evident in table 1.

Tobacco price elasticity by decile
Unlike previous tobacco price elasticity estimates on South Africa, this study uses the NIDS individual-level data sets for relevant estimates by income decile. Tobacco price elasticity estimations using national aggregate time series data on production and sales could face several problems. First, it is difficult to differentiate among the number of cigarettes sold, the number consumed, the number coming from illicit trade, or the price paid. Moreover, estimates are typically produced with a small number of observations and often lack granularity. A longitudinal panel reflecting repeated observations of individuals in both purchases and the prices paid over time for tobacco products is the ideal data set. The NIDS allows us to conduct such an analysis and control for unobserved heterogeneity among individuals for four waves from 2008 to 2015. Another advantage of using 17 Cigarette price data were provided by Statistics South Africa; because of data restrictions, product brands are not referenced in the analysis. Year

NIDS
Stats SA this survey is that one may detect the price paid by consumers and account for promotions, sales, or even the purchases of illicit cigarettes.
Nonetheless, the NIDS suffers from nonresponse and attrition as most surveys do. Similarly, the sample size is also reduced when using a balanced panel. 18 In addition, Kacker (2016) finds that the NIDS data do not match information from other South African surveys and appear to oversample rural areas, the less well educated, and college-educated individuals. Despite these limitations, Kacker (2016) recognizes that the survey is clearly internally consistent and that there is no reason to doubt the reliability of the data. Caution is recommended in drawing conclusions from the survey because the NIDS appears to be biased toward the part of South Africa that has improved the most. 19 Although several models have been tested to estimate decile-level elasticities, the random effects in a near balanced panel is the preferred model. (See annex A for more on the methodology and the various iterations of the estimation of tobacco price elasticities by decile.) The estimated average tobacco price elasticity in South Africa is −0.25, which is lower in absolute terms than the elasticities found in the literature for developing countries (−0.4 and −0.8), as well as those previously estimated for South Africa (−0.5 and −0.87) (Boshoff 2008;Chaloupka et al. 2000;Reekie 1994;van der Merwe and Annett 1998;van Walbeek 2000). 20 Nonetheless, all models tested for this study estimate an average elasticity between −0.23 and −0.28 (see annex A, table A2). Moreover, as noted above, to the best of our knowledge, previous estimates of tobacco price elasticities in South Africa have not relied entirely on data on households or individuals, and few rely on income groups. In addition, most of these estimates have been calculated in years prior to the more profound changes in tobacco policy in South Africa. Previous elasticities were also estimated at a time when smoking prevalence was significantly higher in South Africa; as of 2015, fewer than a fifth of the households interviewed reported spending money on cigarettes, relative to over 40 percent in 1995. 21 As expected, lower income deciles exhibit higher elasticities relative to richer deciles. For instance, the poorest decile has a medium-bound elasticity of −0.36, whereas the richest has an elasticity of −0.22 (table 2). The standard error of these estimates is approximately 0.10, producing a 95 percent confidence interval of −0.20, +0.20. 22 To show the effect of a tax increase under different scenarios, we simulate a lower-and an upperbound elasticity for each decile. The former tends to reflect income groups that would not change consumption patterns, such as rural residents or older individuals, while the latter tends to show a longer-term scenario, reflecting the effect the tobacco tax would have on younger individuals. After a few decades, only these would still be alive; the total average effect of the price increase would therefore be approximated more accurately by the upper-bound price elasticity. Lastly, in South Africa, because the wealth distribution does not vary substantially across the bottom deciles, which largely rely on government subsidies and transfers, price elasticities were also estimated for quintiles as a robustness check (Inchauste et al. 2017) (figure 2, panel b).

Figure 2 -Tobacco Price Elasticity, South Africa
Source: Estimates based on the National Income Dynamics Study, waves 1-4, 2008-15. Note: Deciles and quintiles were created using per capita household expenditure. Lower-and upper-bound elasticities show differences of −0.2 and +0.2, respectively, with the medium-bound elasticity. Random effects estimates reported using near balanced panel data, where only individuals present in all four waves are kept but some are dropped due to missing observations in the relevant tobacco expenditure questions.

Mortality and morbidity
Statistics South Africa has information on mortality and causes of death in South Africa in 2015. The data are disaggregated according to the smoking status of the deceased (table 3), where a smoker is defined as someone who had smoked any form of tobacco on average for more than six months a year or four or more days a week five years previously. Approximately 36 percent of all deaths in 2015 (173,241) were associated with tobaccorelated diseases. 23 Of these, over 7,000 were women and nearly 25,000 were men (tables 3 and 4). Though a large share of deaths in South Africa (46 percent) are recorded with the smoking status unknown, the incidence of tobacco-related deaths appears to be higher among smokers. For instance, in 2015, even though there were roughly the same number of cases of chronic obstructive pulmonary disease among smokers (4,348) and nonsmokers (4,270), smokers represented fewer than 20 percent of the South African population. Pillay-van Wyk et al. (2016) find that mortality by disease is severely underreported in the South African data; the results presented using these numbers will thus provide lower-bound estimates of tobacco-related deaths. 23 If only individuals with known smoking status are considered (32,767), the share drops to 7 percent.

Table 3 -Tobacco Related Deaths by smoking status, 2015
Source: Stats SA 2017. Note: A smoker is defined as someone who smoked any form of tobacco on most days (an average of more than six months a year, or four or more days a week), five years previously. If smoking started less than five years before death, the answer is No. If smoking started more than five years before death, but for less than six months, then the answer is No. Smoking status is only reported for individuals ages 15 or older. Note: A smoker is defined as someone who smoked any form of tobacco on most days (an average of more than six months a year, or four or more days a week), five years previously. If smoking started less than five years before death, the answer is No. If smoking started more than five years before death, but for less than six months, then the answer is No. Annex C reports mortality by smoking status and gender.   Note: Incidence is defined as the number of new cases of a given disease during a given period in a specified population. It is also used for the rate at which new events occur in a defined population. It is differentiated from prevalence, which refers to all cases, new or old, in the population at a given time.
Data on morbidity are obtained from the Global Burden of Disease Study (table 6). Chronic respiratory diseases, ischemic heart disease, and tuberculosis are among the most prevalent diseases among men and women in South Africa. In 2015, approximately 1.3 million cases of tobacco-related disease were reported.

Tobacco-related medical costs
The most recent study that analyzes the medical costs of tobacco consumption in South Africa occurred in 1988. Yach, McIntyre, and Saloojee (1992) estimate the cost of smoking-related diseases at R 3.64 billion (US$1.3 billion in 2015 prices), equivalent to 1.82 percent of GDP. The cost of health care and lost productivity because of admission to hospital and premature mortality the cost was estimated at between R 1.39 million and R 2.45 million in 1988. Even though these estimates are specific to tobacco-related illnesses, they are far too outdated to represent the current reality in South Africa and cannot be used in the analysis.
Another option is to use cost data from private hospitals in South Africa. To the best of our knowledge, there are no recent studies that have estimated tobacco-related costs in South Africa for the public sector or administrative data that would allow for these estimations. While a middle income country, South Africa's health system is associated with higher private health spending (51.8 percent) than most high income countries of the Organisation for Economic Co-operation and Development; yet, only 17 percent of the population can afford private insurance. 24 Lorenzoni and Roubal (2016) estimate the average price of 28 case types using data from private hospitals from several large medical schemes in South Africa in 2011-13. 25 Of these 28 case types, four may be categorized as tobacco-related: heart failure, malignant neoplasm of bronchus and lung, pneumonia, and acute myocardial infarction, and, for two, we also have information on the number of cases in 2015: bronchus and lung cancer and ischemic heart disease (table 7). 26 Nonetheless, using only private sector information for two tobacco-related diseases provides severe underestimates of the effects on reduced medical expenses. To mitigate some of these limitations, the analysis could be complemented by using medical costs from countries with a health system similar to South Africa's health system. However, few countries have available medical cost information for tobacco-related diseases.
An alternative is to use the aggregate cost estimates of Goodchild, Nargis, and Tursan d'Espaignet (2018), who apply a cost-of-illness approach to estimate the economic cost of smoking-attributable diseases in 152 countries in 2012. 27 For the purpose of this study, we limit this to direct health care expenditures and exclude indirect 24 As of 2017, government expenditures on health amounted to R 170.9 billion (8.8 percent of GDP) or US$1.2 billion. Public expenditure on health rose from 3.4 percent of GDP in 1995 to 4.2 percent in 2014, similar to private expenditures on health (4.6 percent of GDP). Moreover, 48.2 percent of health expenditure in South Africa comes from public funds. Out-of-pocket health expenditure has been decreasing over the years and was at 6.5 percent of total health expenditure in 2014. See WDI (World Development Indicators) (database), World Bank, Washington, DC, http://data.worldbank.org/products/wdi. 25 "'Case type' refers to categories of hospital services that are similar from a clinical perspective and in terms of their consumption of resources. . . . The term 'price' relates to the amount paid to health care providers from risk pools, savings accounts, and out-of-pocket rather than the amount claimed. Payments to all private provider types are included" (Lorenzoni and Roubal 2016, 12). 26 Ischemic heart disease is also referred to as coronary artery disease or coronary heart disease. It occurs if arteries are narrowed and increases the chances of heart failure (Cleland and McGowan 1999). Myocardial infarction is one of the manifestations of ischemic heart disease; the cost of myocardial infarction is thus used for ischemic heart disease (Manfroi et al. 2002). 27 Goodchild, Nargis, and Tursan d'Espaignet (2018)  costs such as productivity losses from morbidity and mortality. 28 Ideally, we would want disaggregated cost estimates by disease to have a more accurate estimate of these costs. Nonetheless, given the data available, the aggregate estimate of Goodchild, Nargis, and Tursan d'Espaignet (2018) is the best alternative. The estimated medical costs in South Africa are R 18.9 billion, equivalent to US$1.4 billion in 2015 prices. More research is needed to obtain more accurate estimates of the economic costs of treating each tobacco-related disease in South Africa. Total medical costs assume that private medical costs represent 51.8 percent of total tobacco-related medical costs. To estimate the economic cost of treating tobacco-related diseases, we multiply the average price of each treatment by the number of events related to tobacco (assuming that most individuals were treated at some point). Table 8 summarizes the most important indicators, including total monthly household expenditures and the share of expenditures on tobacco products. The share of household expenditures on tobacco rises with income until the eighth decile but decreases for the last two deciles. Thus, tobacco consumption prevalence is concentrated in South Africa's middle class. Meanwhile, poorer households are less likely to have smokers: 18 percent do among the poorest decile, and fewer than 30 percent do in the top two deciles.

Results
To analyze the distributional effects of an increase on tobacco taxes, we estimate the effect on prices and medical expenditures, aggregating these two into a single measure. The price elasticities estimated in table 2, including the lower-and upper-bound elasticities, will allow us to understand how the results could change under different assumptions.

Tobacco price increase
Income changes that arise from an increase in tobacco prices are estimated for each decile based on low-, medium-, and upper-bound elasticities. With equation (4), the price elasticities, and the share of household expenditure on tobacco by decile, we can simulate the effects of an increase in tobacco prices. To show the effect of the elasticities on prices, table 9 also includes estimates of a complete pass-through scenario, whereby the increase in prices is completely transferred to consumers without a reduction in consumption. For instance, if we assume that the prices for tobacco products rose by 25 percent, given the medium-bound elasticity (−0.36) in table 2 and the proportion of tobacco expenditures for the bottom decile (1 percent) in table 8, the expected decrease in household expenditures would be 0.20 percent (table 9). This represents a loss in welfare because consumers would devote a higher share of their incomes to purchase the same amount of tobacco, thereby reducing the consumption of other goods. These results hold for all analyzed scenarios. Nonetheless, the effect of the price increase is relatively progressive, that is, it affects upper-income groups in a larger proportion up to the ninth decile, though the top decile is affected less than the poorest one (figure 3).

Figure 3 -Income Gains: Direct Effect of Tobacco Taxes (Increase in Expenditure because of tobacco taxes)
Source: Estimates based on data of the National Income Dynamics Study, wave 4, 2014-15. Note: Estimates assume a price shock of 25 percent.

Medical expenses
Table 10 and figure 4 report the income effect of a reduction in medical expenses. As noted above, the model assumes that the health effects of tobacco-related diseases will immediately diminish with the reduction in tobacco consumption. 29 Even though this assumption is implausible in the short term because changes in the effects of tobacco-related diseases take some time to materialize, it provides an upper-bound estimate of the effects of tax increases. Moreover, because the costs of tobacco-related diseases are likely underreported given the available data for the estimates, the results are also likely to underestimate the potential benefits of reduced medical expenditure. The overall results indicate that the reduction in medical expenditures is highly progressive, disproportionally benefiting lower-income groups. This derives from two factors: (1) the higher price elasticity and (2) a lower income base that massively benefits from the reduction in medical costs.
A potential concern with these results comes from South Africa's health financing architecture, where most health services are free for the poor. Although the poor might not pay large amounts of out of pocket health care, there are still intangible costs associated with a household member being sick due to tobacco related use. If we assume that the overall health budget remains the same, the fiscal saving from lower tobacco-related illness would allow for improved health care services to all households, either through lower wait times or through better services. 30 As a robustness check we've allocated a uniform compensation-due to the reduction in total medical costs-as an improved future benefit to all households. This implies that the long-run impact of tobacco taxation remains progressive since higher health spending will be progressive in relative terms, and the estimates on productive life will continue to be progressive (see annex D).

Income gains deriving from an increase in working life years
We estimate the cost of working life lost because of tobacco consumption, assuming that the impact of lower tobacco use on health and work-generated income is direct. The 1.3 million deaths attributed to tobacco 29 Other studies have estimated the pass-through between the decline in tobacco consumption and the effect on medical expenditures. These estimates may also differentiate the effect associated with people who stop consuming tobacco versus people who do not start at all because of the tax policies. Because of data restrictions, these assumptions are not included in the analysis. 30 As noted earlier, we were not able to obtain out-of-pocket health expenditures by income level or specific costs of treating tobacco related diseases for the public sector.
consumption are distributed using the occurrence of mortality profile. For each death, working years lost are divided across deciles proportionately to the number of households that consume tobacco in each income group. Using equation (6) and table 5, we can calculate the results of the tax increase. The results show that the reduction in tobacco consumption and the expected reduction in work years lost have positive impacts on welfare. Overall, the gains are evenly distributed across income groups; however, elasticities vary across deciles, generating an important impact on lower-income groups ( figure 5; table 11). 31

Net effects: Total distributional impact
The aggregate effect of an increase on tobacco taxes is highly progressive; in the long run, poorer deciles benefit more than richer ones from the tax increase (table 12; figure 6). The positive effect of reduced medical expenses and years of life gained, diminish the negative price effect across all deciles in all three elasticity scenarios. Moreover, the benefits are amplified if we compare the lower-bound elasticity with the upper-bound elasticity as in the latter case total effects are positive for all deciles.  The same analysis was conducted for income quintiles to mitigate potential bias given South Africa's wealth distribution (see section 4). The results are robust for different groupings (see annex E).

Discussion
Despite the wealth of research on the negative effects of tobacco consumption and on the benefits of various public policy mechanisms aimed at reducing tobacco use, questions remain about the progressivity or regressivity that these entail. The implementation of tobacco taxes, is considered one of the most effective ways to discourage tobacco use. Nonetheless, this policy has a direct impact on household incomes, especially among low-income households that are more likely to smoke, have limited access to health insurance and adequate health care. Moreover, the net effect of an increase in tobacco taxes depends on the price elasticity of this product across different sectors of the population. The price elasticity determines the magnitude of the income shock and the benefits gained from the decline in tobacco consumption.
To assess the net welfare gains from this policy, one must look beyond the direct impact on household income and consider other benefits of lower tobacco consumption, including a reduction in medical costs and an increase in the potential working years associated with good health. Thus, it is critical to justify the maintenance or intensification of the use of tobacco taxes by means of a demonstration of the aggregate monetary gains or losses generated. Moreover, the policy should focus on low-income households that are more likely to smoke and, hence, tend to be the most affected by consumption taxes. One of the main motivations of this study is to weigh the main costs and benefits of tobacco taxation to determine if, in the end, the policy is regressive.
The results indicate that the aggregate effect of an increase in tobacco taxes is highly progressive. If we include the benefits through lower medical expenses and an increase in working years, the negative effect from an increase in prices is eliminated. Overall, the net effect shows an aggregate welfare gain among the bottom five deciles relative to the medium-bound elasticity and among all income groups relative to the upper-bound elasticity scenario.
These results are partly driven by lower tobacco price elasticities relative to what has been previously estimated for developing countries (−0.25 estimated elasticity for South Africa; −0.4 and −0.8 for developing countries) (Chaloupka et al. 2000). Earlier estimates for South Africa have relied on national aggregate time series data on the production and sales of tobacco products. With such data, it is difficult to differentiate between the number of cigarettes sold, how many were consumed, the price paid, or how many come from illicit trade. Instead, this study uses a longitudinal panel whereby there are repeated observations on individuals on both purchases and the prices paid over time for tobacco products.
South Africa's low price elasticities may also be explained by the country's experience with tobacco control policies and the changing structure of the tobacco market in 2010. Moreover, these estimates likely capture the illicit cigarette market. Following the profound changes produced by tobacco legislation, smoking prevalence decreased by over 10 percentage points. People who continued to smoke are probably more addicted and possibly less sensitive to price changes. 32 To test our hypothesis, we applied the tobacco price elasticities by income quartiles estimated by van Walbeek (2002b) to these simulations. 33 These elasticities are much higher than the ones used in this study (−0.25 versus −1.1) and were also estimated at a time when smoking prevalence was significantly higher in South Africa. As of 2015, less than a fifth of households report spending money on cigarettes, compared with over 40 percent in 1995. 34 As expected, the results of a tobacco tax increase using higher elasticities, such as that of van Walbeek (2002b), indicate a positive and highly progressive effect of an increase in tobacco prices, one that, in magnitude, is larger than the results presented in table 12 (see annex F for simulation results). Overall, tobacco tax increases have a small effect in the presence of a low tobacco price elasticity. Thus, a population that is not as sensitive to tobacco price changes, as is the status quo in South Africa, will not reduce consumption sufficiently to experience even more gains from the health and work benefits.
In Chile, Moldova, and Ukraine, the authors find evidence that tobacco price increases are also a progressive policy in favor of low-income groups (Fuchs and Meneses 2017a, 2017b. Nonetheless, these countries present much higher smoking prevalence rates and higher tobacco price elasticities than South Africa and do not have the same history of substantial tobacco control policies. 35 These low price elasticities indicate that a uniform increase in tobacco taxes should not be the only policy in place to reduce tobacco consumption in South Africa further. In addition, we hypothesize that limited data availability on the medical costs of tobacco-related diseases accounts for a lower-bound estimate of the potential benefits of reduced medical expenses. The medical costs of treating tobacco-related diseases must be investigated further in South Africa to have more accurate estimates of the distributional effects of tobacco taxation in the country. Similarly, Pillay-van Wyk et al. (2016) indicate that mortality by disease is severely underreported in the South African data, presenting yet another lowerbound indicator for estimating aggregate wealth effects. Moreover, most low-income households in South Africa still have very limited access to specialist care; when they get cancer, heart attacks or strokes, they simply perish with little or no care. Thus, the impact of increased life expectancy and reduced morbidity would probably be much more significant than the impact of medical costs reductions for the poor.
These three factors combined highlight that more research is needed if we are to understand the distributional effects of tobacco taxation in South Africa. 32 Since the paper uses reported (expenditure on tobacco) prices, total elasticity for the different deciles cannot be estimated. Thus, the price elasticities in this paper are conditional elasticities. 33

A. Tobacco Price Elasticity by Decile 36
Let be defined as the average quantity smoked per day by individual i in income decile d; P the average price per cigarette (unit value of tobacco use); the real household per capita income per adult equivalent of individual i; and the individual level characteristics. Then, the smoking intensity equation is written as follows: The empirical analysis of equation (A1) assumes a log-log relationship among smoking intensity, price, and income.
is observed if and only if the individual from a given decile d is a current smoker. The corresponding Hausman test does not reject the null hypothesis that the differences between the random effects model and the fixed effects model estimates are not systematic [Chi 2 (15) = 105:34; p = 0:000]. Therefore, the fixed effects model is the consistent model because it controls for unobserved heterogeneity. However, if timeinvariant variables such as gender, race, and religion are regarded as significantly important in explaining the outcome variable (smoking intensity) and if the degree of variability of our main variable(s) (within variation) is lower, then the random effect model is preferred to the fixed effects model, that is, using the fixed effects model will then produce less efficient estimates (Plümper and Troeger 2007).
The NIDS uses the following questions to measure individual smoking behavior. For current smokers, do you smoke cigarettes? For nonsmokers, did you ever smoke cigarettes regularly? Both smokers and ex-smokers were asked the age at which they first smoked cigarettes, but only ex-smokers were asked when they last smoked cigarettes regularly. Finally, individuals were asked to indicate, on average, the number of cigarettes they smoke per day. Only individuals who smoke cigarettes remain in the sample for estimating smoking intensity, while nonsmokers and former smokers are excluded. A smoker is defined as someone who consumes some positive amount of cigarettes at the time of the interview. Smoking intensity is defined as the average number of cigarettes an individual smokes per day. Cigarette prices (unit value) are real household per capita expenditure on tobacco per day. These values are deflated using the consumer price index so that each cigarette price is in 2010 prices.
Since the paper uses reported (expenditure on tobacco) prices, total elasticity for the different deciles cannot be estimated. Thus, the price elasticities in this paper are conditional elasticities.
Several models were tested before deciding the final elasticities to use in the model (tables A3-A14). Both a balanced and unbalanced panel were considered. Despite the large number of observations that are dropped in using a balanced panel, this allows us to control for individual time invariant effects (table A1). Moreover, coefficients for all deciles remain statistically significant at the 95 percent level if a balanced panel with random effects is used (table A3).

D. Simulation Applying Uniform Compensation: Reduction in Medical Expenditure
In South Africa most health services are free for the poor. Although they might not pay large amounts for outof-pocket health care, there are still intangible costs associated with a household member being sick due to tobacco related use. If we assume that the overall health budget remains the same, the fiscal saving from lower tobacco-related illness would allow for improved health care services to all households, either through lower wait times or through better services. As a robustness check we have allocated a uniform compensation of 0.1due to the reduction in total medical costs-as an improved future benefit to all households. The long-run impact of tobacco taxation remains progressive (see table D1-D2; figure D1-D2).