The Effectiveness of Alcohol Control Policy in South Africa

THE EFFECTIVENESS OF ALCOHOL CONTROL POLICY IN SOUTH AFRICA: ESTIMATING THE PRICE ELASTICITY OF DEMAND FOR ALCOHOL ABSTRACT: Alcohol control policy is a topical issue in South Africa, as new measures to further restrict the consumption of alcohol are being considered. This study examines the possible effectiveness of increases in excise taxes on alcohol as one such measure. Using an empirical analysis, the study considers the price elasticity of demand for alcohol among different groups of consumers in South Africa. The results suggest that demand for alcohol is inelastic, particularly for heavy drinkers. While alcohol is found to be an inferior good, different income groups are not found to have a significantly different price elasticity of demand. Furthermore, low­income, heavy­drinking households are found to be the least responsive to changes in the price of alcohol. THE EFFECTIVENESS OF ALCOHOL CONTROL POLICY IN SOUTH AFRICA: ESTIMATING THE PRICE ELASTICITY OF DEMAND FOR ALCOHOL 1. Introduction The effectiveness of alcohol control policy is a discussion worth noting in South Africa’s present context. As efforts are made to re­evaluate the role that alcohol abuse plays in our society, it is equally important to evaluate the potential effectiveness of any policies that would seek to discourage the abusive consumption of alcohol. The damaging effect that alcohol plays in South Africa is a broadly discussed issue. It is continually noted that alcohol plays a significant role in the number and the severity of the crimes committed in our country; alcohol is a major catalyst in domestic violence in our homes; a concerning proportion of the accidents that take place on South African roads are attributed to alcohol, and many diseases are caused by heavy drinking, for which South Africa is widely noted (Parry, 2012; World Health Organisation, 2011). Because of the social and medical problems that alcohol presents to our society, there have been calls to increase the scope and stringency of alcohol control policy, in order to discourage heavy drinking. Among other suggestions for discouraging heavy consumption, such as imposing restrictions on the advertising of alcohol and reviewing the age limit for drinking in the country, one popular suggestion is raising excise taxes to increase the price of alcohol. Economic theory suggests that raising the price of alcohol would cause a fall in consumption of alcohol; however, this relies on the critical assumption that the demand for alcohol among drinkers is sufficiently sensitive to changes in price. Another assumption that this policy suggests is that an increase in the price of alcohol will affect all drinkers similarly enough such that even heavy drinkers, at whom the policy is aimed, will be induced to consume less alcohol. This paper seeks to investigate the relationship between the price of alcohol and its consumption; in particular, it seeks to investigate whether a price increase will induce a substantial decrease in alcohol consumption among heavy drinkers. To do this, this paper seeks to analyse whether there are differences in the price elasticity of demand among different groups of the South African population. In particular, it will analyse whether persons of lower incomes have the same sensitivity to the price of alcohol as persons of higher incomes. This will help in answering the question of whether the increase in excise taxes is an effective policy move, and if it exhibits the same effectiveness for all groups of the population. This paper will proceed as follows. Firstly, it will explore the existing literature on alcohol policy and the effectiveness of various tools in reducing alcohol consumption. It will then present the methodology which will be followed in the present study, as well as a discussion on the data set that will be used for the study. Thereafter, the findings of the study will be presented, followed by conclusions drawn from these findings. 2. Alcohol Control Policy In The Literature The problems associated with alcohol consumption are well­documented in the literature. These economic costs of alcohol consumption can be classified either as the social costs that abusive drinking produces through its negative externalities, or the individual costs borne by drinkers themselves. Odejide (2006) discusses the societal cost of abusive drinking in the African context, particularly when undertaken by the youth, as involving an increase in vehicular accidents, reckless sexual behaviour and violence. Other social costs associated with problem drinking include loss of employment, marital disharmony and domestic violence (Odejide, 1978). Therefore, excessive alcohol consumption is argued to impose a burden not only to those who engage in abusive drinking, but to society in general. Alcohol also presents a disease burden to society that is equal to that imposed by tobacco globally. Room et al. (2005) find that alcohol is causally related to more than sixty medical conditions. This includes cirrhosis of the liver, haemorrhagic stroke, breast and liver cancer, and neurological conditions such as epilepsy. Alcohol abuse is also associated with unintentional injuries such as poisoning and homicide. Moreover, Rehm and his colleagues (2009b) find that approximately 6.3% of all disability­adjusted life years were lost due to alcohol­attributable causes. Of these deaths, 16.12% were due to HIV/AIDS. Indeed, Parry, et al. (2010) review the relationship between alcohol consumption and the progression of HIV infected patients and show through the literature that there is evidence suggesting that alcohol consumption worsens the disease course of HIV/AIDS, shortening the lifespan of HIV positive individuals. The various problems associated with the excessive consumption of alcohol have prompted the development of an extensive body of literature on the demand for alcohol and how different measures of public policy can affect this demand to reduce abusive drinking. In particular, Fogarty (2006) suggests that in order for public policy to be effective in reducing the economic cost of alcohol abuse, the nature of alcohol demand must be understood. That is, in order to reduce the economic cost of alcohol abuse, the levels of drinking among consumers and the sensitivity of the demand for alcohol must be understood. One view of alcohol demand among consumers holds that there is a fundamental difference between normal drinking behaviour that may be associated with general levels of consumption, and problematic drinking behaviour that presents a cost to society. In their study of the nature of alcoholism, Taylor and Helzer (1983), Vaillant and Hiller­Sturmhöfel (1996) and Vaillant (1983) all treat alcoholism as a psychological disorder with genetic and environmental causes, which implies a qualitative difference between normal drinking behaviour and the heavy drinking that characterises alcoholism. An opposing view holds that all social drinking behaviour may be modelled on a continuum of increasing levels of consumption. Ledermann’s hypothesis (1956), the basis for a significant portion of research on the distribution of alcohol consumption, models alcohol consumption as a lognormal distribution. This model relies only on mean consumption; it views excessive drinking as closely connected to the mean consumption, such that a high prevalence of excessive drinking is associated with a relatively high mean level of consumption. Therefore, Ledermann’s model views the entire population as being at risk of exposure to the negative consequences associated with alcohol. Lemmens (1993) studied this suggested regularity in the distribution of alcohol consumption, and its implication that a decrease in the average per capita consumption would result in lower economic costs associated with alcohol consumption. Lemmens finds that the regularity in the distribution of consumption is supported by findings in empirical studies, and suggests that this has implications for who in the distribution is targeted by public policy to discourage excessive drinking in the population as a whole. In particular, this implies that a policy that successfully reduces the mean level of alcohol consumption will effectively reduce consumption among excessive drinkers and thus reduce the economic costs of alcohol abuse. Duffy (1993) also investigated Ledermann’s hypothesis of a lognormal distribution of the consumption of alcohol. He finds that the literature based on Ledermann’s contribution suggests that since the proportion of heavy drinkers is an increasing function of mean consumption, policy aimed at reducing mean consumption will effectively reduce the number of heavy drinkers and thus alcohol­related problems. Duffy suggests that if consumption does not have a lognormal distribution, this effect will not hold and any blanket policy that seeks to decrease per capita consumption will be ineffective in reducing the economic costs of alcohol abuse. This assertion points to the development of an alcohol control policy that will target the demand for alcohol among heavy drinkers in particular, rather than a blanket policy which will prove to be ineffective if consumption does not have a lognormal distribution. Various policy measures have been suggested in order to control the consumption of alcohol. These include restrictions on the availability of alcohol, education programmes on its harmful effects, restrictions on advertising and increased excise taxes on alcohol (Dejong, 1997). Toomey and Wagenaar (1999) also discuss some policy measures that can be used to reduce the risk of alcohol problems by changing the way in which alcohol is marketed and sold. These measures can be classified either as individual­level programmes or institutional policies; the policy implemented would depend on the type of drinking behaviour. Many have called for the restriction on the advertising of alcohol, in particular, as alcohol marketing has been argued to cause an increase in the consumption of the youth (Kessler, 2005), a group at high risk of alcohol abuse. An increase in excise taxes in order to increase the real cost of alcohol has been of particular interest in the literature, and this policy measure faces less opposition from the alcohol industry than restrictions on alcohol marketing (Nelson, 2010; Kessler, 2005). The effectiveness of policy implementation in the reduction of alcohol intake has been a subject of interest in the literature. In particular, many have studied the impact of excise taxes on the demand for alcohol among different groups in the population (e.g. Gallet, 2007). Manning et al. (1995), for instance, investigated the price sensitivity of light, moderate and heavy drinkers in order to evaluate the impact of excise taxes on alcohol among these groups. Manning et al argue that since the economic costs associated with alcohol come from heavy drinking in particular, and not from light or moderate drinking, an effective policy mix will discourage excessive alcohol use rather than all consumption of alcohol. They find that the demand for alcohol of heavy drinkers is less responsive to a change in price than the demand of lighter drinkers. Therefore, a policy that seeks to reduce alcohol consumption will not necessarily have the same effects on all members of the drinking population. Saffer and Chaloupka (1998) also noted the variation in response to alcohol policy according to demographic differences, and they analysed the demand for alcohol in China, to see whether the price sensitivity of drinkers would be similar to that of drinkers in developed countries. They find that price elasticity in China is very small compared to the elasticity in developed countries, and suggest that alcohol tax increases alone would be ineffective as a control policy in that country. Further study in this field may seek to analyse the own­price elasticity of demand for alcohol among South African drinkers in particular, and the variation among beer, wine and spirits. Furthermore, it may analyse the variation in the own­price elasticity of demand for alcohol among types of drinkers, as this may have serious consequences for the policy mix implemented to reduce alcohol abuse. The purpose of this paper is to analyse this variation in demand among income groups in South Africa. 3. Methods The methodology we proceed with is an adaptation of the study on the demand for alcohol as conducted by Manning, Blumberg and Moulton (1995). We maintain the essential procedures that they undertake with their data, allowing for certain variations that will make the analysis more appropriate to South Africa, and observe whether South African data will provide similar results to those of the Michigan data used by Manning et al (1995). We make a further departure from the Manning et al. study to accommodate our use of unit values to estimate elasticities, whereas Manning et al. used Tobit regressions. The main concern in this study is the sensitivity of alcohol consumption to changes in price. In particular, we want to compare price elasticities of alcohol consumption among different groups of consumers. In order to do this, we define four distinct income groups from the data as well as three different levels of drinking. We use Cox and Klinger’s (1988) definition of light drinking as up to 432 standard drinks per annum, moderate drinking as between 432 and 792 drinks per annum, and heavy drinking as at least 792 drinks per annum. These definitions of the different levels of drinking are particularly appropriate for use in this study, as the data set measures consumption in standard drinks rather than alcohol content. Before proceeding with defining a model with which to analyse the data, we recognise that the study involves the two­fold issue of the initial decision to drink and then the quantity of alcohol to consume once the individual has decided to drink. This is of particular relevance in this present study of South African data, as about two­thirds of the adult population are not drinkers (World Health Organisation, 2011). Ignoring the determinants of the decision to drink will lead to unreliable results. We run a logistic regression equation to determine the main determinants of the initial decision to drink. Of particular interest are the socio­economic differences that occur between those who choose to drink and abstainers. We then turn our attention to the issue of the sensitivity of alcohol consumption among the subset of the population who chooses to drink. NIDS data poses a challenge for the present study in that it does not contain any information on prices faced by households and individuals. Instead, households are requested to answer questions regarding their monthly expenditure on different household consumable goods, amongst others, alcohol. Furthermore, individuals are asked questions regarding their weekly consumption of alcohol. We therefore implement a method to impute a proxy price of alcohol based on the quantity and expenditure data, in order for us to derive meaningful elasticity estimates. In order to tackle the matter of missing price information in survey data, Deaton (1988) proposes a method of estimating the unit value of goods that respondents face. The method requires that the survey data is gathered in several clustered households such that we control for measurement error, and so that there is sufficient variation in unit values between the clusters. If this is the case, then Deaton suggests that we compute the unit value of household purchases as follows: v= total expenditure physical quantity The problem with the unit value is that it is not the same as price. Unit values are affected by the actual market prices that consumers face and the quality of purchases. Because consumers who purchase high­quality items will also have higher unit values, unit values are not exogenous, as they are chosen, and they are therefore likely to be correlated with income. Furthermore, because they are computed from data reported by respondents, unit values are prone to measurement error. Deaton proposes a methodology that accounts for the endogeneity of unit values and thus allows for the use of unit values in regressions. First, we estimate the unit value v; we then compute an average unit value across each cluster in order to control for measurement error, such that each household in a given cluster faces the same unit value. We write the unit value and alcohol quantity functions as follows: ln vhc = α1 + 1 lnxhc + lnqhc = α0 + n 1 1 lnnhc + ∑ ζj( nj ) + uhc xlnxhc + j 0 Z hc + f c + uhc where vhc is the unit value of alcohol purchased by household h in cluster c; qhc is the quantity of alcohol consumed by household h in cluster c; xhc is the total household expenditure of household h in cluster c; nhc is the household size of household h in cluster c, and the ratio nj n is a control for the proportion of adults in the household; Z hc is a vector of the socio­demographic characteristics of the household; f c is the cluster­specific fixed effect and the uhc is the idiosyncratic error of the household. When we estimate the above equations we obtain the coefficients 1 and x, elasticity of expenditure with respect to the household unit value, and 1 where x is the is the elasticity of household expenditure on alcohol with respect to the quantity of alcohol consumed. We then construct the following equations: y0c = lnqhc − y1c = lnvc − xlnxhc 1 xhc where we have taken averages such that the equation y0c represents the average cluster quantity and equation y1c represents the average cluster unit value. These equations allow us to estimate the effect of market price on consumption because they contain quantities and unit values that have been stripped of the effect of household expenditure. Therefore, if we regress y0c on y1c we will obtain estimate of the price elasticity of demand for alcohol that we can use for the intended purposes in this study (Deaton, 1997). Before we use the above methodology to obtain elasticities, we perform a one­way analysis of variance test to determine the variation of unit values between the clusters. TABLE I : Analysis of Variance Test for Household Clusters Source of Variation Within Groups SS 18067620193 5 10432164585 7 1 3165 4 Total 28499784779 2 3165 5 Between Groups df MS 18067620193 5 F 5482 2 P­valu e F crit 0 4 3295686 Table I reports the results of the ANOVA test for spatial variation between household clusters. The F­score for the test is highly significant, such that we can proceed using Deaton’s unit values in our analysis. Another matter of interest is the regression that will be used in the study. Because many observations in the sample will report a zero quantity value if they do not consume alcohol, the alcohol consumption problem will have a corner solution. This results in a violation of the assumption of normality in the distribution of the data, as the variable quantity will take on the value zero with a positive probability. An appropriate model to use in this case would be the Tobit model developed by Tobin (1958), as it deals well with truncated or censored data, and has been used widely in studies of limited dependent variables (Mroz, 1987; Austin, Escobar & Kopec, 2000; Bellemare & Barrett, 2006). However, as Deaton (1997) explains, the unit values we use in this study can only be derived in the specific method outlined by Deaton (1988), such that we are restricted to the use of OLS in our study. Although this represents a problem for the unbiasedness of our estimates, it is still arguably our best alternative to derive reliable estimates of demand elasticity. Having obtained a method of estimating meaningful price elasticities, we proceed with the following regressions. First, we run a conditional OLS regression for those respondents who are drinkers, which we compare with Tobit regressions featuring similar regressors; this will allow us to say something about the effect that the initial decision to drink has on elasticity estimates of those respondents who choose to drink alcohol. Secondly, we run another conditional OLS regression, this time including interaction terms between the unit value of alcohol and income; this will allow us to say something about whether there are differences in the elasticities of alcohol demand among different income groups. Thirdly, we run a conditional OLS regression including the interaction terms between the level of drinking of a consumer and the unit value of alcohol that they face. Finally, we include interaction terms between the level of drinking of the consumer and their income, to establish if something can be said about the relative elasticities of these groups. 4. Data The primary data source is Wave 1 of the National Income Dynamics Study (NIDS) of 2008. This survey was completed by 7 503 households from different parts of the country. It features a broad set of information on each household’s expenditure, income, and demographic composition, as well as information on the individuals in each household. This makes NIDS data particularly suitable for the present use of estimating various elasticities, as it features information on much of the behaviour we would like to observe. Although Wave 2 of 2012 of the study is also available, we make use of Wave 1 because it is more representative of the population, as it is not subject to attrition like Wave 2. Table II reports the summary statistics of selected variables in the Wave 1 data set, where the base group that has been selected is white males from an urban area. We use survey weights that are specific to the data set to ensure that the sample is representative of the population. We construct the quantity, family size, rural, log of unit value and monthly household income variables at a household level. The remainder of the variables are defined at an individual level. TABLE II MEANS OF THE DATA Of particular importance is the construction of the quantity variable. In the survey, respondents are asked to respond to questions concerning their drinking behaviour, such as how often they drink, and the number of drinks they consume when they drink alcohol. These variables about the frequency and intensity of consumption were aggregated at a household level. Although this results in a loss of precision concerning the behaviour of each respondent, it allows us to construct the unit value of alcohol that the household faces. As is expected from the World Health Organisation’s report on alcohol consumption in South Africa (World Health Organisation, 2011), only approximately 24% of the respondents in the survey state that they consume alcohol. We find that the average number of drinks consumed per household is 368 per annum, which may be considered as light consumption of alcohol; however, among the respondents who are drinkers, the average number of drinks consumed per annum is 1 013, which is considered heavy consumption. Another variable of interest is religion. The survey asks respondents what their religious affiliation is, if any, and the importance of religious activities in their lives. This is used to construct the religion variable, which is unity if religious activities are either important or very important to the respondent, and zero otherwise. In the sample, 89% of respondents consider religious activities as significant to them, which corresponds to previous findings that religion is a significant social factor in South Africa (Garner, 1998, 2000). The high number of abstainers in South Africa can be partially attributed to the importance of religion in society as Islam and certain Christian denominations prohibit the consumption of alcohol. Because NIDS data is survey data, however, it is subject to certain imperfections that present challenges to the present analysis. Firstly, the data contains missing information for some variables where respondents either do not know, cannot recall or refuse to answer the question about the matter at hand. This results in fewer observations, which is a problem of varying degree across the variables. Furthermore, the missing information limits the extent to which we can make manipulations to the data in order to say something interesting about certain aspects of the respondents’ behaviour, as it affects the reliability of the study. For instance, although the data set contains 3 982 households where at least one resident is a current drinker, only 1 495 households indicate that they spend money on alcohol. Additionally, because the data collection process relies on responses from the participating households, NIDS data is prone to measurement error. Information on subjects such as the respondents’ incomes and expenditures rely on the recall of the respondents, which is prone to recall bias. Nevertheless, because of the large sample size and the use of methods such as Deaton’s unit values, NIDS data will still allow us to analyse consumer behaviour in a comprehensive manner. NIDS data features information on whether an individual drinks alcohol, and in the event that they are a current drinker, how much they usually drink. This allows us to observe the two dependent variables of interest, namely the decision to consume alcohol and the quantity of alcohol consumed. Furthermore, the data allows us to create a comprehensive set of independent variables that allow us to explain what determines the decision to drink, and the sensitivity of alcohol consumption to various characteristics of the individual. This will enable us to make comparisons between the sensitivity of alcohol consumption of individuals with different characteristics. 5. Empirical Results 5.1 The drinking decision We first analyse the determinants of the decision to drink, and secondly the determinants of the quantity of alcohol consumed, given that an individual drinks. To do this, we first run a logistic regression in which the dependent variable is consumption, and we use a set of regressors that is similar to that of Manning et al. We adjust the education variables such that they better represent the education levels in South Africa; we adjust the race groups to reflect South African racial groups and we change the geographical regions to make them relevant to South Africa. We also add an indicator variable, religion, that is equal to unity when the respondent considers themselves religious, and zero otherwise. The Logit column of Table III reports the results of the logistic consumption equation. We run this regression at the individual level, such that the female, age, race and religion variables are all at the individual level. The family size, rural, and household income variables are all at the household level, as well as the unit value variable, which is common within each cluster. We report the marginal effects of the equation, where the base group that has been selected is white males from an urban area with a tertiary education and no religious affiliation. We note that female respondents are significantly less likely to drink alcohol than male respondents, while the coefficient on the age variable is positive such that older respondents are more likely to consume alcohol. The quadratic variable for age is significant and has a negative coefficient, such that the effect of age on the probability of drinking is reversed at the age of 50. Interestingly, the possession of some high school education until at least Grade 11 causes a significant reduction in the probability of drinking, more so than the possession of a Grade 12 education. Black respondents are less likely to consume alcohol than white respondents, as are Indian and Coloured respondents. We also find that the religion variable is negative and statistically significant, such that religious respondents are 14% less likely to consume alcohol than those who are not religious. We note that the Christian and Muslim affiliations result in respondents being less likely to drink. TABLE III THE DRINKING DECISION AND CORE REGRESSION Independent Variable Base Female Male Age Black Coloured Indian Religion Christian Muslim Hindu Jewish Traditional ln (Income) Family Size ln(Income).family size Grade 7 Grade 11 Grade 12 Rural Logit ­0.166*** (0.021) 0.014*** (0.001) White ­0.329*** (0.025) White ­0.145*** (0.012) White ­0.157*** (0.17) No religion ­0.140*** (0.17) No ­0.048** affiliation (0.015) No ­0.179*** affiliation (0.013) No 0.065 affiliation (0.067) No 0.067 affiliation (0.082) No ­0.067*** affiliation (0.017) ­0.058* (0.028) ­0.059*** (0.012) 0.006*** (0.002) Tertiary ­0.050** (0.016) Tertiary ­0.066*** (0.014) Tertiary ­0.029 (0.015) Urban ­0.038*** Conditional OLS Tobit ­0.168 (0.253) ­0.001 (0.006) ­0.094 (0.112) ­0.198 (0.13) ­0.08 (0.27) ­0.188* (0.083) ­0.498* (0.227) 0.005 (0.005) 0.062 (0.114) ­0.069 (0.131) 0.112 (0.289) ­0.195* (0.087) ­ ­ ­ ­ ­ ­ ­ ­ ­ ­ ­0.832* (0.025) ­0.076 (0.076) 0.020* (0.01) 0.206 90.123) ­0.075 (0.114) ­0.071 (0.117) ­0.262*** ­0.113* (0.052) ­0.071 (0.059) 0.015* (0.008) 0.141 (0.124) ­0.094 (0.118) ­0.061 (0.123) ­0.231*** Western Cape Limpopo Eastern Cape Limpopo Northern Cape Limpopo Free State Limpopo Kwazulu­Natal Limpopo North­West Limpopo Gauteng Limpopo Mpumalanga Limpopo log of Unit Value ln(Income).ln(unit value) N * p<0.05, ** p<0.01, *** p<0.001 (0.009) 0.167*** (0.028) 0.105*** (0.024) 0.150*** (0.03) 0.179*** (0.03) ­0.035 (0.019) 0.175*** (0.026) 0.095*** (0.025) 0.105*** (0.027) ­0.056* (0.027) (0.057) 0.382** (0.128) 0.236 (0.135) 0.283* (0.137) 0.206 (0.143) 0.112 (0.126) 0.467*** (0.133) 0.119 (0.133) 0.148 (0.145) ­0.832** (0.17) (0.048) 0.293** (0.105) 0.142 (0.109) 0.099 (0.111) 0.092 (0.115) 0.072 (0.098) 0.248* (0.105) 0.04 (0.11) 0.076 (0.115) ­0.463** (0.147) 0.007 0.029 0.017 (0.004) 13489 (0.022) 3484 (0.019) 6241 We note that the coefficient on the unit value of alcohol is negative and significant. This suggests that the price of alcohol that the individual is faced with in the market has a negative effect on their decision to drink, such that a 10% increase in the market price will result in the individual being 0.58% less likely to consume alcohol. We also note that the coefficient on the income of the household is negative, such that individuals with higher incomes are less likely to consume alcohol. This finding is of value for the present study, because it suggests that individuals with lower incomes may indeed be more reliant on alcohol than those with higher incomes. Finally, the coefficient on the family size variable is negative, such that respondents from larger households are less likely to drink. The interaction term between family size and household income, however, is positive. This means that an increase in income for an individual from a larger household will reduce the probability of drinking less than for a respondent from a smaller household, all else equal. This effect of household size on the probability of drinking could be a result of poverty: larger households tend to have lower incomes, such that individuals from larger households may not be able to afford to drink. However, for those individuals who come from larger households, an increase in income will have a smaller effect in reducing their probability of drinking, as the income must be spread across more household members. While this finding suggests that poor individuals are more likely to consume alcohol, a regression of the quantity of alcohol consumed on income is necessary in order to draw inference on whether poverty results in higher consumption of alcohol. 5.2 Core OLS Regression In order to analyse the effect that variables such as income have on the demand for alcohol, we estimate the alcohol consumption equation. In this equation, the variable of interest is the log of alcohol quantity consumed, while a set of regressors similar to those used by Manning et al. is used as explanatory variables. Unlike Manning et al., however, the consumption equation is analysed at the household level instead of the individual level. This is done because NIDS data contains most of the variables of interest, such as expenditures and household income, at the household level rather than the individual level. For the individual­level characteristics such as gender and age, we assign the entire household with the characteristics of the household head. We define the base group in the regression as white households with a male head who has a tertiary education, residing in an urban area in Limpopo. In Table III we compare the findings of the conditional OLS regression with that of the Tobit regression. We find that, although the partial effects on the quantity of alcohol consumed differ between the two methods, both regressions have the same significant variables and the same signs on their coefficients. If the partial effects are significantly different then the OLS coefficients suffer from sample selection bias, such that they will be unreliable estimates. In order to test for sample selection bias, we use the Heckman two­stage model, and find that the coefficient on λ is highly significant. This suggests that sample selection bias indeed is a problem in this equation; however, because Deaton specifies the unit value method using OLS, we proceed using conditional OLS rather than the Tobit model. In the core regression of the consumption equation, we note that although gender and age are significant determinants in the initial decision to drink, they have an insignificant effect on the quantity of alcohol consumed once the respondent chooses to drink. In the Tobit model of the core regression, however, the presence of a female household head has a significant, negative effect on the quantity of alcohol that the household consumes. The geographic location of the household is found to have a significant effect on the quantity of alcohol that it consumes. Rural households are found to drink significantly less than urban households. While other provinces are found to have no significant effect on consumption, households from the Northern Cape, the Western Cape and North­West Province are found to drink significantly more than the base group. This is an unsurprising, as there is evidence suggesting particularly high levels of alcohol abuse in the Western Cape and Northern Cape (Freeman, Parry, 2006). Household income is found to have a significant effect on the level of household alcohol consumption. The income elasticity of demand is negative and inelastic, such that a given increase in income will result in a smaller reduction in the quantity of alcohol consumed. This result differs from that of the study done by Manning et al. (1995), who find that income elasticity is inelastic and positive. At the mean log of unit value and family size, a 10% increase in income will result in a 6.93% reduction in alcohol consumed. This suggests that alcohol is an inferior good. The interaction between income and unit value, which measures whether the price elasticity of demand differs with the level of household income, is found to be insignificant in the core regression. This suggests that the responsiveness of alcohol demand to changes in the price of alcohol does not significantly depend on the level of household income. Interestingly, while family size has a negative effect on the probability of drinking is no longer a significant determinant on the quantity consumed once the respondent chooses to drink. Income is also found to be an insignificant determinant of the quantity consumed. When the two variables are interacted with each other, however, we find the variable to be significant. For two respondents who start off with the same household income, with one respondent belonging to a large household and the other belonging to a small household, a given increase in household income will increase the quantity consumed more for the respondent from a larger household than the one from a smaller household. As with the effect on the initial decision to drink, this may suggest that respondents from poorer households appropriate an increasing share of their income to alcohol consumption, however, this hypothesis would need to be tested directly. We also note that the coefficient on unit value is negative and significant, such that an increase in the unit value of alcohol will result in a decrease in the quantity of alcohol consumed. We compute the price elasticity of demand at the mean log of household income and find that a 10% increase in unit value will result in a 5.98% reduction in quantity consumed, such that the demand for alcohol is price inelastic. Of particular interest in the study will be whether this elasticity differs among light, moderate and heavy drinkers, as well as across different levels of household income. 5.3 Demand elasticity by level of drinking We now consider whether the price elasticity of demand for alcohol differs among different levels of consumption. We define two indicator variables for moderate and heavy consumption, such that the base group is light drinkers. In order to test whether the price elasticity of demand is significantly different across the different levels of consumption, we interact the unit value of alcohol with the indicator variables for level of consumption, and include these in the core regression. The sum of the coefficients on the interaction term and the log of unit value variable is then the price elasticity of demand for that level of consumption. When included in the core regression, the interaction terms are found to be significant and positive. Because the coefficient on the log of unit value is negative while the coefficient on these variables is positive, moderate and heavy drinkers are less responsive to changes in the price of alcohol, such that their demand is more inelastic. These findings are similar to those of Manning et al.’s study, where moderate and heavy drinkers were also found to have more price inelastic demand than light drinkers. Specification 1 in Table IV reports the results for the regression that includes all the variables in the core regression, and the interaction terms between level of consumption and the log of unit value. We note that household income remains a significant determinant of the quantity of consumption, such that an increase in income will result in a decrease in quantity consumed. Family size is also significant and negative, such that an increase in household size has a similar effect to an increase in household income. Furthermore, the interaction between family size and household income remains significant. When we include the interaction terms between the log of the unit value of alcohol and level of consumption, the coefficient on the unit value of alcohol is now the price elasticity of demand for light drinkers. We find that this coefficient is negative and highly significant; the demand for alcohol among light drinkers is found to be elastic, such that at the mean level of household income, a 10% increase in the unit value of alcohol will still result in a 15.9% reduction in the quantity of alcohol consumed. The interaction between the indicator variable for moderate drinkers and the log of unit value gives us the price elasticity of demand among moderate drinkers. This variable is also highly significant and negative, but the elasticity is less elastic, with ɛ=­1.213 at the mean level of household income. TABLE IV DIFFERENT SPECIFICATIONS OF THE DEMAND FOR ALCOHOL Independent Variable ln (Income) Family Size ln (Income).family size Rural ln (Unit Value) Ln (Income). ln(unit value) Moderate. ln(unit value) Specificatio n1 Specification 2 Specification 3 ­0.176*** (0.047) ­0.152* (0.061) 0.029*** (0.008) ­0.152*** (0.046) ­ 0.81*** (0. 025) ­0.097** (0.041) 0.380*** (0.050) ­0.137* (0.058) ­0.1 (0.076) 0.023* (0.01) ­0.274*** (0.057) ­0.833 (0.025) ­0.0987 (0.003) ­0.203*** (0.051) ­0.188** (0.067) 0.034*** (0.009) ­0.178*** (0.05) ­0.824*** (0.025) ­0.098*** (0.003) Heavy. ln(unit value) Midlowincome.ln(unitvalue ) Midhigh income. ln(unit value) High income. ln(unit value) Low.moderate.ln(unit value) 0.695*** (0.041) ­ 8 . 5 9 e ­ 0 6 ­9.15e­06 (1.28e­06) (1.27e­06) ­2.197e­06 (1.72e­06) Low.heavy. ln(unit value) Midlow.light. ln(unit value) Midlow.moderate. ln(unit value) 6 . 8 5 e ­ 0 6 * Midlow.heavy. ln(unit value) Midhigh.light. ln(unit value) Midhigh.moderate. ln(unit value) 6 . 3 5 e ­ 0 6 * 0.477* (0.134) 0.619*** (0.088) ­2.3e­05 (1.48e­06) (2.88e­06) 1.71e­05*** (2.19e­06) ­2.22e­05 (1.48e­06) (2.88e­06) Midhigh.heavy.ln(unit value) 1.72e­05*** (2.18e­06) ­4.34e­04*** (1.92e­05) High.light. ln(unit value) High.moderate.ln(unitvalu e) High.heavy. ln(unit value) R­squared N * p<0.05, ** p<0.01, *** p<0.001 0.42 3494 0.118 3494 2.71e­04*** (4.46e­05) 2.0e­04*** (3.4e­05) 0.324 3494 The elasticity among heavy drinkers is unsurprisingly the most inelastic, with ɛ=­0.898 at the mean level of household income. This suggests that heavy drinkers will adjust their quantity of alcohol consumed less following a change in price, such that any increase in price will result in alcohol becoming a larger proportion of their expenditure. Of particular concern then is whether lower­income households tend to be heavy­drinking as well, as this will imply that poorer households will be hurt most by increases in the price of alcohol. We find that the interaction between household income and unit value is significant and positive, such that an increase in the price of alcohol will result in a smaller reduction in the quantity of alcohol consumed for households with larger incomes. This would suggest that lower­income households are more sensitive to changes in the price of alcohol than higher­income households. Of particular interest is whether this finding is similar across different levels of income. 5.4 Demand elasticity by income group In order to analyse whether the sensitivity to the price of alcohol differs across different levels of household income, we define four income groups: low, midlow, midhigh, and high income households, where the groups are defined according to the income quartiles within the data set. We compare the elasticity of demand for alcohol among the different income groups by interacting the unit value of alcohol with indicator variables for the different income groups, where the base group is low­income households. Specification 2 in Table IV reports the results of a regression with variables from the core regression, and the interaction terms between income and the log of unit value. We find that income still has a significant negative effect on the quantity of alcohol that a household consumes, although the income elasticity of demand is now less elastic than in prior regressions. The interaction between household income and unit value, however, is no longer significant. We also note that the coefficient on the unit value of alcohol is now insignificant. Furthermore, the interaction terms between the income group of the household and unit value are all highly insignificant, such that an increase in the price of alcohol will not have a significantly different effect on the quantity of alcohol consumed across the different income groups. We therefore cannot conclude that lower­income households are more sensitive to changes in the price of alcohol than higher­income households. 5.5. Demand elasticity by income group and level of drinking Although the above analysis suggests that households do not differ in their sensitivity to changes in price across the income groups, a question of interest is whether the price elasticity of demand for alcohol differs between households with different levels of consumption across the different income groups. That is, we are interested in whether two households that are both heavy drinkers will have different sensitivities to changes in the price of alcohol if they belong to different income groups. In order to test this hypothesis, we define interaction terms between the log of the unit value of alcohol, indicator variables for the level of consumption, and indicator variables for the different income groups. Our base group here is low income households that are light drinkers, such that the coefficient on the unit value variable reports the price elasticity of demand for this group. Specification 3 in Table IV reports the results for the regression with all the variables in the core regression, and the interaction terms between the log of unit value, level of consumption and income group. We note that household income is a significant determinant of the quantity of alcohol consumed by the household. The income elasticity of demand is now even more elastic, such that an increase in income will now result in a larger decrease in the quantity of alcohol consumed. The interaction term between household income and the log of unit value of alcohol is also significant; a given increase in the price of alcohol will have a more significant effect on consumption among higher­income households. Of particular interest is the price elasticity of demand among the income groups across the different levels of consumption. We find that the coefficient on the unit value of alcohol is negative and highly significant. We also find that all the interaction terms between the log of unit value, income and level of consumption are highly significant, with the exception of the midlow­income, light drinking and midhigh­income, light drinking households. Although economically minor, the coefficients on midlow­income, moderate drinking; midlow­income, heavy drinking and high­income, light drinking households are all negative, such that their demand for alcohol is more elastic than the base group. The coefficients on the other interaction terms are all positive, such that these groups have more inelastic demand than the base group. Because we interact the log of unit value with level of drinking as well as income, the price responsiveness of alcohol will now depend not only on the level of consumption, but also on the level of income at which we evaluate the price elasticity. That is, for a given level of drinking, the price elasticity of demand will change as income changes. We therefore need to choose some meaningful level of income at which we may compare the magnitudes of the price elasticities of the different types of consumers. TABLE V PRICE ELASTICITIES OF DEMAND FOR DIFFERENT INCOME GROUPS AND LEVELS OF DRINKING Income 750 1200 2000 5250 Income category Low income Midlow income Midhigh income High income Level of Drinking Light Light Light Light Ԑp ­1.473 ­1.519 ­1.569 ­1.664 750 1200 2000 Low income Midlow income Midhigh income Moderate Moderate Moderate ­0.996 ­1.519 ­1.569 5250 High income Moderate ­1.66 750 1200 2000 5250 Low income Midlow income Midhigh income High income Heavy Heavy Heavy Heavy ­0.854 ­1.519 ­1.569 ­1.66 Table V reports the price elasticity of demand for each type of consumer for the median household income of that particular income group. At the median level of household income, the price elasticity of demand for all light drinkers is found to be highly elastic; a 10% increase in the price of alcohol will result in approximately a 15% reduction in the quantity of alcohol consumed by all light drinkers. All moderate drinkers are also found to have a highly price­elastic demand for alcohol, although low­income, moderate drinkers are less price­elastic in their demand. Heavy drinkers are also found to have a similar price elasticity of demand when the income group is controlled for. The coefficients we are most interested in, however, are those on low­income, heavy drinking households. The coefficient on low­income, heavy drinking households is the least price­elastic in the sample. A 10% increase in the price of alcohol will only result in a 8.5% reduction in the quantity of alcohol consumed by these households, such that their demand is inelastic. Low­income households that are heavy drinkers are thus two times less elastic in their demand for alcohol than higher­income households that consume the same quantity of alcohol. 6. Conclusions The above finding supports the hypothesis that alcohol control policy that is based on changing the price of alcohol will affect consumers in different ways. Heavy­drinking consumers are unsurprisingly found to be less elastic in their demand for alcohol, such that an increase in the price of alcohol will result in them appropriating a larger portion of income to expenditure on alcohol. While we find no significant difference in the elasticities of lower­income and higher­income households, alcohol is still found to be an inferior good when the level of drinking is not controlled for, such that lower­income households appropriate a larger portion of their income to alcohol expenditure. Furthermore, among the group of households in the study that may be classified as heavy drinkers, low­income households have a highly inelastic demand for alcohol, while the elasticity of high­income households is not significantly different from that of light drinkers. This implies that increases in the price of alcohol will not reduce the quantity of alcohol consumed by low­income, heavy drinking households, but will rather result in them reallocating their income towards expenditure on alcohol. The same policy, however, will be significantly more effective in other groups of consumers. This suggests that any policy based solely on the increase of excise taxes will have no significant impact on reducing the social and economic cost of alcohol abuse among lower­income areas in society. In particular, since many of the problematic behaviours associated with alcohol use is specifically associated with alcohol abuse, increasing the price of alcohol will not effectively reduce the abuse of alcohol and will thus fail to alleviate this burden to society. We note that there are several limitations to the study. Firstly, as Manning et al. (1995) noted in their analysis, respondents may underreport their alcohol consumption, which would lead to biased results. Respondents may also incorrectly report their expenditures on alcohol or household income, which would further lead to biased estimates of price and income elasticities. Secondly, the National Income Dynamics Survey data does not contain information on the market prices faced by households. This drawback imposes the use of estimation methods in order to obtain price elasticities, resulting in the loss of precision in the analysis. Despite these limitations, however, the results are importance to alcohol control policy, as they suggest the possible impact of current policy considerations. REFERENCES Abel, E.L., Kruger, M.L. & Friedl, J. 1998. How Do Physicians Define ?Light,??Moderate,? And ?Heavy? Drinking? Alcoholism: Clinical And Experimental Research. 22(5):979­984. DOI:10.1111/J.1530­0277.1998.Tb03692.X. Austin, P.C., Escobar, M. & Kopec, J.A. 2000. The Use Of The Tobit Model For Analyzing Measures Of Health Status. Quality Of Life Research. 9(8):901­910. Bellemare, M.F. & Barrett, C.B. 2006. An Ordered Tobit Model Of Market Participation: Evidence From Kenya And Ethiopia. American Journal Of Agricultural Economics. 88(2):324­337. Casswell, S. & Maxwell, A. 2005. Regulation Of Alcohol Marketing: A Global View. Journal Of Public Health Policy. 26(3):343­358. Chaloupka, F.J., Grossman, M. & Saffer, H. 2002. The Effects Of Price On Alcohol Consumption And Alcohol­Related Problems. Alcohol Research And Health. 26(1):22­34. Chow, G.C. 1960. Tests Of Equality Between Sets Of Coefficients In Two Linear Regressions. Econometrica. 28(3):591­605. Chung, C., Dong, D., Schmit, T.M., Kaiser, H.M. & Gould, B.W. 2005. Estimation Of Price Elasticities From Cross­Sectional Data. Agribusiness. 21(4):565­584. DOI:10.1002/Agr.20065. Colditz, G.A., Dejong, W., Emmons, K., Hunter, D.J., Mueller, N. & Sorensen, G. 1997. Harvard Report On Cancer Prevention, Vol 2, Prevention Of Human Cancer. Cancer Causes & Control. 8:S1­&. Cox, W.M. & Klinger, E. 1988. A Motivational Model Of Alcohol Use. Journal Of Abnormal Psychology. 97(2):168­180. Deaton, A. 1987. Estimation Of Own­And Cross­Price Elasticities From Household Survey Data. Journal Of Econometrics. 36(1):7­30. Deaton, A. 1997. The Analysis Of Household Surveys: A Microeconometric Approach To Development Policy. Johns Hopkins University Press. Dee, T.S. 2001. Alcohol Abuse And Economic Conditions: Evidence From Repeated Cross­Sections Of Individual­Level Data. Health Economics. 10(3):257­270. DOI:Http://Www3.Interscience.Wiley.Com/Cgi­Bin/Jhome/5749. Duffy, J.C. 1993. Alcohol Consumption And Control Policy. Journal Of The Royal Statistical Society.Series A (Statistics In Society). :225­230. Fogarty, J. 2006. The Nature Of The Demand For Alcohol: Understanding Elasticity. British Food Journal. 108(4):316­332. Fogarty, J. 2010. The Demand For Beer, Wine And Spirits: A Survey Of The Literature. Journal Of Economic Surveys. 24(3):428­478. DOI:10.1111/J.1467­6419.2009.00591.X. Gallet, C.A. 2007. The Demand For Alcohol: A Meta‐Analysis Of Elasticities*. Australian Journal Of Agricultural And Resource Economics. 51(2):121­135. Garner, R.C. 1998. Religion And Economics In A South African Township. Garner, R.C. 2000. Safe sects? Dynamic religion and AIDS in South Africa. The Journal of Modern African Studies. 38. pp41­69 Ground, M. & Koch, S.F. 2008. Hurdle Models Of Alcohol And Tobacco Expenditure In South African Households1. South African Journal Of Economics. 76(1):132­143. Gruenewald, P.J., Ponicki, W.R., Holder, H.D. & Romelsjö, A. 2006. Alcohol Prices, Beverage Quality, And The Demand For Alcohol: Quality Substitutions And Price Elasticities. Alcoholism: Clinical And Experimental Research. 30(1):96­105. DOI:10.1111/J.1530­0277.2006.00011.X. Heather, N. & Robertson, I. 1997. Problem Drinking. Oxford University Press Oxford:. Hentz, J.J. 1996. Democratization In South Africa: The Elusive Social Contract By Timothy D. Sisk Princeton University Press, 1995. Pp. Xvi + 342. $39.50. £29.95. The Journal Of Modern African Studies. 34(04):739. DOI:10.1017/S0022278X0005597X. Jones, S.C., Hall, D. & Munro, G. 2008. How Effective Is The Revised Regulatory Code For Alcohol Advertising In Australia? Drug And Alcohol Review. 27(1):29­38. DOI:10.1080/09595230701499175. Kasolo, F. 2011. Socio­Economic Context: South Africa, 2011. World Health Organisation. Kessler, D.A. 2005. Alcohol Marketing And Youth: The Challenge For Public Health. Journal Of Public Health Policy. 26(3):292­295. Kitchin, P.D. 1981. Some Socio­Economic Determinants Of Alcohol Consumption: A Research Note. International Journal Of Social Economics. 8(5):31­35. DOI:Http://Www.Emeraldinsight.Com/Ijse.Htm. Ledermann, S. 1964. Alcool, Alcoolisme, Alcoolisation. Presses Universitaires De France. Lemmens, P.H. 1993. Regularity In The Distribution Of Alcohol Consumption. Journal Of The Royal Statistical Society.Series A (Statistics In Society). :231­235. Lemmens, P.H. 1993. Regularity In The Distribution Of Alcohol Consumption. Journal Of The Royal Statistical Society.Series A (Statistics In Society). :231­235. Manning, W.G., Blumberg, L. & Moulton, L.H. 1995. The Demand For Alcohol: The Differential Response To Price. Journal Of Health Economics. 14(2):123­148. DOI:10.1016/0167­6296(94)00042­3. Mroz, T.A. 1987. The Sensitivity Of An Empirical Model Of Married Women's Hours Of Work To Economic And Statistical Assumptions. Econometrica: Journal Of The Econometric Society. :765­799. Nelson, J.P. 2010. Alcohol Advertising Bans, Consumption And Control Policies In Seventeen OECD Countries, 1975–2000. Applied Economics. 42(7):803­823. Odejide, O.A. 2006. Alcohol Policies In Africa. African Journal Of Drug And Alcohol Studies. 5(1):27­39. Odejide, A.O. 1978. Alcoholism: A Major Health Hazard In Nigeria? Nigerian Medical Journal : Journal Of The Nigeria Medical Association. 8(3):230­232. Parry, C., Rehm, J. & Morojele, N. 2010. Is There A Causal Relationship Between Alcohol And HIV? Implications For Policy, Practice And Future Research. African Journal Of Drug And Alcohol Studies. 9(2). Parry, C.D. 2005. South Africa: Alcohol Today. Addiction. 100(4):426­429. Parry, C., Burnhams, N.H. & London, L. 2012. A Total Ban On Alcohol Advertising: Presenting The Public Health Case. SAMJ: South African Medical Journal. 102(7):602­604. Paton, A. 1985. The Politics Of Alcohol. British Medical Journal (Clinical Research Edition). 290(6461):1­2. Room, R., Babor, T. & Rehm, J. 2005. Alcohol And Public Health. The Lancet. 365(9458):519­530. Saffer, H. & Chaloupka, F. 1998. Demographic Differentials In The Demand For Alcohol And Illicit Drugs. Saffer, H. & Dave, D. 2002. Alcohol Consumption And Alcohol Advertising Bans. Applied Economics. 34(11):1325­1334. DOI:10.1080/00036840110102743. Skog, O. 1985. The Collectivity Of Drinking Cultures: A Theory Of The Distribution Of Alcohol Consumption.*. British Journal Of Addiction. 80(1):83­99. DOI:10.1111/J.1360­0443.1985.Tb05294.X. Taylor, J.R. & Helzer, J.E. 1983. The Natural History Of Alcoholism. In The Biology Of Alcoholism. Springer. 17­65. Tobin, J. 1958. Estimation Of Relationships For Limited Dependent Variables. Econometrica: Journal Of The Econometric Society. :24­36. Toomey, T.L. & Wagenaar, A.C. 1999. Policy Options For Prevention: The Case Of Alcohol. Journal Of Public Health Policy. :192­213. Tremblay, V.J. & Okuyama, K. 2001. Advertising Restrictions, Competition, And Alcohol Consumption. Contemporary Economic Policy. 19(3):313­321. DOI:Http://Www.Blackwell­Synergy.Com/Loi/Coep/. Vaillant, G.E. 1983. The Natural History Of Alcoholism. Harvard University Press Cambridge, MA. Vaillant, G.E. & Hiller­Sturmhöfel, S. 1996. The Natural History Of Alcoholism. Alcohol Health & Research World. Wilsnack, R.W., Vogeltanz, N.D., Wilsnack, S.C. & Harris, T.R. 2000. Gender Differences In Alcohol Consumption And Adverse Drinking Consequences: Cross­Cultural Patterns. Addiction. 95(2):251­265. DOI:10.1046/J.1360­0443.2000.95225112.X. World Health Organization 2011. Global Status Report On Alcohol And Health. Geneva: World health organization. 286. Yen, S.T. & Jensen, H.H. 1996. Determinants Of Household Expenditures On Alcohol. Journal Of Consumer Affairs. 30(1):48­67. DOI:10.1111/J.1745­6606.1996.Tb00725.X.