Tobacco consumption and policy in the United Kingdom

Abstract

In recent years, several structural changes and sequences of shocks have occurred in the market for tobacco in the UK, including an explosive growth in smuggling. This study examines whether it is still possible to estimate a reliable, plausible tobacco demand equation from time series data for the UK whilst allowing for all of the various shocks and structural changes. A second objective of this study is to use the estimated demand function to evaluate tobacco policy in the UK, including the scope for using tax increases to achieve reduced consumption and increased revenues. It is found that smuggling has diminished the revenue yield of higher rates of duty, but total consumption of tobacco has been reduced. In addition to the introduction of further measures to control smuggling, policy may need to place more emphasis in the future on health campaigns and smoking restrictions.

Acknowledgements

I should like to thank an anonymous referee and the editor for their helpful comments, which have led to considerable improvements in the paper. I am responsible for any remaining errors and shortcomings.

Notes

¹ See the survey by Chaloupka and Warner (2000). The econometric study of consumer demand is perhaps the oldest element of applied economics. It is, therefore apposite that many of the excellent modern contributions to the literature on the demand for tobacco and other consumer goods have appeared in the Applied Economics journals: see the references cited in footnote 4 and elsewhere in this paper.

² See, for example, Galbraith and Kaiserman (1997) and Gruber et al. (2003).

³ This study employs aggregate time series data rather than micro data for a major concern is to estimate the effects of the huge growth in smuggling in recent years of smuggling on the aggregate consumption of tobacco, market price and income elasticities of demand and tax revenues. Data on smuggling is not published at the micro level, so it would be much more difficult to shed light on the above issues using household budget data.

⁴ For examples of applications of the Becker-Murphy model to the empirical study of demand for different types of habit forming goods, see Chaloupka (1991, 1992), Keeler et al. (1993), Becker et al. (1994), Conniffe (1995), Waters and Sloan (1995), Duffy (1996a), Olekalns and Bardsley (1996), Cameron (1997, 1999), Grossman et al. (1998), Bardsley and Olekalns (1999), Bentzen et al. (1999), Escario and Molina (2001) and Fenn et al. (2001). Of these papers, only Duffy (1996a) relates to tobacco consumption in the UK. Chaloupka et al. (2000), Chaloupka and Warner (2000) and Gruber and Kőszegi (2001) contain critical reviews of the RA hypothesis and other economic theories of addiction.

⁵ The number of smokers is calculated as the product of the ‘smoking prevalence (or participation) rate’ and the UK population aged 15 and over. The smoking prevalence rate is measured as the proportion of the total population aged 15 and over who smoke cigarettes (manufactured or hand rolled). This series is calculated as an unweighted average of the separate figures for males and females (Source: Statistics of Smoking in the UK (Tobacco Research Council); Social Trends (Stationery Office), various issues]. The population data are taken from the Annual Abstract of Statistics, various issues. The smoking prevalence and population series are published at annual frequency. In this project, prior to their use in estimation, they are interpolated to quarterly frequency using the procedures described in Boot et al. (1967), which have been programmed in FORTRAN by the author.

⁶ See Chambers (1999). The official estimate in 2000 was that ‘currently around 18 per cent, nearly one in five cigarettes smoked, is smuggled’, and the share of hand rolling tobacco (HRT) is about 75 per cent: HM Customs and Excise (2000). However, later estimates of HM Customs and Excise (2005) put the estimated value of the illicit market share for cigarettes in 2002–2003 at 15%, whilst the share for smuggled HRT in that period is estimated (by new ways of calculating cross-border shopping) to be ‘only’ 57%.

⁷ As noted by a referee, the decline in real tobacco expenditure in the late 1990s may reflect to some extent product substitution, such as increased consumption of cheaper hand rolling tobacco. Since a separate time series on HRT expenditure is not available, no further analysis of this issue could be undertaken here, but it certainly needs to be borne in mind when evaluating the results of this and other studies.

⁸ The residuals in STAMP are the standardized one-step ahead prediction errors or innovations.

⁹ Part (c) of Fig. 3 is a graph of the estimated seasonal component after re-arrangement into four different series, one for each season. In other words, this part of Fig. 3 comprises annual plots for the individual seasonal effects.

¹⁰ The test is a joint test of significance of the s − 1 seasonal effects. Under the null hypothesis of no seasonality, the large sample distribution of the test statistic ST is χ²(3). In the case of a stochastic seasonal, the test may not be appropriate and it must be interpreted carefully. It may indicate an insignificant seasonal effect at the end of the sample period, but that does not preclude significant seasonal effects at earlier dates. For stochastic seasonality, it is important to evaluate the test in conjunction with a full plot of the seasonal. This point applies to tobacco consumption where the seasonal effect has diminished over time with market demand becoming less dependent on seasonal events and behaviour.

¹¹ For this study, advertising on ‘other non-durables’ covers the following Media Expenditure Analysis Ltd categories: Beer; Spirits; Wine; Food; Soft Drinks; Cosmetics and Toiletries; Pharmaceuticals; Clothing; and Household Stores and Services.

¹² Analysis of the auxiliary residuals in the estimated models of demand which included stochastic trends failed to detect outliers that needed to be modelled through the use of interventions or dummy variables (which can take various forms over time): see Koopman et al. (2000). This is not altogether surprising. The UK market for tobacco has not been affected by a few, isolated health shocks. Rather, it has been affected by myriad shocks over the years, including numerous health reports at home and abroad, the introduction and repeated redesign of warning labels on tobacco products, a ban on TV advertising, a complete ban on all advertising in the media, and the development of a general, anti-smoking culture through health education, campaigns and articles and programmes in the media. As Jones (1989) writes, ‘health scares have a cumulative impact, perhaps with the diffusion of new attitudes and behaviour driven by a process of social interaction, and that the initial impact will be amplified as time goes on’ (p. 140). Rather than trying to identify and estimate the effects of shocks with a collection of dummy variables and various time trends in a huge data mining exercise, the author decided that it would make for better econometric practice to represent the on-going, cumulative effects of this multitude of exogenous shocks through the inclusion of a single, variable stochastic trend in the model.

¹³ If the true model ought to include a stochastic trend, but estimation is based on a specification that uses instead a deterministic trend, then tests of the effects of various variables on tobacco consumption may yield misleading conclusions. See Nelson and Kang (1984).

¹⁴ It must be emphasized that a stochastic trend component needs to be introduced into a dynamic regression model only when the trend in a non-stationary dependent variable cannot be fully explained by other explanatory variables. The stochastic trend acts as a proxy for other immeasurable influences upon the dependent variable. Thus, in the demand equation for UK spirits estimated by Koopman et al. (2000), the stochastic trend is interpreted as accounting for changes in tastes and habits that cannot be measured explicitly. Another latent variable, the state of technology, has been represented by a stochastic trend in some factor demand studies: see Harvey et al. (1986) and Slade (1989).

¹⁵ Harvey and Scott (1994) argue that incorrect modelling of seasonality may have a serious adverse effect on an equation's performance. They attribute the breakdown in the late 1970s and 1980s of the error correct specification for the UK aggregate consumption function to the misspecification of seasonality as a deterministic process. In that study, the consumption function appeared to be rendered stable by the inclusion of a stochastic seasonal component.

¹⁶ Advertising studies often include amongst the right hand side variables a cumulative advertising stock variable, calculated as a weighted average of recent flow advertising expenditures. The use of this variable permits testing for the possible build up over time of a cumulative advertising influence upon demand. In this case, the estimates that employed the stock term did not lead to conclusions that differed from those based on a flow specification.

¹⁷ The potential for this problem to occur is seen most easily in the case of the least-squares regression estimates of a relationship . Working from the normal equations, it is easily shown that each of the estimates of θ₁ and will be equal approximately to 0.5 when the variable C follows a random walk (proof available on request). In practice, this tendency may be masked by the presence of other correlated regressors in the relationship. Even so, the results in Table 2 clearly bear out the prediction that time-series based estimates of the coefficients θ₁ and on C_{t −1} and C_{t+ 1}, respectively, may be close to 0.5. This tendency is also manifest in many other time series studies of the Rational Addiction Hypothesis, although those studies also indicate that the estimates can deviate somewhat from the 0.5 values if other explanatory variables are included or estimation methods other than OLS are used: see, for example, Bardsley and Olekalns (1999, Table 1), Becker et al. (1994, Table 2), Bentzen et al. (1999, Tables II, III, and IV), Fenn et al. (2001, Table 1), and Olekalns and Bardsley (1996, p. 1103).

¹⁸ A more appropriate test for residual serial correlation than the Box-Ljung statistic in a model that contains lagged dependent variables would be the Lagrange Multiplier statistic in its modified form as an F-statistic. This affects only tests for models M1 and M2, not M3. The LM statistic is not available in STAMP. However, the Box-Ljung statistic can still be used with some confidence when establishing the result that serial correlation is at least considerably reduced when moving from M1 to M2, and eliminated quite definitely when moving M2 to M3.

¹⁹ In their meta-analysis of 45 published estimates, Andrews and Franke (1991) calculate a weighted mean price elasticity of −0.363. In their survey paper, Chaloupka and Warner (2000) write of a distribution of price elasticity estimates that centres closely on −0.4. Chambers (1999) estimates the long-run price elasticity in the UK to be −0.29. Chaloupka (1991) estimates the long-run US price elasticity to be in a range from −0.27 to −0.48. Duffy (2001) estimates a UK tobacco price elasticity of about −0.4. The estimate in Duffy (2002) is −0.5 and in Duffy (2003) it is −0.407. Conniffe (1995) reports a price elasticity estimate for Ireland of −0.31.

²⁰ Andrews and Franke (1991) calculate a weighted mean income elasticity of +0.360. Duffy (2001) estimates a variety of complete demand systems, with the associated income elasticities for tobacco falling in the range +0.41 to +0.57. In Duffy (2003), the long-run income elasticity for tobacco is put at +0.403. Conniffe's (1995) estimate for Ireland, though, is much lower at +0.18. Chambers (1999) estimated the income elasticity for tobacco in the UK to be +0.275.

²¹ The stability condition for the difference Equation 5 is : see Becker et al. (1994, p. 413).

²² Since the Wu F-statistics in Table 4 fail to reject the null (that the explanatory variables are orthogonal to the equation disturbances), it seems worthwhile considering OLS as well as IV estimates.

²³ Results were obtained by employing the package PcGive, version 9.10 (Hendry and Doornik, 1999).

²⁴ Similar results and conclusions were reported recently for Canada by Galbraith and Kaiserman (1997). See also Gruber et al. (2003).

²⁵ It is possible, however, that the official estimates of smuggling understate the scale of the problem. Then the behaviour over time of the price (and income) elasticity estimates in part (b) of Table 5 may be biased by the influence of the omitted component of smuggled consumption.

²⁶ The product term is a measure of the legitimate, duty-paid tobacco sales in volume terms.

²⁷ Real tobacco tax revenues resumed their trend decline in the financial year 2003/2004: see HM Customs & Excise (2005).

²⁸ In some published studies, authors discuss the relationship between tax rates and revenues where both are measured in nominal terms. That seems inappropriate. Just as the theory of consumer behaviour seeks to explain demand in real terms by reference to real prices and real income, so it seems more appropriate to analyse fiscal policy in real terms. As the UK data show in Table 6, nominal tax revenues may not decline significantly (and can even rise) as the rate of tobacco duty is increased, but this does not preclude a fall occurring in the much more relevant real tax revenues (which has occurred in recent years as real tax rates have bitten more deeply).

²⁹ The official campaign against smuggling may have achieved some success in 2002. The reversal in that year of the trend rise in the market share of smuggled goods was, however, short-lived. The market share of smuggled tobacco goods rose again between 2002 and 2003 from 16.7% to 19.1%.

³⁰ H.M. Customs and Excise (2000) estimates that smuggling costs £2.5 billion per annum in lost tax revenue, about 25% of all tobacco revenue (excise duty and VAT) due.

³¹ The RA hypothesis has been criticized on both empirical and theoretical grounds. A commonly encountered problem, for example, is the presence of highly implausible discount rates implied by the estimates: see, for example, Cameron (1997, 1999). Baltagi and Griffin (2001) note that ‘before it [the rational addiction model] can be widely accepted, plausible and statistically significant estimates of the implied discount rate are needed. Based on [Becker et al. (1994)] and our results, aggregate panel data do not seem likely to provide sharp estimates of the discount rate. The most promising approach appears to be microdata as discussed by Chaloupka (1991)' (p. 454). In his time series study of the Irish consumpton of tobacco, Conniffe (1995) tested and rejected the RA model.  Conniffe (1995) is amongst those who have criticized the RA hypothesis for assuming that smokers have perfect foresight. Gruber and Kőszegi (2001) have also questioned the validity of a key assumption of the RA model, that individuals’ preferences are time consistent, which is at odds with strong evidence from psychological experiments on the nature of choice over time. The authors allow consumers to be time inconsistent, so that they repeatedly cheat on their professed intentions to kick the habit. These sorts of criticism remind us that analysis of addiction is never likely to be easy, least of all when using aggregate data.  The view of the addicted consumer as a rational, far-sighted individual with exponentially discounted, time consistent preferences may have appeared to some to be an attractive theoretical paradigm with interesting implications, but the theory's practical relevance is something that has not yet been finally determined by researchers. As Viscusi (1992) noted, the population of smokers may not be homogeneous but comprise instead different types of individuals located on a spectrum ranging from the perfectly rational to the opposite extreme. Then the characteristics of the average smoker, and the specification of the best model to analyse the behaviour of that individual, becomes an empirical matter. This suggests, in other words, that the RA model may not be the best model to apply in all circumstances.

³² Chaloupka and Warner (2000) draw attention to the drawbacks of using collinear time series in investigating addiction models in their excellent survey of the literature.