A Comparative Demand System Analysis of Alcoholic Beverage Sector in Japan

Transcription

1 A Comparative Demand System Analysis of Alcoholic Beverage Sector in Japan Makiko Omura Faculty of Economics, Meiji Gakuin University Shirokanedai, Minato-ku, Tokyo , Japan Telephone: This article examines the effects of tax policies on the alcoholic beverage sector in Japan through demand analysis, utilizing data from 1948 to Three types of demand analysis are conducted, namely, almost ideal demand system (AIDS), quadratic AIDS (QUADS), and dynamic AIDS (DAIDS) with/without demographic characteristics, for five types of alcoholic beverages. The analysis suggests that preferential tax rates may be beneficial in boosting the sectoral performance of certain types of alcoholic beverages. In both tax policy and demand system analysis, Japanese hard liquor shōchu, and in DAIDS, beer, are found to be the most ideal taxable items in the Ramsey sense. While model comparisons suggest that the DAIDS formulation is the most theoretically coherent approach with robust results in our case, the sensitivity of results according to the type of price data and model specifications is also reported. Keywords: alcohol tax; demand system analysis; AIDS; QUAIDS; DAIDS; demographic characteristics; Japan 1. Introduction In many countries, government policies have played a major role in the alcoholic beverage industry. The importance of liquor taxes has been examined by numerous researchers worldwide, largely to investigate two issues: (1) the role of these taxes in mitigating the adverse effects and social costs of 1

2 alcohol consumption, such as health problems and vehicle accidents (Cook and Moore 2002; Chaloupka, Saffer, and Grossman 1993), and (2) their role in raising government revenues (Grossman et al. 1993). Such studies appear to provide sufficient justification for the existence of liquor taxes and suggest that rate increases may be appropriate, although some studies cast doubt on the claimed effectiveness of such taxes (Kenkel 1996; Mast et al. 1999). The major rationale for liquor taxes in Japan has always been on the tax revenue side. 1 However, regardless of the government s intentions in taxing alcohol, it is a matter of concern as to whether the tax has any effect on the production and consumption of alcohol, as significant effects would suggest the possibility of using the liquor tax for various policy purposes. It is also argued that excise taxes, especially taxes on items such as alcohol and tobacco, are less distortionary. According to the optimal consumption tax model proposed by Ramsey (1927), tax rates on goods should be inverse to the price elasticity of demand for such goods thus, inelastically demanded goods should be taxed more heavily. Some studies suggest that alcohol consumption is price inelastic, particularly for heavy drinkers (Manning et al. 1995). By contrast, other studies suggest that alcohol consumption responds well to price changes, with negative own-price elasticities varying in degree depending on the type of alcoholic beverage considered, as observed by Cook and Moore (2002). Estimates of price elasticities vary widely, with beer typically found to have the lowest elasticity. According to research conducted by Elder et al. (2010), who compiled past studies on the effects of alcohol taxes, the price elasticity of demand for alcohol has median values of for beer, 1 To mitigate the adverse effects of alcohol consumption, Japan has implemented other regulations, such as increased severity of punishment for drunk driving in terms of both criminal charges and social sanctions, rather than using taxes as a tool to curb consumption. Few studies estimating the social costs of alcohol-related problems in Japan find considerable costs to society, although these studies use base estimates from studies in the U.S., which can affect estimation results drastically (Nakamura et al. 1993; see also Kaji 2013, for more information on various studies). 2

3 -0.79 for spirits, and for wine, although they are measured in different ways. However, some studies show variations in elasticity estimates. Applying static and dynamic Almost Ideal Demand System (AIDS) models, Eakins and Gallagher (2003) estimate the own-price elasticity of beer to be ~ -0.42, depending on the model applied, whereas the values for spirits and wine are found to be ~ and ~ -0.36, respectively. The authors also compile past studies showing wider ranges of elasticities, such the own-price elasticity of beer varying from to However, no standard errors (SE) or statistical significance for the elasticity estimates are provided in the studies of Elder et al. (2010) and Eakins and Gallagher (2003). Andrikopoulos and Loizides (2010) apply a dynamic AIDS (DAIDS) model to data on Cyprus and find that beer is price elastic with statistical significance, although wine and brandies are not. While estimates can vary depending on the data and analysis methods, the estimation of elasticities continues to be an important topic as it can have specific policy implications. In Japan, the alcohol industry was once a major contributor to tax revenues. The Japanese government has implemented several significant legal changes regarding alcohol production and consumption, including changes in tax rates. There are, however, few studies of the effects of liquor taxes. One study estimates the price elasticity of saké at 0.58, shōchu at -0.15, beer at -0.63, whisky at -0.35, and low-malt beer (referred to here as fizzy drinks) at 0.61 (Takahashi et al. 2009), although the method used is rather ad hoc in that elasticities are calculated based on the differences in consumed quantities of goods at currently prevailing prices and at hypothetical prices excluding the liquor tax. 2 Given the gaps in the past literature and the differing methods applied, this study examines the 2 Positive and negative signs are added by the author from their results table because these were not specified. No SE are provided. 3

4 effects of tax policies on alcohol sector and provides a comparative analysis of demand systems for alcohol beverages, using Japanese data from 1948 to The article consists of three major parts: first, a brief overview of the Japanese liquor tax system; second, the analysis of tax effects applying panel and time-series estimations; and third, the analysis of demand systems through the double-log model and the AIDS model with its several variants, namely quadratic AIDS and DAIDS with/without demographic characteristics. 2. A Brief Overview of the Liquor Tax 3 There have been several major changes in the liquor tax in recent decades. Currently, there are 10 primary types of alcoholic beverage classified by liquor tax law: (1) saké, (2) synthetic saké, (3) shōchu (Japanese spirits), (4) beer, (5) whisky and brandy, (6) wine, (7) spirits, (8) liqueur, (9) fizzy drinks (low/no-malt beer), and (10) other alcoholic beverages. 4 The evolution of liquor-specific (volumebased) consumption is shown in Graph 1. The liquor tax in Japan has a long history and has undergone numerous changes since The existing tax law was created in 1953 (S28), 5 with a significant revision in 1962 (S37) establishing the base for the current tax structure (Japan Cabinet Office (JCO) 2000). The general tendency of liquor taxes was for a higher tax rate to be applied to expensive alcohols via ad valorem and class-specific taxes. The ad valorem tax applicable to expensive saké, whisky, and wine was abandoned in 1989 (H1). The specific tax rate in nominal terms generally decreased during the 1950s until this trend reversed 3 The information here is based on the liquor tax evolution table ( ), which is available from the National Tax Agency. 4 Japanese categorization of liqueur includes various sweet/sour cocktails and shōchu cocktails. Note also that the category of wine includes wine-like beverages made from other fruits, while spirits includes gin, vodka, and rum, as well as distilled alcohol. 5 The expressions S# and H# in parentheses signify the year according to the Japanese-era name. We note this because all legal and official systems in Japan utilize these expressions. 4

5 in 1968 for most alcohol types. The specific tax was then increased in several stages until a legal revision in April 1989 (H1) that implemented significant decreases in the tax rates for wine, whisky, beer, and first-class saké, along with the abrogation of the special saké classification and the class system for wine and whisky. According to the Japanese government, tax revisions, especially since 1989, have aimed to achieve neutrality, simplicity, and fairness in taxes across different alcohol types (JCO 2000). The H1 revision in 1989 was followed by a unified tax rate system for saké in 1992 (H4) and for shōchu in May The reduction and simplification of liquor tax seems reasonable given the introduction of a consumption tax in April 1989, initially set at 3. 6 [Graph 1] 3. The Tax Data The actual amount of taxes levied on each alcohol type-class is complex because it depends on the actual ethanol content and because there have been different rules and exemptions characterizing the applicable tax rates. There is a liquor-specific base tax rate for each category or sub-category/class of alcohol, and the actual tax rate is increased according to the ethanol content above the base. This means that a tax rate per 1º of ethanol is not necessarily uniform, even for the same type of alcohol. Further, no data is available on class-wise production and consumption for each type of alcoholic beverage. We thus utilize a type-wise average tax rate across classes as a proxy variable for the tax rate, derived by dividing the taxed value by the taxed quantity of each type. 7 An aggregate taxed value encompasses 6 The ad valorem tax can be viewed as being replaced by the sales tax. The sales tax, which was made applicable to all commodities in principle, was subsequently increased to 5 in Consistent with the argument of Chetty et al. (2009), the effect of a sales tax may have been smaller than that of a price increase because the sales tax was not initially included in the price tag. The inclusive sales tax, the salient tax rule, was introduced in 2004 (H16) after 15 years. 7 Average tax rates for whisky, wine, spirits, and liqueur during ( ) are extrapolated from the corresponding ratio data in 1963, as only aggregate figures are available for these types. 5

6 an ad valorem tax and a specific tax. In Graph 2, the average tax of major beverage types in real terms are shown, namely for saké, shōchu, beer, wine, spirits, and whisky. They show largely decreasing trends, except in the case of whisky, which exhibits a large increase roughly from 1970 to Although the tax rate is highest for whisky per kilolitre, the rate is highest for beer for 1º of alcohol and in terms of the proportion of tax to average commodity price (Table 1). By far the lowest tax rate in all respects is that for wine. [Graph 2] [Table 1] Reviewing the average prices of alcoholic beverages, we find nominal prices generally increasing with economic development, whereas real prices are decreasing for most alcohols, as shown in Graph 3. 8 During the period of economic growth until the bubble burst around 1993, expensive commodities were sought and the observed trends for whisky and wine are understood to reflect increased imports of high-priced items. Whereas the real price of whisky decreased fairly constantly after the bursting of the bubble through 2005, the real price of wine did not decrease as much, but reverted to its increasing trends around the mid-1990s, likely boosted by the polyphenol boom of [Graph 3] Several features are worth noting with regard to the tax-price ratio, calculated as a ratio of average real tax value to real retail price per liter for saké, shōchu, beer, whisky, and wine as shown in Graph 4. Most taxratio values are relatively stable or slightly decreasing, except for those of shōchu and 8 We utilize two data sets for household consumption expenditure for the period for saké, shōchu, beer, whisky, and wine. Expenditures are used to derive the average price per liter for each of these alcohols. 6

7 whisky. The tax rate for shōchu was raised considerably in 1997 and 1998 (H9 and H10) as a result of foreign criticism that it was too low compared with those for other, often imported, spirits, such as whisky and brandy. For whisky, we observe a steep increase in tax-price ratio during the mid-1960s and then a steep decrease in the mid-1990s, both resulting from tax revisions. Two particularly high tax-price ratio values are those for whisky and beer, where that for beer can be traced to an outdated view from the prewar period, when beer was regarded as an imported luxury commodity. In addition, the fact that beer is produced by large companies and consumed in large quantities makes the collection of liquor taxes relatively easy for the government. Indeed, beer shows the highest tax-price ratio in most periods. [Graph 4] 4. The Effects of Tax Policy As we have observed, government policy appears to have significantly influenced supply and demand in alcoholic beverages. We therefore attempt to estimate the likely effects of tax policy on the production and consumption of different types of alcohol, using nine types of alcoholic beverage: (1) saké, (2) synthetic saké, (3) shōchu, (4) beer, (5) whisky and brandy, (6) wine, (7) spirits, (8) liqueur, and (9) fizzy drinks. A summary of the variables used is provided in Table 1. We define two dependent variables: taxed quantity (taxq) and consumption (cons). 9 The taxq figures are essentially the quantities of domestically produced alcohol supplied in the market. Because of the complicated system of production measurement and evolving regulations, published production figures diverge from the actual quantities supplied or traded in the market. The consumption quantities also include imported 9 The correlation coefficients for these three variables are high: that for taxq and cons is

8 alcohols. 10 The JCO s stated objective of achieving neutrality, simplicity, and fairness in liquor taxes, particularly in recent policy revisions, may indicate the liquor tax policy as historically being a revenue-generating tool for the government. This implies the possibility of tax policy having been influenced by market performance factors, such as sectoral growth, rendering such policy as endogenous to the system, although not contemporaneously. While alcohol tax revisions do not present any clear evidence of sector growth-based tax strategies, 11 past sectoral growth, calculated as the change in the logged three-year moving average of the type-wise tax values (ΔlnMA3taxv i ), is added to the estimation in order to avoid possible endogeneity arising from an omitted variable problem. Additionally, we consider the final consumption expenditure per capita growth (Δlnfcepc) to capture general economic trends, as well as polyphenol and shōchu boom dummies. To ensure stationarity, numerical variables are converted into a log difference, in annual growth rate form. Given serially correlated and/or heteroskedastic error terms, panel estimations across different types of alcohols are conducted using a feasible generalized least squares (FGLS) estimator, allowing for panel-specific autocorrelation and heteroskedasticity across panels. We estimate the following basic model with/without various independent variables. (1) Y it = α + β 1 ΔlnMA3taxv it + β 2 Δlnaveragetax it + β 3 Δlnfcepc t + γδd taxchng + θδd boom + u it, where u it = ρ u it-1 + ε t, ε t ~ IID(0,σ 2 ε ), ρ <1, 10 For most alcohol types, the proportions of imported alcohols are not large, with wine and whisky as exceptions. Wine has the highest share of imports, a proportion that has been constantly increasing, with imports accounting for more than 50 of wine consumption since For whisky, the proportion has been approximately 20, but it has exhibited a continuously decreasing trend. 11 Perhaps a possible exception is the tax rate increase for fizzy in 2003 (H15) and 2006 (H18) after its introduction in 1992 (H4), although an equitable tax concern was also likely to occur as its tax rate remained considerably lower compared with that of beer. A similar argument can be applied to the third or new genre beer, categorized either in liqueur or in other. Regressing ΔlnMA3taxv i on to tax variables with fixed effect estimator suggests that H6tax may be affected by past sectoral growth at the 5 significance level, but not other tax revisions. 8

9 and where Y it is Δlntaxq it or Δlncons it. Two dependent variables Y it are considered: Δlntaxq i as the growth rate of the taxed quantity (taxq); and Δlncons i as the growth of consumption (cons) for alcohol type i, including imported alcohols. The following independent variables are considered: Δlnaveragetax i as the growth of average real tax rate (averagetax) for alcohol type i; ΔlnMA3taxv i (sector) and Δlnfcepc t (expenditure) as described above; D taxchng as a vector of dummy variables denoting the year of important tax changes (S37 (1962), H1 (1989), H6 (1994), H15 (2003), and H18 (2006)); and D boom as a vector of dummy variables for the red-wine polyphenol boom ( ) and the shōchu boom ( ). Taking into account the possibility of tax policy being affected by sector, we also conduct estimations with an interactive term sector*d taxchng. Note that tax-price ratio (taxratio) is not included here because of its consistent nonsignificance in the estimation and its possible multicollinearity problem. Given that different genres of alcohol may respond differently to policy changes, group-wise panel estimations are conducted for the following groups: (G all ) all alcohol types; (G dinner ) dinner alcohol (saké, synthetic saké, shōchu, and wine with 12º~15º of alcohol), where shōchu is normally consumed in diluted forms; (G hard ) hard liquor (shōchu, spirits and whisky, with approximately 25º~50º of alcohol); and (G light ) light alcohol (beer, fizzy drinks, and liqueur with roughly 5º of alcohol). However, the grouping may not be appropriate since tax rates vary within the same genre and each alcohol type is expected to have a different response to a tax policy. Hence time-series estimations for each alcohol type are also conducted, applying the Prais-Winston GLS estimator, given the nature of serial correlation and nonevidence of lagged dependent variables Serial correlation is tested via the Wald test of the serial correlation hypothesis proposed by Wooldridge (2002) and Drukker (2003). The non-existence of lagged dependent variables is robustly accepted by a series of serial correlation tests for an auxiliary regression 9

10 4.1 Estimation Results: The Effects of Tax Policy The panel estimation results are presented in Table 2 for four groups for taxq and cons, where the even numbered models employ interactive tax dummies, sector*d taxchng. Across groups and across taxq and cons, we observe significant positive effects of sector robustly apart from G light. The expenditure exhibits positive effects for taxq in G all and G dinner and for cons in G all and G light without interactive terms at the 5 significance level, with higher magnitudes for taxq. This indicates higher impacts of economic growth on production vis-à-vis consumption, with variability in impacts across alcohol groups. The coefficient for averagetax is negative and significant at the 5 level only in G all, although the coefficients are mostly negative as expected. A possible reason for not finding this variable statistically significant for other subgroups may be a lack of variability or opposing forces within each group. With regard to tax policies, S37tax is positive and significant only in its interactive form, in G hard taxq. More robust findings are the negative impacts of H1tax and sector* H1tax for G dinner and G hard both in taxq and cons. While H1tax reduced the tax rate for most alcohols, it raised the rate for shōchu and thus the negative effects may be attributable to reduced supply and demand for shōchu, or to the introduction of a consumption tax in Japan of 3, or both, rather than the shōchu tax increase. Another tax policy with statistically significant effects is H6tax, which increased the tax for saké, synthetic saké, shōchu, beer, and wine, while it exhibits negative effects for cons of G dinner, and taxq and cons of G hard. Finally, H15tax, which raised the tax for synthetic saké, wine, fizzy drinks, and other alcohols, shows significant positive effects on G hard taxq with a magnitude of 0.73, suggesting a for residuals that also includes a lagged dependent variable. The variables applied are in the first-differenced log form. The theoretical discussion of Prais-Winston GLS is provided in Davidson and MacKinon (1993). Applying the Autoregressive Moving Average Model (ARMA) produces highly similar results. 10

11 possible production shift towards hard liquor that was unaffected by this tax revision. Although not presented here, estimations without sector produce highly similar results. There is no evidence that booms had any significant effect for these groups. [Table 2] The results of the time-series estimations for each alcohol type are provided in Table 3. Two taxq estimation results and one cons result are presented for each type of alcohol, with/without sector and uniform D taxchng. 13 Significant and positive effects of sector are found for beer taxq and cons, whisky taxq, and wine taxq and cons. The expenditure for saké and beer are found significant and positive for both taxq and cons with large magnitudes, suggesting that a 1 increase in the economic growth rate leads to around a 1.5 increase in the taxq/cons growth rates. In particular, beer production and consumption, both of which rank at the top in terms of quantities, appear to correspond well with general economic performance. On the other hand, averagetax has significant negative effects on cons and taxq for saké and beer, and on whisky cons. Although not significant, the averagetax coefficients are unexpectedly positive in all shōchu estimations. This finding for shōchu may appear to be anomalous; however, we observe in the graphs that both production and consumption increased for shōchu along with the rise in its tax rate during the 1990s (consumption shown in Graphs 1). Given the historically low tax rate for shōchu, the rate increase itself may not have affected its performance, as it was still the least expensive means of obtaining ethanol (see Table 1). Turning to tax policies, the significant negative effects of H1tax are seen for saké cons, shōchu taxq and cons, and whisky taxq. 13 The tax dummy in the panel analysis varies depending on the alcohol type ( 1 if applicable to that type, 0 otherwise); here in time series analysis, the tax dummy was uniformly 1 if that tax change was applicable in a given year, regardless of alcohol type. The estimated results utilizing growth_sector*d taxchng are highly similar to the ones here, both in terms of magnitudes and significance. 11

12 Although tax rates are essentially reduced for higher class saké and whisky, these more expensive alcohols might have been hit by the new universal consumption tax of 3. On the other hand, the nominal tax rate for shōchu was substantially increased by 30~44 depending on its classification, amounting to a real tax rate increase of 20.4 between 1988 and 1989, on top of the consumption tax. This caused its actual figure for taxq to decline by a remarkable 51 in With respect to the H6tax of 1994, which increased the tax rates on saké, synthetic saké, shōchu, beer, and wine, significant negative coefficients are estimated for saké cons and shōchu taxq at the 5 significance level. Another tax revision with a significant positive coefficient is that of H18tax on wine taxq, despite its tax rate increase. This is likely due to wine s steady popularity and its growth in domestic production, and the fact that its tax rate still remains the lowest of all the alcohols even after this tax rise. Another feature of wine is the significant impact of the polyphenol boom both in taxq and cons. However, the shōchu boom is not found to have produced any similar effects on shōchu. The estimation results provide strong evidence that saké and beer are responsive in both production and consumption to tax rate rises, while shōchu and wine are not. In terms of general economic circumstances, strong positive effects are exhibited particularly for saké and beer. These results confirm that different alcohols respond differently to tax policies, despite the general findings of negativity in response to tax rate and positivity in response to sectoral growth and general economic growth in panel estimations. [Table 3] 5. Demand System Analysis: Expenditure Elasticity and Price Elasticity of Demand The estimation results presented in the previous section suggest that tax policies have significant and varied effects on the production and consumption of alcoholic beverages. The differing effects of 12

13 economic growth, tax rates, and tax policies suggest that the income and price elasticities of demand are likely to vary among alcohol types. Given that both the liquor and sales taxes are invisible to consumers, we investigate the effects of price change in alcohol on alcohol expenditures. 14 As the theory of optimal consumption taxation proposed by Ramsey (1927) suggests that welfare loss is minimized if the tax rate is set higher for inelastically demanded goods, we estimate the expenditure and price elasticities of demand for alcoholic beverages by applying the double-log model and, following Deaton and Muellbauer (1980), the AIDS model and its variants, namely, AIDS, QUAIDS and DAIDS with/without demographic characteristics. We utilize annual household expenditure data for the period. 5.1 Double-Log Estimations To estimate elasticities, we first calculate the expenditure and constant price elasticities of demand in β log-log multiplicative form, q it = α X t 1 β p t 2 e γ t 1, translated into log-linear form, utilizing aggregate longitudinal data on commodity-wise household expenditures across alcohol types (panel estimation) and per alcohol type (time-series estimation). Thus, the estimable equation is as follows: (2) lnq it = lnα + β 1 lnx it + β 2 lnp it + γ 1 t = α + β 1 lnx it + β 2 lnp it + γ 1 t, where lnq is the log of the quantity purchased in milliliters, lnx is the log of the current total household consumption expenditure (as a proxy for income), lnp is the log of the current price of each alcoholic beverage, and t accounts for time effects. As price variable, we apply three different types, (1) average purchase prices (p_average) and (2) representative retail prices within given alcohol classes (p_retail), 14 Note, however, that historical tax rates have not necessarily translated well into alcohol prices in the cases of some items. For available data between 1963 and 2011, partial correlations between tax rates and the real average prices of saké, shōchu, beer, whisky, and wine are 0.358, 0.697, 0.980, 0.871, and 0.498, respectively. 13

14 and (3) extrapolated prices using the consumer price index (CPI) with retail prices in 2000 as the base (cpi) (see Table 4 for the detailed derivation of each price variable). 15 While p_average essentially represent unit values of goods and are different from prices because they reflect consumers choice of quality, Deaton (1997) states their usefulness in indicating price variation. As for p_retail, these are prices for selected representative products chosen by the government statistics bureau and it is not certain how well they reflect actual prices of different products in the same alcohol groups that consumers face. 16 The estimated coefficients β 1 and β 2 are the partial expenditure elasticity and price elasticity of demand, respectively. The model is estimated with either time effects or one of the demographic characteristics, given their high degree of correlation. The demographic factors are the number of household members (h_num) and the age of the household head (hh_age), both available in average terms across households in any given year. As noted above, the panel estimation is performed with FGLS, permitting panel-specific autocorrelation and heteroskedasticity across panels, and typewise time-series estimation is performed with the Prais-Winston GLS estimator. Summaries of the variables and the estimated results are provided in Table 4 and Table 5, respectively. [Table 4 & Table 5] Based on consumer demand theory, income/expenditure elasticities are expected to be positive for normal goods, whereas own-price elasticities are expected to be negative. The results in Table 6 show significant positive expenditure elasticities in all panel estimations and for beer, whisky, and wine except for p_average, in time-series estimations. Shōchu, although only one is significant at the 5 15 Deaton discusses the issues concerning the application of unit values and regional prices in a developing country context (1997, p ) 16 The retp are those of central Tokyo which are not bad representation of national price variation. 14

15 level, has negative coefficients in all cases, suggesting the possibility of being an inferior good. In terms of magnitudes, we observe that beer, whisky, and wine except for p_average case are fairly elastic, with respect to total household consumption expenditure. With respect to own-price elasticities, we find statistically significant negative inelastic coefficients for panel estimations, negative fairly elastic coefficients for saké p_average, beer p_average and cpi, whisky p_average, and negative elastic coefficients for wine p_retail and cpi. On the other hand, saké has two nonsignificant positive coefficients, and shōchu has all positive price elasticities, one significant at the 1 and another significant at the 5 level. This result suggests that while beer, whisky, and wine are normal goods, shōchu may be a Giffen good, although it could be that the quality of shōchu has improved, accompanied by increases in both price and demand. Another possible reason for this finding is that shōchu still remains inexpensive relative to other alcohol types. The magnitudes of coefficients and in some cases the statistical significance can differ depending on which price variable is applied, indicating that caution in its choice is required. For household characteristics, h_num is found to significantly reduce expenditure on shōchu and wine and significantly increase that on whisky measured in p_retail, while hh_age is found to significantly reduce the expenditure on saké, beer, and whisky, while increasing that on shōchu and wine. The time trend exerts positive effects on shōchu and wine expenditure, and negative effects on saké, beer, and whisky, in all cases at the 1 significance level. 5.2 AIDS, QUAIDS, and DAIDS Estimation Models Although the results for the double-log model appear to be plausible, this model has been criticized as crude and inconsistent with utility theory except in special cases (Deaton 1997). We thus apply AIDS, 15

16 QUAIDS, and DAIDS with/without socio-demographic factors. Brief descriptions of each model are provided below, followed by the estimation results. In AIDS models, demand systems are specified in terms of expenditure shares of different types of commodities, in this case, alcoholic beverage expenditure shares of different types of alcoholic beverages. A household s expenditure share for good i is defined as w i p i q i X -1, where p i is the price of alcohol type i, q i is the quantity of alcohol type i purchased or consumed, and X is total expenditure on all alcoholic beverages in the demand system. With this definition of X, we have "&' w " = 1, where K is the number of alcoholic beverages in the system. Using an indirect utility function, with utility expressed in terms of price p and total expenditure X, we can express the expenditure share equations as follows: (3) w " = α " + γ "+ lnp + + β " ln { 4 +&' }, 5 p where p is the vector of all prices and P(p) is a price index defined as follows: (4) lnp p = α 9 + "&' α " lnp " + 1/2 "&' +&' γ "+ lnp " lnp +. Because the expenditure function is linearly homogeneous of degree zero in prices and total expenditure, the following conditions must hold: (1) adding-up: " α " = 1, " γ "+ = 0, " β " =0; (2) homogeneity: + γ +" = 0 ; and in addition, (3) Slutsky symmetry: γ "+ = γ +", for any i j. As stated by Deaton and Muellbauer (1980), α 0 is generally difficult to estimate and is thus assigned an a priori value as the minimum level of expenditure required for subsistence when all prices are unity. Accordingly, α 0 is set at 4.9 throughout the analysis. 17 Based on the estimates, the expenditure (as a 17 Noting that alcohol may not be a necessity and that some households may not consume it at all, our data are only in aggregate form, with above-zero alcohol consumption. Additionally, AIDS does not allow for corner solutions in which all commodities are consumed in positive amounts (Deaton 1997, p. 304). 16

17 proxy for income) and the price elasticities are calculated. The expenditure elasticity is given by: (5) e i =1+ β i w i -1, and the own/cross price elasticity is given by: (6) η ij = δ ij + { γ ij β i (α j + "&' γ "+ lnp " )} w -1 i, where δ ij is the Kronecker delta, with δ ij =1 if i=j and δ ij =0 otherwise. 18 These elasticities can be derived in a straightforward manner by differentiating the log of the purchased quantity of item i with respect to the log of expenditure (dlnq i /dlnx) and by differentiating the log of the purchased quantity of item i with respect to the log of the price of item i (dlnq i /dlnp i ), respectively, applying the chain rule in both cases. These forms of elasticity have been presented by numerous authors, including Ray (1980) and Green and Alston (1990). An AIDS model with socio-demographic factors incorporates demographic characteristics using the scaling technique introduced by Ray (1983). Application of a scale allows for incorporation of household characteristics (z) into an expenditure analysis across households that vary. The scaling function m 0 is composed of two multiplicative factors, a basic component and a price- and utilityvarying component, m 0 (p, z, u) = m 0 (z) φ(p, z, u), where the first component captures increases in a household s expenditures as a function of a vector of household characteristics z and the second component φ represents the dependence of the general scale on the structure of relative prices and utility, capturing changes in consumption patterns as a function of z. The estimable equation of the expenditure share takes the following form: 4 +&', D E (z)5 p (7) w " = α " + γ "+ lnp + + (β " + θ " z) ln 18 The price elasticities presented are Marshallian or uncompensated elasticities, in which Hicksian or compensated elasticities, which solely represent price/substitution effects, can be obtained directly from the Slutsky equation, η ij C = η ij + X i_mean w i. 17

18 where ϑ i indicates the effects of price and utility variations on scale, for which the summation condition requires that Σ k i=1 ϑ ri = 0 for column r = 1 s. The basic scale of the household characteristics vector z is set at m 0 (z) = 1 + σ z, where σ is a vector of estimable parameters representing a basic equivalent scale (Ray 1983). The incorporated demographic factors are either those of h_num or hh_age. The QUAIDS model was devised by Banks et al. (1997) to make the AIDS model consistent with a more realistic Engel curve. An additional term for the quadratic log of expenditure enables the demand function to differentiate responses to goods based on income level, such that goods can be luxuries or necessities, depending on income level. The functional form is as follows: (8) w " = α " + γ "+ lnp + + β " ln 4 + G H +&' [ln 4 5 p I p 5 p ]L, where the final term, added to the original AIDS equation (3), is the quadratic log of expenditure O N" divided by the price index, with the Cobb-Douglas price aggregator b p = "P' p ", and the additional term λ i which requires that Σ k i=1λ i = 0 (see Banks et al. (1997) for details). Checking for likely model fit through nonparametric kernel regressions with Epanechnikov functions, as well as Gaussian specification as in Banks et al. (1997), the results indicate that the shares of item expenditures vis-à-vis the log of alcohol expenditures take a quasi-linear form for beer and whisky (increasing), a quasi-concave form for saké, shōchu and wine, although the estimated functions do not fit well the observed data. For QUAIDS with demographic characteristics, the expenditure share equation, with additional terms, is as follows (Poi 2012): 4 +&' + D E z 5 p (9) w " = α " + γ "+ lnp + + (β " + θ Q "z) ln G H [ln 4 I p R p,z D E z 5 p ]L. The expenditure and own/cross price elasticities for this functional form in the QUAIDS version are 18

19 as follows: (10) e i = 1+[ β i + θ " z + LG H ln 4 ] w i -1 I p R p,z D E z 5 p (11) η ij = δ ij,+ [γ ij [β i + θ " z + LG H ln 4 ](α j + γ I p R p,z D E z 5 p "+ lnp " ) "&' (N H TUQ H z)g H I p R p,z [ln 4 D E z W p ]L ] w i -1. Finally, we consider two variants of the DAIDS model; the first, presented by Ray (1984), incorporates past expenditure terms, following the linear habit formation models of Phlips (1972) and Pollak (1970). The estimated expenditure share equation and the price index become the following: +&' }, 5 \ p (12) w "X = α " + (γ "+ + θ "+ X XP' )lnp +X + (β " + η " X XP' )ln { 4 \ (13) lnp X p = θ 9 + "&' δ " X ",XP' + "&' α " lnp "X + 1/2 "&' +&' (γ "+ + θ "+ X XP' )lnp "X lnp +X, where adding-up restrictions require that for all j: " α " = 1, " γ "+ = " θ "+ = 0 and β " " = " η " = 0. The homogeneity restrictions require that for all i: " γ "+ = " θ "+ = 0. The symmetry restrictions require that for all i and all j: γ "+ = γ +", θ "+ = θ +". Here, θ ij and η i capture the degree to which past total group expenditure affects current expenditure. Whereas θ ij is defined in subsistence utility terms, η i is defined in additional utility terms or bliss à la Deaton-Muellbauer. By contrast, δ i, called the memory coefficient by Pollack (1970), captures the effects of previous purchases of individual items. 19 Allowing for autocorrelated disturbances with autocorrelation coefficients ρ i, where the estimation equation s disturbances are assumed to take the form, u it =ρ i u it-1 + ε t, with ε t ~ IID (0, σ 2 ), the expenditure share equation and the elasticity formulas become: +&' (14) w "X = ρ " w "XP' + α " 1 ρ " + (γ "+ + θ "+ X XP' )lnp +X + (β " + η " X XP' )ln { 4 \ 5 \ p } 19 To achieve stability, Pollack assumes that δ i is the same for all i and that δ i =[ 0,1 ). However, if such restrictions are applied, then certain parameters become inestimable within the equation systems used in the present article. 19

20 ρ " (γ "+ + θ "+ X XPL )lnp +XP' ρ " (β " + η " X XPL )ln { 4 \`a +&', 5 \`a p } (15) e i = 1+( β i +η i X t-1 ) w i -1, (16) η "+ = δ "+ + (γ "+ + θ "+ X XP' ) β " + η " X XP' α + + "&'(γ "+ + θ "+ X XP' )lnp " w P' ". Another DAIDS is proposed by Lewbel (1989) and elaborated by Yen and Chern (1992). This is a flexible demand system allowing for serial correlations, where the expenditure share function and the price index take the following forms: +&' (17) w "X = [α " + γ "+ lnp +X + β " lnp X p { γ "+ + β " 1 + γ "+ lnp "X +&' "&' +&' }lnx X ] ( 1 + "&' +&' γ "+ lnp "X ) P', (18) lnp X p = α 9 + "&' α " lnp "X + 1/2 "&' +&' γ "+ lnp "X lnp +X, and with serially correlated errors u it =ρ i u it-1 + ε t, we have: (19) w "X = ρ " w "XP' + F( x X, p X, β ) ρ " F(x XP', p XP', β ), where F( ) is essentially the right-hand side of (17). In this form of equation, demographic characteristics are incorporated by specifying the α i parameter as a function of vectors of these variables z: (20) α " = a "9 + f d "f z f. The corresponding expenditure and price elasticities are calculated as: (21) e " = +&' γ "+ + β " 1 + "&' +&' γ "+ lnp "X 1 + "&' +&' γ "+ lnp "X P' w -1 (22) η "+ = δ "+ + {(γ "+ + β " (α + + "&' γ "+ lnp "X "&' γ "+ lnx X "&' γ "+ w "X } 1 + "&' +&' γ "+ lnp "X ) P' P' w ". The estimation of all AIDS functions fits a system of nonlinear equations using iterative feasible generalized nonlinear least squares (IFGNLS), imposing the set constraints. Note that we do not 20

21 present the results of the first version of DAIDS with lagged expenditure in this paper, as the estimated elasticities are of implausible magnitudes, likely suggesting the need for further specifications of the autoregressive nature of log expenditure/log prices. 20 The basic DAIDS model based on the second specification is estimated with item-wise autocorrelation coefficients (ρ i ) while the DAIDS with demographic characteristics is estimated with one autocorrelation coefficient (ρ) due to constraints on the estimable number of parameters vis-à-vis the number of observations. 5.3 AIDS, QUAIDS, and DAIDS Estimation Results Tables 6, 7, and 8 present the estimation results and the corresponding expenditure and own price elasticities of AIDS, QUAIDS, and DAIDS models, without demographic variables, and with h_num and with hh_age, for each set of price data (p_average, p_retail, and cpi). All elasticities are calculated at the mean values of the concerned variables. As seen earlier, there are some degree of divergences in estimated coefficients depending on which price data are used as well as on the model specifications. From a broad perspective, the most striking features comparing across models are the differences between static AIDS/QUAIDS and DAIDS results. There is also variability in the results of p_average and those of the other two prices in static AIDS/QUAIDS. Looking at the estimates for β i, which is supposed to be negative for necessities and positive for luxuries in AIDS, we observe mixed findings; among them, negative coefficients for saké and shōchu, and positive coefficients for beer are robust in AIDS and DAIDS. As for QUAIDS, the estimated β i and λ i produce mixed results which coincide the poorly fit kernel regressions of raw data noted earlier. The only fairly robust results are those of negative β i and negative λ i for beer p_retail and cpi. 20 With further restrictions such as applied in Phlips (1972), it may be estimable. 21

22 Focusing on expenditure elasticities, both static models produce only nonsignificant results apart from one beer estimation each in AIDS and QUAIDS both with hh_age, with positive elastic results and significant at the 5 level. There are several saké and beer coefficients found to be significant at the 10 level, although only beer produces coherent results with positive expenditure elasticities. Ignoring statistical significance, expenditure elasticities for shōchu are mostly negative while those for beer, whisky and wine are robustly positive. 21 On the other hand, all expenditure elasticities in DAIDS models are estimated to be positive and statistically significant at the 1 level. The expenditure elasticities for saké, shōchu, and wine are robustly elastic, being 0.97~1.35, 0.89~1.73, and 1,10~1.95, respectively, while beer is moderately inelastic being 0.75~0.91, and whisky varies from moderately inelastic to elastic being 0.70~1.28. For own-price elasticities, none is found to have statistically significant coefficients in AIDS or QUAIDS estimations. Ignoring statistical significance, 12 saké and 12 shōchu results out of 18 ownprice elasticity estimates are positive, contrary to theoretical predictions. These findings again resonate with those found in the double-log estimations. The AIDS and QUAIDS own-price elasticity estimates for beer, whisky, and wine are mostly negative as predicted by the theory, although none is significant, with seven estimates for beer barely significant, at the 10 level. Turning to DAIDS estimates, majority of own-price elasticity estimates are statistically significant, and all statistically significant results are negative including those for saké and shōchu. Apart from shōchu, the results are fairly robust. The most price-elastic alcohol is wine with its figure varying from to Granting variability, the least price elastic alcohol is generally beer mostly being around -0.79~-0.34 with one 21 These results match fairly well with the results of double-log estimations utilizing the total expenditure data, despite the differences in model justifications and applied data. 22

23 extreme estimate of -2.68, or is possibly shōchu, with four statistically nonsignificant cases. There are a fair number of statistically significant cross-price elasticities, where positive coefficients indicate substitutes and negative ones indicate complements. As we have seen there are essentially two demographic variables in AIDS and QUAIDS estimations, ϑ i and σ, where the former indicates the effect of the relative price of i on the scale effect of the latter. While σ-h_num and σ-hh_age are found to be statistically significant with negative coefficients in AIDS with p_average, SE could not be estimated in any of the p_retail and cpi estimations apart from σ-hh_age in QUAIDS cpi which is positive and nonsignificant. The missing SE, together with incoherent estimates of ϑ i suggest the likely inadequacy of model specification. In d-h_num are statistically significant except for beer, where the effects are positive for saké and whisky, negative for shōchu, and mixed for wine depending on the price variable. As for d-hh_age, there are variability in statistical significance, yet its effects are significant and negative for all saké and two wine estimates. In both d-h_num and d-hh_age applications, p_average model produces estimates with different signs vis-à-vis p_retail and cpi models. A robust finding of negative effects on saké in DAIDS is the only observation that match those of double-log estimates. Comparing the results of AIDS, QUAIDS, and DAIDS, DAIDS estimation produces the most coherent results with the theoretical predictions on elasticities. DAIDS estimates are also the most consistent across estimations with different price variables and demographic specifications. Additionally, the appropriateness of the dynamic structure is supported by the fact that the serial correlation coefficient ρ is highly significant in all the estimations, while the appropriateness of the static quadratic structure is not supported by the nonparametric kernel regressions. The demographic 23

24 variables do not seem to have much explanatory power. We have also seen that the choice of price variable, as an average purchase price, retail price, or CPI-extrapolated price, affects the estimation results particularly in double-log and static AIDS models, while affecting the magnitude of coefficients in DAIDS. [Table 6, 7, 8] 6. Conclusions Utilizing data from 1948 to 2011, we observed that liquor tax rates were once discriminative towards expensive alcoholic beverages through ad valorem taxes and class systems, although this system was abolished by early The liquor tax policy revisions have had differing implications for each alcohol type, as observed in regressions investigating the effects of tax policies on alcoholic beverage production and consumption through panel and time-series analysis. In the tax policy analysis, sectoral growth is found to have significant positive effects on production and consumption in general, and individually on those of beer and wine. The effects of final consumption per capita are found to be positive and statistically significant for production and consumption in general across all alcohol types and for saké and beer individually, while the effect is negative for shōchu although statistically nonsignificant. Rather crudely translating general economic performance into the household total expenditure and alcohol expenditure, the results of double-log and static demand analysis using data from 1963 to 2011 roughly coincide with these findings, where expenditure elasticities are generally positive except for frequent cases of shōchu and some cases of saké having negative elasticities, although without statistical significance. In DAIDS, expenditures elasticities are all positive and statistically significant where wine is the most expenditure elastic and beer is generally the most 24

25 expenditure inelastic alcohol. In terms of price, liquor tax rates in panel/times-series regressions are found to have negative effects in general and across all alcohol types apart from shōchu where the effects are positive, yet nonsignificant or significant only at the 10 level, a finding that is corroborated by the nonsignificant positive own-price elasticity estimates for shōchu frequently evidenced in static AIDS/QUAIDS. Nonetheless, for saké, the results of tax policy analysis are only partly confirmed by those of static AIDS/QUAIDS. In DAIDS estimation, which is the most theoretically coherent model in our case, wine is the most price elastic and beer or possibly shōchu is the least price elastic alcohols. We must note that past literature has often failed to present or consider the statistical significance of the results of AIDS estimations, which could have affected their conclusions drastically. Both tax policy and demand system analysis suggest that preferential tax rates may boost the sectoral performance of certain alcohols and that beer and shōchu are relatively ideal items to be taxed. These alcohols happen to be two of the lowest real price per liter among the five alcohol types considered and the recent tax policies seem to be adequate in the Ramsey sense. Although we have seen the sensitivity of results to the model specifications as well as to the type of price data, the reasonably consistent results given by DAIDS are assuring. The study highlights the importance of price data and model selection which does not necessarily receive scrutiny in empirical investigations. In addition, these models do not properly reflect the variability in qualities as such data are not available, and even if available, they can overburden the demand system estimations with an enormous number of estimable parameters. A further accumulation of empirical work from different angles with theoretical endorsement will be necessary to improve this type of analysis and facilitate relevant policy 25