Beverage consumption patterns in a new market economy

: Beverage consumption refl ects consumer behavior shaped by social and economic factors. Observations of beverage consumption patterns provide unique insights of interest to many decision-makers, from the marketers and distributors to public health experts. We apply count data modeling technique to data on beverage consumption frequencies and show the importance of selecting the appropriate model. Th e models are estimated using survey data collected from more than 2,000 households in Bulgaria. Results from nested and non-nested tests and the ability to predict zero observations suggested that the zero-infl ated Poisson with heterogeneity and negative binomial models performed best. Only results of these models are discussed leading to the overall conclusion that age, employment, education and income are the core characteristics of each beverage consuming group, while gender and household characteristics such as the number of adults in the family or the presence of children together with the religion and the geographical location helped to fi ne tune the profi le of beverage consumers.


Introduction
A fundamental change in the economic and political system results in a cascade of changes, each calling for major adjustment on the part of all members of a society.Th e transition from a centrally planned economic system controlling the distribution and consumption of goods and services to a system based on markets as an allocative mechanism represents such a fundamental change.Individual responses to changes vary.Th e nature of the experienced change aff ects the system of values and preferences, and emerging preferences are refl ected in consumption choices.
In this paper we investigate factors infl uencing beverage choice in Bulgaria using consumption frequency data.Bulgaria has undergone a major transition Volume 6 Number 1 2006 from a consumption-restricting centrally planned system emphasizing heavy industry growth to an economy based on the market mechanism.Liberalization of the economy led to a rapid increase in the variety of food items available, including beverages.Th e pattern of beverage consumption captures phenomena associated with the dramatic socio-economic change induced by transformation of the country's political and economic system.Depending on the type of beverage, the profi le developed by the consumer is expected to mirror several diff erent trends that may occur simultaneously but are diffi cult to discern.Trends refl ect lifestyle changes, uncertainties brought about by structural transformation of the economy and employment insecurity, concerns about adequate nutrition, or changes in consumption resulting from varying incomes.Beverage consumption choices provide a synthetic measure of societal reaction to a major change in the socioeconomic environment and allow multiple recommendations ranging from specifi c marketing strategies to nutrition implications and public health concerns to be made.Th e deliberate purpose of this study is to investigate beverage consumption in an economy subjected to a major transformation aff ecting the population.Factors aff ecting the individual beverage consumption frequencies were analyzed.Specifi cally, we distinguish among eight beverage categories and examine demographic, economic, and geographic factors responsible for frequency of their consumption.Th e eight beverage categories are milk, coff ee, tea, soft drinks, lemonade, beer, wine, and spirits.Nutritional concerns are refl ected in the profi les of milk consumers, while soft drink consumption off ers an immediate and relatively inexpensive way to demonstrate a fashionable lifestyle.Liberalization of trade and investment led to a rapid development of the soft drink market, and foreign and domestic manufacturers supply a variety of soft drinks.Black tea has been commonly available for years, but choices of brands and fl avors increased during the 1990s.Lemonade was widely consumed prior to adoption of a market economy.Coff ee is the most common hot beverage, and the Bulgarian coff ee market has experienced a major growth in recent years.
Transition changed the structured and regulated lives of Bulgarians, and increased fears of unemployment because many industries in Bulgaria proved to be unable to compete on world markets and closed many plants.It has been observed that in stressful situations some increase in alcohol consumption occurs (Cooper et al., 1995;Peirce et al., 1996).Increased alcohol intake can negatively aff ect health of individuals and cause a loss of productivity.Earlier studies suggested that the most educated Bulgarian experienced a decline in health larger than that of the less educated (Wnuk-Lipinski, 1990).Th e disproportional decline was attributed to greater sensitivity and feelings of frustration on the part of the highly educated.We will examine, among other associations, the link between the educational attainment level and the frequency of consumption of various alcoholic beverages.
Th e available data are self-reported records of beverage consumption frequency.Th e integer nature of the data calls for an appropriate estimation technique.A count-data model is suitable, but within the family of count-data models a specifi c, more accurate alternative may be preferred to others.If the results are to be used for the purpose of developing a policy, the degree of model robustness matters.We apply four diff erent count-data models to show the relevancy of selecting a proper model.Selection of the best performing model is based on the penalized consistent Akaike Information Criteria (CAIC), Vuong test and the ability of the model to predict beverage choices made by consumers.

Econometric models
We adopt the standard demand model in which an individual maximizes his utility subject to a budget constraint.Th e goods entered in the individual's budget are indivisible and consumed in multiples of some basic unit (Pudney, 1989).We also assume that the choice of each beverage product, q 1 is a nonnegative integer.In principle, each individual chooses q 1 as well as the quantity of a vector of other goods, q 2 , some point in a given time.Furthermore, we don't know whether the individual has other beverage products or a subset of them on his choice set for q 2 because the way it is defi ned, e.g., composite goods.
Let the individual's utility function be u = V (q 1 , q 2 ; β), where q 1 ∈ I is the quantity of the beverage in question, where I is the set of {0, 1, 2, . ..}, q 2 is the quantity vector for all other goods and β is a vector of unobservable preferences shift variables.Each individual maximizes max , ; | max max , ; , q I q q I V q q P Q p q p q Y V q q where P is the vector of prices and divided into p 1 , the price of the beverage in question, q 1 , and p 2 , a vector of prices of all other goods, q 2 , Y is income, and H is the indirect utility.Usually detailed information on q 2 is unavailable in cross-section studies and thus the utility must be defi ned as a function of the quantity of the bev-erage in question, its price, aggregate price (p 2 ) for other consumption good q 2 and income.In such a case, the individual's discrete choice problem is defi ned as max , ; where P 2 is the price aggregate for a composite good q 2 .Prices of individual beverages are not available, but due to the competitive nature of the economy, likely show little variation in retail outlets where they are purchased most oft en.Th erefore, the assumption of equal prices across individuals is plausible and they are captured in the constant term in the equation.Morever, regional variables are included because these variables may also serve as proxies for regional price variations as well as social behaviour diff erences.( Su and Yen, 2000).Th e objective functions defi ned in equations ( 2) and (3) depend on the integer value of q 1 as the unique characteristics distinguishing each j (j =1, 2, ...) alternative corresponding to quantities q j =0, 1, 2, ... units of good q 1 .In this case, the individual will make a decision by comparing the utilities which give him the highest satisfaction u 1 =V(0, y P 2 ; β), u 2 =V(0, (y-p 1 ) P 2 ; β) etc.It is worth noting that some of the elements of q j will not vary across individuals and other elements may vary because they are not specifi c to alternative j.V(.;.) is known to all individual and all alternatives displaying all common properties of a utility function for every estimable β (Pudney, 1989).
Th e discrete non-negative nature of the beverage consumption frequency-dependent variable suggests the use of the simple Poisson count data model.Let y i , i = 1, ..., N represent the discrete beverage consumption frequency and x i be a (k x 1) vector of independent variables.In the benchmark model, we assume with λ i = mean (y i |x i ) = variance (y i |x i ).Setting a common specifi cation for the mean parameter as λ i = exp(x i β), where the exponential form ensures non-negativity of counts and β is a vector of unknown parameter estimates, yields the compound Poisson regression model.Th e benchmark model has the restrictive assumption equalizing the conditional mean of each count-dependent variable with its corresponding conditional variance.However, in most economic applications, empirical count data are characterized by the presence of overdispersion: the conditional variance exceeds the conditional mean.Unobserved heterogeneity and positive conta-gion are main sources of overdispersion (McCullagh and Nelder, 1989).In addition, another frequently encountered issue in consumption studies is a relatively higher frequency of zero observations (Gurmu and Trivedi, 1996;1997;1998).Th us, the compound Poisson model is not appropriate for data with a large portion of new observations and the potential presence of overdispersion.
To account for the unobserved heterogeneity and excess of zeros in the model, we use a zero-infl ated variant of the standard Poisson and negative binomial count data models.Th e zeros may come from two diff erent data-generating processes.Th e fi rst zero-generating process results from the pervasive abstention decision in which an individual does not have an inherent preference about the beverages.Th e second zero-generating process follows the usual corner solutions in the underlying count data distribution.Th e high consumption frequency of zeros is thus accounted for by introducing an extra probability mass at zero, say with the probability π and reducing the probability mass for other non-zero frequencies with probability 1 -π.
Th e link between the economic decision and statistical analysis for the probability mass at zero is that the binary binomial procedure addresses the abstention decision, while the standard count data analyzes the corner solution along with positive counts.Because zeros are not a certainty in the event count model (i.e., zeros are assumed to occur only if the corner solutions are present in the standard count models or otherwise), the compound count data model is not truncated at zero.Th us, cases which make the transition to the standard count stage may nevertheless have zero counts; crossing the hurdle in zero-infl ated count data models does not guarantee a positive realization of beverage consumption frequencies.To specify the zero-infl ated Poisson (ZIP I) and the zero-infl ated negative binomial (ZINB) models, let where F(.) is either the cumulative normal probability, Φ(.), for the probit or the cumulative logistic probability, ∧(.), for the logit model, τ will be subsequently outlined.Let f (.) denote either the Poisson (λ i ) or the negative binomial (λ i, θ) probability density function, then the probability density function of observing a count, y i , is where Γ and α are the gamma function and the dispersion parameter, respectively.And the corresponding log-likelihood is An important distinction of the Poisson with normal heterogeneity model (ZIP II) is that the log-likelihood is maximized by numerical integration using the Hermite quadrature method to obtain parameter estimates since the heterogeneity term introduced into λ = x ' β+ε (Greene, 2002;2002).
Th e model is a mixture of observations from two independent sources.Th ey may either result from the compound Poisson and negative binomial distributions with probability (1 -π) or occur independently by the extra probability π in underlying binomial distribution (logit or probit).For this model, the expected mean and variance of the zero-infl ated model is then Th e expected conditional mean of the ZIP I, ZIP II and ZINB models are equal to each other.Th e dispersion parameter, α, determines the shape of the negative binomial distribution.As α tends to zero, the conditional negative binomial in the infl ated model tends to a Poisson compound distribution.Th us, the ZINB is nested within the ZIP I when α tends to zero.Th e nice thing about testing the null hypothesis α=0 is the simultaneous testing whether the ZINB reduces to ZIP I and if it does not then it shows the presence of overdispersion.As α increases, the conditional negative binomial model becomes more skewed with a heavier tail and observing higher probability of a zero consumption of the product.Notice when π is one, the distribution is a probability mass at zero compared to when π is zero, the probability distribution becomes strictly compound count.Th e restriction π = 0 is not a simple parametric restriction because it depends on the covariates.When π does not depend on the covariates then a simple score test developed can be applied to see whether the zero-infl ated reduces to its compound version.However, when it depends on the covariates and to make π = 0 , it is necessary for some parameter needed to be +∝ or -∝.Th erefore, the zero-infl ated Poisson is not nested within either the Poisson or the negative binomial models.We use Vuong (1989) test distinguishing non-Poissonness due to the overdispersion of the negative binomial model (Greene, 2002;2003).We use a Likelihood Ratio (LR) and the penalized consistent Akaike Information Criterion (CAIC) log-likelihood to discriminate among variants of nested zero-infl ated models (e.g., ZIP I versus other zero-infl ated models such as the zero-infl ated negative binomial (ZINB) and the zero-infl ated Poisson with heterogeneity (ZIP II)).Th e ZIP II is nested within ZIP I when the variance, σ 2 , is equal to zero.Th e preferred model has the lowest value in the CAIC test.For the use of a regression model, Lambert (1992) considers where X i and Z i are the observable vectors of two diff erent covariate sets and β and γ are vectors of unknown parameter estimates.Th e basic assumption implied in the zero-infl ated model is that the two sets of covariates in both the Poisson or negative binomial and the binary logit models may or may not coincide.However, when the same set of exogenous variables is used in both models, then more parsimonious models can be developed by assuming that linear predictors from both models are related in some way.Indeed, no economic theory shows a guideline for what set of variables should be included in each of the models (Su and Yen, 2000).We, therefore, use the same set of variables and thus the zero-infl ated-tau count data models (ZIP I (τ), ZIP II (τ), ZINB (τ)) are deemed appropriate when the same set of variables coincides in both probability and frequency models.Th e simplest model is the zero-infl ated Poisson-tau, ZIP I (τ), which is a multiplicative function of the covariates used to explain the standard count where tau, τ, is a scalar parameter and implies that π i = (1+λ i -τ ) -1 .Individuals at the corner solutions may display the same preferences as those who are already users of a beverage as compared to individuals who do not derive an inherent utility from the product at all.Th us, specifi c merits of zero-infl ated count models are their ability to combine the eff ects of individuals who are both at the corner solutions and users on the count density function while identifying a diff erent probability function for the eff ects of those individuals who are assumed to be abstentionists.If the sub-segment of abstentionists is ignored, the model resembles a Tobit-type censoring count data technique.Th is inherent feature diff erentiates zero-infl ated count models from their rival, the standard hurdle count models, which assumes that observations at the corner solution and observations of those who do not have any desire to consume a product come from the same data-generating process.

Data
Th e data are from a survey of Bulgarian households collected in May 1997.Following a pretest, the Bulgarian National Statistical Service distributed questionnaires to a panel of 2,500 households scattered throughout the country.Th e survey instrument was hand-delivered and responses were gathered by enumerators four weeks later, yielding 2,133 completed questionnaires.Questions inquired, among others, about daily, weekly and monthly consumption frequencies of many food items, including eight beverages.A series of questions requested information about demographic characteristics of the respondent and other household members.In addition, respondents shared details about their education level, household income, and employment status.
Table 1 shows descriptive statistics of all items found to be relevant to the study.Consumption frequencies and individual characteristics have been expressed as variables, which are included in the empirical models.Th e majority of respondents were females, a fairly typical occurrence in food consumption studies.Th is is also consistent with the respondents who were primarily in charge of meal preparation in the family.Respondents who were the primarily food shopper in the family were mix of gender.Th e average age is relatively high, but Bulgaria's population is characterized by a stable population with a large segment of retirees.Th e average income level refl ects a compression of incomes that resulted from the need to balance the government budget and freeing of prices.Furthermore, periods of infl ationary fi scal policies, resulting from inconsistencies in the implementation of economic reforms, further depressed incomes prior to 1997.Th e low number of children per household refl ects the general demographic trend, suggesting a declining population size.In contrast, the relatively large number of adults in each household is refl ected in the high average age and refl ects the frequent occurrence of multi-generational families.A high number of respondents were from married households and reported at least a high school level of educational attainment.Also, two out of fi ve respondents were employed, still a high percentage given the large share of pensioners and increasing incidence of unemployment.
Patterns of beverage consumption are infl uenced by the religious background of the respondent.Although the majority declared themselves as Christians, about six percent classifi ed themselves as Muslims.Religious background plays a role in food consumption and can infl uence the consumption of alcoholic beverages.Finally, the region where the respondent resided could infl uence beverage consumption because of climatic or cultural reasons.Respondents from the metropolitan region were likely to be more exposed to the new beverages, leading the rest of the population in changing beverage consumption patterns.It is not unusual for large, highly urbanized areas to diff er distinctly from small towns or rural areas in consumption of foods, including the incidence of beverage consumption.
Th e consumption frequency of beverages diff ered across eight beverages considered in this study.Some consumption frequency indicate a wide variation across respondents, e.g., soft drink consumption.Also, alcoholic beverage consumption frequency was highly variable.However, as stated earlier, reasons for the observed variation in, for example, soft drinks versus alcoholic beverages, are dissimilar and likely result from diff erent motives for consuming each type of beverage.

Results
Each equation was estimated using four techniques.Th e purpose was to verify the usefulness of all techniques in an empirical study based on the data set that contains a large number of zeros, i.e., respondents who did not report the consumption of a specifi c beverage.Table 2 shows the results of nested and non-nested tests of model performance as a series of pair-wise comparisons.A careful indiction for the preferred model shows that the ZIP II model is of the choice for milk, tea and spirits beverages.Th e ZINB model fi ts best to coff ee and soft drink beverages.All three models, ZIP II, negative binomial and ZINB remain equal for lemonade and wine beverages and the ZIP II and ZINB were not distinguishable for beer beverage.Th e models gaining equal weights for beverages were further distinguished by the power of zero observations predictions.In doing so, we credit the model which has the more power of the zero observations prediction than the other models.In this case, the ZINB was determined for lemonade and wine beverages whereas the ZIP II was chosen for beer beverage frequency consumption.Results discussed in the following sections refer to the ZINB for coff ee, soft drink, lemonade, and wine beverages and the ZIP II for milk, tea, beer, and spirits beverages.

Milk
Milk product consumption in Bulgaria increased from 80 kg in 1952 to 193 kg in 1989 (Bresch, 1999), but the milk production declined aft er 1990.Milk consumption frequency was positively infl uenced by the age-squared, cooker, number of adults in household, employment, marital status, and negatively infl uenced by the income, age, shopper, number of children in household, gender and education of respondents (Table 4).Muslims are more likely and Christians are less likely to consume milk than other religious group.People residing in southern, coastal, and northern regions are less likely to consume milk than the people residing in metropolitan area.Th is result can be as expected because people residing in the metropolitan area may have more excess to markets selling varieties of milk such as skim and non-fat milk and it could be associated with the location of high-value farm production near large accessible urban centers.Th ese results may further indicate the price diff erences in regions.A metropolitan area may have higher prices than all other regions.Because reduced-fat milk was one of a few items subject to government pricing guidelines, its price was not expected to be a barrier to consumption.Th is result is in line with consumption trends observed in Western societies, where milk consumption has either stagnated or declined and the only major change has been shares in the volume consumed according to the fat content (Bunch, 1985).
Th is contradicts the observed tendency of declining ability to digest milk with advanced age, but supports the observation that older consumers, especially those living off fi xed incomes (e.g., pensions or disability payments), might substitute milk for other animal protein sources.Assuming that most of the respondents who were primarily in charge of family cooking are female, the sign for that variable is consistent with that of gender variable.In other words, females are more likely to consume milk than males.Respondents from married households consumed milk more frequently than those from other households.Th is suggests that although respondents may not have consumed milk when children were present in the household, they were still more likely to drink milk if married.Less frequent consumption of milk by the highly educated refl ects a diff erent food consumption pattern.Despite expectations that education would improve knowledge of nutrition and the role of milk in a diet, highly educated Bulgarians choose to drink milk less oft en than those with less education.Also, those employed consumed milk more frequently than those without obligations resulting from having a job.A plausible explanation is that those without jobs either had their own cow or bartered, doing odd jobs in exchange for milk.Th e latter situation was more likely in villages where, aft er the collapse of the state farm system, many Bulgarians took advantage of having access to grazing land, allowing ownership of a cow.

Hot beverages
Two beverages served hot in Bulgaria are considered in this study.Coff ee appears to be the preferred hot beverage, but tea is commonly available and inexpensive.Aft er the adoption of a market economy, the variety of coff ee and tea brands and fl avors noticeably increased.Both tea and coff ee contain caff eine, although in dif-ferent amounts.Adverse eff ects of heavy caff eine consumption on human health and the feeling of well-being are a concern (Snel and Lorist, 1997).However, moderate coff ee drinking also has been found to improve some measures of job performance (Streufert et al., 1997).Black tea has been implicated in cancer chemoprevention (Dreosti, Wargovich and Yang, 1997;Hakim et al., 2000) and may protect against coronary heart disease and stroke due to antioxidant vitamins and fl avonoids (Hertog et al., 1995;Keli et al., 1996).Tea ingestion has been reported to improve alertness and information processing capacity (Hindmarch et al., 1998).Th e amount of caff eine consumed by a typical Bulgarian consumer is expected to have a more therapeutic rather than detrimental health eff ect because the average monthly consumption frequencies indicated that coff ee was consumed about once a day, while tea consumption was infrequent.
Coff ee.Income, education, the number of adults in a household, and full time employment positively infl uenced the frequency of coff ee consumption.Income is important as it aff ects purchase frequency of coff ee, which is more expensive than tea in Bulgaria.Th e signifi cance of income in coff ee consumption was expected because many households experienced a decrease in real incomes and likely consumed coff ee with a lower frequency than prior to the transition.Th e more adults in the household, the higher the consumption frequency.Clearly, households with several adults have a diff erent consumption pattern, possibly pool resources for coffee purchase, and exercise their preference for coff ee more oft en than households with fewer adults.Respondents with more education are likely follow a diff erent consumption pattern and have a diff erent lifestyle leading to a higher coff ee consumption frequency than those with less education.Full-time employees reported drinking coff ee more oft en than others.Psychosocial job environment may be conducive to coff ee consumption.Other studies have revealed that those who experienced high job demands drank coff ee more oft en on days they felt distressed (Steptoe and Wardle, 1999).
Several variables negatively infl uenced the dependent variable.A large number of children in the household negatively infl uenced coff ee consumption frequency, but was statistically insignifi cant.Th ere were no gender diff erences in the frequency of coff ee consumption.Excessive coff ee consumption during pregnancy has been linked to health problems (Torfs and Christianson, 2000) of the newborns, but the average reported frequency was far below that implicated in causing any harm.
Tea.An increased tea consumption frequency was expected as respondents were advancing in age or had a higher level of educational attainment (Table 4).Older consumers may prefer tea because it contains less caff eine, oft en blamed for its relation to well being.Tea consumption has been reported to lead to improvement of blood circulation in people suff ering from coronary heart disease (Hertog et al., 1993;Duff y et al., 2001), most common among the elderly.Highly educated con-sumers search for variety and choose tea as an alternative hot beverage.Education has been linked to demand for food variety in previous studies in Western economies and economies in transition.Furthermore, tea consumers are more likely to be females than males.Gender diff erence supports fi ndings of improved mood among women who drank tea and enjoyed high social acceptance at work (Steptoe and Wardle, 1999).Although a large dosage of caff eine from tea, among other beverages, is not recommended (Food Standards Agency, 2001), the consumption frequency of tea in our study was generally very limited.Respondents identifying themselves as Muslims consumed tea more oft en than those of other religions or non-believers, while residents in the Northern region tended to drink tea less oft en than those in other regions.Th ese confi rmed diff erences in tea consumption suggest that religious or regional preferences in beverage consumption frequency exist and should be recognized by marketers.

Non-carbonated and carbonated beverages
Consumption of non-carbonated (still) and carbonated beverages has been rapidly growing world-wide.In Bulgaria, the average soft drink consumption was relatively low (Table 1) but highly variable, suggesting that a small percentage of people consume soft drinks oft en.Th e consumer profi le allows marketers to target their promotion and advertising eff orts and also provides public health offi cials with guidelines concerning a segment of the population that may develop health problems associated with high sugar consumption.
Th e soft drinks consumed include commercially manufactured products and home-made drinks such as lemonade.Although homemade drinks are inexpensive, they do not enjoy the same image as branded soft drinks, which are a part of the new, trendy lifestyle among young Bulgarians.Th e low production cost of most soft drinks, combined with a growing demand, off ers exceptional earning opportunities for manufacturers, distributors, and retailers.Furthermore, the long, hot summers in many parts of Bulgaria are conducive to high soft drink consumption.
Lemonade.Th e number of children and adults positively infl uenced the consumption frequency (Table 4).Lemonade can be purchased, but it can be easily prepared at home and, therefore, income may have no signifi cant eff ect on lemonade consumption.Males or married respondents were more likely than females or those who were not married to drink lemonade more oft en.It appears that lemonade was more likely prepared in married households, but at the request of husbands or children.Th e number of children also positively infl uenced the consumption of lemonade by respondents, suggesting that it was a family-type beverage.It is possible that some lemonade was prepared using ingredients specifi cally purchased for making it.Distributors and retailers, if aware of the strong association between lemonade drinking frequency and family households, may develop marketing strategies en-couraging increased lemonade consumption.Respondents identifying themselves as Christians also consumed lemonade more oft en than non-Christians.Th e lack of regional diff erences in lemonade consumption suggests that there is no need for regionally diff erentiated marketing strategies.
Soft drinks.Results indicate that income, the number of adults in a household, employment, and education drove soft drink consumption frequency (Table 4).Th ese factors are consistent with expectations and the observed behavior in western economies, with one exception.It appears that the number of adults was more relevant than the number of children in the household for consumption frequency of this category.Although one could associate the presence of children as the reason for soft drink purchases, according to survey results, only adult household members significantly aff ected the consumption frequency.Because the age variable had a negative eff ect, soft drinks were consumed by the younger generation, likely motivated by the trendy image of soft drink consumption.Married respondents consumed such beverages less oft en than unmarried respondents.Furthermore, residents in all but the Metropolitan region consumed soft drinks infrequently.Th is strong regional eff ect results from the access of metropolitan region residents to soft drinks, exposure to promotional and advertising campaigns, and the preference of foreign companies to target large, urban centers at the onset of entering a new market.

Alcoholic beverages
Information on consumption of alcoholic beverages was expected to refl ect an increased stress associated with transition to the new socio-economic system in general and uncertainties associated with employment in particular.Public health experts argued that if macroeconomic policies led to unemployment which, in turn, promoted alcohol use, then indirect costs measured in the loss of health, premature death and crime may require a policy change (Ettner, 1997).Studies from post-industrial economies suggest that unemployment reduces alcohol consumption and attribute this reduction to decrease in incomes (Ettner, 1997;Claussen and Aasland, 1993).However, involuntary unemployment had a mixed eff ect (Ettner, 1997).Results of estimating the consumption frequency of beer, wine and spirits help to identify population segments likely to consume each of the three alcoholic beverages, providing information not only for alcohol manufacturers and distributors, but also public policymakers.
In a rare study of alcohol demand in a centrally-planned economy during the pre-transition period, Florkowski and McNamara (1992) showed that income had a signifi cant infl uence on the consumption of spirits and the observed shortage of spirits had a signifi cant dampening eff ect on consumption.Anecdotal evidence, however, indicated that prolonged shortages led to illegal alcohol production and the offi cial consumption fi gures may have been underestimated.Such local condi-tions as accessibility to alcoholic beverages stemming from the legality of production of alcohol suggest that the self-reported consumption frequency may more accurately refl ect consumption patterns than the offi cial sales fi gures.In Bulgaria, production of beer and wine for home use is permitted.
Beer.Age of the respondent was the only signifi cant variable that negatively affected the beer consumption frequency (Table 4).Older respondents were less oft en consuming beer, possibly because it was not widely available throughout their lives until very recently when the economic liberalization led to increased production and improved quality of beer.Income and employment positively infl uenced beer consumption frequency, suggesting that as incomes increase, many will drink beer more oft en.Th e positive infl uence of employment may be associated with a lack of supervision and the limited alcohol content of beer, making it more acceptable to drink on the job in contrast to drinking other alcoholic beverages.Th ese attitudes seem to coincide with two household characteristics that stimulate beer consumption frequency, i.e., the number of adults and marital status.Beer was consumed more oft en when more adults lived together or in married households as compared to households with few adults or residents not married.Socializing and low alcohol content of beer may help explain these positive relationships.Males more oft en consumed beer than females, a trend observed in many other countries and characteristic of the ways of socializing typical of this gender.Th e more frequent consumption of beer in the Coastal region than other regions results possibly from the location of a traditionally well known brewery in that region, making quality beer more accessible to residents of this region.
Wine.Wine consumption frequency declined with age and the level of education (Table 4).Decreasing alcohol consumption frequency with advancing age was expected.Feelings of well-being and possible health implications lead to less consumption.Th e negative eff ect of education on wine consumption suggests a likely country-specifi c eff ect.Domestic wines are well known but do not off er variety often sought by highly educated consumers.Also, the quality of domestic Bulgarian wines varies, while the best quality wines are exported and less oft en available for domestic consumption.Th e more children in the household, the lower the wine consumption frequency.Th is may be a positive development from the standpoint of public health.
However, those who were employed, married or males consumed wine more often than those not employed, not married or who were females.Reasons behind the observed consumption patterns are likely similar to those responsible for beer consumption.Also, respondents identifying themselves as Muslim or Christian were more likely to drink wine more oft en than those of other religions or non-believers.Among regions, residents of the Northern and Coastal regions consumed wine more oft en than those from the Metropolitan or Southern regions.Some of the premier commercial wineries are located in those regions, but the increased consumption frequency may result also from deeper traditions of making wine in general and refl ects drinking of home-made wine.Noticeably, the income eff ect was statistically insignifi cant in this equation, supporting the notion that some wine may have been home-made and, therefore, the income eff ect was neutralized.
Spirits.Respondents with high incomes, high levels of educational attainment, working full time or male consumed spirits more oft en than their counterparts.Liquor tends to be expensive because of the imposed excise tax and, therefore, the positive income eff ect was expected.Also, males are generally more likely to consume beverages with high alcohol content.Th e eff ect of education was not expected, but it may result from a variety-seeking behavior.Employment, which led to more frequent consumption of spirits, likely refl ects the style of socializing with co-workers.Note that, in general, the consumption of spirits was very low, suggesting that this type of alcohol was consumed on special occasions.Muslim respondents were less likely to consume spirits than respondents of other religious affi liations.Older respondents were less likely than those of younger age to drink spirits.Advancing age leads to less frequent consumption of alcohol in general, a trend supported by results of this study.Finally, regional diff erences were confi rmed, indicating that spirits were less oft en consumed in Coastal and Northern regions compared to Metropolitan and Southern regions.

Implications
Although the consumption frequency data of the beverages measures the number of times a respondent consumes a beverage, it lacks a perfect measure for the consumption quantity.However, when the lack of the consumption quantity exists, the frequency data may be an alternative way to look for individual behaviors aff ecting the beverage consumption.Many researchers have begun to use count data model in modifying the pattern of consumption where the dependent variable is an integer (Moon et al., 1999;2001).Examination of beverage consumption patterns was done using the count data procedure.Among the four approaches tested, the zeroinfl ated Poisson with heterogeneity and negative binomial models performed consistently better than other models.Without comprehensive testing, other models might have been selected, given the statistical signifi cance test of individual coeffi cients.For example, in the case of milk consumption, the ZIP I model showed all coeffi cients as signifi cant, while the preferred ZINB version did not confi rm these results.In general, the ZIPs an ZINB models diff ered in terms of statistical significance and, sometimes, the direction of the eff ect.Diff erences occurred with greater frequency for coeffi cients accounting for regional and religious characteristics, followed by respondent demographic features such as gender, presence of children and number of adults.No diff erences were detected in terms of age or employment eff ects, while income and education infl uences diff ered in single cases.
From the standpoint of developing a marketing plan or public health messages, identifi ed diff erences across various models are very important.Th ey suggest that not all factors are equally important and that some of the essential variations are related to the location, religious beliefs, and the family unit living within a single household.Income, age and education aff ected beverage consumption frequencies in a similar manner, regardless of the applied modeling technique.Th ese characteristics may establish core profi les of each beverage consumption group, while factors which diff ered in their infl uence enable to diff erentiation among groups.
Older consumers are a potentially lucrative group of buyers for milk products.Health concerns, including adequate calcium intake, may improve sales, especially because no gender diff erences were identifi ed, while women are more likely than men to suff er bone loss.Marital status positively infl uenced milk consumption frequency, suggesting these respondents who are either primarily shoppers or in charge of cooking in the family are more likely to purchase milk, but also supporting the general fi nding of traditional attitudes towards health maintenance within family households (Watson, 1995).In our study, the traditional view of the family and its views seems to be more responsible than the education-based knowledge of nutrition for milk consumption.
Coff ee consumption increased in Bulgaria between 1997 and 2001 by about 28 percent, reaching 1.9 kg per capita (Anonymous, 2002), placing Bulgaria among the top ten growth markets.Th is growth is consistent with the strong preference for coffee found in our survey.Improving incomes probably increased the demand for coffee, and the trend towards increased consumption was strengthened by an increased consumption frequency, education and full time employment.Th e strong preference for coff ee in Bulgaria is a country-specifi c phenomenon because, compared to other countries in the region, the relative share of food expenses of 58 percent in a household budget is very high (Kolev, 2002).Assuming continuous growth in income, coffee consumption is likely to increase, creating new opportunities for coff ee exporters as well as for domestic roasters, distributors, and retailers.Th is increase in activity in the coff ee market in Bulgaria may call for policy changes due to the link between coff ee consumption and health (Ming et al., 1995).Although moderate consumption appears to benefi t drinkers, excessive consumption may aggravate the existing health condition.In sharp contrast to coff ee consumption, despite the benefi cial eff ects of tea, its consumption is small in Bulgaria and likely to increase slowly.Increased incomes are unlikely to increase consumption frequencies and tea marketers may focus on older consumers, females, or the highly educated as potential buyers.Given the already infrequent tea consumption among respondents in the survey, the small growth rate will only marginally increase the volume sold.However, increased promotion of health benefi ts of tea may change the observed trend.
Soft drink consumption frequency can be expected to increase due to a strong positive income eff ect.Such a response in Bulgaria will be consistent with a worldwide increase in consumption of soft drinks.Results showed that soft drink consumption frequency was adversely aff ected by age of respondent.Such a relationship was found for American consumers in the early 1980s (Wood, 1983).According to our survey, consumption was particularly high in the Metropolitan region, but improvements in distribution probably made soft drinks more accessible throughout the country.Th e soft drink market, which includes a mix of domestic and international companies, remains very competitive, registering a high volume growth with only a moderate growth in sales value.Consequently, soft drinks have become available to a wider group of consumers, leading to potential for developments observed in other countries where soft drinks replace nutritionally more desirable beverages (Moyer and Mayell, 1981).From the public health standpoint, the growing consumption of soft drinks in Bulgaria may aggravate the existing problems of maintaining proper body weight, lowering calcium intake by replacing milk, and treating chronic diseases induced by improper diet.In contrast to soft drinks, lemonade consumption is unlikely to grow in response to rising incomes.Manufacturers of dry ingredients for homemade lemonade can expect increased consumption frequency in households with children, among female consumers, and in households with many adults.Th ree characteristics driving the consumption of most other beverages, i.e., employment, education and income, lacked any signifi cant eff ect in the case of lemonade.
Earlier studies concluded that unemployment in Eastern Europe can be expected to be more damaging to health than in Western Europe because of the severe decline in incomes during transition to a market-oriented economy, coping mechanisms impaired by lower incomes, increased feelings of uncertainty and underdeveloped civil institutions (Watson, 1995).According to our results, employment led to an increased consumption of all alcoholic beverages (Table 4).Despite the low average consumption frequencies, the link between employment and alcohol consumption may call for increased monitoring of the trend.Further investigations may distinguish between occupation and the frequency of consumption of diff erent types of alcoholic beverages.Th e importance of a respondent's family may discourage consumption of spirits.Gender diff erences in consumption frequency of all alcoholic beverages were found in this study.Men always drink more oft en than women and gender eff ects of diff erent alcoholic beverages vary (Epstein, 1989), but the results of our study also support earlier fi ndings showing that men diff er from women in using alcohol in times of stress and the alcohol tension reducing eff ect on men (McNair, 1996;Cooper et al., 1997).Four characteristics stimulating alcohol intake, i.e., gender, employment, income, and number of many adults in the respondent's family, provide guidelines for alcohol manufacturers and distributors.From the public health standpoint, results seem to support a correlation with high incidence of stroke and heart diseases among Bulgarian men (Bresch, 1999), despite sometimes inconclusive evidence of linkages between alcohol consumption and health complications.Furthermore, residents of the Metropolitan region revealed divergent preferences for beverages not evident among those living in other parts of the country.
, f 1 (.) and f 2 (.) are densities for the competing models, and m bar and s m are the sample mean and standard deviation for the sample of m i s.Asymptotically, the V statistic is distributed as standard normal, so the computed value is comparable with the critical value from the standard normal distribution.
LR= -2(Log-Likelihood of ZIP I-Log-Likelihood of ZIP II) and LR=-2(Log-Likelihood of ZIP I-Log-Likelihood of ZINB) distributed as χ 2 with number of restrictions imposed as degree of freedom.Th e consistent Akaike information criterion: CAIC=-2(Log-Likelihood)+(number of free parameters)*(Log(N)+1), where N is the number of observations.Th e CAIC is consistent with the LR test and both the ZIP II and ZINB models outperform the ZIP I model since the disparities between th log-likelihood value of the ZIP II and ZIP I and between the log-likelihood value of the ZINB and ZIP I are enormous for each corresponding beverage .
the predicted zero observations are over or under estimated compared to the actual zero observations in the sample.Interestingly, all zeroinfl ated count models have a nice prediction about zero observations as compared to the compound count data model, negative binomial model.

Table 1 . Descriptive statistics of variables
Note: N = 2007 and number in parentheses are the percentage of zero counts and age is scaled by 100.

Table 4 . Parameter estimates for beverage consumption frequencies from zero-infl ated count data models
for the zero-infl ated Poisson model; ZIP II stands for the zero-infl ated Poisson model with heterogeneity; Negbin stands for compound negative binomial model; and ZINB stands for the zero-infl ated negative binomial, b Signifi cant at α=.10, c Signifi cant at α=.05 and d Signifi cant at α=.01.
a ZIP I stands Note: t-values in parentheses.