Alcohol Monopoly in Sweden and Wine Properties That Drive the Wine Sales

(1)

INOM

EXAMENSARBETE TEKNIK, GRUNDNIVÅ, 15 HP

STOCKHOLM SVERIGE 2020,

Alcohol Monopoly in Sweden and Wine Properties That Drive the Wine Sales

ROUMI ROUMI

KTH

(2)

(3)

Alcohol Monopoly in Sweden and Wine Properties That Drive the Wine Sales

Roumi Roumi

ROYAL

Degree Projects in Applied Mathematics and Industrial Economics (15 hp) Degree Programme in Industrial Engineering and Management (300 hp) KTH Royal Institute of Technology year 2020

Supervisor at KTH: Mykola Shykula, Julia Liljegren Examiner at KTH: Sigrid Källblad Nordin

(4)

TRITA-SCI-GRU 2020:114 MAT-K 2020:015

Royal Institute of Technology School of Engineering Sciences KTH SCI

SE-100 44 Stockholm, Sweden URL: www.kth.se/sci

(5)

Abstract

This research study in applied mathematical statistics, and industrial engineering and management aims to determine the wine properties that drive the wine sales in Sweden, while also examining how the monopoly in Sweden affects the alcohol sales and demand.

Data were collected from Systembolaget, which is the only alcohol retail store chain in Sweden. A multiple linear regression analysis was performed on 4 931 observations from Systembolaget’s sales statistics of 2019. The covariates were the type of wine, taste, price, size, country of origin, ecologically classified and ethically certified. The final model contained 18 variables where the taste parameters dominated the model.

The results conclude that wines from Australia, Italy and Portugal have an advantage when it comes to the wine sales in Sweden, while wines from Sweden have a negative correleation with the volume sold. Other results suggest that wine producers should make ethically produced wines. From an economic perspective, the alcohol monopoly in Sweden is not a regular monopoly. Moreover, the monopoly is considered to has had a negative impact on the alcohol sales by lowering the demand through a strict alcohol law, among other things. The positive aspects of Sweden’s alcohol monopoly is the overall better health of the Swedish society and the in-store consultation that is offered. This study has contributed to previous research and also given the wine producers a better overview of the factors that drive the wine sales in Sweden.

(6)

(7)

1 Introduction

1.1 Background

The world’s wine industry has seen a stagnation or decline in consumption in the last couple of years,¹ where only US and Asia have been on a rise.² Even though Sweden accounts for merely 1 % of the world’s wine consumption, the Swedes are in the top of the world when it comes to consumption per capita.² It is even more remarkable considering Sweden’s alcohol retailing monopoly which limits the wine consumption.³ Today the production of wine in the world is higher than the world’s consumption.⁴ The market is therefore considered competitive and the consumers’ choice of wine is of high importance to the producers. Hence, the wine properties that boost the sales are the main focus of this study.

1.2 Previous Research

1.2.1 Wine Industry & Wine Sales

The largest producing country of wine is France followed by Italy.⁴ The wine production in Italy is also dominating the Italian agricultural economy.¹ Italy is also one of the worlds largest exporting countries together with countries such as Australia, Argentina, Chile, Portugal and South Africa.⁴ According to Lecat et al. the alcoholic drink market in the world is estimated to be 1397 billion dollars.³ The largest segment is the beer and cider market with a value of 613 billion dollars in 2015. On the other hand, the wine market is one of the smallest segments accounting for only 287 billion dollars. The beer industry is dominated by the ten largest brewers that have a market share of 60 %. In contrast, the wine industry is considered highly fragmented and the top ten largest wine companies only hold 14 % of the total wine market. This makes the wine industry highly intense and therefore, the wine sales is crucial for producers.

1Rinaldi, A. (2015). Wine Global Trends - Traditional Leaders and New Markets. Rivista di Scienze del Turismo. 6: 5-10.

2Anderson, K., Nelgen, S. and Pinilla, V. (2017). Global wine markets, 1860 to 2016: a statistical compendium. Adelaide: University of Adelaide Press.

3Giesbrecht, N. and Österberg, E. (2012). Alcohol Retailing in Canadian and Nordic Contexts:

Challenges and Opportunities in Balancing Trade and Prevention Agendas. Contemporary Drug Problems. 39 (1): 107-145.

4Lecat, B., Amspacher, W., Higgins, L., Ferrara, A. L., Wolf, M. M. (2018). Wine sector: Definitions and nuances from global to country analysis—A comparison between Old World, New World, and emerging wine countries from 2005 to current. Case Studies in the Wine Industry. 2019: 7-32.

(12)

In recent years the sustainability characteristics of wines have become more important for consumers.⁵ As a result, consumers often have a positive view of for example organic, local and ethical wines. The literature review of Schäufele and Hamm shows that customers associate sustainability labels with quality and are also ready to pay a premium for such wines.⁵ In conclusion, the results indicate that producers should invest in producing wines with sustainability characteristics to stand out from the competition. However, this research paper is solely based on scientific articles that do not use real world data of purchase behavior. Hence, the wine sales need to be studied further with statistics of consumers’ wine purchases.

Previous research regarding the wine sales have mainly focused on wines in restaurants.

This type of wine sales is associated with the look and content of the wine list according to Yang’s and Lynn’s study.⁶ Thus, it is unclear why a specific wine gets chosen. Another problem is that different restaurants have various prices on wines which can affect the customer’s choice, even if the restaurants have the same wines on the menu. This is not the case in a monopoly where a single actor can set the price.

1.2.2 Monopoly

In 1985 the state of Iowa in the United States saw an increase of 79.2 % in wine sales following the privatisation of wine sales.⁷ Similarly there was an 9.5 % increase in distilled spirits sales in 1987 when Iowa eliminated the state monopoly on spirits sales. However, the privatisation of distilled spirits caused a long-term decline in wine sales of 13.7 %.

In summary, the results suggest that the removal of the monopoly increases the alcohol sales, which is also emphasised in the research study of Giesbrecht and Österberg.⁸ Lai et al. studied the alcohol monopoly in Norway which is controlled by the Norwegian state.⁹ All alcoholic beverages above 4.75 % alcohol in Norway are either served in pubs, restaurants etc. or sold at Vinmonopolet. The results of the study indicate that the wine market is competitive and therefore, the taste is one of the key attributes that matter the

5Schäufele, I. and Hamm, U. (2017). Consumers’ perceptions, preferences and willingness-to-pay for wine with sustainability characteristics: A review. Journal of Cleaner Production. 147: 379-394.

6Yang, S. S. and Lynn, M. (2009). Wine List Characteristics Associated with Greater Wine Sales.

Cornell Hospitality Report. 9 (11): 6-14.

7Holder, H. D. and Wagenaar, A. C. (1990). Effects of the elimination of a state monopoly on distilled spirits’ retail sales: a time-series analysis of Iowa. British Journal of Addiction. 85: 1615-1625.

9Lai, M., Cavicchi, A., Rickertsen, K., Corsi, A. and Casini, L. (2013). Monopoly and wine: the Norwegian case. British Food Journal. 115 (2): 314-326.

(13)

most for the customers. Norwegian customers are fond of small wine producers who have high quality wines and the wines’ country of origin plays a crucial role when making wine decisions. The red wines are also more preferable than any other type of wine. However, the research paper focuses mainly on perceptions of people that work in the wine industry, and not on the wines or on ordinary consumers. Furthermore, the study was qualitative based on interviews which is not reliable in this case, since perceptions of what wine experts think are sold do not need to align with the real wine choices of customers in the stores. Hence, a better study would be to look at the wine sales’ statistics of the stores.

The alcohol monopoly in Sweden is comparable to other alcohol monopolies in the Nordic, such as Finland, Iceland, Norway and the Faroe Islands.¹⁰ Wines and other alcoholic beverages must be bought at Systembolaget, which is the only physical retail store chain in Sweden that is allowed to sell alcoholic beverages above 3.5 % alcohol. Moreover, the alcohol commercials are almost non-existent due to the strict alcohol law in Sweden.¹¹ However, wine and beer advertisements are allowed as long as they are not intrusive or encourage the use of alcohol, which means that sales discounts or contests, among other things, are not allowed.¹¹ Thus, the alcohol law in Sweden makes it hard for winemakers to promote their wines. As a consumer, the wine properties become very important in the choice of wine. Hence, the wine characteristics and their influence on the wine sales need to be studied.

1.3 Research Questions

Based on the problematisation above, the research questions are the following:

- Which wine properties drive the wine sales in Sweden?

- How does the alcohol monopoly in Sweden affect the alcohol sales and, in turn, demand?

1.4 Purpose & Contribution

The purpose of this study is to find the most important wine properties that drive the wine sales in Sweden, in combination with how the sales and demand are affected by the monopoly system. The main goal is to contribute to previous research since they were mainly focused on wine sales in restaurants. This research study will also be relevant to wine producers who are looking to either establish themself in the Swedish market or

11SFS 2010:1622. Alkohollag. Stockholm: Socialdepartementet.

(14)

want to expand their market share.

1.5 Demarcation

This quantitative and qualitative study is limited to consumption, that is not export or import of wines. Factors such as graphical design, media coverage, reviews etc. are not possible to integrate in this study. There are also a few online retailers of wine on the Swedish market, such as Vinoteket, from which no data will be collected. However, as Systembolaget is by far the main distributor of wine and the only alcohol store chain in Sweden, this should not be a problem. The delimitation is methodological and will therefore be justified in more detail in section Method.

2 Theory

2.1 Mathematical Theory

2.1.1 Multiple Linear Regression Analysis

A multiple linear regression analysis is needed in order to find a relationship between the sales volume of wines and the wines’ properties. The general model can be expressed as the following:

y_i =

k

X

j=0

x_ijβ_j + ε_i , i = 1, · · · , n (1) In equation (1), y_i denotes the response variable of the i:th observation.¹² The predictors are represented by x_ij, β_j denotes the unknown coefficients that must be estimated from the observations and εi are the residuals (random errors). Equation (1) can also be expressed in matrix form:

Y = Xβ + ε (2)

Y =







y₁ ... y_n







, X =







1 x₁₁ x₁₂ · · · x_1k 1 x₂₁ x₂₂ · · · x_2k ... ... ... ... 1 x_n1 x_n2 · · · x_nk







, β =







β₁ ... β_n







, ε =







ε₁ ... ε_n







12Montgomery, D. C., Peck, E. A. and Vining, G. G. (2012). Introduction to Linear Regression Analysis. Hoboken, New Jersey: John Wiley & Sons, Inc.

(15)

2.1.2 F-test

An F-test of total significance for the linear model needs to be performed in order to see if a linear model should be fitted or if other models are suitable.¹³ It is an approach to test for the null hypothesis that a number r of estimators are equal to zero.¹³ If the p-value is lower than 0.05, the null hypothesis is rejected and the estimators are implied to be significant to the model, meaning that there is a significant relationship between the covariates and the response variable. The F-test can be formulated as the following:

F = 1 r

βˆ^TVˆ⁻¹βˆ (3)

p-value = Pr(F (1, n − k − 1) > F ) (4) The covariance matrix of the estimators is denoted by ˆV⁻¹, n is the number of data points and k denotes the number of variables in the model.

2.1.3 Theoretical Assumptions

The following five basic assumptions must be satisfied in order to be able to perform a linear regression analysis:¹⁴

1. A linear dependence should exist between the response variable and the regressors.

2. The mean of the error terms is zero, E [εi] = 0.

3. Homoscedasticity meaning that the variance of the residuals is constant, Var[ε_i] = σ². This also implies no or little autocorrelation.

4. The residuals are uncorrelated which assumes that there is no or little multicollinearity.

5. The error terms are normally distributed.

These assumptions should also be analysed in order to validate a model or else it will lead to model inadequacies.¹⁴ This study will use a residual analysis to evaluate the assumptions.

13Lang, H. (2015). Elements of Regression Analysis. Stockholm: KTH.

(16)

2.1.4 Ordinary Least Squares

The Ordinary Least Squares (OLS) estimator of β guarantees that the estimator is optimal if the assumptions above are fulfilled.¹⁵ The Gauss-Markov theorem shows that the least-squares estimator ˆβ is the best linear unbiased estimator (BLUE) of β.¹⁵ The estimator of β is given by the following formula:

β = (Xˆ ^TX)⁻¹X^TY (5)

The aim is to minimise the sum of the squared residuals, ˆε^Tε =| ˆˆ ε |², by solving X²ˆε = 0 where ˆε = Y − X ˆβ.¹⁵

2.1.5 Influential Observations

A preferable way to handle influential observations is by examining both the location of that point and the dependent variable.¹⁵ This can be achieved by calculating the Cook’s distance. All points exceeding Cook’s distance should be considered influential points and therefore, excluded from the dataset.¹⁵ However, outliers that are not considered influential for the model should not be removed. The formula for Cook’s distance can be expressed as the following:

D_i = ( ˆβ_(i)− ˆβ)^TX^TX( ˆβ_(i)− ˆβ)

p · M SE (6)

MSE = 1 n

n

X

i=1

(Y_i− ˆYi)² (7)

In equation (6), ˆβ_(i) denotes the estimated coefficients excluding observation i, ˆβ are the estimated coefficients, X denotes the covariate matrix with ones in the first column and p is the number of coefficients in the fit including the intercept β₀. In equation (7), MSE denotes the mean squared error, Y_i are the observed values and ˆY_i stands for the fitted values.

2.1.6 Multicollinearity

Problems with multicollinearity arise when there is a dependence between the predictors.¹⁵ As a result, the final model’s coefficients will be highly unreliable and the model will have redundant covariates. Both principal component regression (PCR)

(17)

and ridge regression can be used to fix multicollinearity but the final model can be hard to interpret.¹⁶ Thus, the variance inflation factor (VIF) as suggested by Montgomery et al. can be used instead to detect multicollinearity. The VIF value for each covariate is calculated by allowing each predictor to be the response variable and at the same time the other independent variables try to define the model. The following formula can be used for the calculations:

VIF = 1

1 − R²_i (8)

R² denotes the coefficient of determination.¹⁶

A high (≥ 10) VIF value for a predictor indicates a strong correlation with the other regressors.¹⁶ When an independent variable is removed due to a high VIF value, it will affect the VIF values of the other variables. Therefore, the VIF values were used in combination with a correlation matrix to determine which predictors to remove. It is sufficient to only remove one of two highly correlated variables to fix multicollinearity according to Montgomery et al.¹⁶

2.2 Economic Theory

2.2.1 Consumption & Demand

Krugman and Wells describe goods as products that give satisfaction when consumed.¹⁷ Consumers try to maximise their satisfaction from consumption. This can be measured as utility which a rational person tries to maximise. Moreover, the marginal utility is an increase or decrease in utility when consuming one more unit of a specific good.

Krugman and Wells suggest that the marginal utility for each successive consumed unit of good is decreasing and adding less to the total utility.¹⁷ This is called the principle of diminishing marginal utility and holds for a majority of goods, and for a majority of people. An example is alcohol where the first unit of alcohol has a high marginal utility since it gives a large increase in utility. However, the marginal utility is decreasing for each new unit of alcohol consumed. In the worst case of consuming too much alcohol, the marginal utility will be negative, which will affect the total utility negatively if a new unit of alcohol is consumed.

In economics, the demand will rise for a normal good if the income increases.¹⁷ On the other hand, the demand for a luxury good such as champagne (sparkling wine) will increase even more as income increases. In a competitive equilibrium market there is

17Krugman, P. and Wells, R. (2015). Economics. New York: Worth Publishers.

(18)

a balance between supply and demand. This is reached if the market contains a lot of buyers and sellers of the same goods or services and none can affect the prices. In other words, an equilibrium quantity of goods or services are traded at an equilibrium price. In theory the price will move the market back to equilibrium if the market is experiencing a shortage or surplus of goods or services.¹⁸

2.2.2 Monopoly Systems

Originally, the word monopoly represented the right to exclusively sell an item.¹⁹ Today, Krugman and Wells explain a monopoly as a market where the monopolist is the only producer of a good and there are no close substitutes for that good.¹⁸ In theory, the monopolist is said to posses a market power to set the price, which is referred to as a price-maker.¹⁹ However, there are two types of constraints that the monopolist faces when determining the price and quantity output; technological constraints and consumers’

behaviour.¹⁹ According to Varian, the technological constraints can be denoted as a cost function and the consumer’s behaviour as a demand function.¹⁹ This yields the following problem for the monopolist in order to maximise profit:

maxp,y py − c(y) (9)

such that D(p) ≥ y (10)

The variable p denotes the price, y is the quantity, c(y) denotes the cost function and D(p) is the demand function. In a normal monopoly, the rational strategy is to sell the quantity that the buyers demand, meaning that the constraint can be rewritten as y = D(p). However, Systembolaget in Sweden is not a regular alcohol monopoly and therefore, the constraint can not be rewritten as y = D(p). The conditions for the maximisation problem in a normal monopoly are the following:

r⁰(y) = c⁰(y) (11)

2p⁰(y) + p⁰⁰(y)y − c⁰⁰(y) ≤ 0 (12) r(y) denotes the revenue function and the inverse demand function is denoted by p(y), which is a price function dependent on the amount y.¹⁹ These conditions are fulfilled when the marginal revenue curve crosses the marginal cost curve from above as the following figure:

19Varian, Hal. (1992). Microeconomic Analysis. New York: W. W. Norton & Company.

(19)

Figure 1: The optimal quantity output for a monopolist is where the marginal revenue is equal to the marginal cost.²⁰

In contrast, Systembolaget’s marginal revenue is constant and equal to the demand which can be seen in the figure below:

Figure 2: Systembolaget acts as if it was in a perfectly competitive market.²¹

As previously mentioned, Systembolaget does not act like a traditional monopolist. In fact, Systembolaget operates as if it was in a perfectly competitive market, while also

20Varian, Hal. (1992). Microeconomic Analysis. New York: W. W. Norton & Company.

(20)

trying to limit the demand in different ways. Systembolaget as a non-profit monopolist does not want to raise prices or maximise sales because they are regulated by the Swedish alcohol law and overseen by the Public Health Agency of Sweden.²² Research findings show that alcohol is a world problem lined with diseases, injuries and deaths.²³ Therefore, the Public Health Agency of Sweden together with Systembolaget have an interest to control the total consumption of alcohol by limiting the consumers’ demand. There are seven main areas that are highlighted by scientific studies in which strategies can be implemented to reduce and prevent alcohol-related damage.²³ Two of the most important strategies are higher taxation on alcohol which leads to higher prices, and regulating the physical availability of alcohol. As a result, Systembolaget has short opening hours on weekdays, even shorter opening hours on Saturdays and the stores are closed on Sundays.²⁴ Furthermore, there is no marketing or sales prices inside the stores as previously mentioned. However, Systembolaget’s employees will help customers match alcohol beverages to different types of meals.²⁴

3 Method

3.1 Research Design

3.1.1 Data Collection & Selection Process

Sales statistics of alcohol in Sweden are publicly available information. The dataset, containing the wine sales among other alcoholic drinks, was accessible on Systembolaget’s website. The latest annual report (”2019”) was downloaded as an Excel file and then imported into RStudio for analysis. The imported file contained 34 407 observations of alcoholic beverages which was narrowed down to 14 851 wine sales observations. All wines that had been returned or had zero volume sold, were excluded from the set of data in order to perform a logarithmic transformation if needed. Observations that lacked taste notes were eliminated from the study because it would be time consuming to complete these observations from Systembolaget’s website. This narrowed down the number of observations from 14 499 to 5 221. Moreover, this research study is based on the assumption that wines are alcoholic beverages as defined in section 2 Theory and

22SFS 2010:1622. Alkohollag. Stockholm: Socialdepartementet.

23Alcohol and Public Policy Group. Alcohol: No Ordinary Commodity - a summary of the second edition. (2010). Addiction. 105: 769-779.

(21)

therefore, non-alcoholic wines were removed from the dataset. Niche wines such as port wines, vermouth and blue wines were also excluded since they do not fall under the four main categories of wine: red wine, white wine, rosé wine and sparkling wine. The final dataset contained 4 931 observations with volume sold in litres as the response variable and the following independent variables: type of wine, taste, price, size, country of origin, ecologically classified and ethically certified. Due to the limited time of this study, it was not possible to include the type of grapes as a covariate by manually extracting the grape data of thousand of wines from Systembolaget’s website.

3.1.2 Indicator Variables

As previously mentioned, the response variable in this study is the volume sold in litres with a range from 0.375 to 2 493 003. Furthermore, the qualitative explanatory variables were transformed into categorical variables (dummy variables). A dummy variable takes the value 1 if it belongs to a specific category and the value 0 otherwise. If there are n number of categories for a regressor, then there will be n − 1 dummy variables created to avoid perfect multicollinearity.²⁵ The following explanatory variables were used:

Covariate Explanation No. of dummy variables

Type of wine Red, white, rosé or sparkling 3

Taste 15 different taste categories 14

Price Ranges from 25 to 79 989 SEK Not applicable Package size Content of 0.187 to 15.0 litres Not applicable Country of origin 42 different categories of origin 41

Ecological Ecologically certified 1

Ethical Fairtrade or Fair for Life certified 1

Table 1: Quantitative covariates had zero dummy variables and qualitative regressors had n − 1 indicator variables.

The following reference group of the dummy variables was used:

(22)

Covariate Reference

Type of wine Red

Taste Dry

Price Not applicable

Package size Not applicable Country of origin France Ecological Not ecological

Ethical Not ethical

Table 2: The response variable will equal the reference group if the dummy variables are set to zero.

3.2 Model Assessment & Verification

3.2.1 Best Subset Regression

There were 60 regressors after the multicollinearity treatment, which meant that an exhaustive approach where all possible regressions are evaluated would be computationally time and memory consuming even with a fast computer. Furthermore, all possible regression evaluates 2^kmodels,²⁶meaning that 2⁶⁰ = 1 152 921 504 606 846 976 models would have been assessed in this case. Forward selection, backward elimination and stepwise regression are the three stepwise regression methods that can be used instead.²⁶ Sequential replacement was chosen since it is a combination of forward selection and backward elimination. The method starts with the intercept and no predictors, then the covariate that has the largest simple correlation with the response variable is added to the model (like forward selection). For each added variable, the variable that does not improve the model is removed (like backward elimination).²⁶

The set of observations was sampled randomly into ten different groups. Then a 10-fold cross validation was used considering that a higher fold number gives a less biased model.²⁶ The cross validation is a type of model validation that tests one group at a time while using the other groups for training. The criterions used were adjusted R², the Bayesian information criterion (BIC) and Mallows’ C_p.

(23)

3.2.2 Adjusted R²

Adjusted R² is a better measurement than R² due to the fact that R² keeps increasing when the number of regressors increases. This will result in overfitting.²⁷ However, the enhanced measurement adjusted R² takes into account the number of predictors and will penalise if too many predictors are used. Adjusted R² can be expressed by the following formula:

R²_adj = 1 − (1 − R²)(n − 1)

n − k − 1 (13)

In equation (13), n is the number of observations, k denotes the number of covariates and R² is the coefficient of determination. It follows that 0 ≤ R²_adj ≤ 1 and a higher R_adj² value implies a better model.

3.2.3 Bayesian Information Criterion

Both Akaike information criterion (AIC) and Bayesian information criterion (BIC) applies a penalty term for the number of predictors to solve the problem with overfitting.²⁷ However, BIC was used as one of the criterions for the model selection, since it will penalise additional regressors more than AIC. The Bayesian information criterion is given by the following formula:

BIC = k ln (n) − 2 ln ( ˆL) (14)

The variable k denotes the number of coefficients including the intercept, n are the observed data and ˆL is the maximum likelihood estimate of the model. Lower BIC value implies a better model.

3.2.4 Mallows’ C_p

Mallow’s C_p is primarily based on variance and bias.²⁷ Montgomery et al. showed that if the C_p value equals the number p of covariates in a model, it means that the model is unbiased.²⁷ However, some bias is acceptable in order to have a simpler model.²⁷ Mallow’s C_p is defined with the following formula:

C_p = SS_Res(p) ˆ

σ² − n + 2p (15)

The sum of squared residuals is denoted by SS_Res(p), n is the sample size and p denotes the number of regressors. Smaller C_p value implies a better model.

(24)

3.2.5 Bootstrap Confidence Intervals

Bootstrap-based confidence intervals for the final model’s coefficients were used in order to validate the final model.²⁸ The steps are described by Montgomery et al. which start with the residuals, ε^∗.²⁸ These bootstrapped residuals were generated from the model’s residuals by sampling with replacement. This yielded a new model that was composed of the original linear regression model and the bootstrapped residuals, Y^∗ = X ˆβ + ε^∗. The estimated regression coefficients, ˆβ^∗, were then computed from the original dataset (X) and the new model (Y^∗). These steps were then repeated from the start as many times as the number of bootstrap samples. This study used 10 000 bootstrap samples to ensure accurate estimations. Furthermore, the bootstrap distributions were expected to be Gaussian distributed. 95 % bootstrap confidence intervals were constructed for each parameter by taking the 2.5^th and 97.5^th percentiles of the bootstrap distributions.

3.3 Economic Perspective

A qualitative approach will be used in order to answer the second research question.

The results will take into account previous research and also the theoretical framework regarding consumption, demand and monopoly systems. The main focus will be how the alcohol monopoly affects the alcohol sales and what conclusions can be drawn regarding the demand. This will be discussed in section 5 Analysis & Discussion and linked to the final regression model.

4 Results

4.1 Initial Model

4.1.1 F-test & Adjusted R²

The initial model had 62 regressors including 60 dummy variables. The calculated p-value in the F-statistic was < 2.2e-16 which implies that there is a significant relationship between the covariates and the response variable. In other words, a linear model can be fitted to the data.

Only 12 regressors had a p-value < 0.05 and contributed significantly to the initial model.

(25)

Furthermore, the adjusted R²for the initial model was 0.2318. These results show a highly ineffective model that can be improved.

4.1.2 Residual Analysis

The residual analysis showed that the basic assumptions of linear regression were not satisfied.

Figure 3: The plot indicates heteroscedasticity and a fluctuating variance of error terms (Var(εi) 6= σ²).

The points in Figure 3 are not randomly and equally spread out. Therefore, this does not indicate homoscedasticity and there exists no linear dependence between the response variable and the regressors. In a similar way, the other scatter plots of the residual analysis showed that the basic assumptions of linear regression were not fulfilled. They can be found in section Appendix under Analysis - Initial Model. A variable transformation was needed in order to fix the model inadequacies.²⁹

(26)

4.1.3 Transformation & Cook’s Distance

One way to fix heteroscedasticity is to logarithmically transform the response variable.³⁰ This yielded the following model:

log (V olumeSold) = β₀+ β₁x_1,i+ . . . + β₆₂x_62,i+ ε_i , i = 1, . . . , 4931 (16) The F-test for the transformed model showed a p-value of < 2.2e-16 which indicated that there is still a significant linear relationship between the predictors and the response variable. Moreover, the residual analysis of the transformed model satisfied the basic assumptions of linear regression better than the initial model but there were influential points that had to be removed. The plots of the residual analysis are in section Appendix, under the title Residual Analysis - Transformed Model. Cook’s distance was used to decide which influential points to exclude:

Figure 4: Observation 4636 is outside the Cook’s distance which indicates a potential influential point.

Observation 4636 was removed from the dataset due to the fact that it was larger than Cook’s distance. A better figure than Figure 4 of all datapoints regarding Cook’s distance can be found in section Appendix, under the title Cook’s Distance - Transformed Model.

(27)

4.2 Reduced Model

4.2.1 Basic Assumptions

After the response variable transformation and the handling of influential points, Figure 5 below shows that the residuals are equally spread around the red line.

Figure 5: The scatter plot shows that the points are almost equally spread and random around the red line.

There are still minor signs of heteroscedasticity but it is assumed that the error terms have constant variance. In conclusion, Figure 5 shows a linear dependence between the response variable and the regressors.

(28)

Figure 6: The quantile-quantile plot shows that the residuals follow a linear line.

It can be concluded from Figure 6 that the residuals are normally distributed since they follow a linear line even though there is a slight skew.

Figure 7: The plot shows equally spread points around the red line with some minor clustering.

The Scale-Location plot is another plot that shows homoscedasticity where the error terms are equally distributed around the predictor range.

(29)

4.2.2 Multicollinearity Treatment

Only the predictors with a VIF value ≥ 10 are shown in the table below:

Predictor VIF value

Sparkling wine 152.8

White wine 122.5

Fruity & Flavoursome 100.8 Spicy & Full-bodied 78.8 Fresh & Fruity 40.6 Tight & Nuanced 38.3 Soft & Berry 21.6 Rich & Flavoursome 20.3 Fresh & Berry 15.2

Table 3: The table shows the predictors with large VIF values, which implies multicollinearity in the model.

Sparkling wine and white wine are two different types of wine, but the sparkling wine is produced from white wine. Therefore, it is reasonable to assume that if one of these regressors are removed, the model could still manage to fit the data points with the remaining covariates. Thus, the sparkling wine was removed as a predictor. Furthermore, a correlation matrix was used in order to determine the correlation between the regressors.

The highest correlation, 0.735, between two predictors was between the white wine and the taste Fresh & Fruity. Thus, the taste Fresh & Fruity was excluded from the model.

The VIF values were calculated once more. All the VIF values of the reduced model were lower than 2.5. The mean of all VIF values was calculated to be 1.19, down from 10.6 in the beginning. Thus, the multicollinearity after the treatment was negligibly small and could be ignored.

4.2.3 Variable Selection

The 60 covariates were reduced to 19 regressors based on the criterions adjusted R², BIC and Mallows’ Cp. The results of the sequential replacement are shown in the figure below:

(30)

Figure 8: The graphs show that a model with 19 variables give the best model based on the three criterions.

The adjusted R², the BIC and Mallows’ C_p suggest that 19 variables will give the model with the best goodness of fit.

4.3 Final Model

4.3.1 Predictors

The 19 coefficients corresponding to each regressor and the intercept of the best model were extracted and can be seen in the table below:

(31)

Coefficient Value p-value βˆ₀ (Intercept) 4.062646085 < 2e-16

Price -0.001765178 < 2e-16

Size 1.652850235 < 2e-16

White wine 2.670170815 < 2e-16 Grapy & Floral 1.055117571 0.000931 Fresh & Berry 2.497966066 < 2e-16 Fruity & Flavoursome 2.713369014 < 2e-16 Rich & Flavoursome -1.206467736 4.49e-09

Semi dry 2.220541095 0.000160

Spicy & Full-bodied 2.206940970 < 2e-16 Soft & Berry 3.054808439 < 2e-16 Other tastes 2.317739915 1.35e-05

Sweet -2.117796817 2.60e-10

Tight & Nuanced 1.295412596 7.73e-13

Australia 0.837795787 3.23e-05

Italy 0.526015172 1.33e-08

Portugal 0.993008994 0.008890

Sweden -1.684616273 1.21e-08

Ethical 1.465143110 8.37e-05

Czech Republic 0.000000000 0.931367

Table 4: The table shows the parameters of the final model before removing Czech Republic.

The Czech Republic parameter at the bottom of Table 4 did not contribute to the final model and therefore, it was removed based on the negligibly small value and the high p-value. The final model with 18 regressors had an adjusted R² of 0.3316 based on a log-transformed response variable.

4.3.2 Multicollinearity

The VIF values of all the predictors are very low and therefore, acceptable without any signs of multicollinearity in the model. The VIF mean of the final model is 1.23 and the VIF values are in the table below:

(32)

Predictor VIF value

Price 1.823563

Size 1.705047

White wine 1.862026

Grapy & Floral 1.118044 Fresh & Berry 1.061135 Fruity & Flavoursome 1.444747 Rich & Flavoursome 1.353695

Semi dry 1.010562

Spicy & Full-bodied 1.268587 Soft & Berry 1.085124

Other tastes 1.020139

Sweet 1.025724

Tight & Nuanced 1.145732

Australia 1.062739

Italy 1.122517

Portugal 1.028521

Sweden 1.031622

Ethical 1.046348

Table 5: The low VIF values indicate no presence of multicollinearity in the model.

4.3.3 Residual Analysis

The final model contained 18 variables as shown in Table 4, except the Czech Republic.

(33)

Figure 9: The four plots show that the five basic assumptions of linear regression are satisfied.

In the Residuals vs Fitted plot the majority of the points are equally spread out around the red line with the variance of the error terms being constant (homoscedasticity). The line is close to 0 for most of the points which indicates that the mean of the error terms is zero. These signs show that there is a linear dependence between the response variable and the regressors. Thus, the theoretical assumptions 1 to 4 are satisfied. Furthermore, the quantile-quantile plot shows that the residuals are normally distributed since they follow a linear line even though there is a slight skew. This satisfies the fifth basic assumption of linear regression. The Scale-Location plot confirms homoscedasticity because the majority of the points are equally spread out around the predictor range. The bottom right plot shows no signs of influential points since all of the points are within the Cook’s distance.

(34)

4.3.4 Validation of Model

The 95 % confidence intervals are given by the 2.5^thand 97.5^thpercentiles of the bootstrap distribution of each coefficient in the final model. The bootstrap confidence intervals for the intercept and the first three parameters are shown below. The other bootstrap confidence intervals can be found in section Appendix, under the title Bootstrap Confidence Intervals.

Figure 10: The figures show that the values for the first four coefficients in Table 4 are within the range of the 95 % confidence intervals.

The bootstrap confidence intervals are 95 % confident that the parameters in the final model are between the following values:

Percentile Intercept Price Size White wine 2.5^th 3.881825 -0.001978107 1.504690 2.433300 97.5^th 4.225780 -0.001641774 1.809803 2.937806

Table 6: The table shows the range of the 95 % confidence intervals.

(35)

The range of the confidence intervals for the other coefficients are under Bootstrap Confidence Intervals in Appendix. These results are in line with the final model’s coefficients’ values since the values are within the range of the confidence intervals.

5 Analysis & Discussion

5.1 Model Assessment

The final model is based on three different criterions; adjusted R², BIC and Mallows’ C_p. The adjusted R² improved to 0.3316 in the final model compared to 0.2318 in the initial model. However, this could have been improved even more if each type of wine had its own model. In conclusion, the final model is valid, showing no signs of multicollinearity and satisfying the five basic assumptions of linear regression as seen in section 4.3.3 Residual Analysis. Furthermore, the values of the fitted parameters are within the range of the 95 % confidence intervals of the bootstrap distributions as shown in section 4.3.4 Validation of Model.

5.2 Theoretical Findings

Since most of the parameters in Table 4 belong to dummy variables, it is interesting to take a look at the intercept. The reference group of the dummy variables consists of red coloured, dry taste, French and non-ecological wines according to Table 2. The high value of the intercept should be interpreted that there is a strong positive linear correlation between the response variable and the reference group of the dummy variables. However, it is not possible to determine which reference variable that contributes the most to the volume sold at Systembolaget.

The white wines also have a strong positive correlation with the dependent variable.

Sparkling wine that was removed from the model would probably also have had a strong positive correlation with the response variable because they are made from the same base. However, it is not possible to decide which type of wine contributes the most to the volume sold at Systembolaget, since the intercept is also high which might indicate that red wine has a positive correlation with the dependent variable.

In Table 4, the size coefficient is a positive value larger than one, which indicates a positive correlation with the predicted variable. This is not rational since larger wine packages usually lead to fewer packages being sold and, in turn, less volume sold. However, a

(36)

rational parameter is the price, which is negatively correlated with the volume sold. This is coherent, since humans are rational beings that are less likely to buy a wine if the wine is expensive. Individuals, as previously mentioned, try to maximise their utility when consuming goods.³¹ The satisfaction is clearly higher for cheaper wines, meaning that the utility is greater for cheaper wines, all things being equal. Theoretically, the demand will increase when the prices go down.²⁵ This is the case in a competitive market as described by Krugman and Wells. Consequently, the alcohol monopoly in Sweden has a negative effect on the alcohol prices since there are no competition. Hence, the alcohol sales is affected negatively from Systembolaget’s monopoly. However, even though wine is consumed all the time, the demand can not be infinite even if the price is zero. This is a result of the principle of diminishing marginal utility, that the marginal utility for each successive consumed bottle of wine is decreasing and contributes less to the total utility.³⁰ Even if Systembolaget as a monopoly can be perceived as restricting, the staff at Systembolaget for example are very helpful when it comes to matching wines with different types of meals.³² The in-store consultation should be seen as something positive for the wine producers.

The taste coefficients are highly correlated with the response variable, since half of the parameters in the model are related to the taste of a wine. Moreover, it is assumed that the taste of a wine depends on the wine’s grapes. The taste covariate was included in the regression analysis in order to try to compensate for the missing grape data. The values of the taste parameters in Table 4 show that wines with a Soft & Berry taste have the strongest positive correlation with the sales volume of wine at Systembolaget.

In contrast, the wines with a Sweet taste have the most negative correlation with the dependent variable, in other words, they contribute negatively to the volume sold.

The final regression model shows that Australia, Italy and Portugal as a wine’s country of origin are positively correlated with the predicted variable. However, Sweden as a country of origin of a wine contribute negatively to the sales volume of wine at Systembolaget.

These results are in line with previous research, considering that Italy is the second largest producer of wine in the world after France³³ and the wine production is a big contributor to the Italian agricultural economy.³⁴ Both Australia and Italy are in the top five of exporting countries in the world.³² Australia is also growing its export both in value and

34Rinaldi, A. (2015). Wine Global Trends - Traditional Leaders and New Markets. Rivista di Scienze del Turismo. 6: 5-10.

(37)

volume.³⁵

Wines at Systembolaget that are labelled as ethical are Fairtrade or Fair for Life certified.

This predictor has a positive correleation with the response variable as seen in Table 4.

The sustainability aspects of wines are discussed in the research paper of Schäufele and Hamm.³⁶ They conclude that consumers are positive when it comes to ethical wines and wines with sustainable characteristics in general. Those results are in line with the results in Table 4.

5.3 Practical Findings

Wine producers in Australia, Italy and Portugal should establish themself in the Swedish wine market since there is a positive correleation that wines from these countries are sold more at Systembolaget. This is also an opportunity for producers that already are established in the Swedish wine market, because there is a probability that they can expand their market share. Producers in Sweden should stop selling their wines in Sweden based on the negative correleation with the sales volume. However, this might not be the case if the cost of transportation and taxes of changing the production region are taken into account.

Producers of wine should invest in larger wine boxes or bottles instead of small bottles.

Furthermore, the producers should avoid to produce wines with a Sweet or Rich &

Flavoursome taste. In contrast, the wine producers should make wines with, for instances, a Soft & Berry or Fruity & Flavoursome taste. It is also preferable if the wines are ethically produced, which is supported by both Table 4 and the research study of Schäufele and Hamm.³⁶

The in-store consultation at Systembolaget that helps customers match alcoholic beverages with different types of meals can be exploited by wine producers. They can offer the staff at Systembolaget sales training, wine tastings and other activities to create a brand awareness that can spill over on the customers. However, there is a risk that Systembolaget can see through these methods depending on how they view their policy about responsible selling of alcohol.

36Schäufele, I. and Hamm, U. (2017). Consumers’ perceptions, preferences and willingness-to-pay for wine with sustainability characteristics: A review. Journal of Cleaner Production. 147: 379-394.

(38)

6 Conclusions

This research study has determined the wine properties that drive the wine sales in Sweden together with how the alcohol monopoly in Sweden affects the alcohol sales and demand. 4 931 observations of wine sales data from Systembolaget’s annual report of 2019 have been studied. The multiple linear regression analysis concluded that a model with 18 variables gives the best goodness of fit. The model was validated by checking that the five basic assumptions of linear regression were satisfied through a residual analysis.

Furthermore, the bootstrap-based confidence intervals for the regression coefficients were consistent with the results.

The wines’ country of origin had a positive correleation with the volume sold if the wines were from Australia, Italy or Portugal. In contrast, the wines from Sweden had a negative impact on the wine sales at Systembolaget. Half of the variables in the final model are associated with different taste categories. The most notable results are that wines with a Sweet or Rich & Flavoursome taste are negatively correlated with the sales volume of wine. However, tastes such as Soft & Berry and Fruity & Flavoursome have a positive effect on the sales. Furthermore, wine producers should make ethically produced wines to improve their chances of increased sales.

The alcohol monopoly in Sweden is not considered a traditional monopoly from an economic perspective. Systembolaget is non-profit based and does not want to maximise sales. Instead they try to restrict the demand of alcohol by for example limiting the opening hours. In addition, alcohol advertising is extremely limited due to Sweden’s strict alcohol law. All of these measures’ goal is to limit the general demand of alcohol.

Therefore, the alcohol sales in Sweden have been affected negatively by the government’s monopoly. However, there are some elements that can be utilised by the wine producers.

They can for example try to influence the in-store consultation by increasing the brand awareness even though it can be hard considering Systembolaget’s selling policy.

This research study has created a foundation that can help wine producers make smart decisions when they want to establish themself in the Swedish market or expand their market share by boosting their sales. In addition, this subject has contributed to previous research since earlier studies has been focused on wine sales in restaurants. Future research should study the sales statistics of more than one year to see if there is a difference amongst wine properties that boost the sales compared to this study’s results. Another topic that should be studied is which wine colour that contributes the most to the wine sales. In addition, instead of looking at different taste properties of wines, future studies can examine different types of grapes. An interesting study would include both taste characteristics and different grapes. From an industrial engineering and management

(39)

perspective, future research studies can explore factors such as the graphical design of the labels, media coverage, tests and reviews.

(40)

References

Alcohol and Public Policy Group. Alcohol: No Ordinary Commodity - a summary of the second edition. (2010). Addiction. 105: 769-779.

Anderson, K., Nelgen, S. and Pinilla, V. (2017). Global wine markets, 1860 to 2016: a statistical compendium. Adelaide: University of Adelaide Press.

Giesbrecht, N. and Österberg, E. (2012). Alcohol Retailing in Canadian and Nordic Contexts: Challenges and Opportunities in Balancing Trade and Prevention Agendas.

Contemporary Drug Problems. 39 (1): 107-145.

Holder, H. D. and Wagenaar, A. C. (1990). Effects of the elimination of a state monopoly on distilled spirits’ retail sales: a time-series analysis of Iowa. British Journal of Addiction.

85: 1615-1625.

Krugman, P. and Wells, R. (2015). Economics. New York: Worth Publishers.

Lai, M., Cavicchi, A., Rickertsen, K., Corsi, A. and Casini, L. (2013). Monopoly and wine: the Norwegian case. British Food Journal. 115 (2): 314-326.

Lang, H. (2015). Elements of Regression Analysis. Stockholm: KTH.

Lecat, B., Amspacher, W., Higgins, L., Ferrara, A. L., Wolf, M. M. (2018). Wine sector:

Definitions and nuances from global to country analysis—A comparison between Old World, New World, and emerging wine countries from 2005 to current. Case Studies in the Wine Industry. 2019: 7-32.

Montgomery, D. C., Peck, E. A. and Vining, G. G. (2012). Introduction to Linear Regression Analysis. Hoboken, New Jersey: John Wiley & Sons, Inc.

Rinaldi, A. (2015). Wine Global Trends - Traditional Leaders and New Markets. Rivista di Scienze del Turismo. 6: 5-10.

Schäufele, I. and Hamm, U. (2017). Consumers’ perceptions, preferences and willingness-to-pay for wine with sustainability characteristics: A review. Journal of Cleaner Production. 147: 379-394.

SFS 2010:1622. Alkohollag. Stockholm: Socialdepartementet.

Varian, Hal. (1992). Microeconomic Analysis. New York: W. W. Norton & Company.

Yang, S. S. and Lynn, M. (2009). Wine List Characteristics Associated with Greater Wine Sales. Cornell Hospitality Report. 9 (11): 6-14.

(41)

Appendix

Residual Analysis - Initial Model

Figure 11: The four plots show that the basic assumptions of linear regression are not satisfied.

(42)

Residual Analysis - Transformed Model

Figure 12: The four plots show that the basic assumptions of linear regression are satisfied.

(43)

Cook’s Distance - Transformed Model

Figure 13: The datapoint 4636 is clearly an influential point that exceeds Cook’s distance.

(44)

Bootstrap Confidence Intervals

Figure 14: The bootstrap distributions are Gaussian distributed which coincide with the theory. The coefficients of the final model are within the range of the confidence intervals.

Percentile Grapy & Floral Fresh & Berry Fruity & Flavoursome

2.5^th 0.4153642 1.81684 2.469925

97.5^th 1.6146188 2.93430 2.953667

Percentile Rich & Flavoursome

2.5^th -1.5232995

97.5^th -0.7592219

(45)

Percentile Semi dry Spicy & Full-bodied Soft & Berry Other tastes

2.5^th 1.007633 1.953757 2.420878 1.365280

97.5^th 3.228483 2.460017 3.355269 3.582965

(46)

Percentile Sweet Tight & Nuanced Australia Italy 2.5^th -2.980690 0.9365118 0.4678296 0.3830545 97.5^th -1.572088 1.6310050 1.3224248 0.7734204

(47)

Percentile Portugal Sweden Ethical 2.5^th 0.198276 -2.445942 0.7225029 97.5^th 1.443878 -1.204559 2.1592155

(48)

(49)

(50)

TRITA 2020:114

Alcohol Monopoly in Sweden and Wine Properties That Drive the Wine Sales

Alcohol Monopoly in Sweden and Wine Properties That Drive the Wine Sales

ROUMI ROUMI

Alcohol Monopoly in Sweden and Wine Properties That Drive the Wine Sales

Roumi Roumi

Abstract

Contents

1 Introduction

1.1 Background

1.2 Previous Research

1.3 Research Questions

1.4 Purpose & Contribution

1.5 Demarcation

2 Theory

2.1 Mathematical Theory

2.2 Economic Theory

3 Method

3.1 Research Design

3.2 Model Assessment & Verification

3.3 Economic Perspective

4 Results

4.1 Initial Model

4.2 Reduced Model

4.3 Final Model

5 Analysis & Discussion

5.1 Model Assessment

5.2 Theoretical Findings

5.3 Practical Findings

6 Conclusions

References

Appendix

Residual Analysis - Initial Model

Residual Analysis - Transformed Model

Cook’s Distance - Transformed Model

Bootstrap Confidence Intervals