Nutri-scores: The Impact on healthy purchasing behavior
by
COLIN BROER
1
MASTER THESIS
Nutri-scores: The Impact on healthy purchasing behavior
by
COLIN BROER
University of Groningen
Faculty of Economics and Business
MSc Marketing Intelligence
MSc Marketing Management
Supervisor: Ittersum, K. van
2nd Supervisor: Bijmolt, T.
8th of January 2020
Van Heemskerckstraat 38a 9726 GM Groningen
2
ABSTRACT
The number of people coping with health issues such as obesity and overweight is growing at an alarming rate. One of the more significant causes of this problem is an unhealthy dietary pattern. Consumers are victim of heuristic purchasing behavior, caused by information and choice overload, resulting in consumers purchasing groceries that are unhealthy and harmful for their overall well-being. Moreover, the growth in product assortments of supermarkets produces an overload of choice. Additionally, the information that food packaging provides can be overwhelming for consumers, resulting in an overload of information. One method which partially reduces information overload is the Nutri-score. The Nutri-score in a summary of multiple healthy and unhealthy nutritional values within groceries. Moreover, the overload of choice can be reduced by using interactive decision aids, such as filtering and sorting. The first part of this paper uses online sales data from one of the largest Dutch supermarkets to measure the effect of Nutri-scores on the purchasing behavior of customers. A Multiple Linear Regression model was used to measure the effects of the Nutri-score, interactive decision aids and other control variables. However, this paper did not find any significant change in behavior which proves that the Nutri-score positively improves the healthiness of consumer’s purchasing behavior. But, this study did find evidence that the usage of interactive decision aids positively influences the healthiness of customers purchasing behavior.
The second part of this paper used the provided data to determine which online factors influence the basket size of consumers. A Truncated Negative Binominal model was used to investigate which factors increase or decrease the basket size. This paper found that the usage of a predetermined shopping list has a substantial influence on the increase of shopping baskets size.
3
PREFACE
Dear reader,
Two and a half years ago, after finishing my bachelor at the NHL Stenden University of Applied Sciences, I decided that my next goal would be to achieve a master’s degree. Now, after two and a half years of working hard and studying many hours at the University of Groningen, I hereby present my master’s thesis.
I must admit, I had imagined my final year to be much different to what I have experienced. Studying and working on assignments in the current society has been particularly difficult. The transition to online education and the decrease of social interaction, both on University and in personal life has been an obstacle. However, even though the world was and still is in vast turmoil, studying at Groningen University was one of the only things that continued, even though in a different format.
I am proud of the definitive version of my master’s thesis. A lot of hard work and effort were needed to sometimes continue working on it. But I was fortunate to experience guidance from my supervisor van Dr K. van Ittersum, who gave advice whenever I got stuck on certain elements of my thesis, for this I can only express my utmost gratitude. Additionally, I would like to thank PhD candidate D. Olk, who gave additional statistical advice and helped with the merging of the datasets.
I hope that you will enjoy reading my thesis. Colin Broer
4 Table of contents
1. Introduction ... 5
2. Literature review and hypothesis development ... 8
2.1 Effects of choice overload ... 8
2.2 Effects of information overload ... 9
2.3 Effects of Nutri-scores ... 10
2.4 Effects of sorting and filtering ... 12
2.5 Conceptual model ... 14 3. Methodology ... 15 3.1 Data collection... 15 3.2 Plan of Analysis ... 16 3.2.1 Testing of hypotheses ... 16 3.2.2 Additional analysis ... 17 3.3 Data preparation ... 18
3.3.1 Merging and aggregating ... 18
3.3.2 Recoding and outliers ... 19
3.4 Model formulation... 21
4. Results ... 22
4.1 Multiple Linear Regression assumptions ... 22
4.1.1 Linearity... 22 4.1.2 Multicollinearity ... 23 4.1.3 Homoscedasticity... 24 4.1.4 Normality ... 25 4.1.5 Regression outcomes ... 27 4.2 Poisson regression ... 30 4.2.1 Multicollinearity ... 30 4.2.2 Poisson Assumptions ... 31
4.2.4 Poisson regression results ... 32
5. Conclusion & discussion ... 35
5.1 Conclusion ... 35
5.2 Future Research ... 36
5.3 Limitations ... 37
References ... 39
5
1. INTRODUCTION
During the past century, the importance of living by healthy standards has become exceptionally clear. Non-communicable and chronic diseases have exceeded the mortality rates of infectious or communicable diseases in western countries, such as for example the United States (Neuhouser, 2019). With primary chronic diseases being heart attacks, cancer, cerebrovascular disease, and diabetes (Dwyer-Lindgren et al., 2016). Furthermore, a large majority of the previous stated chronic diseases are related to overweight and obesity (Billington, 2000). Obesity is a chronic disease on its own but does also increase mortality risks of complementary diseases such as cerebrovascular diseases (Billington, 2020). The study of Hruby et al, (2016) investigated common determinants and consequences of obesity. They discovered that one of the major and well-known causes of overweight and obesity are dietary factors. They discovered that ‘unhealthy’ food intakes, such as desserts, sweets and junk food led to an increase in weight gain. Individuals who instead increased ‘healthy’ food intakes, for example, vegetables, whole grains, and nuts experienced in general less weight gain. Additional literature from Smethers and Rolls (2018) found during the past years, that eating behaviors around excessive energy intake have increased, resulting in an increase in overweight and obesity.
One of the main culprits for this unhealthy behavior is the promotional tools such as marketing campaigns, which are used by big corporations, to make the high energy consumer goods seem more appealing (Smethers & Rolls, 2018). Moreover, the study of Chandon and Wansink (2012) found that due to the nature of the food market, a lot of emphasis is put on profitability. They found that on the short term, consumers want food that is tasty, economical, diverse, accessible, and healthy, ordered by importance. Thus, companies actively promote food that is mostly tasty and economical, which often results in unhealthy foods such as snacks and junk food. This overrepresentation of unhealthy food in marketing is one of the main reasons for unhealthy dietary (Chandon & Wansink, 2012).
6 consumers can be convinced to purchase healthy alternatives over unhealthy snacks (Koese, Marchiori & Ridder, 2016; Hanks, Smith & Wansink, 2012). Current literature has studied which forms of promotional tools are the most and least suitable for promoting healthy eating and raising nutritional knowledge. The study of Brambila-Macias et al. (2011) found that by using social marketing through public information campaigns, healthy consumption of fruits and vegetables could be increased. However, one of the major drawbacks from a social marketing campaign is the long period of time needed to achieve changes in eating patterns. Moreover, supermarkets have experienced a transformation in assortment size. During the previous decades, supermarket assortments continuously expended in both quantity and diversity. For example, grocery stores initially only sold a few varieties of sodas, whereas currently there is a huge diversity of various sodas. Differing literature has either advocated or opposed the vast variety of product assortments. The study of Chernev and Hamilton (2009) states that a large assortment size positively influences store attractiveness compared to competitors. Furthermore, the study of Oppewal and Koelemeijer (2005) states that large assortments, offer a better chance of fulfilling the needs of consumers. However, more recent literature has proposed that having too much choice is disadvantageous. When consumers are overloaded with information and choice, they tend to become frustrated since they are not able to decide on which products to purchase, resulting in making a less than optimal decision, for example unhealthy purchases (Ketron, Spear & Dai, 2016). Furthermore, the study of Katron, Spears and Dai (2016) also found that due to information overload, consumers are in general less satisfied with their purchase.
7 This study will look at the possibilities of product labeling by adding Nutri-scores, to ultimately reduce the information overload and eventually lower extensive energy consumption and increase conscious purchasing. Nutri-scores are graded, color coded labels which help consumers guide in making healthy purchases. The Nutri-score is meant to encourage consumers into purchasing healthier food and ease the decision-making process. The Nutri-score is calculated by determining a ‘healthy’ and ‘unhealthy’ Nutri-score. Based on the number of points of the unhealthy and healthy scores, the Nutri-score is being given a colored label.
8
2. LITERATURE REVIEW AND HYPOTHESIS DEVELOPMENT
2.1 Effects of choice overload
As mentioned before, one of the biggest changes in the supermarket industry has been the increase in assortment size and variety. As explained by Schreibehenne, Greifeneder and Todd (2010), choice overload is seen as a special case of information overload, mainly since only the number of options is immense. The topic of information overload due to nutritional labeling and packaging will be discussed later. However, as will be discussed later on, information overload can hinder the decision making of consumers, which could result in purchasing unhealthier products.
Even though choice overload has many similarities with information overload, choice overload is important enough to deserve an additional explanation. Both information and choice overload share the similarity that they heavily burden the cognitive capabilities of consumers information processing (Schreibehenne, Greifeneder & Todd, 2010). The average consumer would state that having more product choice would be a good thing, as the chance of finding a product closer to their requirements would increase (Kahn & Lehmann, 1991). However, product assortments which are too large can feel overwhelming consumers and sometimes result in consumers not making a choice at all (Townsend & Kahn, 2014; Schreibehenne, Greifeneder & Todd, 2010). Furthermore, similarly to information overload, an overload in choice sometimes results in a lower satisfaction on selection or purchase of products (Haynes, 2009; Polman, 2012).
9 2.2 Effects of information overload
Consumers are being overwhelmed by the sheer amount of information available in supermarkets, and whilst doing their grocery shopping (Ketron, Spears & Dai, 2016). The study of Ketron, Spears and Dai (2016) states that when consumers are overloaded with information in brick-and-mortar stores, they can become frustrated and can choose suboptimal products. Suboptimal products are products which the consumer did not want to buy initially, and result in a lower purchase satisfaction in the long term (Ketron, Spears & Dai, 2016). This would imply, that when consumers want to purchase healthier, they are generally unable to stick to their wanted purchasing behavior, due to information overload. In reality, whenever a consumer has decided to buy healthier products, but is overloaded with information, they often fall short on their determination. Their frustration takes over and they tend to fall back on their heuristics, often going for less healthy alternatives.
10 Moreover, food producers sometimes induce information overload themselves. Hoping that consumers fall back on their heuristics, to not discourage consumers from purchasing unhealthy snacks and beverages. This has been one of the leading reasons for stricter labeling regulations (Persson, 2018). Moreover, shoppers with low-level literacy and numeracy skills are easily misled by incomprehensible labeling and are more likely to give up understanding food labels. Around 38% of all consumers are unable to comprehend information provided on packaging (Sørensen, Clement, Gabrielsen, 2012). Nevertheless, consumers have shown to partly protect themselves from information overload. When confronted with an information overload, consumers tend to create subsets of the information. This is done by mainly focusing on only three to five product dimensions, instead of all present information (Heroux, Laroch & McGown, 1988).
2.3 Effects of Nutri-scores
In recent years, both researchers and regulators have become more interested in food labeling systems. One of the food labeling tools that is gaining a lot of popularity is the Nutri-score. The Nutri-score is a new labeling system, developed in France throughout the last decade. One of the primary goals of Nutri-scores is to ease the consumers information processing and decrease information overload. As mentioned before, due to the increase in assortment sizes and an increase in information on packaging, consumers are experiencing an overload of information and choice. The Nutri-score supports consumers in their ability to properly compare products, which can improve their decision-making process.
Governments all over the world have been demanding more information on food packaging, with the social aim of improving health and safety (Golan, Kuchler, Mitchell, Greene & Jessup 2001). Both consumers, films, third parties and governments influence in what food attributes are being displayed on food packaging. Initially, the labeling systems were meant to simply inform consumers about nutritional values within products. However, recent years have seen a change in the purpose of food labeling. Government intervention has changed, to the point of influencing consumer choices to line up with social goals (Golan, 2001).
11 ‘healthy’ and ‘unhealthy’ score. The unhealthy score is calculated by measuring the amount of unhealthy nutrition and categorizing the results on multiple 10-point scales. The unhealthy nutrition that are considered are energy (KJ), saturated fatty acids, sugars, and sodium. Then the combined 10-point scale becomes an overall unhealthy score ranging between 0 and 40 points. This unhealthy score is then being subtracted by a healthy score. The healthy score is calculated by measuring the percentage of fruits and vegetables in the product, the fibers, and the protein, which are defined on a 5-point scale. The healthy score ranges between 0 and 15 points (Chauliac, 2018). Based on the number of points, the Nutri-score is being given a colored label. There are a few adaptations for three different food groups, namely added fats, cheeses and beverages (Chauliac, 2018).
Literature has revealed that Nutri-scores can positively influence purchasing behavior and raise product awareness. The study of Mora-García, Tobar, and Young (2019) found that by adding Nutri-scores, consumers were more likely to purchase healthy products compared to unhealthy products. Furthermore, they found that by adding Nutri-scores, the overall profit of a store could be increased, since consumers are more likely to purchase healthy alternatives which may be more costly. The study of Hagmann and Siegrist (2020) found that by using Nutri-scores, consumers became more aware of the unhealthiness of certain products, compared to when presented with the regular nutritional values on a package. Furthermore, the Nutri-scores helped by determining the healthiness evaluations of food by consumers. Consumers had an easier time to correctly decide which products were healthy and which were unhealthy. They also found that the effectiveness of the Nutri-scores strongly depends on the persuasiveness of the label. Indicating that if only a portion of the products use it, it will have a weaker effect than when all products are presented with Nutri-scores (Hagmann & Siegrist, 2020).
12 investigated the effect of nudging towards healthy food choices based on literature published between 2015 and 2018. They found that by nudging towards healthy food choices is effective and even increases awareness of healthy consumption. By adding Nutri-scores, the amount of nutritional data is being simplified and information overload is being partly reduced. This study expects that by adding Nutri-scores the consumer will be nudged towards making more conscious and healthier purchases, thus formulating the following hypothesis:
H1: When products are being shown with Nutri-scores, consumers will increase the overall healthiness of their online baskets.
Furthermore, this study expects that by adding Nutri-scores, consumers will gradually increase their awareness of healthy products. By adding Nutri-scores, consumers could decrease the tendency to fall back on their heuristics. Expecting that over time their purchase intention becomes healthier. Thus, formulating the following hypothesis:
H2: An addition of Nutri-scores will result in a more conspicuous and healthier purchase intention over time.
2.4 Effects of sorting and filtering
13
Even though interactive decision aids such as sorting are becoming increasingly more important, the current state of literature around product sorting is in the early stages. Whereas some studies have tried to measure the effect of efficient product sorting, an extension of current literature could help to further understand the actual effects and benefits of product sorting. The study of Cai and Xu (2008) investigated the different effects of designing product lists for E-commerce. Their study made the distinction between two types of product sorting, the first being sorting by price, the latter being sorting based on the quality of the product. Whilst product quality can be specified on varying attributes, this paper will use Nutri-scores as a substitute for quality. The study of Cai and Xu (2008) looked at the different types of sorting, either ascending, descending and random sorting and their effects on purchase. Based on their research, they found that ordering based on quality in an ascending format does not positively influence purchase intentions. People did not necessarily purchase more or higher quality products. Which does make sense, since in an ascending format, the first products showed are of a lower quality. However, they did find that ordering products by quality in a descending format does influence purchase intentions. Resulting in higher quality purchases and more frequent purchases. This study will use Nutri-scores as a substitute for quality, this study proposes that sorting and filtering based on healthiness results in more healthy purchases. And ultimately lead to a positive effect on wellbeing. Thus, the consumer can sort products based on the Nutri-score, ranging from dark-green A (healthiest) up to dark-orange E (unhealthy). Furthermore, ordering tools such as sorting, enables consumers to discover more suitable products (Häubl & Trifts, 2000).
14
+
+
+
+
proposes that by using interactive decision-aid tools such as sorting and filtering, consumers are more likely to purchase healthier products. Thus, the following hypotheses are formulated:
H3: When consumers use sorting as an interactive decision aid, they are more likely to purchase healthier products.
H4: When consumers use filtering as an interactive decision aid, they are more likely to purchase healthier products.
2.5 Conceptual model
Based on the literature, interactions between our predictors and predicted variable have been suggested. An overarching model will be used to test whether the formulated hypotheses are appropriate. Figure 1 signifies the expected relation between seeing the Nutri-score and healthiness of consumer purchases. Furthermore, it incorporates the expected effect that throughout time, the implementation of the Nutri-score increases the healthiness of purchases. Lastly, the model focusses on whether the interactive decision aids sorting, and filtering have a positive effect on the healthiness of customers items.
Figure 1 Conceptual framework
15
3. METHODOLOGY
This chapters focuses on the set up of the research design, the data collection and how the data was analyzed. Furthermore, this chapter will indicate how the data has been recoded and how certain outliers have been dealt with. It also gives an overview of how the data was merged for further analysis.
3.1 Data collection
16 3.2 Plan of Analysis
The primary focus of this study was to measure whether by adding Nutri-scores in an online grocery shopping environment, has a positive effect on the healthiness of purchases done by customers. Additionally, we tried to discover whether a reduction of information, and choice overload has a positive effect on the healthiness of purchases done by consumers. Furthermore, the richness of the given data was used, to model the online features, that either positively or negatively influence the basket size whilst doing online groceries.
3.2.1 Testing of hypotheses
Firstly, the main objective of was to understand whether an addition of the Nutri-score, in an online grocery shopping environment has a positive influence on the healthiness of consumers purchases. However, working with the current data had a major weakness. The current data does not indicate which products were actually displayed online with their corresponding Nutri-score. Thus, it is hard to account for the actual effect of the Nutri-scores. The healthiness of the basket was to be calculated by taking the overall average of products added to basket. Potentially including products that were not displayed with their corresponding Nutri-score. This research design is quite unfavorable but considering the availability of the data the only option.
As mentioned in the literature review, this study expects to find that by offering interactive decision aids such as sorting and filtering, the overall healthiness of consumer purchases will increase. Moreover, it is expected that over time the average healthiness of online grocery shopping will increase due to the addition of the Nutri-score. Because all of the formulated hypotheses share the same dependable variable, namely the average Nutri-score, we are able to incorporate all of the expected effects within one overarching model. Furthermore, the expected relationships were investigated through the use a Multiple Linear Regression (MLR) model. Multiple Linear Regression has some important assumptions which should be satisfied (Williams, Grajales & Kurkiewicz, 2013).
17 Goldfield-Quandt test (Leeflang, et al., 2015). Lastly, the multiple linear regression requires that all variables be normally distributed. This will be tested through the use of the Lilliefors (Kolmogorov-Smirnov) Test For Normality.
3.2.2 Additional analysis
As mentioned above, the richness of the data will be used to evaluate which factors in an online grocery shopping environment influence the online basket size. The results of this additional analysis could be used to benefit the supermarket in basket size optimalization. By anticipating which factors either decrease or increase the average basket size, online supermarkets are able to potentially alter basket size reducing aspects. This could lead to more conversion and higher revenues.
To appropriately measure how online factors, influence the basket size, a count model analysis was to be conducted. Through the usage of a count model, we will be able to distinguish which factors have either a positive or negative impact on the basket size of consumers. Working with count models requires the dependent variable, in our case products added to basket, to satisfy a few important conditions. The dependent variable has to be negative, discrete, and non-continuous (Leeflang, et al., 2015). To further elaborate, a Poisson model will be used to measure which online factors influence the size of the shopping basket. A Poisson model is appropriate to function with count data. The Poisson model assumes a that the dependent variable follows a non-normal distribution.
18 3.3 Data preparation
3.3.1 Merging and aggregating
The data provided by the Dutch supermarket consisted of a series of datafiles, which had to be merged based on their unique Session_ID and WI_nummer. The merging of the datafiles was done by PhD candidate D. Olk. Figure 2 indicates how the datafiles were merged and aggregated.
Figure 2 Dataset merging and variable creation
Following the merging of the datasets, some additional variables were made based on existing variables. The variable Page_Path contains the user’s behavior during a session. It indicates what the user did on the website of the supermarket during a session. Every new action and added product creates a new page path. Numerous variables were extracted out of the variable
Page_Path by PhD candidate D. Olk. The newly created variables provide some meaningful
insights. For example, it indicates whether an online shopper added any products to the basket that were on sale. It also indicates whether a user either filtering or sorting, based on the Nutri-score, sorted based on price, or sorted based on purchase frequency. Furthermore, it also created variables that indicate whether the user searched for any specific products, added products based on the website’s recipe lists or indicates whether the user had previously purchased any of the added products. Moreover, the variable Unique_Pageviews indicates how many pages were seen during the session. This variable could be used as a proxy for time spent
Page info •Session_ID •Date •Days_diff •Version •Treatment •Page_path •Sale •FilterNS •SortNS •SortPrice •SortPurchaseFreq •Search •Recipes •My list •PreviouslyBought •Unique_Pageviews Product info •Session_ID •Date •WI_Nummer •Control_Variant
•Products added to basket •Products seen online
Products Nutri-score •WI_Nummer
19 on the website during the session. Lastly, the variable Date was used to create the dummy variable Date_Diff, which is a variable that is used to indicate how long the experiment has been online. As mentioned previously, there were two distinct versions of the experiment. The original adaptation revealed that only a low number of customers in the treatment group became aware of the existence of the Nutri-score. The second adaptation tried to make the existence of the Nutri-score more salient, by adding an additional pop-up, which indicated that a Nutri-score has been added. The variable Version indicates which adaptation was present during a session. The variable Treatment shows whether the user was in the treatment group or the control group.
The dataset ‘Product info’ also provides some highly important information. This dataset was merged based on the Session_ID of the Page info dataset. This data indicates which products were added to the basket during the session. It also indicates how many products were seen during the session. Finally, the dataset with the Nutri-scores of the products were added to the dataset based on the WI_Nummer. This variable is a unique code for every specific product on the website of the supermarket. As mentioned before, the Nutri-score is a color graded score ranging from A (healthy) to E (unhealthy). For the upcoming analysis, the Nutri-score has been recoded, ranging from 1 (healthy) to 5 (unhealthy). By aggregating the data based on the session id, an average Nutri-score (NSavg) of the basket was calculated. Meaning that an increase in NSavg indicates an increase in unhealthiness. By merging the datasets into one covering datafile, analysis and model estimations are easier to be calculated and more understandable.
3.3.2 Recoding and outliers
One important limitation of the presented dataset is the nature of the sessions. As mentioned above, all the actions of online shoppers are tracked through the usage of the Page_Path. While aggregating the data on session level, all actions within the session are summed up into variables such as FilterNS. This variable now indicates how many times an online grocery shopper filtered based on the Nutri-score. This paper is interested in whether the usage of filtering and sorting has an effect on the average health of the basket. Thus, for the following variables: Sale, FilterNS, SortNS, SortPrice, SortPurchaseFreq, Search, Recipes, My_list and
PreviouslyBought will be recoded into dummy variables 0 and 1. Indicating whether the event
20 After a certain time of being inactive on the website, the website will create a new session. However, all previously placed products in basket will remain within the basket. Therefore, it is possible that multiple session added products towards one singular basket. Similarly, it is possible that due to the nature of the session, the number of products added to the basket can have a negative value. This occurs when an online user added products to their basket within one session, created a new session and removed the products from the basket. Since this is an irregularity within the dataset, all observations with a negative value for products added were removed. Resulting in an alteration of the total amount of observations within the dataset, changing from 2.638.620 to 2.618.989 observations.
Additionally, the numeric variables: unique_pageviews, products_added_to_basket and
products_viewed have been examined for outliers. Outliers are observations which do not
represent the average behavior of online shoppers. As explained previously, this data suffers from some irregularities due the fact that multiple session can be used for one singular basket. This also results in a difficulty for calculating outliers. Normally outliers are decided based on an interquartile range (IQR) outlier calculation. However, since multiple sessions could be used for one basket, calculating outliers through a traditional way results in a high amount of data loss. Thus, all observations with a value of unique_pageviews over 200 will be regarded as outliers. Similarly, for the variable products_added_to_basket, all observations with a value higher than 200 are classified as outliers. Lasty, for the variable products_viewed, all observations with a value higher than 1000 are viewed as outliers. Leaving out these outliers results in a change of observations, from 2.618.989 to 2.617.917.
21 online shopping behavior, the minimal number of products added to basket has been chosen to be 5. This will change the total number of observations for the MLR model from 528.953 to 265.527. Admittedly, this is a loss of information, but working with a smaller dataset that closely represents actual session behavior has a statistical advantage. Heavy programming statistics such as bootstrapping benefit from estimating with a smaller dataset.
3.4 Model formulation
After cleaning the dataset and recoding some of the variables into dummy variables. The following mathematical formula has been formed for the MLR model and the Poisson model respectively:
𝑁𝑆𝑎𝑣𝑔 = 𝛼 + 𝛽1𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡_𝐷1+ 𝛽2𝑈𝑛𝑖𝑞𝑢𝑒_𝑝𝑎𝑔𝑒𝑣𝑖𝑒𝑤𝑠2+ 𝛽3𝑆𝑎𝑙𝑒_𝐷3 + 𝛽4𝐹𝑖𝑙𝑡𝑒𝑟𝑁𝑆_𝐷4 + 𝛽5𝑆𝑜𝑟𝑡𝑁𝑆_𝐷5+ 𝛽6𝑆𝑜𝑟𝑡𝑃𝑟𝑖𝑐𝑒_𝐷6 + 𝛽7𝑆𝑜𝑟𝑡𝑃𝑢𝑟𝑐ℎ_𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦_𝐷7
+ 𝛽8𝑅𝑒𝑐𝑖𝑝𝑒_𝐷8 + 𝛽9𝑀𝑦_𝐿𝑖𝑠𝑡9+ 𝛽10𝑃𝑟𝑒𝑣𝑖𝑜𝑢𝑠𝑙𝑦_𝑏𝑜𝑢𝑔ℎ𝑡10 + 𝛽11𝑃𝑟𝑜𝑑𝑢𝑐𝑡_𝑉𝑖𝑒𝑤𝑠11 + 𝛽12𝐷𝑎𝑦𝑠𝑑𝑖𝑓12+ 𝛽13𝑉𝑒𝑟𝑠𝑖𝑜𝑛_𝐷13+ 𝜀 Figure 3 Mathematical formulation of the MLR model
𝑦𝑏𝑎𝑠𝑘𝑒𝑡 𝑠𝑖𝑧𝑒𝑖 = exp( 𝛼 + 𝛽1𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡𝑖1+ 𝛽2𝑈𝑛𝑖𝑞𝑢𝑒_𝑝𝑎𝑔𝑒𝑣𝑖𝑒𝑤𝑠𝑖2+ 𝛽3𝑆𝑎𝑙𝑒_𝐷𝑖3+ 𝛽4𝐴𝑣𝑒𝑟𝑎𝑔𝑒_𝑁𝑢𝑡𝑟𝑖𝑠𝑐𝑜𝑟𝑒𝑖4+ 𝛽5𝑀𝑦_𝐿𝑖𝑠𝑡_𝐷𝑖5+ 𝛽6𝑃𝑟𝑒𝑣𝑖𝑜𝑢𝑠𝑙𝑦_𝑏𝑜𝑢𝑔ℎ𝑡_𝐷𝑖6+
𝛽7𝑃𝑟𝑜𝑑𝑢𝑐𝑡_𝑉𝑖𝑒𝑤𝑠𝑖7+ 𝛽8𝑅𝑒𝑐𝑖𝑝𝑒_𝐷𝑖8 + 𝛽8𝑆𝑜𝑟𝑡𝑃𝑟𝑖𝑐𝑒_𝐷𝑖8 + 𝛽9𝑆𝑜𝑟𝑡𝑃𝑢𝑟𝑐ℎ_𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦_𝐷𝑖9) Figure 4 Mathematical formulation of the Poisson model
Figure 3 indicates which predictors are included in the MLR model. It is expected that the average Nutri-score (NSavg) is being influenced by all of the variables showcased in figure 3. Variables that are coded with an “_D” indicates dummy variables, which are binary variables with a value of 0 or 1. The other variables are coded as continuous variables. Moreover, figure 4 indicates the expected relation between predictors and predicted variable
products_added_to_basket. This model functions based on the same coding, with “_D”
22
4. RESULTS
4.1 Multiple Linear Regression assumptions
As mentioned before, a multiple linear regression has to satisfy several assumptions before insightful results can be interpreted. Myers (1990) states that a violation of assumptions could have huge implications for the validity of the results. Additionally, this is further emphasized by Osborne and Walters (2002). They state that by satisfying the assumptions, multiple linear regressions become more trustworthy. The first important assumption is the assumption of linearity. Linearity presumes that the right linear function has been chosen, and that the variables are correctly transformed and follow a linear trend. The second assumption is the absence of underlying multicollinearity between the predicting variables. Whensoever multicollinearity is discovered within the model, the estimates of predicting variables cannot be interpreted (Schmidheiny, 2013). Furthermore, the residuals within the model should not suffer from underlying heteroscedasticity, the residuals should follow a homoscedastic distribution. Osborne and Walter (2002) state that a slight violation of homoscedasticity is generally accepted but should be prevented if possible. Lastly, the residuals should be normally distributed. This is verified at the final step, primarily due to the fact that remedies of previous named assumptions could cause a non-normal distribution within the residuals (Leeflang, et al., 2015). When all the assumptions are satisfied, the coefficients can be properly interpreted (Myers, 1990).
4.1.1 Linearity
23 Implying that the current
functional form does not follow a linear form. One remedy for this problem is to modify the model. By plotting the martingale residuals against each predictor variable, we are able to see which
predictors follow a non-linear behavior (Lin & Zelterman, 2002).
The variables: FilterNS, SortNS, SortPrice and SortPurchaseFreq were changed from dummy variables into continuous variables, which is their original continuous form. Figure 5 indicates that after modifying the model, predicting variables such as FilterNS have a much better fit. Residuals of predictors that were previously unaccounted for are being predicted considerably better. After modification, there seems to be a more linear relation within the model. A secondary RESET test has been used to test the linearity of the modified model. The modified model was highly insignificant F(1, 265511) = 0.87, p = .3504. Implying that the modified functional form satisfies the linearity assumption. This is somewhat expected, since some of the dummies where changed into continuous variables, and continuous variables follow a more linear behavior compared to dummy variables.
4.1.2 Multicollinearity
The next assumption which will be tested is the underlying multicollinearity. This assumption assumes that there is no underlying correlation between predictors. Multicollinearity can be detected through various ways. One of the most popular ones is through calculating the variance inflation factor (VIF) statistics for each predictor. Generally, when the VIF statistics has a value above 5, multicollinearity can be assumed (Paul, 2006; Field, Miles & Field, 2012).
Table 1 shows the VIF statistic for all Table 1 VIF statistics of predictors
Independent variables VIF
Treatment 1.003 Unique_Pageviews 1.917 Sale_D 1.045 FilterNS 1.012 SortNS 1.006 SortPurchaseFreq 1.004 SortPrice 1.057 Recipe_D 1.083 MyList_D 1.062 PreviouslyBought_D 1.026 Product Views 1.697 Difdays 3.575 Version 3.568
Figure 5 Residuals after model modification
24
predictors. Based on the VIF statistics it can be concluded that there is no multicollinearity present within the current model. All predicting variables seem to be highly uncorrelated, with the exception of Difdays and Version. However, this is reasonable since the variable Version was derived from the variable Difdays. So, a slight correlation between these two predictors is expected. Based on the VIF statistic, the assumption that no multicollinearity is present within the model has been satisfied.
4.1.3 Homoscedasticity
The assumption of homoscedasticity requires that the error term has a variance that is constant. If this is not the case, heteroscedasticity is assumed, which is the opposite of homoscedasticity. Heteroscedasticity is different from the previous two assumptions. It is a less serious problem, since heteroscedasticity only affects the efficiency of the MLR model (Leeflang, et al., 2015). However, this paper still strives to revolve any potential heteroscedasticity problems. The Breusch-Pagan test was used to measure whether the assumption of homoscedasticity is violated. The Breusch-Pagan test measures whether the variance of the residuals differs across observation (Breusch & Pagan, 1979). The outcome from the Breusch-Pagan test for the formulated model is highly significant (nχ2(14, 265511) = 8804.4, p < .001). This indicates that our model suffers from significant heteroscedasticity.
Knowing that heteroscedasticity is present, our model should be transformed to accommodate for the assumption. Multiple remedies to resolve heteroscedasticity within our model were attempted. Firstly, to resolve heteroscedasticity we need to find out what causes it. Similarly, to resolving the functional form, heteroscedasticity can be visually interpreted by plotting the
residuals against each
predictor variable (Leeflang, et al., 2015). However, due to the high number of observations, detecting heteroscedasticity
visually can be quite
ambiguous. Heteroscedasticity is often caused by observations that differ over time (Breusch & Pagan, 1979). Periodical
changes in consumers lives Figure 6 Residuals plotted over variable Difdays
Sp re ad o f r es id u als
25
can every so often impact the uncertainty of the error term. Figure 6 indicates that the residuals up until day 22 are evenly distributed. However, after day 22, a slight change in distribution of the residuals can be observed. Followed by a decrease in distribution spread of residuals after day 22. A Goldfeld–Quandt test will be used to accurately measure whether the period before and after day 22 contains a different distribution of residuals. The Goldfeld-Quandt test divides the dataset in two different groups. The test then compares whether the two distinct groups differ in residual distribution. The Goldfeld-Quandt test uses a F statistic to statistically measure the difference between groups. After splitting the dataset in two groups, the Goldfeld-Quandt test found significant proof of heteroscedasticity F(13, 265512) = 0.978, p < .001. This
indicates that the distribution of residuals significantly differs between time periods. This could be caused due to different advertisement periods or other fluctuations over time. A Box-Cox transformation of the dependent variable (NSavg) was used to try resolving the problem of heteroscedasticity. Using a Box-Cox transformation can sometimes resolve the issue of heteroscedasticity within the residuals (Nwakuya & Nwabueze, 2018). However, the problem of significant heteroskedasticity was still present after transforming the data with the Box-Cox transformation (nχ2(14, 265511) = 9895.7, p < .001). Another well-known remedy for heteroscedasticity is the application of Generalized least squares (GLS) (Leeflang, et al., 2015). Generalized least squares is a statistical method to resolve heteroscedasticity, in both time series data and cross-sectional data. Since cross-sectional data is used, GLS is suitable to resolve the problem of heteroscedasticity. GLS uses both the datasets before day 22 and after day 22 to run two distinct MLR models. It then calculates the residuals of both models, followed by a transformation of both the independent variables and constant of the model, based on the model’s residuals. The newly transformed independent variables will be used to proceed in following statistical tests. However, there is one limitation when using GLS. It cannot be ensured that the problem of heteroscedasticity has been resolved entirely. Other variables could potentially still cause a heteroscedasticity within the residuals. However, by reapplying GLS over two new different datasets will likely nullify the effects of the first application of GLS. Thus, this paper will continue using the transformed variables by the first application of GLS, assuming that the problem of heteroscedasticity has been resolved.
4.1.4 Normality
26
when calculating the p-values for coefficients. When the error term follows a non-nominal distribution, the p-values cannot be trusted (Leeflang, et al., 2015; Osborne & Waters, 2002). Meaning that the significance of the estimates cannot be interpreted with confidence. Non-normality can be detected through plotting the distribution of residuals of the LMR model (Leeflang, et al.,2015). To statistically measure whether non-normality of the residuals is present, the Lilliefors (Kolmogorov-Smirnov) Test For Normality was used. When dealing with a lower number of observations, a Shapiro-Wilk test could also be used. However, since this dataset contains a high number of observations, consequently the Shapiro-Wilk test cannot be applied. The Lilliefors test is an EDF omnibus test, which tests whether normality of residuals is violated. The test can only be used reliable when the p-value is smaller than 0.1. After transforming our model through GLS, the Lilliefors test was conducted. The Lilliefors test indicated that the current model’s residuals are not normally distributed (D(13, 265512) = 0.05, p < .001).
27
4.1.5 Regression outcomes
After accommodating for all of the assumptions of a MLR model, the coefficients can be properly interpreted. Myers (1990) stated that by violating the assumptions, the validity of the results could be untrustworthy. However, given that all assumptions have been satisfied, we are now able to answer the hypotheses formulated within chapter 2.
Coefficient Estimate Std.Error t-Statistic Pr(>|t|) P(bootstr apped Constant -2.4341*** 0.0161 -171.391 0.000 - Treatment -0.0007 0.0035 -0.248 0.804 0.386 Unique_pageviews -0.0032*** 0.0000 -39.367 0.000 0.000 Sale_D -0.1263*** 0.0033 -36.431 0.000 0.000 FilterNS -0.0690*** 0.0067 -8.266 0.000 0.000 SortNS -0.0413*** 0.0072 -5.746 0.000 0.000 SortPurchaseFreq -0.0128*** 0.0025 -4.005 0.000 0.000 SortPrice -0.0061*** 0.0006 -8.653 0.000 0.000 Search_D -0.0111* 0.0051 -2.267 0.023 0.041 Recipe_D -0.0333*** 0.0057 -5.753 0.000 0.000 My_list_D -0.0189 0.0147 -1.456 0.145 0.177 Previously_bought_D -0.1625*** 0.0039 -44.690 0.000 0.000 Product_views -0.0009*** 0.0000 -45.434 0.000 0.000 Difdays -0.0004 0.0002 -1.800 0.071 0.134 Version -0.0487*** 0.0067 -7.539 0.000 0.000 R-squared 0.02238 F-statistic 434.1
Adjust R squared 0.02232 Prob(F-statistic) 0.000
Signif. Codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Table 2 Bootstrapped coefficients ranging the healthiness from 1 (healthy) to 5 (unhealthy)
As mentioned before, the Nutri-score has been recoded ranging from 1 (healthy) to 5 (unhealthy). An increase in NSavg indicates an unhealthier average online basket. Thus, predictors with significant negative estimates are positive in terms of healthiness.
28 products. Which is indicated by the predictor Treatment ( = –.0008, p = .38). This is a contradiction to the literature presented in paragraph 2.3, where research has shown that displaying products with a Nutri-score increases the healthiness of purchases. However, our contradicting finding could possibly be explained by the two distinct versions of the Nutri-score on the website. The variable Version indicates that consumers in version 2 purchased significantly healthier compared to version 1 ( = –.049, p < .001). However, the difference between versions is still relatively small. Being a consumer within the second version of the experiment lowers the Nutri-score by 0.049. Additionally, there is another possible explanation why the variable Treatment is not significant. This could be caused due to the fact that products that were not shown with scores were included when calculating the average Nutri-scores. The data did not offer any method to account for this irregularity within the treatment. As explained in paragraph 3.2.1, not all products within the treatment group were shown with Nutri-scores. However, all products were taking into account when calculating the NSavg, not just the products within the treatment group that were actually shown with Nutri-scores on the website.
Hypothesis 2 stated: An addition of Nutri-scores will result in a more conspicuous and healthier purchase intention over time. The variable Difdays failed to discover an increase in healthy purchasing behavior over time ( = .0004, p = .13). This is unexpected, since version 2 of the experiment performed significantly healthier than version 1. However, it is hard to fully incorporate what might have occurred beyond the scope of the variable Difdays. Since sales and other external factors were not accounted for. Furthermore, even though the variable
Difdays is insignificant, there is some indication that the effect over time could be present.
Mainly due to the fact that there is a low but insignificant p-value. A model that incorporates non-tested variables could indicate that an improvement in Nutri-score over time does exists. Additionally, the time period that was used to measure the effect of the variable Difdays only consists of 7 weeks.
29 consumers sorted their products based on price, the average healthiness of their basked became significantly unhealthier ( = .006, p < .001). One hypothetical explanation for this could be due to unhealthier products potentially being cheaper in comparison to healthy products. This is also supported by the fact that when consumers added products which were on sale to their basked, the average healthiness of the basked became significantly unhealthier ( = .126, p < .001).
Hypothesis 4 stated: When consumers use filtering as an interactive decision aid, they are more likely to purchase healthier products. This hypothesis is confirmed based on our model. In general, every time a customer filtered based on the Nutri-score, the average healthiness of their basket improved ( = -.069, p < .001). An interesting finding is that the effect of filtering based on Nutri-score is stronger compared to the effect of sorting based on the Nutri-score. This could be caused due to the large number of products being shown when using sorting compared to filtering. As mentioned before, the study of Diehl (2005) found that whenever products are shown in a with a large quantity of options, the quality of decisions will decrease. Thus, explaining that filtering, which shows less options in comparison to sorting, has a stronger effect on the healthiness of the purchasing behavior.
Additional findings
Our MLR model indicates that an increase in Unique_pageviews slightly improves the average healthiness of the basket ( = -.003, p < .001). Implying that if customers open more pages during a session, they tend to purchase healthier products. This could be caused due to pages other than products overview pages playing a role in the healthiness of baskets. In example, when products were directly added to the shopping basket from a recipe list, the average healthiness of the shopping basket improved ( = -.0333, p < .001). This implicates that pages such as recipe pages can have a positive effect on the healthiness of baskets.
-30
.162, p < .0001). This too is slightly surprising, since consumers who usually fall back on previously bought products often act based on heuristics. Additionally, when products are added based on previously bought lists, we would generally expect no change in basket healthiness. Furthermore, customers acting based on heuristics often fall back into routine picking of products, generally ending up choosing for less healthy alternatives (Garcia & Barreiro-Hurlé, 2019). Lastly, by viewing more online grocery products, the average basket becomes unhealthier, as an increase in Product_views increases the NSavg ( = .0009, p < .0001). This implies that for every product a customer sees online, the average healthiness of the basket worsens by 0.0009. This is consistent with the literature presented in paragraph 2.1. When consumers are overloaded with choice, embodied by a high number of products viewed, they tend to prefer unhealthy products over healthy products.
4.2 Poisson regression
As briefly explained in paragraph 3.2.2, this paper will additionally look at the richness of the data presented by the Dutch supermarket. The data will be used to get a general understanding, about which online factors in a grocery shopping environment, either increase or decrease the customers shopping basket. This data will be analyzed through the use of a Poisson model, which is a widely used form of regression, particular for count data (Lawless, 1987).
Similarly, to how the Multiple Linear Regression model has its assumptions. The Poisson regression also has some key assumptions which have to be accounted for before interpreting results. It needs to be excluded that predictors are mutually correlated among themselves. Similarly, to the procedure in paragraph 4.2.2, this study will make use of the VIF statistic to determine whether predictors are correlated. As mentioned in paragraph 3.2.2, different alterations of the Poisson regression can be used, depending on a violation of the assumptions.
4.2.1 Multicollinearity
31
3 indicates that all the VIF statistics are below a value of 2. Hence, within our model, no correlation between predictors has been found.
4.2.2 Poisson Assumptions
Besides multicollinearity, this study will verify two other important assumptions for the Poisson model. Firstly, as described in paragraph 3.2.2, the mean, and the variance of the number of products placed in basket have to be equal. If the mean and variance are of equal value, and the Lambda value is smaller than 10, a Poisson distribution can be assumed (Berk & MacDonald, 2008). The next assumption emphases on the number of baskets having a value of 0. The value of 0 can be interpreted as observations where 0 products were added to the basket, which can be derived from the variable: products_added_to_basket. There could be 3 distinct conditions that focus on the number of zeros. There could either be too many zeros, the number of zeros cannot be accounted for, or the number of zeros is suitable for a normal Poisson regression.
Variance mean and Lambda
As mentioned before, the first assumption of a Poisson model assumes that the equal and the variance are of equal value. This could either be checked through estimating the variance and mean traditionally. However, this paper will use a Dispersion test to statistically measure whether the variable products_added_to_basket follows a Poisson distribution. Loukas and
Kemp (1986) state that a Dispersion test accurately indicates whether the mean is larger than the variance (Underdispersion) or when the mean is smaller than the variance (Overdispersion). After running the Dispersion test, our dependent variable products_added_to_basket seems to suffer from Underdispersion Z = 76.754, p < .0001. The Dispersion test indicates that instead
Independent variables VIF
Treatment 1.000 Unique_Pageviews 1.911 Sale_D 1.063 NSavg 1.009 SortPurchaseFreq_D 1.004 SortPrice_D 1.047 Recipe_D 1.048 MyList_D 1.057 PreviouslyBought_D 1.041 Product Views 1.721
32 of a regular Poisson model, our data is more appropriate to run a negative binominal regression model. A negative binominal regression accommodates for a mean which is larger than the variance (Berk & MacDonald, 2008).
Observation of zeros
Besides the topic of Under, - and Overdispersion, the Poisson model differs based on the number of zeros within the observations. If the event of zero is occurs within the dataset, but is not over-represented, a normal negative binominal Poisson regression could be used. Another scenario is when there are too many zeros within the dataset. If this is the case, one could choose to proceed with a Zero-inflated or Zero-altered negative binominal Poisson regression (Ridout & Hinde, 2001). The last scenario is when the observation of zero cannot be observed. This happens when the variable products_added_to_basket, has no observations with a value of 0. If this is the case, a Truncated negative binominal Poisson model would be most suitable (Liu, Saat, Xin & Barkan, 2013). As explained in paragraph 3.3.2, this study focusses on observations that have a value larger than 0 for the variable _to_basket. Primarily due to the fact that observations with a value of 0 products_added-_to_basket do not have an average Nutri-score value. Thus, a Truncated negative binominal
Regression is most suitable, based on the assumptions.
4.2.4 Poisson regression results
The Truncated negative binominal model will be used to investigate which online factors influence the online basket size of customers. Table 4 gives an overview of the coefficients within the Truncated negative binominal Regression model. One important feature of a Truncated negative binominal Regression model are the two intercepts. The first intercept is the traditional intercept for any regression model. The second intercept indicates that dispersion parameter.
Coefficient Estimate Std.Error Z-value Pr(>|z|)
33 My_list_D -0.5001*** 9.318e-03 -53.66 0.000 Previously_bought_D -0.2043*** 3.429e-03 -59.57 0.000 Product_views -0.0048*** 2.342e-05 -207.25 0.000 Recipe_D -0.0251*** 5.632e-03 -44.60 0.000 SortPrice_D -0.1996*** 8.275e-03 -24.11 0.000 SortPurchaseFreq_D -0.2285*** 2.243e-02 -10.18 0.000 Log likelihood -14977728 Signif. Codes: ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
Table 4 Coefficients Truncated negative binominal Poisson regression
One of the first important findings of the Poisson regression on the basket size, there was no significant difference between the treatment groups ( = 0.0030, p = .35). This indicates that whether customers could see the Nutri-score or not, does not affect the basket size. Furthermore, both Unique_pageviews and Product_views increase the average basket size ( = 0.0034, p < .0001) and ( = 0.0048, p < .0001) respectively. This suggests that when customers see more products and pages during a session, the overall basket size increases. At face-value this seems to be an expected behavior. A slightly unexpected finding is the effect of Sale_D, this variable indicates whether a customer added products that where on sale to its shopping basket. The data suggest that when a product is added during a discount, the average basket size decreases by 0.7% (keeping all other variables in the model constant), which can be derived by taking the exponent of -0.0068 ( = -0.0068, p = .041). Additionally, the data shows that an increase in average Nutri-score results in a larger average basket ( = 0.0285, p <
.0001). This indicates that unhealthy products are purchased in a higher quantity than healthy products. This could be partially explained by our findings in paragraph 4.2.5. Where we found that when products are added to the basket during a sale, the average healthiness worsens. Implying that unhealthy products are added more often during a sale than healthy products. Which could be a result of unhealthy products being potentially more frequent on sale and might be promoted more often.
34 using personal lists, either made by customers or made by the grocery store, significantly increased the basket size of customers. However, the data suggest that when customers add products based on the recipes on the grocery store’s website, the basket size decreases ( = -0.0251, p < .0001).
35
5. CONCLUSION & DISCUSSION
This goal of this study was to discover whether the addition of Nutri-scores results in a healthier purchasing behavior, achieved by decreasing the amount of information. Making consumers more aware of the degree of healthiness of purchases. Which suggestively results in an increase in healthiness of grocery shopping. Furthermore, a supplementary goal of this study was to see whether interactive decision making tools such as filtering and sorting reduced choice overload. Resulting in a more contemplated purchasing behavior of customers, focused on healthy purchasing.
5.1 Conclusion
Surprisingly, no evidence was found that indicates an increase in healthiness of consumer purchasing by implementing the Nutri-score. This is a conflicting finding based on the literature of paragraph 2.3. However, as mentioned in paragraph 3.2.1, not all products within the treatment group were displayed with Nutri-scores. The data does not accommodate for us to indicate which products were included with a Nutri-score and which products were not. This could be one of the reasons why the effect of the Nutri-score does not significantly influence the healthiness of grocery purchases. Furthermore, Hagmann and Siegrist (2020) state that when only a fraction of products is showcased with corresponding Nutri-scores, the effect is weaker compared to when all products are showcased with Nutri-scores. This could be another explanation why this paper failed to find any significant evidence that proves the effect of Nutri-scores on healthy purchasing behavior. However, this paper found find that during the second phase of the experiment, customers on average had significant healthier baskets compared to phase one. This indicates that by making the Nutri-score more salient on the website, consumers do increase the healthiness of their purchases. Nevertheless, it is unknown whether the difference between phase one and phase two is caused by the change on the website of the supermarket. It could potentially be caused by uncontrolled external factors.
36 improvement of healthiness of grocery purchases. Another reason for the absence of a learning effect could be due to an impact of promotions and discounts. It could be one possibility that unhealthy brands and products are more heavily advertised compared to healthy alternatives. Since our study does not include promotions, it is hard to find statistical evidence that a learning effect is present.
Lastly, this study found that both filtering and sorting based on the Nutri-score increases the average healthiness of customers groceries. This shows that generally whenever customers are aware of the existence of the Nutri-score, they improve the healthiness of their groceries. This is in line with our literature, which states that a decrease in choice overload can be achieved through filtering and sorting (Shi & Zhang, 2014). Which would ultimately lead to the acquisition of products that are in line with customers goals to eat healthier (Häubl & Trifts, 2000). Another interesting finding is that the effect of filtering seems to be more helpful than the effect of sorting. This could be explained by the fact that with filtering, the customer only sees products of for example Nutri-score ‘A’. With sorting the products are ranged assumably from A to E. This would mean that whenever a customer scrolls down far enough they eventually see products with a B instead of an A. In the case of filtering, the customer will only be presented products that are based on their set criteria. Additionally, when customers sort based on price, the average healthiness seems to worsen. This could potentially be explained by the assumption that unhealthy products are apparently lower in price. Which would result in unhealthy products being shown first when sorting based on price.
5.2 Future Research
37 and filtering is in the early stages. This study has shown that sorting and filtering play an important role in the reduction of information overload. Future research could further investigate the subjects of filtering and sorting.
Moreover, future research could potentially investigate whether the period of day influences the healthiness of grocery shopping. The study of Harris, Bargh and Brownell (2009) states that during the day, consumers are continuously depleting their cognitive resources. This continuous usage of cognitive resources and self-control results in a shortage during the later hours of the day. Making consumers more susceptible to commercials and increasing the effect of the seen advertisement. Additionally, the study of Powell, McMinn and Allan found a similar finding. They state that during the day the ability to access self-control fluctuates over time, resulting in a bigger possibility of self-regulation failure. This is also in line with the fact that in general, consumers snack more during the evenings than during other times periods of the day (Masterson, Kirwan, Davidson & LeCheminant, 2016).
5.3 Limitations
This study offers some new informative insights in the field of nutritional food labeling in the form of information reduction and a reduction in choice overload. However, there are some important limitations that have to be formally addressed.
First of all, one of the main limitations of this study is the inability to account for the fact that some products were, and some products were not showcased with a Nutri-score. This is a major limitation, as it has resulted in a non-robust statistical design. Since products that were not displayed with scores were also taken into account when calculating the average Nutri-score per basket, the statistical measurements might have been influenced by the existence of these products. This might be one reason why this study deviates from earlier literature in terms of effectiveness of the Nutri-score. Furthermore, it can be concluded that based on the way how the Nutri-score is presented on the website, the effectiveness of the Nutri-score fluctuates. In phase two, a significant difference in the impact that the Nutri-score has compared to phase one can be observed. This indicates that by properly displaying the Nutri-score, and making it more salient, the total effect of the Nutri-score might increase.
38
consequential role in the buying behavior of consumers, promotional activities might have influenced the average Nutri-score, due to a higher volume of sales. Our model does not incorporate specific discounts that vary over time.
39
REFERENCES
Alba, J., Lynch, J., Weitz, B., Janiszewski, C., Lutz, R., Sawyer, A., & Wood, S. (1997). Interactive home shopping: consumer, retailer, and manufacturer incentives to participate in electronic marketplaces. Journal of marketing, 61(3), 38-53.
Bauer, H. H., Heinrich, D., & Schäfer, D. B. (2013). The effects of organic labels on global, local, and private brands: More hype than substance? Journal of Business Research, 66(8), 1035-1043.
Berk, R., & MacDonald, J. M. (2008). Overdispersion and Poisson regression. Journal of
Quantitative Criminology, 24(3), 269-284.
Bialkova, S., Grunert, K. G., & van Trijp, H. (2013). Standing out in the crowd: The effect of information clutter on consumer attention for front-of-pack Nutrition labels. Food
policy, 41, 65-74.
Billington, C. J., Epstein, L. H., Goodwin, N. J., Hill, J. O., Pi-Sunyer, F. X., Rolls, B. J., & Wing, R. R. (2000). Overweight, obesity, and health risk. Archives of Internal Medicine, 160(7), 898-904.
Brambila-Macias, J., Shankar, B., Capacci, S., Mazzocchi, M., Perez-Cueto, F. J., Verbeke, W., & Traill, W. B. (2011). Policy interventions to promote healthy eating: a review of what works, what does not, and what is promising. Food and nutrition bulletin, 32(4), 365-375.
Breusch, T. S., & Pagan, A. R. (1979). A simple test for heteroscedasticity and random coefficient variation. Econometrica: Journal of the Econometric Society, 1287-1294. Cai, S., & Xu, Y. (2008). Designing product lists for e-commerce: The effects of sorting on consumer decision making. Intl. Journal of Human–Computer Interaction, 24(7), 700-721.
40
Chauliac, M (2018) “NUTRI SCORE THE FRONT OF PACK NUTRITION LABELLING SCHEME RECOMMENDED IN FRANCE”, ECEuropa, April 23, available from:
https://ec.europa.eu/food/sites/food/files/animals/docs/comm_ahac_20180423_pres4.p df, (accessed October 13, 2020)
Chernev, A., & Hamilton, R. (2009). Assortment size and option attractiveness in consumer choice among retailers. Journal of Marketing Research, 46(3), 410-420.
De best, R (2020) “Share of consumers who purchased food or groceries online in selected countries in Europe from 2006 to 2019”, Statista, June 4, available from:
https://www.statista.com/statistics/915391/e-commerce-purchase-rate-of-food-or-groceries-in-europe-by-country/#statisticContainer, (accessed October 5, 2020)
de Edelenyi, F. S., Egnell, M., Galan, P., Druesne-Pecollo, N., Hercberg, S., & Julia, C. (2019). Ability of the Nutri-Score front-of-pack nutrition label to discriminate the nutritional quality of foods in the German food market and consistency with nutritional recommendations. Archives of Public Health, 77(1), 28.
Diehl, K. (2005). When two rights make a wrong: Searching too much in ordered environments. Journal of Marketing Research, 42(3), 313-322.
Diehl, K., Kornish, L. J., & Lynch Jr, J. G. (2003). Smart agents: When lower search costs for quality information increase price sensitivity. Journal of Consumer Research, 30(1), 56-71.
Dwyer-Lindgren, L., Bertozzi-Villa, A., Stubbs, R. W., Morozoff, C., Kutz, M. J., Huynh, C., & Flaxman, A. D. (2016). US county-level trends in mortality rates for major causes of death, 1980-2014. Jama, 316(22), 2385-2401.
Eppler, M. J., & Mengis, J. (2008). The concept of information overload-a review of literature from organization science, accounting, marketing, mis, and related disciplines (2004). In Kommunikationsmanagement im Wandel (pp. 271-305). Gabler.
