Abstract
Online grocery shopping has risen rapidly over the last years. Consumers, however, still have a lack of trust in online grocery shopping and perceive it to be risky. This, together with the better marketing possibilities that online shopping has to offer, indicates that there might be an alteration of shopping behaviour by consumers when grocery shopping online compared to offline. The purpose of this research was to investigate if doing online grocery shopping leads to an extension of the total number of purchased products via more purchased brand- and product-types.
Preface
My academic career at the Rijksuniversiteit Groningen started almost 2.5 years ago as a Pre-Master Marketing student. Because my previous study was a study at a university of applied sciences, this Pre-Master year was beneficial for me in getting used to more academic courses. After finishing the Pre-Master year, I decided to follow both the Intelligence track and the management track of the Master Marketing.
During the past 1.5 academic year, I was able to gain knowledge in both Marketing Management courses as; Retail and Omnichannel Management, Customer Management, Purchasing and Business to Business Marketing. Next to that, I was able to gain knowledge in Marketing Intelligence courses as; Data Science & Marketing Analytics, Market Models, Digital Marketing Intelligence and Customer Models. Following both courses provided me thoroughgoing knowledge of the overall field of Marketing.
Although at first, I was a bit uncertain whether I was capable of doing the Marketing Intelligence courses, looking back at the Master, these courses interested me the most. Hence, I was thrilled the topic of my first choice, “generating insights in a data-rich environment,” turned out to be the topic for my thesis. The part of my thesis that I enjoyed the most was working with the data, and this suited the topic entirely. Furthermore, researching a relevant topic like online grocery shopping was exciting to do, and analysing this using big-data gave me the feeling that I was doing a meaningful research.
1. Introduction
1.1 The rise of online grocery shopping
The improving connectivity and speed of the internet have risen the worldwide usage of the internet drastically throughout the last years (ITU, 2017). This digital revolution has increased the ease for people to obtain available information, engage in social- or economic-interactions and participate in online platforms. Also, because of the increasing use of the internet, online shopping has evolved into one of the most popular internet-activities. Worldwide e-commerce sales grew 28% in 2017, 22.9% in 2018 and forecast are that this growth keeps going on for the next five years, although these growth rates are expected to go below 20% by 2020 (Lipsman, 2019).
Traditionally the e-commerce categories fashion, travel and books contribute to the majority proportion of online sales. Although these are still the most popular categories among online consumers, the categories restaurant delivery-services and online grocery shopping show the most significant growth (Nielsen, 2018). The reason for this is that fashion, travel and books are generally the first online purchase for consumers. However, after multiple online transactions, consumers become more familiar with online shopping. Hence, they tend to expand their online purchases into other areas like beauty and personal care and eventually to an even wider field like online grocery shopping (Nielsen, 2018).
The frontrunners to shop online groceries were the Millenials in Europe, as in 2016, almost 50% of the European consumers between 16 and 24 years old had shopped online for groceries (Mintel.com, 2016). Currently, the heaviest European online grocery shopper is between 25 and 44 years old, with an above-average household income, higher educated and lives in an urban environment (Forbes, 2019). Within Europe, the UK has had the most online orders for food and grocery as a percentage of the total ordered online goods or services for almost a decade now (Forbes, 2019).
Scientific research can be useful inputs for grocery retailers to form an online strategy. Due to the emergence of online grocery shopping, already much academic literature regarding online grocery shopping has been released over the past years (Cebollada, Chu, Jiang, 2019; Elms, Kervenoeal, Hallsworth, 2016; Kureshi & Thomas, 2019: Richards & Rabinovic, 2018). A common theme in researches regarding online grocery shopping is the lack of trust and perceived risk that consumers have towards online grocery shopping (Driediger and Bhatiasevi, 2019; Ramus & Nielsen, 2005; Pauzi et al, 2017).
This lack of trust and perceived risk might also result in different types of purchase behaviours when consumers shop their groceries online compared to offline. These different behaviours could come in the form of possible online-offline differences in the repertoires of consumers’ shopping baskets or may restrain consumers from major online expenditures. Research about the online-offline differences in the repertoires of consumers shopping baskets and the development of online expenditure regarding grocery shopping is still limited (Chu, Chintagunta, Cebollada, 2008; Chu et al, 2010). This, while the risk and trust issues that consumers have with online grocery shopping, might indicate that there is an alteration of consumers' shopping behaviours when shopping for groceries online. Hence, this research aims to analyse if these indications are correct.
1.2 Online-offline differences in repertoires
Certain product types are more acceptable to purchase online than others (Sam & Sharma, 2015), resulting in online-offline differences in the repertoire of shopping baskets at online retailers. This is not different for online grocery retailers because Chu et al (2010) found that consumers are more brand loyal and size loyal but less price-sensitive when grocery shopping online. Also, these online-offline differences are more substantial for food products and sensory products. Indicating that when shifting their grocery shopping practices from off- to on-line, consumers are likely to alter their brand- and product-choices. Furthermore, this indicates that consumers will purchase certain brand- and product-types more online.
be one of the primary drivers for consumers to rely more on particular brand- and product-types (Moore & Andradi, 1996; Png & Reitman 1995; Wernerfelt 1988 in Chu et al, 2008).
Additional knowledge about the online-offline differences in brand- and product-repertoires could be useful for grocery retailers in setting their online strategy and determining which products they should offer more prominent online to attract and retain customers. Moreover, this is especially interesting because grocery retailers can fully control their on- and off-line marketing strategies and hence further optimize them based on these findings. Being on time with implementing such a fact-based strategy will give retailers a first-mover advantage. Such an advantage will be crucial for grocery retailers in capturing a significant market share in the e-commerce market (Galante, Lopez, Monroe, 2013).
1.3 Effects on customers’ online expenditure
Lack of trust and the perceived risk towards online grocery shopping could decrease the likelihood of major online consumer expenditures. However, when online shopping becomes more habitual, consumer trust in online shopping tends to increase (Chou & Hsu, 2016). It would be interesting to see if this also applies to online grocery shopping and if the expenditure level of online grocery consumers increases over time, as they are getting more familiarized with online grocery shopping. Also, this would provide further insights between the on- and off-line differences in the repertoires of shopping baskets. Since an increase in online grocery shopping expenditure could lead to the cannibalization of offline grocery expenditures.
Furthermore, downsides of the e-grocery business are the additional costs that are coming with the delivery service. Mainly because these costs can be higher than the fees consumers are willing to pay for delivery (Galante et al, 2013). Hence, it would be, additionally, useful for online grocery retailers to know whether the expenditure level of online grocery consumers increases over time.
grocery shoppers could also be interesting to study because when online grocery shopping experience increases, consumers' tend to shop at more than one chain (Melis et al, 2015).
1.4 Research question
As discussed, there is still little research available on the alteration of consumers' shopping behaviour in terms of brand- and product-choices, and expenditures, when shifting from off- to on-line grocery shopping. This research will contribute to the current academic literature by providing a more comprehensive view of this given topic. Next to that, this research will contribute to practice by providing insights on how grocery retailers should shape their online strategy to enhance online performance.
Given the limited research on the possible online-offline differences in the repertoires of shopping baskets and developments of online expenditures, this research considers the following research question:
How does online grocery shopping alters consumer behaviour in terms of the number of purchased products, purchased brand- & product-types and expenditure-patterns, compared to offline grocery shopping? To answer the research question, several sub-questions have been prepared: Are certain brand types more likely to be purchased at online grocery stores than offline grocery stores? Are certain product types more likely to be purchased at online grocery stores than offline grocery stores? Is online grocery shopping leading to an increase in the number of purchased products, via particular brand- and product-types?
How does online grocery shop experience affect the level of online grocery expenditure, also in terms of cross-channel expenditure?
Because the UK is the leading country in online grocery shopping within Europe (Forbes, 2019), the scope of this research will be on the British food market. Since this will give a good representation of one of the front running countries in online grocery shopping and hence can make clear in which direction the e-grocery business of other countries will go.
2. Theory
In this section, more elaboration will be given on each topic and definitions will be given to unknown words and expressions. Further, there is a review of prior researches on the online-offline differences in the repertoires of shopping baskets and consumers' online expenditure. Since particular brand- and product types are reasoned to gain more preference online, it will be expected that this will lead to an increase in the total number of purchased products. Based on all of this, hypotheses are formulated.
A recurring theme in this theoretical section is the lack of trust and risks consumers still perceive towards online grocery shopping (Nittala, 2015). Because of this, consumers tend to rely more on brands and specific product types online compared to offline. As a result, this will alter the repertoire of their shopping baskets when grocery shopping online. Further, the perceived risk towards online grocery shopping withholds consumers from high expenditures, especially when they are starting with online grocery shopping.
2.1 Altered repertoires, online compared to offline
Digitalization has resulted in the possibility for the consumer to shop groceries not only in physical stores but also to shop groceries online. Kurnia and Chien (2013) refer to online grocery shopping as the use of retailers’ web sites by consumers to purchase grocery products by simply clicking the mouse button for the required items. The retailer will then make home delivery.
Although the same grocery products can be purchased both on- and off-line, multiple studies indicate that consumers prefer to buy certain types of products via the online channel (Chu et al, 2008; Chu et al 2010; Dawes & Nenycz-Thiel, 2014; Degeratu, Rangaswamy, Wu, 2000). In the next section, further elaboration will be given on the reasons behind these preferences.
Online loyalty
The available features on grocery retailers’ websites can explain why consumers tend to buy more similar brands and sizes online because these features enable consumers to use saveable shopping lists and to search for products by price, brand or size. This makes it more likely for consumers to purchase similar brands and product-sizes repeatedly. Furthermore, if e-retailers are allowed to use customer-cookies, then they can personalize the web design to the wishes of the customer (Wills & Zeljkovic, 2011), this allows the retailer to place the customers' preferred brands and products more prominent on the website. Because online grocery shoppers mostly buy their products from the first category page displayed on the website and are consistently using the default options (Anesbury et al, 2016), personalized websites will likely stimulate consumers to purchase certain brands and products repeatedly.
Moreover, because the online features allow priming these repurchased brands
more, after priming, mental associations will be activated and consumers will retrieve all their brand knowledge from a particular brand (Keller, 1993). Hence, they will identify the primed brand as one of their essential routine-purchases and are likely to add it to their shopping basket after priming or exposure. Such behaviour is likely to occur even though consumers initially did not plan to purchase this product. Hence, consumers will purchase their more preferred repurchased brands more frequently online, leading to an increase in the total number of purchased products.
Further, trying new products is generally associated with risks (Popielarz, 1967). This, together with the perceived risk towards online grocery shopping (Nittala, 2015), implicates that purchasing a new brand or product online will be associated with even higher risk. Hence, to encounter both the risks of online grocery shopping and purchasing a new product, it likely that consumers rather repurchase familiar brands when grocery shopping online than buying a brand that is new to them.
Furthermore, online, consumers are searching for the optimal trade-off between search cost and search performance (Kumar, Lang, Peng, 2005). Repurchasing brands often will keep the search costs low while getting the best available product for the effort put in.
H1A. Online grocery shopping will increase the number of products purchased by consumers via more repeatedly purchased brands. H1B. Online grocery shopping will increase the number of products purchased by consumers via more repeatedly purchased product sizes. Brands Because it is expected that consumers are more likely to purchase similar brands online, this might also indicate that they rather buy certain types of brands online compared to offline. Meaning that the composition of brand repertoire for consumers online differs from offline. Multiple researches indicate that such an online-offline difference in types of brands might exist. However, multiple types of brand differences could be possible.
Firstly, Arce-Urriza and Cebollada (2018) found that when grocery retailers become active online, this will generally improve the performance of their private label brands. Private label brands, or store brands, are product labels used by retailers who are the owner of the brand (Bruhn, 1996 in Bruhn, 1996). Private label brands are found to have an increased market share online (Arce-Urriza & Cebollada, 2012). The reason for the better performance of private label brands online could be because grocery retailers that are operating both on- and off-line are customizing their websites in such a manner that it is easier for consumers to purchase private label brands (Arce-Urriza & Cebollada, 2012). These findings would indicate that it is expected that online grocery shopping will increase the number of products purchased by consumers via more purchased private label brands.
H2A. Online grocery shopping will increase the number of products purchased by consumers via more purchased private label brands.
Because consumers are less price-sensitive when grocery shopping online, they could also be more likely to purchase national brands or premium own brands. National brands are brands not owned by retailers but by a manufacturer (Arce-Urriza & Cebollada, 2012). A premium own brand is an extended retailer-own brand known for its high value-added products, innovative design and sometimes even higher quality than national brands (Ahlert et al, 2009). Premium own brands and national brands are more expensive and given the lower price sensitivity of online consumers, these brands are more likely to be purchased online. Thus, it is expected that online grocery shopping will increase the number of products purchased by consumers via more purchased premium own- and national-brands. H2B. Online grocery shopping will increase the number of products purchased by consumers, via more purchased premium- & national-brands. From origin, high share brands have more buyers and are tend to be bought more often (Fader & Schmittlein, 1993). Furthermore, Danaher, Wilson and Davis (2003) observed that for a grocery brand with a high market share, the brand loyalty is significantly higher online. Moreover, they also found that for brands with a low market share, brand loyalty is significantly higher offline. The reason for this could be that consumers are not able to assess the quality, look and feel of products when shopping online. As a result, consumers might rely more on well-known brands. Well-known brands generally are brands with a high market share, next to that, they also have an above-average quality and are considered a less risky purchase (Moore & Andradi, 1996).
Further, high market share brands are more preferred by consumers and with customer cookies, e-retailers can observe this (Wills & Zeljkovic, 2011). Which after they will place these brands more prominently on grocery retailers’ websites. By doing this, the website will be more convenient for consumers and it will require less effort to find the desired product. Next to that, it also means that high market share brands are more likely to be purchased online. Hence, it is expected that online grocery shopping will increase the number of products purchased by consumers via more purchased high market share brands.
Product types Next to preferring certain brands, consumers might also prefer to buy certain product types when grocery shopping online. Firstly, it might be more likely for consumers to buy non-food products when grocery shopping online because they are not able to assess the quality of food products. It must be noted, however, that for certain product types, consumers can rely on brands to which they relate certain quality standards. Hence, consumers are more brand loyal when grocery shopping online for food products than for non-food products (Chu et al, 2010). Also, consumers generally tend to assess products of high quality when they are priced above average (Valentin & Grazin, 1990).
Nevertheless, product types like vegetables and fruit are displayed in grocery stores in a way that these can be physically examined pre-purchase. Also, fruit and vegetables, just like meat and dairy, have a limited expiration date. Online, consumers cannot physically examine products or see the expiration dates of the products. For this reason, online vegetable shopping is less likely to be done by consumers that find health and taste important (Kang et al, 2018). When grocery shopping online for fruit and vegetables, consumers’ need to rely that grocery retailers will deliver on their quality standards. Nevertheless, not all consumers likely trust that the employees of these retailers will send the products that fulfil their wishes. For non-food products, however, there is no expiration date and physical examination is less applicable. Hence, there is less of a barrier to buying non-food products online than food products and it is expected that consumers will more likely grocery shop online for non-food products.
found a long tail effect at online grocery shopping and highlighted the importance of niche categories. Non-food categories are not core products for grocery retailers and can thus be considered as niche products. So, together with the previous reason, it is also likely that the more available non-food products will make it more likely for consumers to buy non-food products when grocery shopping online. Thus, it is expected that online grocery shopping will increase the number of products purchased by consumers via more purchased non-food products.
H3A. Online grocery shopping will increase the number of products purchased by consumers via more purchased non-food products.
Just like food products, sensory products are also hard to verify online (Hansen, 2005). Sensory products are products characterized by its sensory attributes. These are attributes that can be directly examined through our senses before purchasing (Degeratu et al 2000). Because these attributes cannot be verified online, it is hard for consumers to assess the quality and overall feel of sensory products (e.g. the smell and look of a mango). Thus there is a discrepancy between the possibilities of online shopping and the ability to assess quality, which makes it less likely for consumers to grocery shop sensory products online (Raijas, 2002). Hence, Chu et al (2008) found that sensory grocery products are purchased more offline.
Because of this, consumers might separate these two product categories in that
sensory products being more suited for offline and non-sensory products more for online grocery shopping. Hence, it is expected that online grocery shopping will increase the number of products purchased by consumers via more purchased non-sensory products.
H3B. Online grocery shopping will increase the number of products purchased by consumers, via more purchased non-sensory products.
budget of FMCG brands usually goes to price promotions, making it their primary promotion tool (Bolton, Shankar, Montoya, 2010 in Krafft & Mantrala, 2010).
If e-retailers are allowed to use customer-cookies, then they can personalize online promotions and web-shops to the wishes and needs of the individual customer (Wills & Zeljkovic, 2011). Because of this, consumers will be exposed to much more relevant promotions online compared to offline since offline promotions are generally aimed at the mass-consumer. So, because the promotions online are more relevant for consumers than in-store, it is expected that online grocery shopping will increase the number of products purchased by consumers, via more purchased promoted products. H3C. Online grocery shopping will increase the number of products purchased by consumers via more purchased promoted products. So, it is expected that the repertoires of the shopping baskets from consumers will be altered from off- to online in a sense that consumers will purchase more similar brands and product sizes. Also, consumers will purchase more private label brands, premium- & national-brands and more high market share brands. Lastly, consumers will purchase more non-food products, non-sensory products and promoted products. Because these brands and products are expected to gain more preference online, it is expected that online grocery shopping will lead to an extension of the repertoire of consumers’ grocery baskets, and hence increase their total number of purchased products.
2.2 Effects on customers’ online expenditure
Next to differences in the repertoires of shopping baskets, it is also likely that grocery expenditures online differ compared to offline. Because when consumers begin with online grocery shopping, low-trust and high-perceived risk will likely decrease the chance of substantial consumer expenses (Nittala, 2015; Rafiq, Fulford, Lu, 2013) Nevertheless, these risk- and trust-issues will diminish over time and hence this will result in a change in online grocery expenditure.
attitude toward an online business, resulting in repeat purchasing behaviour’ (Anderson & Srinivasan, 2003).
General level of online expenditure
Consumers who stop with online grocery shopping likely had a lack of satisfaction with the online purchasing process, product assortment, delivered products, relationship quality or trustworthiness of the online grocery retailer (Chou and Hsu, 2016; Rafiq et al, 2013). Because consumers were presumably not satisfied with the online grocery retailers, their purchase intentions most likely also were not high when they were still online grocery shopping (Chou and Hsu, 2016). Hence, it is expected that when consumers are about to stop with online grocery shopping, this will harm online grocery expenditure. H4A. Consumers that are about to stop with online grocery shopping, negatively affect their current online grocery expenditure. Trust towards online grocery retailers has a strong effect on the e-loyalty of consumers in online grocery shopping (Rafiq et al, 2013). The reason for this is that consumers generally perceive online grocery shopping to be risky (Nittala, 2015). However, trust in online shopping significantly increases when consumers have developed online shopping habits. Moreover, online shopping habits increases consumers’ repurchase intentions (Chou and Hsu, 2016). Hence, it is likely that when consumers are more experienced online grocery shoppers by repeatedly grocery shopping online, their trust towards online retailers increases and also their loyalty- and expenditure-level will increase. Meanings that the expenditure of consumers on online grocery shopping will increase the more experienced consumers are with online grocery shopping.
H4B. Online grocery shop experience has a positive effect on consumers’ online grocery shop expenditure.
Chain-level online expenditure
not pleasant experiences with the retailer at which they are currently shopping (Singh, 2019).
2.3 Conceptual model
Based on the formulated hypotheses, the following conceptual model (figure 1) has been developed. Firstly, the independent variable indicates whether consumers have shopped their groceries on- or offline. This will lead to different buying behaviour in terms of repurchase behaviour, brands and products. The preferred brands and products will lead to an increase in the total number of products purchased by households.
3. Methodology
In this chapter, the research approach will be described step by step. By explaining what research type will be done, on which population the research will be tested, the sampling size, the representativity of the sample, the required data preparation steps and lastly the different analysing techniques. From here on out, the research question and hypotheses related to online expenditure will no longer be discussed. The reason for this is that adding this topic would seem too extensive for a single topic paper. Further explanation on this follows in the discussion-chapter. 3.1 Quantitative method
To analyse a broad number of observation that will help generalize findings, a quantitative research method was selected for this study.
The data that will be used comes from Kantar, one of the world-leading Consultancy Companies in the field of data insights. Kantar has over 30,000 employees and more than half of the companies in the Fortune 500 list use their services. The panel-data that is being used is data from two years and contains over 67 million observations from over 35,000 households shopping at primarily nine supermarket chains in the United Kingdom. The UK is the leading country in online grocery shopping within Europe (Forbes, 2019) and this dataset will thus give a good representation of one of the front running countries in online grocery shopping.
From five of the nine supermarket chains, there is data available on whether consumers have purchased their products online. These supermarket chains are Tesco, Asda, Sainsbury’s, Waitrose and Marks & Spencer. Further, there is data available from unknown other supermarkets from whom it is also known whether purchases have been made online. Lastly, there is data available on expenditure, brands, product-types and whether or not there was a promotion present for a product. Meaning that the available data makes it possible to test all the formulated hypotheses.
3.2 Data preparation Dummy-coding procedure
share brands) and product types (non-food products and non-sensory products). This makes it possible to test the effect on all the selected mediation variables. Regarding the brand variables, there is data available of 9229 brands from 3738 manufacturers. Hence, it would not be feasible given the time frame to code every brand. Because of this, the product categories cola, peanut butter and instant coffee are chosen to test the brand-related hypotheses on. These product categories were chosen because these are everyday groceries characterized by having both a broad range in national brands and private label brands.
All the manufacturers and brands of the product categories cola, peanut butter and instant coffee have been researched online. Based on this, the brands found to be manufactured and solely being sold by one retailer are coded as private label brands. Further, the manufacturer-owned brands and brands that are being sold for an above-average price are coded as premium- & national-brands. The brands that combined have over 50% of the market share for each product category are coded as high market share brands. Both private label brands and national brands are included here because even though the top-selling private label brands are highly dependent on supermarket size, these are the brands with the highest market share. Meaning that, as discussed in the theoretical section, these top-selling private label brands, just like the top-selling national brands, will be better known among consumers. For a list of all private label brands, premium- & national-brands and high market share brands, see appendix 1.
To evaluate whether a product is a sensory-product or a non-sensory product, a classification technique from Chu et al (2010) has been used as a reference. According to them, a product category is defined as sensory when transparent or semi-transparent packaging allows shoppers to examine the product pre-purchase in-store. In addition to that, a product category is defined as non-sensory when products are packaged in such a way that consumers cannot examine the product both on- and off-line.
coded correctly. For a list of all non-sensory products (non-sensory is used for dummy-coding) and non-food products (non-food is used for dummy-coding), see appendix 2. All product types that are not present in the list are sensory products or food products.
Variable creation
The next step in data preparation is transforming the data into a tidy format, which allows for a more effective and simplified analysis.
By first creating the dependent count variable that indicates the number of products purchased by each household during the observation period. This is not the total number of observations of purchases from a household because an observation indicates the type of product that has been purchased by a household at a certain point of time. However, during this observation, the consumer could have purchased one unit but also multiple units of this product type. So, by calculating the total number of products purchased, the number of units purchased per observation has been taken into account.
Next, the total number of repurchased brands, repurchased sizes, promoted, non-food, non-sensory, private label brand, premium- & national-brand and high market share brand products that are being purchased by each household are calculated. A repurchased brand, in this case, is a brand that has been purchased more than one time by a household. For a repurchased product size, a consumer has also purchased, more than once, the same product size from this repurchased brand. As when calculating the dependent variable, at this step, there has also been accounted for the number of units purchased per observation.
These total-numbers are transformed into a percentage number of the total number of products purchased per household, calculated at the previous step. The share of the purchased brand- and product-types, relative to consumers’ total purchased products, are now known for each household. Hence, this step has provided all the mediation variables needed for testing the formulated hypotheses.
Creating and transforming all these variables has made it possible to collapse over 67 million rows into 36,420 rows. Each row represents one household and the information needed to test the hypotheses can be found in the columns of this row. So, preparing the data has resulted in a tidy dataset that facilitates effective hypothesis testing. Control variables
Lastly, several consumer demographic factors will be taken into measurement as a control variable. General demographic factors like gender, age, household size, number of children, social class, region, working status, household income, ethnicity, house ownership and life stage will all be added.
3.3 Analysing techniques
This research aims to see if online grocery shopping leads to more purchased total products via several mediation variables, each representing a different type of brand or product.
Mediation
To see if there is a mediation effect, the three basic steps for mediation suggested by Baron and Kenny (1986) will be followed. Step one is estimating a direct model for the effect of online grocery shopping on the number of products purchased by a household. This relationship needs to be significant. Step two will be to estimate an indirect model for the effect of online grocery shopping on the mediation variables. This relationship also needs to be significant. The third and last step is to estimate a mediation model for the effect of online grocery shopping on the number of products purchased with one extra explanatory variable, which is the mediation variable. For full mediation, the effect of online grocery shopping on the number of purchased products disappears. For partial mediation, the effect of online grocery shopping on the number of purchased products has weakened.
intervals. Hence, bootstrapping will be used to estimate the confidence intervals that indicate if there is a significant mediation effect.
Generally, as long as computer-processing power is not an issue, a bootstrapping approach with at least 1000 resamples should be performed (Tingley et al, 2014). The R package mediation will be used to do the significance testing via bootstrapping and the paper by Tingley et al (2014) will be used as guidelines in doing this. Because these authors provide an example in which the independent variable on the mediation variable is analysed using OLS and effects on the dependent variable are analysed using a general linear model (GLM). This approach is similar to this research.
Count models
For the first and third step in mediation, a count variable needs to be estimated. Generally, for regression-type models where the dependent variable is a count variable, a Poisson regression would be most applicable. However, if λ is equal or larger than 10, the probability density function of a Poisson distribution looks like a normal distribution and then normal regression analysis can be used. Meaning that a normal regression can be used for data with large counts, as the dependent variable in this situation.
However, a violation of the Poisson model would be a mean that is not equal to the variance, meaning that there is over- or under-dispersion. The mean (2808) and variance (4954997) of the dependent variable here indicate clear overdispersion. Furthermore, a dispersion test, when running a Poisson model, shows that there is indeed overdispersion (𝛼 = -0.41, p = 1) and that a Poisson distribution would not be adequate for modelling.
The heterogeneity across cases in the dependent variable is larger than expected by the Poisson model. A negative binomial regression is, generally, used in case of overdispersion. Because with a negative binomial regression, the long tail in the dependent variable can be modelled by a low mean and high variance.
The equation to derive the probability density function for the negative binomial distribution is as followed (Hilbe, 2011): 𝑃 𝑌! = 𝑦! = (𝑓 𝑦! 𝜆, 𝑢 = 𝑒!!!!(𝜆!𝑢!)!! 𝑦!! 𝜆! depends on all the characteristics of the households. The negative binomial model can be seen as a Poisson model with gamma heterogeneity. The gamma noise has a mean value of 1, and this gamma mixture fits overdispersed Poisson counts (Hilbe, 2011).
Another violation of both the Poisson model and the negative binomial model is if there are no counts with a zero-value. For the dependent variable in this situation, there are no zeroes and there cannot be any zeroes because then the households could not be labelled as grocery shoppers. A truncate model can overcome this limitation. Since truncate models are used to model dependent variables for which some of the observations are not included in the analysis because the dependent variable has a zero-value.
that a linear model also is possible, there does not seems to be one clear model that applies to this situation. Hence, a linear model (OLS) a negative binomial model and a truncated negative binomial model will be estimated, after which the best performing model will be chosen, based on the information criteria.
Linear model
For the second step in mediation, the effects of online grocery shopping on the share of purchased brand- and product-types will be estimated. Both variables are continuous percentage numbers, making it possible to perform OLS regression. However, before these linear models can be interpreted a set of assumptions needs to be tested. The first assumption is that variables cannot be correlated, so there cannot be any multicollinearity. The next assumption for OLS is that the disturbance term is normally distributed. The equation for a simple linear relation is as followed (Leeflang et al, 2015): 𝑌 = 𝑎 + 𝛽𝑥 + 𝜀 Datasets Model-estimation for most of the hypotheses will be done based on all the available data. However, for the different brand-types, there is only data available for the product categories colas, peanut butter and instant coffee. Because of this, model-estimations for the brand-types hypotheses will lead to invalid model and bootstrap-estimates when using data from all the product categories. Hence, these hypotheses (H2) are estimated based on a subset of the dataset containing data for colas, peanut butter and instant coffee solely.
4. Results
4.1 Descriptive statistics
Looking at the demographic statistics (appendix 3.1), it looks like online grocery consumers differ from general consumers in that they are younger, have more children and bigger household sizes. This all reflects in the life stage in which online grocery shoppers are because young families dominate this category. While for the general grocery consumer, there is no dominating group in the life stage category. Further, females seem to prefer online grocery shopping more than men. There are no differences between online grocery shoppers and general grocery shoppers in terms of household income, working status and social class.
In terms of product categories (appendix 3.3), the most noticeable differences seem to be that household and cleaning products, canned goods and frozen prepared foods are all more than one percent more preferred online than offline. Noticeably, all these three product categories are sensory products.
Tesco seems to pre-eminently be the biggest supermarket chain in Great Britain, followed by Asda, Morrisons and Sainsburys (appendix 3.4). Especially online, Tesco seems to be performing exceptionally well, as online it has a market share of 52%.
Further looking at the data required for analysis (appendix 3.5), it seems that the average consumer repurchases more than half of the times the same brand from the total of 2808 on average purchased products. Also, more than half of the purchased products are non-sensory products. Among all households on average, 6% of their purchased grocery products are done online. For the product categories colas, peanut butter and instant coffee, high market share brands and premium- & national-brands seem to substantially more preferred that private label brands.
Lastly, examining the detected outliers of the data required for analysis, there do not seem to be any extreme values due to incorrectly entered or measured data. Hence, outliers are not removed from the analysis because this would exclude a particular group of observations and lead to biased estimates. Missing values are excluded from the analysis. Implicating that for each row in which a variable is missing, this row is not used for analysis. Each row represents one specific household.
4.2 correlations
Results of the Pearson correlation test indicated that there were multiple significant positive and negative associations between the mediation variables and between the independent variable and mediation variables. For the results of the Pearson correlation test, see appendix 4. However, the significant associations between the independent variable and the mediation variables are all very weak positive or negative associations. Indicating that multicollinearity probably will not be an issue when testing the hypotheses in the next section.
Concerning correlations between mediation variables, interesting results of the Pearson correlation were that there was an almost perfect significant negative association between premium own- & national-brands and private label brands, (r = -.99, p < .01). This result indicates that a product is either a private label brand or a premium own- & national-brand. The brands preventing the two variables from correlating perfectly likely are the hard discount brands that are being sold at more than one supermarket.
Furthermore, results of the Pearson correlation test showed that there was a significant moderate positive association between premium- & national-brands and high market share brands, (r = .59, p < .01) and significant moderate negative association between private label brands and high market share brands (r = -.58, p < .01). Indicating that premium- & national-brands have the highest market share for the product categories cola, peanut butter and instant coffee.
Further, the results of the Pearson correlation indicated that non-food products and non-sensory products have a weak positive association (r = .39, p < .01). Indicating that most non-food products likely have less sensory attributes or that these attributes are harder to verify when displayed in supermarkets.
Promoted products have a weak positive association between premium- & national-brands (r = .23, p < .01) and high market share brands (r = .25, p < .01), but a weak negative association with private label brands (r = -.23, p < .01). These results indicate that premium- & national-brands and high market share brands are more likely to be promoted than private label brands.
Lastly, the results of the Pearson correlation test showed that there is a significant strong positive association between repurchased brands and repurchased product sizes (r = .95, p < .01). The results implicate that when consumers repurchase a particular brand, in most situations, they also choose to purchase the same product size.
4.3 Adjustment conceptual model
Because there is a very strong or almost perfect correlation between some mediation variables, some formulated hypotheses discussed during the theoretical section can be excluded from the analysis. The reason for this is that due to the high correlations, some hypotheses will result in almost identical outcomes.
This implies that the perfect negative correlation between premium- & national-brands and private label brands, allows for excluding private label This implies that the perfect negative correlation between premium- & national-brands from the analysis. The reason for this is that the results of premium- & national-brands automatically generates the effect of private label brands since this will be the exact opposite effect.
Further, because of the very strong correlation between repurchased brands and repurchased product sizes, repurchased product sizes will be excluded from the analysis. Because a repurchased brand almost always results in the same repurchased product size from this brand.
The new adjusted conceptual model is displayed below (figure 3). Note that, as
4.4 Hypothesis testing
To test the formulated hypotheses, the three steps for mediation analysis from Baron and Kenny (1986), discussed in the previous section, will be followed. So, a direct model, indirect model and mediation model will be estimated.
When discussing the first model, the most appropriate modelling type, for the direct model and the mediation models, will be chosen for estimating the effects on the dependent variable.
Effect of online grocery shopping on total purchased products
A linear model, negative binomial model and truncated- or zero altered- negative binomial model were estimated to test the effect of online grocery shopping on total products purchased. After examining the VIF-scores, all three models showed signs of multicollinearity. This was resolved after deleting the life stage variable, which is one of the control variables. It must be noted that every model that is being estimated in the next steps will not contain the life stage variable due to multicollinearity-issues. Next, the three models were compared based on their information criteria (table 1). Also, all three models were compared to the null-model. All models outperform the null-model. Further, Both the negative binomial (NB) model and the truncated negative binomial model, or zero altered negative binomial model (ZANB), clearly outperform the linear model (LM). However, the AIC and BIC scores from the negative binomial model compared to the zero altered negative model show contradicting results. Because the BIC indicates better performance by the negative binomial model, while the AIC indicates that the truncated negative binomial model performs better. The BIC penalizes model complexity more heavily and because a simple model is preferred, the model with the best BIC is chosen. Hence, the analysis continues, using the negative binomial model.
LMnull NBnull ZANBnull LM NB ZANB
AIC 664806 657654 650559 637106 623772 623580
BIC 664823 657662 650585 637580 624238 624521
Table 1. Information criteria
Step 1 for mediation/direct model
As discussed, the negative binomial model has been adjusted for multicollinearity and the estimates of the parameters can now be interpreted. Negative binomial models produce estimates that are difficult to interpret, which is why incidence rate ratios (IRR) are calculated by taking the exponents from the coefficients. The IRR represents the change in the expected number of the dependent variable in terms of a percentage increase or decrease each time the independent variable goes up by one, with the remaining predicting variables held constant. The percentage number is determined by the number the IRR is either above or below 1 (Hilbe, 2011; Piza, 2012).
The hypothesized mediation effects are only possible if there is a direct effect of online grocery shopping on the total number of products purchased. The parameter for the effect of online grocery shopping on the number of products purchased is significant (p < .01) and positive. The IRR has a value of 1.0015, indicating that when all else is equal and the percentage of someone’s online purchases goes up with one, then the expected number of purchased products goes up by .15%. This result is in line with the formulated hypotheses stating a positive relationship between these two variables and all the mediation variables. Mediation thus makes sense as well as continuing to step two and three for mediation testing for all the formulated hypotheses.
For the estimations and VIF-scores of the direct model, see appendix 5.1. The exponents still need to be taken of these estimates to get the IRR’s. This applies to the estimates of all negative binomial models.
Effect of online grocery shopping on total purchased products, via repurchased brands.
Step 2 for mediation/indirect model
The linear model, estimating the effect of online grocery shopping on the share of repurchased brands, did not show any signs of multicollinearity after examining the VIF-scores.
will provide significant evidence on whether the model is normally distributed or not. A Kolmogorov-Smirnov and Jarque-Bera test are used to test for normality. Both tests are significant (p < .01), meaning that the disturbance term is not-normally distributed. To overcome this non-normality issue, bootstrapping was applied. After bootstrapping, the model checks for all assumptions needed for OLS and the model can be interpreted. Plotting the residuals to check for outliers and checking the distribution, will not be done for the following linear models because the Kolmogorov-Smirnov- and Jarque-Bera-test provide enough statistical evidence to test for normality. Figure 4. Distribution of the residuals Figure 5. Normal probability plot The parameter that shows the effect of online grocery shopping on repurchased brands is significant (p < .01) and positive. It is implying that a 1% increase in online grocery shopping leads to .08% increase in repurchased brands. The independent variable thus affects the mediation variable, making it possible to continue to the next step for mediation analysis.
Step 3 for mediation/mediation model
one, the expected number of purchased products decreases by .10% instead of increasing by .15%. This means that there is a causal mediation effect of .25%. For the estimations and VIF-scores of the indirect model and the mediation model, see appendix 5.2. Figure 6 gives a representation of the tested relationship. Figure 6. Mediation repurchased brands Bootstrapping with 1000 resamples was performed to test if the mediation is significant. After bootstrapping, the average causal mediation effect is significant (p < .01). Hence, it can be concluded that there is indeed a significant positive effect of online grocery shopping on total products purchased via more repurchased brands. Hence, H1A is
accepted. Effect of online grocery shopping on total purchased products, via premium- and national-brands. Step 1 for mediation/direct model As discussed in the previous section, model-estimations for the brand types hypotheses need to be based on a different dataset. Hence, for hypothesis 2B and 2C, a new direct model was estimated to test the effect of online grocery shopping on the number of products purchased. This negative binomial model is solely based on data for the product categories colas, peanut butter and instant coffee. No issues of multicollinearity were found after examining the VIF-scores. Hence, the estimates could be interpreted.
national brands and high market share brands. It thus makes sense to continue to step two and three for mediation testing for both hypotheses H2B and H2C.
For the estimations and VIF-scores of the direct model, see appendix 5.1.
Step 2 for mediation/indirect model
The linear model, estimating the effect of online grocery shopping on the share of purchased private label brands, did not show any signs of multicollinearity.
The disturbance term of the model is non-normally distributed since the Kolmogorov-Smirnov and Jarque-Bera test both had significant results (p < .01). To overcome this non-normality issue, bootstrapping was applied. After bootstrapping, the model checks for all assumptions needed for OLS and the model can be interpreted.
The parameter that shows the effect of online grocery shopping on the share of purchased private labels is significant (p < .01) and positive. Because a 1% increase in online grocery shopping leads to .05% decrease in the share of purchased private labels. The independent variable thus affects the mediation variable, making it possible to continue to the next step for mediation analysis.
Step 3 for mediation/mediation model
Figure 7. Mediation premium- & National-brands
Indirect effect of online grocery shopping on total products purchased, via high market share brands.
Step 2 for mediation/indirect model
A linear model was estimated to test the effect of online grocery shopping on the share of purchased high market share brands. The VIF-scores showed no signs of multicollinearity within the model.
A Kolmogorov-Smirnov and Jarque-Bera test did, however, show significant
results (p < .01) that the disturbance term of the model was non-normally distributed. So, bootstrapping was applied. After bootstrapping the model checks for all assumption needed for OLS and the model can be interpreted. The parameter on the effect of online grocery shopping shows a significant (p < .01) positive effect on the share of purchased high market share brands. Because a 1% increase in the share of online grocery shopped products leads to an increase of .02% in the share of purchased high market share brands. So, there is a significant indirect effect and checking the next step for mediation makes sense. Step 3 for mediation/mediation model A negative binomial model with one extra explanatory variable for the effect of the share of purchased high market share brands was estimated. Including the added variable, the effect of online grocery shopping remained significant and positive. Nevertheless, an IRR of 1.0021 showed that the effect did weaken from an increase of .24% to .21%, indicating a causal partial mediation effect of .03%.
For the estimations and VIF-scores of the indirect model and the mediation model, see appendix 5.4. Figure 8 gives a representation of the tested relationship.
Figure 8. Mediation high market share brands
Moreover, bootstrapping with 1000 resamples shows that the average causal mediation effect is significant (p < .01). Hence, it can be concluded that there is indeed a significant positive effect of online grocery shopping on total products purchased via more purchased high market share brands. So, the formulated hypothesis cannot be rejected. Hence, H2C is accepted.
Effect of online grocery shopping on total products purchased via non-food products.
Step 2 for mediation/indirect model
The VIF-scores of the linear model that predicts the effect of online grocery shopping on the share of non-food products purchased, showed no signs of multicollinearity.
A Kolmogorov-Smirnov and Jarque-Bera test also showed significant results (p < .01), indicating that the disturbance term of the model is non-normally distributed. Hence, bootstrapping was applied. After bootstrapping, the model checks for all assumptions needed for OLS and the model can be interpreted.
The parameter on the effect of online grocery shopping on the share of purchased non-food products is significant (p < .01), indicating that a 1% increase in online shopping leads to an increase of .01% in the share of purchased non-food products. Thus, there is a significant indirect effect and it makes sense to continue to the next step of mediation analysis.
Step 3 for mediation/mediation model
shopping, a generated IRR value of .100142 showed that the effect did weaken. As this value implicates that when all else is equal and the percentage of someone’s online purchases goes up with one, the expected number of purchased products, instead of increasing by .146%, now increases with .142%. Implying that there is a causal partial mediation effect of .004%. For the estimations and VIF-scores of the indirect model and the mediation model, see appendix 5.5. Figure 9 gives a representation of the tested relationship. Figure 9. Mediation non-food products
Bootstrapping with 1000 resamples shows that the average causal mediation effect is significant (p < .01). Hence, it can be concluded that there is indeed a significant positive effect of online grocery shopping on total products purchased via more purchased non-food products. So, the formulated hypothesis cannot be rejected. Hence, H3A is
accepted.
Effect of online grocery shopping on total products purchased via non-sensory products.
Step 2 for mediation/indirect model
The linear model that estimates the effect of online grocery shopping on the share of purchased non-sensory products showed no signs of multicollinearity, after inspecting the VIF-scores.
The effect of online grocery shopping on the share of purchases non-food products is insignificant (p = .24). Meaning that there is no indirect effect and thus mediation makes no sense, and H3B is rejected.
For the estimations and VIF-scores of the indirect model, see appendix 5.6.
Effect of online grocery shopping on total products purchased, via promoted products.
Step 2 for mediation/indirect model
A linear model was estimated to test if online grocery shopping affects the share of purchased promoted products. After examining the VIF-scores, no multicollinearity was found within the model.
Furthermore, the Kolmogorov-Smirnov and Jarque-Bera test were significant (p < .01), meaning that the disturbance term within the model was non-normally distributed. After bootstrapping, this non-normality violation was resolved and the model can be interpreted.
The effect of online grocery shopping is significant (p < .01), indicating that a one per cent increase in online shopping leads to an increase of .04% in the share of purchased promoted products. Thus, there is a significant indirect effect and it makes sense to continue to the next step of mediation analysis.
Step 3 for mediation/mediation model
Figure 10. Mediation promoted products
Bootstrapping with 1000 resamples shows that the average causal mediation effect is significant (p < .01). Hence, it can be concluded that there is indeed a significant positive effect of online grocery shopping on total products purchased via more purchased promoted products. So, the formulated hypothesis cannot be rejected. Hence, H3C is
4.5 Hypothesis table
To summarize the results, table 3 shows which hypotheses are supported together with the causal mediation effects. To better interpret the mediation effects, table 2 shows the direct effects and how they differ for two different sets of hypotheses. Since, the brand-type effects are measured on a data subset, limited to the product categories colas peanut butter and instant coffee. Table 2. Direct effects Table 3. Hypotheses table
Hypothesis Result Mediation
Effect H1A. Online grocery shopping will increase the number products
purchased by consumers via more repeatedly purchased brands.
Accepted .25% H2B. Online grocery shopping will increase the number of products
purchased by consumers, via more purchased premium- & national-brands.
Rejected -
H2C. Online grocery shopping will increase the number of products
purchased by consumers via more purchased high market share brands.
Accepted .03% H3A. Online grocery shopping will increase the number of products
purchased by consumers via more purchased non-food products.
Accepted .004% H3B. Online grocery shopping will increase the number of products
purchased by consumers, via more purchased non-sensory products.
Rejected -
H3C. Online grocery shopping will increase the number of products
purchased by consumers via more purchased promoted products.
Accepted .04%
Direct effects All product
categories.
Colas. Peanut butter. Instant coffee.
Direct effect of online grocery shopping on the total number of purchased products
.15% .24%
Relevant for hypotheses H1A. H3A.
H3B. H3C.
5. Discussion and conclusions
In this chapter, the results of the previous chapter will be given together with a possible explanation for these outcomes. This will provide a comprehensive answer to the main research question: “How does online grocery shopping alters consumer behaviour in terms of the number of purchased products, purchased brand- & product-types and expenditure-patterns, compared to offline grocery shopping?”Next, there will be elaborated on the implications of this research and lastly, limitations of the research together with possible future research areas to expand on the findings are given.
This research provides clear answers to the sub-questions related to which brand and product-types are more likely to be purchased online than offline, and if this is leading to more total purchased products. The sub-question related to online grocery expenditure seemed too unrelated to the other sub-questions. Adding this topic would result in a two-topic research paper. Because of this, this question remains unanswered.
5.1 Discussion
The sub-questions of whether certain brand types are more likely to be purchased when grocery shopping online, together with if this is leading to an increase in the total number of purchased products are validated. Since results of hypotheses 1 and 2 showed that online grocery shopping leads to an increase in the total number of purchased products, via more repurchased brands and high market share brands.
Nevertheless, multiple indirect brand type-effects were expected, and hence all these tested hypotheses are discussed below.
Indirect effect. Repurchase behaviour
Hypotheses 1 expected a positive effect of online grocery shopping on the total number of products purchased, through a mediation effect of more repurchased brands and product-sizes.
H1A was supported and repurchased brands had the largest indirect effect out of all mediation effects. Because after including the indirect effect, the IRR has a value of .9990, implying that out of the total direct effect of .15%, there is a causal mediation effect of .25%.
The existing literature suggests a positive effect of online grocery shopping on purchased premium- & national-brands because these brands are more expensive, and consumers are less price-sensitive when grocery shopping online (Chu et al, 2010; Degartu et al, 2000). However, not supporting H2B, this research indicates that there is no indirect effect of online grocery shopping on the share of purchased premium- & national-brands.
The almost perfect negative relationship with private label brands can explain this result. Since the performance of private label brands generally improves when grocery retailers become active online (Arce-Urriza & Cebollada, 2018), a positive mediation effect was also expected via private labels. Thus, because there is ground for better online performance of both contradicting brand types, premium- & national brands will not outperform private labels or vice versa. So, considering the almost perfect negative correlation, it makes sense in hindsight that online grocery shopping does not affect the share of purchased products for both brand types.
The positive indirect effect of online grocery shopping on the total number of purchased products, via more purchased high market share brands (H2C), is supported. Because after including the indirect effect, the IRR has a value of 1.0021, implying that out of the total direct effect of .24%, there is a causal mediation effect of .03%.
This is in line with the expectations based on existing literature. Like, that for a grocery brand with high market share the brand loyalty is significantly higher online (Danaher et al, 2003). Further, encountering the risk associated with online grocery shopping, Moore and Andradi (1996) found that high market share brands are generally associated with a less risky purchase, making them more preferable online. Lastly, high market share brands are more preferred by consumers. Because of this, e-retailers will make their website more convenient for consumers by highlighting these brands, hence making them even more likely to be purchased.
The sub-questions of whether certain product types are more likely to be purchased when grocery shopping online, together with if this is leading to an increase in the total number of purchased products are validated. Since results of hypotheses 3 showed that online grocery shopping leads to an increase in the total number of purchased products, via more purchased non-food products and promoted products.
Indirect effect. Product types
Hypotheses 3, indicated a positive effect of online grocery shopping on the total number of products purchased, through a mediation effect of more purchased product types.
The positive effect of online grocery shopping on the total number of purchased products, via more purchased non-food products (H3A) is supported. Because after including the indirect effect, the IRR has a value of 1.00142, implying that there is a marginal causal mediation effect of .004%.
As discussed in the literature section, the main arguments for expecting this effect were that non-food products do not have an expiration date and that physical examination is less applicable for non-food products. This is indicating that there is less of a barrier for buying non-food products online than food products. Further, the long-tail of e-retailing makes it possible to offer a more extensive assortment of products in which non-food products can gain a significant share online compared to offline (Hinz et al, 2011).
Next, existing literature showed that sensory products are purchased more offline (Chu et al, 2008). Because of this, it was expected that consumers might perceive sensory products being more suited for offline and non-sensory products more for online grocery shopping. Hence, a positive effect of online grocery shopping on purchased non-sensory products was expected. However, no effect was found and H3C is rejected. So, it can be concluded that according to this research, there does not seem to be a preference for sensory- or non-sensory-products online.
In support of H3C, there is a positive mediation effect of online grocery shopping on the total number of products purchased, via more purchased promoted products. Because after including the indirect effect, the IRR has a value of 1.0013, implying that out of the total direct effect of .15%, there is a causal mediation effect of .02%.
