Identification of showrooming/webrooming behavior throughout
the customer journey
An exploratory study using site-centric data
Author: Koen Caspar Molendijk
Department: Faculty of Economics and Business (FEB) Qualification: ‘Master thesis’
Completion date: 17th of June, 2019
Address: Het Spangoor 25, 7232 HA Warnsveld Phone number: +316-40824132
E-mail: Koen.m@hotmail.com
Student number: s3571823 First supervisor: Dr. Frank Beke
2
Management Summary
Show- and webrooming are gaining popularity these days, but the understanding of these phenomena is yet limited. The customer journey is another hot topic extensively researched at this moment. Customers use online and offline channels interchangeably throughout this journey. This research is regarded as a first attempt to link the phenomena of showrooming and webrooming to the customer journey. Using various methods, it is tried to segment customers into a showrooming, webrooming and online purchasing segment. The successful method has been a nested logistic regression. The effect of various variables has been tested on these segments to discover what the actual indicators of showrooming and webrooming are.
The results reveal that the online journey of showroomers could very well be identified by just using site-centric data. Their online journey starts in the ‘Desire’ stage and ends in the ‘Action stage’ with a purchase. The main indicators for showrooming behavior are accessing the site via a search advertisement and viewing few pages. To get a full picture of the whole customer journey for showroomers, additional offline in-store data is required. Then retailers could identify showroomers earlier in the journey, understand their need, and target them with accurate search ads of the product they intend to buy. In this way, showroomers could be prevented from purchasing at a competing retailer offering lower prices.
Furthermore it has been found that webroomers tend to visit more pages and spend more time on these pages, especially pages earlier in the customer journey. Their journey starts online and ends with an in-store purchase. It was expected that pages describing the location of physical stores would be visited more often by webroomers. It has been found that it even has an opposite effect on webrooming. If retailers could classify webroomers based upon site-centric data, they should try to facilitate an easy transition from the online to the offline channel. After identifying the product webroomers intent to buy, they should be guided till the final purchase is made in the store, to ensure that they do not buy the product in a competitors’ store.
3
Table of content
Management Summary ... 2 1. Introduction ... 5 2. Theoretical framework ... 7 2.1. Customer journey ... 72.2. Showrooming & webrooming ... 8
2.2.1. Showrooming ... 8
2.2.2. Webrooming ... 9
2.2.3. Showrooming/webrooming in the customer journey ... 9
2.3. Conceptual framework ... 10
2.3.1. Page views & Time on page ... 10
2.3.2. Store Page ... 12 2.3.3. Channel ... 13 3. Research Design ... 16 3.1. Type of data ... 16 3.2. Methodology ... 17 3.2.1. Data preparation ... 17
3.2.2. Latent Class Cluster Analysis ... 19
3.2.3. Model selection ... 20
4. Results ... 21
4.1. Latent Class Cluster Analysis ... 21
4.2. Multinomial Logistic Regression ... 22
4.2.1. Segment creation ... 22
4.2.2. Descriptives ... 23
4.2.3. Multicollinearity ... 24
4.2.4. Model fit & model selection ... 26
4.2.5. Independence of Irrelevant Alternatives... 27
4.3. Nested multinomial logistic regression ... 28
4.3.1. Model fit & model selection ... 29
4.3.2. Model interpretation ... 31
4.4. Discussion ... 40
4.4.1. Showrooming ... 40
4.4.2. Webrooming ... 42
4
4.4.4. Segment comparison ... 44
5. Conclusion and recommendations ... 45
5.1. Limitations and further research ... 47
6. References ... 47
5
1. Introduction
Understanding the customer journey has become more and more important. Although the first article researching the customer journey has already been published more than 20 years ago, it gained significant popularity in publications since 2009 (Følstad & Kvale, 2018).
Understanding which channels consumers use across all stages of the journey is key for designing marketing strategies (De Keyser, Schepers, & Konuş, 2015). Understanding the dynamics of the customer journey is crucial. The stage in the customer journey the consumer currently is in can help suggest ‘which combination of media and messages might be most appropriate there’ (Batra & Keller, 2016, p. 124).
The customer journey is not simply limited to either online or offline channels. Online and offline channels might be used interchangeably throughout the customer journey (e.g. online advertising might lead to offline sales (Skinner, 2010)). This is also where the phenomenon of showrooming and webrooming comes into place.
Nowadays, it is common to apply showrooming behavior. ‘68% of US internet users indicate
that they showroom at least occasionally.’ (Gensler, Neslin, & Verhoef, 2017, p. 29).
Showrooming is the phenomenon where you view and compare physical products in-store, but purchase it online (Nesar & Sabir, 2016). Next to showrooming, webrooming has gained significant popularity as well ‘with 73% of surveyed U.S. customers having showroomed and
88% having webroomed’ (Lemon & Verhoef, 2016, p. 80). Webrooming is basically the
opposite of showrooming, where you conduct the research online, but eventually purchase the product in-store (Nesar & Sabir, 2016). The popularity of showrooming and webrooming emphasizes the potential of successfully dealing with it.
This research attempts to get a better understanding of the online browsing behavior of consumers, specifically regarding showrooming and webrooming behavior. It does so by trying to answer the following research question: “Is it possible to identify
showrooming/webrooming behavior throughout the customer journey, using site-centric data?”
The sub-research questions are as follows: “What browsing behavior corresponds to show-
and webrooming behavior?” and “How should multichannel retailers respond to the observed showrooming and webrooming behavior?”
6
webrooming behavior according to relevant literature is relied upon as input for the segments. The resulting segments will likely encompass customers applying showrooming and webrooming behavior, and additionally an online purchasing segment. If this is not the case, segments will be created manually based on certain characteristics, after which the effect of the main determinants will be tested by means of a logistic regression. The main focus will be on the showrooming and webrooming segments. For each of these segments, it will be identified in which stages of the customer journey, which specific type of behavior is likely to occur. Additionally, insights will be provided on how marketers should respond to the observed behavior for each of these segments.
The data used for this research concerns site-centric data. Or more explicitly, online browsing behavior of visitors on the website of a Dutch home-furniture store. It would be interesting to see if showrooming and webrooming behavior could be discovered by analyzing their browsing behavior using only site-centric data.
Showrooming and webrooming is still a hot topic in research these days, with researchers stressing the importance of further researching the topic (e.g. Kim, Park, & Libaque-Saenz, 2018; Lemon & Verhoef, 2016). These researchers amongst others emphasize the relevance of studying showrooming and webrooming behavior in order to gain a better understanding of these phenomena. As far as I’m concerned, it would be the first attempt to link show- and webrooming to the customer journey. Next to this theoretical relevance, this research will provide practical relevance as well, as most firms only possess site-centric data (Zheng, Fader, & Padmanabhan, 2012). For those firms lacking user-centric data the potential relevance of this research could be extraordinary. By understanding the way their customers behave online, and by knowing how to respond to it, they could maximize the potential value gains. Moreover, identifying showrooming and webrooming behavior by using just site-centric data is to my fullest knowledge something not studied by others yet.
7
2. Theoretical framework
In this section the theoretical background of customer journeys, showrooming and webrooming will be covered. Furthermore, the typical online browsing behavior in each stage of the journey, and corresponding browsing behavior for showroomers and webroomers will be discussed. This will be elaborated upon by describing the main determinants of the browsing behavior, and the related hypotheses. Finally, the last part of this section contains a detailed description of the conceptual framework.
2.1. Customer journey
Customer behavior throughout the journey is typically analyzed following a pre-defined, structured process (Følstad & Kvale, 2018). The structure used varies across the different researchers, for example Lemon & Verhoef (2016) used the stages pre-purchase, purchase and post-purchase. Whereas Richardson (2010) structured the stages as awareness, research, purchase, OOBE (out-of-box experience). With OOBE being the experience of opening the purchased product and guiding you through the first steps of using the product. He notes that the journey often is non-linear, meaning that someone might move from awareness straight to the purchase phase (Richardson, 2010). Batra & Keller (2016) even took it a step further by dividing the journey into 12 stages. They accentuate that although the consumer likely goes through some sequence of steps in the “path to purchase” it nowadays often is shorter in length, less hierarchical, and more complicated. They stress the dynamics, with the potential to move forward and backward, drop out or enter the different stages throughout the decision process (Batra & Keller, 2016).
8
information about the available products as well as comparisons among the attributes and benefits of the products in the consideration set (Wijaya, 2015). In the context of this research, the AISDA model consists of Attention (A), which is the stage where a consumer first pays attention to the product category, Interest (I), the stage at which the consumer becomes interested in the product category, Search (S), the stage at which the consumer will search for additional information and compare products in the category, Desire (D), the stage where the consumer shows passion towards the product and intends to buy it, and Action (A), where the purchase is made.
2.2. Showrooming & webrooming
Nelson (1970) previously classified home furniture as a search good. Although interaction with the product is not necessary to determine the quality, Huang et al. (2009) found that search goods receive higher ratings in offline environments since it provides the ability to touch and feel the products before purchase. It is therefore highly likely that those who buy search goods online, previously touched and felt the products in-store, implying showrooming behavior.
Huang et al. (2009) further found that before buying search goods customers intensely browse online. This, together with the fact that the quality of search goods can be assessed without interacting with the product, indicates that webrooming behavior could be expected as well.
2.2.1. Showrooming
9
2.2.2. Webrooming
Webrooming is when you conduct your research online, and eventually buy the product you like in the store (Nesar & Sabir, 2016). Consumers do not seem to extensively rely on comparison shopping across retailers (Huang et al., 2009). Suggesting that the product comparison often takes place at one retailer. Next to this the data used consists of only site-centric data, limiting the potential webrooming behavior to be discovered in this research to within-firm webrooming. The definition of webrooming applied in this paper is: Searching for information and comparing products online on the website of the retailer before buying the product in the physical store of the same retailer.
2.2.3. Showrooming/webrooming in the customer journey
‘Showrooming has been reported to take place during product evaluation, where physical product attributes are important.’ (Wolny & Charoensuksai, 2014, p. 324). This implies that
consumers tend to test products in-store before searching for the best deals online. Suggesting that consumers might visit the website after testing, somewhat halfway the Desire stage, where the consumer shows passion towards the product and intends to buy it. Daunt & Harris (2017) further argue that showroomers engage with the product in-store, assess all its aspects and acquire relevant information, before going to the web.
In this research the showrooming behavior to be discovered by the data might thus concern the stages of Desire (D) and Action (A).
10
even take place in the Desire stage, with the shopping cart suggesting a purchase intention. The webrooming journey to be captured by site-centric data thus seems to be starting when they gather information, and ends right before a purchase is to be made. Or more explicitly, it might concern the stages Attention (A), Interest (I), Search (S) and Desire (D).
Appendix A provides a graphical representation of showrooming and webrooming behavior throughout the stages of the customer journey.
2.3. Conceptual framework
2.3.1. Page views & Time on page
Page views and visit duration are often jointly modeled, since together they provide a parsimonious representation of the browsing behavior (Bucklin & Sismeiro, 2003). As such both their effects on browsing behavior should be studied. The number of pages viewed as well as visit duration are positively associated with the purchase decision (Mallapragada, Chandukala, & Liu, 2016). The more pages viewed the higher the probability of a sale. Also, the longer the consumers stay on a page (visit duration or time on page), the higher the probability of a sale.
When browsing for search goods, consumers tend to visit many pages (Huang et al., 2009). The number of pages visited is widely accepted as an indicator of user interest (Hofgesang, 2006). ‘A typical online purchase experience includes multiple web page visits, through which
the consumer processes the gathered information, before eventually making a purchase’
(Mallapragada, Chandukala, & Liu, 2016, p. 21). This implies that the number of pages viewed is related to the customer journey. For this research data is structured per session, therefore visit duration is hypothesized as session duration. It is assumed that both page views as well as session duration might reveal showrooming and webrooming intentions.
11
The hypotheses corresponding to the abovementioned determinants of online browsing behavior according to the literature are formulated as follows:
H1a: Page views is negatively related to showrooming behavior. H1b: Page views is positively related to webrooming behavior. H1c: Page views is positively related to online purchase behavior. H2a: Session duration is negatively related to showrooming behavior. H2b: Session duration is positively related to webrooming behavior. H2c: Session duration is positively related to online purchase behavior.
Visit depth, which is the cumulative number of page views it takes to move a certain page, further affects the browsing behavior (Bucklin & Sismeiro, 2003). Thus the number of page views might have differential effects on browsing behavior across the journey. Additionally, the time spent on webpages might depend on page type (information page, contact form, etc), and is believed to be of great value in identifying user intention (Hofgesang, 2006). Time spent on different pages might thus as well have differential effects on the proceeding browsing behavior.
Therefore the variables page views and session duration will be subdivided based on the type of page into three different levels based on their depth in the journey. The effect of the following levels, in ascending order in terms of depth are tested: product category page (PCP), product overview page (POP) and product detail page (PDP), respectively. A PCP shows the product categories, the types of furniture, such as lounge sets, wooden beds, bookshelves and the like. A POP shows an overview of the products one can buy after deciding on the product category, it generally consists of multiple pages. The PDP is the information page of a specific product, where you can choose the color or size of that specific product. Visiting a PCP corresponds to the Interest stage, visiting a POP to the Search stage and PDP to the Desire stage.
12
H3a: Number of PDPs viewed is negatively related to showrooming behavior. H4a: Time on PDPs is negatively related to showrooming behavior.
Webroomers on the other hand gather their information and do their product comparisons online. This might involve visiting many product category pages, product overview pages and product detail pages. Additionally browsing these pages takes time. Likely various PCPs will be viewed in order to determine the type of product needed. Subsequently they will browse through the assortment (POP) and click on the products that appeal to them. These are the PDPs, which will be compared until the decision is made to buy a certain product in store. The following hypotheses are formulated:
H3b: Number of PCPs viewed is positively related to webrooming behavior. H4b: Time on PCPs is positively related to webrooming behavior.
H5b: Number of POPs viewed is positively related to webrooming behavior. H6b: Time on POPs is positively related to webrooming behavior.
H7b: Number of PDPs viewed is positively related to webrooming behavior. H8b: Time on PDPs is positively related to webrooming behavior.
Lastly, for online purchasers the same process as for webroomers takes place online. The only difference being that the purchase will be made online. Therefore the following hypotheses are formulated:
H3c: Number of PCPs viewed is positively related to online purchase behavior. H4c: Time on PCPs is positively related to online purchase behavior.
H5c: Number of POPs viewed is positively related to online purchase behavior. H6c: Time on POPs is positively related to online purchase behavior.
H7c: Number of PDPs viewed is positively related to online purchase behavior. H8c: Time on PDPs is positively related to online purchase behavior.
2.3.2. Store Page
13
webrooming behavior. Visiting these pages is thus likely to take place during the Desire stage, when there is an intention to buy the product. In contrast, those consumers whom browsing behavior suggests that they are likely to showroom, improbably visit such pages, since they have visited a store already before visiting the website. Therefore, it is reasonable to assume that customers visiting storepages are unlikely to be showrooming. In line with this thought, the relationship between the ‘Store Page’ and showrooming is expected to be negative. A similar reasoning applies regarding online purchasing. A negative relationship between online purchasing and store pages is expected. Not because online purchasers already visited a physical store, but because they tend to purchase their product online, which makes visiting a storepage irrelevant. As such, ‘Store page’ will be included as a determinant for browsing behavior in the conceptual framework. The following hypotheses have been formulated: H5a: Visiting a Store page is negatively related to showrooming
H9b: Visiting a Store page is positively related to webrooming. H9c: Visiting a Store page is negatively related to online purchasing.
2.3.3. Channel
‘Customers differ in their preference and usage of channels across different purchase phases, and specific multichannel segments can be identified that differ in terms of consumer characteristics.’ (Lemon & Verhoef, 2016, p. 80). Thus, consumers tend to use different
channels throughout the phases of the customer journey. As such, the channel via which consumers enter the website could be an indicator of their proceeding browsing behavior. The influence of usage of different channels on online browsing behavior is to be investigated in this section. More explicitly, the effect of display ads and search ads will be covered, since those channels seem to influence showrooming and webrooming behavior mostly. The channels used by consumers for entering the website might affect online browsing behavior and whether consumers showroom or webrooming.
2.3.3.1. Display ads & Search ads
14
By estimating a Hidden Markov Model, Abhishek, Fader, & Hosanagar (2012) found that display and search ads have differential effects on individual consumers based on their states in the customer journey. Kireyev et al. (2016) more explicitly argue that display ads might be more effective in influencing consumers at the earliest stages or halfway the journey, whereas the impact of search ads is higher near the end of the funnel. Search advertising is meant to target those customers who are already in the buying process, and therefore perfectly fits the needs of those consumers in that particular stage of the decision journey (Batra & Keller, 2016).
The studies mentioned above imply that display ads tend to influence online shopping behavior more during the Attention, Interest and Search stages, as opposed to the Desire and Action stages, where the effect of search ads dominates. Thus, there presumably is a positive relationship between display ads and webrooming behavior, as well as between display ads and online purchasing. There is no clear relationship expected between display ads and showrooming behavior, since the Attention, Interest and Search stages for showroomers take place in physical stores. Display ads will be added to the conceptual framework as part of the channel category in determining online browsing behavior.
Additionally, since the purpose of search ads is to target customers in the buying process, the main effect to be expected on online browsing behavior is in the Desire and Action stages. As showroomers and online purchasers eventually buy the products online, the effect of search ads on these segments is presumably positive. In contrast, webroomers eventually buy their products in physical stores (Nesar & Sabir, 2016; Wolny & Charoensuksai, 2014). As such, search ads likely do not appeal to them. Therefore, there is no clear relationship expected between search ads and webrooming. Additionally, search ads will be included in the conceptual framework as part of the channel category in determining online browsing behavior. The corresponding hypotheses have been formulated as follows:
H6a: There is a positive relationship between search ads and showrooming behavior. H10b: There is a positive relationship between display ads and webrooming behavior. H10c: There is a positive relationship between display ads and online purchasing. H11c: There is a positive relationship between search ads and online purchasing.
15
abovementioned stage on either showrooming, webrooming or online purchasing behavior. The + sign implies a positive relationship, whereas the – sign illustrates a negative relationship.
Figure 1a. Conceptual Framework Showrooming.
16
Figure 1c. Conceptual Framework Online Purchasing.
3. Research Design
In this section the type of data used for the analysis will be further exemplified. And the methodology used to analyze this data will be described in more detail.
3.1. Type of data
Clickstream data is a record of a user’s activity on the internet and traces the path visitors take when navigating the web (Bucklin & Sismeiro, 2009). It is a rich resource for researchers and practitioners to better understand the behavior of individuals (Bucklin & Sismeiro, 2009). Clickstream data can be subdivided into user-centric and site-centric data. User-centric data provides the possibility of tracing online behavior across sites with information from each individual user (Bucklin & Sismeiro, 2009). Although user-centric data would be very helpful for understanding the customer journey of individuals, often a firm only possesses site-centric data (Zheng et al., 2012). In this research an attempt is made to discover showrooming and webrooming behavior using only site-centric data. Those firms that lack user-centric data could significantly benefit from this.
The data used for this research consists of site-centric data, which is ‘clickstream data collected at a site augmented with user demographics and cookies to identify users’ (Padmanabhan et al., 2001). Since site-centric data only captures the browsing behavior on a specific site, and lacks information regarding activities on other websites, the online shopping behavior to be analyzed in this research is limited to within-firm webrooming. Information regarding activities and browsing behavior in a physical store neither is available for this research. Which makes it difficult to determine where offline browsing took place. For identification of showrooming behavior it is less relevant whether the offline browsing took place at the same retailer where the eventual online purchase is made, or at a different retailer. For the sake of easiness, the assumption is made that showrooming takes place within the same firm.
Fortunately this site-centric data provides various benefits compared to for example the domain-level browsing data used by Huang et al. (2009). More detailed page-level data is available (Bucklin & Sismeiro, 2009; Huang et al., 2009), such as time spent gathering information about the product and product category, providing different insights.
site-17
centric data, online browsing behavior on the website of a retailer selling search goods will be analyzed.
3.2. Methodology
3.2.1. Data preparation
18
same customer after these 30 days is seen as the start of a new journey. Typically, furniture delivery takes very long, let alone specialized furniture (Tammela, Canen, & Helo, 2008). Therefore the 30 days after purchase limit seems to be reasonable.
The remaining number of observations is 355,220. Within this subset, the required variables were created and made ready for the analysis. Most of the variables have been created based on the observed URL’s. As an example, the URL’s for the creation of type of pages viewed can be found in table 1.
3.2.1.1. Page views
The first variable that has been created is ‘TotPageViews’. This variable indicates how many pages are viewed by a visitor each session. It is subdivided into homepage (HomePageTot), category page (CategoryTot), basket (BasketTot), product category page (PCPTot), product overview page (POPTot) and product detail page (PDPTot). These variables will be used for testing the page view related hypotheses.
3.2.1.2. Time on page
A timestamp indicates when a visitor arrives at a certain page. Timestamps are rather detailed, describing the time of arrival on a page in year, month, day, hour, minute, second and millisecond respectively (e.g. 2018-06-24T07:23:46:314). Since timestamps do not reveal when a visitor leaves the page, the time spent on the last viewed page is unknown. In order to calculate the correct total session duration, the variable ‘sessionDuration’ is created. Which is the time a visitor spends on the website in seconds per session. ‘sessionDuration’ consists of the by default available variable ‘sessionduration’ together with 1 more time the average time on page in that session, to fill in the time spent on the last viewed page. By using timestamps, the time spent on more specific pages are calculated. The resulting variables are time spent on the homepage (TotTimeHomePage), category page (TotTimeCategory), basket (TotTimeBasket), product category page (TotTimePCP), product detail page (TotTimePDP) and product overview page (TotTimePOP). These are all computed as the time difference between two consecutive timestamps and are used for testing the time on page related hypotheses.
3.2.1.3. Store Page
19
well. However, when clicking through on these pages, one would automatically visit the store page next. Therefore they are already included in ‘StorePage’. This variable is used for testing the hypotheses related to the storepage.
3.2.1.4. Channel
The variable ‘ga.channelGrouping’ has been added as indicator for browsing behavior as well. This variable and its corresponding levels were available in the data by default. The levels consist of: Affiliates, Branded Paid Search, Direct, Display, Email, Generic Paid Search, Organic Search, Paid Search, Paid Social, Referral and Social. Especially the effect of the display ads and search ads are relevant for the channel related hypotheses. For these variables dummy variables have been created.
Some more variables have been created. The variable ‘purchase’ indicates whether an online purchase is made or not. ‘LandingPage2’, and ‘ExitPage2’, indicate the webpage on which the visitor arrives and leaves the website respectively. These variables are necessary for determining browsing behavior, without explicitly testing hypotheses. Furthermore they identify where the customer sessions start and end.
Table 1 example URL’s for type of pages viewed
Variable URL
HomePageTot “/”
CategoryTot “/slapen”, “/wonen”
BasketTot “/basket/index”
PCPTot e.g. “/slapen-bedden”, “/wonen-stoelen”
POPTot e.g. “/slapen-bedden-…”, “/wonen-stoelen-…”
PDPTot “/p/”
3.2.2. Latent Class Cluster Analysis
20
Considering the different variable types used for creating the latent clusters, and the more extensive documentation, Latent GOLD is regarded the best option.
Methods such as latent class analysis and cluster analysis can be considered person-centered approaches. The goal is to classify individuals into distinct groups or categories based on individual response patterns (Jung & Wickrama, 2008). This should be done in such a way that there is a low within-group variation as opposed to a high between-group variation (Haughton et al., 2009; Jung & Wickrama, 2008; J. Vermunt & Magidson, 2002). In other words, those customers within a cluster are supposed to browse in similar ways. But when comparing the browsing behavior of an individual in ‘cluster a’ with an individual in ‘cluster b’, it should be rather different. LCCA is model-based and probabilistic, meaning that individuals get assigned to classes with the highest probability of membership (Lawrence & Zyphur, 2011).
Latent Class Clustering is explicitly useful since it is rather easy to deal with variables of different scale types, and there are formal decision making criteria available about the appropriate number of clusters (J. Vermunt & Magidson, 2002). Additionally, there is no need to rescale or manually transform variables of different types (Lawrence & Zyphur, 2011). Since the variables used as input for the clusters are of different types (i.e. continuous, categorical), LCCA seems to be the most appropriate method for cluster creation.
3.2.3. Model selection
The model selection procedure in LCCA aims at choosing the optimal number of latent classes (Bartolucci, Bacci, & Gnaldi, 2014). Various information criteria will be used to determine the appropriate number of clusters. The Bayesian Information Criterion (BIC) is assumed to be the best criterion for this (Nylund, Asparouhov, & Muthen, 2007). The BIC allows for comparison of more than two models at the same time (Fraley & Raftery, 1998). The resulting number of clusters with the lowest BIC value is considered the best cluster distribution (Haughton et al., 2009). The BIC score will be taken into account in the identification of the model with the most appropriate number of clusters. Also additional model selection criteria will be taken into account to ensure that the best performing model is chosen.
21
looking at the significance of the relationship with the clusters. A major issue regarding LCCA is the assumption of local independence. Violation of this assumption can result in spurious latent classes and poor model fit (Zhang, 2004). There are methods available for detecting and modeling local dependence. If local dependence turns out to be an issue, multiple latent variables will be introduced to resolve it. If that does not work, the LCCA model will be reformulated as a loglinear model and then constraints will be imposed on it. Which should resolve local dependence (Zhang, 2004).
With latent class cluster analysis you do not know up front what the resulting segments will be. However, as explained in the conceptual framework, these variables have differential effects on showrooming, webrooming and online purchasing behavior. Therefore, the LCCA is expected to result in clusters reflecting these differential effects.
4. Results
4.1. Latent Class Cluster Analysis
The 10% subset of the data after cleaning is used as input for the LCCA. The explanatory variables included in the model are those defined in the methodology section.
In this section, the best model according to the BIC & AIC scores will be interpreted. 4 models were created using a Latent Class Cluster Analysis, 2-cluster, 3-cluster, 4-cluster and 5-cluster models. The LCCA in Latent Gold identified the 5-cluster model as the best model. This model has the lowest AIC and BIC scores. An overview of the model selection criteria used for determining the best model among the clusters can be seen in table 2 below.
Table 2: Model selection criteria Latent Class Cluster Analysis AIC BIC 2-cluster model 23,905,032.5531 23,906,574.1635 3-cluster model 15,307,399.8780 15,309,717.6839 4-cluster model 15,798,170.2173 15,801,264.2187 5-cluster model 12,252,579.5401 12,256,449.7369
22
led to convergence. The number of iterations were doubled from 50 to 100. The number of starting sets were doubled as well, from 18 to 36. All of these options retrieved the exact same scores. Other methods used to solve the non-convergence were increasing the sample size from 355,220 to 915,469 (25% sample of data after cleaning) observations. This made BIC and AIC scores even worse. Adding more parameters improved the model slightly, but the model still did not converge. Also increasing the tolerance level did not help. No remedy for the non-convergence has been found, meaning that the data likely is not informative enough about the parameter values (J. K. Vermunt & Magidson, 2005). Thus, the resulting clusters are not reliable and should not be interpreted.
The model unfortunately does not reveal the expected showrooming, webrooming and online purchasing segments. Thus, a different method for identifying showrooming and webrooming segments is required. A multinomial logistic regression will be conducted for this.
4.2. Multinomial Logistic Regression
As the LCCA did not reveal the expected segments, a second try is given for identifying showrooming and webrooming segments. Showrooming and webrooming segments will be created based on characteristics of these types of browsing behavior as identified in the literature. Additionally an online purchasing segment will be created. A regression will be run for further testing the hypotheses. Statistical program R will be used for this. The type of regression is a Multinomial Logistic Regression (MNL). This is the best fitting regression, since the dependent variable is nominal with more than two choice options (Leeflang, Wieringa, Bijmolt, & Pauwels, 2016). Also only one regression has to be run in this case.
4.2.1. Segment creation
23
Category / PCP), and that no purchase is made (purchase1 = 0). The online purchasing segment consists of those journeys in which both information gathering took place online and a purchase was made online. One requirement is thus that a purchase has been made (purchase1 = 1). Additionally, the landing page should not be equal to the landing pages used in the showrooming segment, since then it would imply showrooming behavior. Lastly, an “Other” segment has been created, consisting of all sessions not belonging to the before mentioned segments. This is done to ensure that no relevant data is left out of the analysis and it can serve as a base in the regression to compare the other segments to. The segments, and their corresponding sizes in terms of sessions can be found in the table 3. The Showrooming and Onlinepurchasing segments are rather small in comparison to the Webrooming and Other segment, since one of the requirements was that a purchase is made. In the data, as well as in the subset used, there are relatively few purchases in total. The total number of purchases in the subset is 798, which is equal to the sum of the segments showrooming & online purchasing. The webrooming segment does not actually represent true webroomers, due to a lack of data concerning offline purchases. But the behavior of the sessions in this segment corresponds to webrooming behavior. Therefore the assumption is made that the effect of the independent variables on the webrooming segment is similar to the effect on actual webroomers.
Table 3: Size of segments
Showrooming Webrooming Onlinepurchasing Other
Absolute size 323 123,643 475 230,779
Relative size 0.09% 34,81% 0.13% 64,97%
A regression will be run with the four segments as the dependent variables. The independent variables will mostly be the same as used for the LCCA. In this way the hypotheses could still be tested by using a regression. From the independent variables used in the LCCA, LandingPage2 and purchase1 have been removed, since these have been used as input for the segments.
4.2.2. Descriptives
24
Showroomers on the other hand more likely access the website via a search ad. Webroomers tend to visit more homepages and product category pages. Showroomers and online purchasers probably view the shopping basket more often and spend more time on basket pages as well. Probably webroomers and Others view more product overview pages and spend more time on these pages as well. Both showroomers, as well as Others presumably view more product detail pages. Showroomers also tend to spend more time on these product detail pages. It looks like showroomers and online purchasers have higher total page views and their average session takes longer. Online purchasers probably spend more time on the homepage than the other segments. Not much time is spent on category pages and product category pages in general. Lastly, storepages seem to be visited most often by Others.
4.2.3. Multicollinearity
The first regression returned the warning that the data did not converge. To solve this issue the number of iterations was increased to 100, which worked. Afterwards multicollinearity has been checked by means of a correlation matrix. Variables were assumed to have high multicollinearity when either the correlation matrix indicated values >.50 or when the VIF was above 4.0. Simply deleting variables exceeding either one of these values might result in an unnecessary loss of explanatory power. Therefore highly correlated variables will be rewritten in a different way to account for the correlation. As could be seen in table 4, ‘TotPageViews’ passed the threshold 4.0, implying multicollinearity.
25 TotTimeCategory 2.916790 TotTimeBasket 1.935386 TotTimePCP 1.409277 TotTimePOP 2.055780 TotTimePDP 2.360998 StorePage 1.059556
As further could be seen in the correlation matrix (Appendix B, table 1). ‘TotPageViews’ is highly correlated with ‘PDPTot’ (0.79), ‘POPTot’ (0.86). This is logical, since the more detail pages and overview pages viewed, the higher the number of total page views. It is also highly correlated with ‘sessionDuration’ (0.65) and ‘TotTimePOP’ (0.57). Also this is not surprisingly, since having longer sessions might result in viewing more pages. The correlation matrix further reveals high multicollinearity of ‘sessionDuration’ with ‘TotTimePOP’ (0.72) and ‘TotTimePDP’ (.71). This seems to make sense, since these variables were all created using timestamps. The high multicollinearity between ‘sessionDuration’ and ‘POPTot’ (.53) and ’PDPTot’ (.50 ) is likely because again, the longer you browse the web (sessionDuration), the more pages you can view, and the other way around. Also ‘BasketTot’ is correlated with ‘TotTimeBasket’ (0.54) and ‘POPTot’ with ‘TotTimePOP’ (0.68). The time on page related variables and page views related variables thus seem to be highly correlated. In order to resolve this, ‘HomePageTot’, ‘CategoryTot’, ‘BasketTot’, ‘PCPTot’, ‘PDPTot’ and ‘POPTot’ have been calculated relative to the total number of page views (TotPageViews), and the same has been done for the sessionDuration related variables. These relative variables are expressed as percentages to facilitate interpretation purposes. This new correlation matrix illustrated even higher correlations. When a correlation matrix is made with only the relative total number of page views, and the normal time on page variables, most of the correlation has been resolved (Appendix B, table 2). As could be seen in table 5 below, all VIF values are below the threshold of 4.0 as well.
Table 5: VIF scores VIF
DisplayAd 1.004140
SearchAd 1.107996
HomePageTotPer 1.187474
26 BasketTotPer 1.329930 PCPTotPer 1.218460 POPTotPer 1.427431 PDPTotPer 1.362435 TotPageViews 3.038086 sessionDuration 3.323932 TotTimeHomePage 1.277733 TotTimeCategory 1.317114 TotTimeBasket 1.455543 TotTimePCP 1.269879 TotTimePOP 2.980725 TotTimePDP 2.677066 StorePage 1.051832
To deal with the last highly correlated variables, interaction effects have been created between ‘sessionDuration’ and ‘TotTimePDP’, ‘sessionDuration’ and ‘TotTimePOP’ & ‘sessionDuration and ‘TotPageViews’.
4.2.4. Model fit & model selection
The structure of the dataset had to be transformed from ‘wide’ to ‘long’, which multiplies the amount of observations by 4, because each observation gets an individual row per segment. Since the dataset is rather large because of this transformation, a 75% random subset of this dataset is taken. This reduced the computational power necessary to run the multinomial regression, which made the analysis more time-efficient.
27
Various models have been created to test which performs best in explaining the effects of the variables on the segments. The comparison of some these models can be found in table 6 below. A null model has been added as well to see if the chosen model performs better than a model without explanatory variables. These models are subsequently compared by means of AIC, BIC, likelihood ratio test, McFadden R², and log-likelihood.
Table 6: Model selection criteria MNL
AIC BIC LR McFadden R² Log
Likelihood M1segments 169,545.4 169,590.4 183,760 0.52027 (p<.001***) -84,719 M2segments 165,723.6 165,776.1 187,600 0.53114 (p<.001***) -82,799 M3segments 183,055.4 183,090.4 170,220 0.48195 (p<.001***) -91,486 M4segments 178,462.1 178,502.1 174,830 0.49499 (p<.001***) -89,183 M5segments 169,720.1 169,762.6 183,580 0.51976 (p<.001***) -84,809
Mnullsegments 353,201.3 353,203.8 -1.1642e-10 -2.2204e-16 (p=1) -176,598
Indeed the best model turns out to be the model without correlated variables, with the relative page view variables and with the interaction effects (M2segments). The other models either excluded interaction effects, included correlation variables, or contained absolute page view variables. M2segments has the lowest AIC, BIC, and Log Likelihood, and the highest Likelihood Ratio and McFadden R2. McFadden R² indicates the variance in the dependent variable explained by the independent variables. The higher this value in comparison to the other models, the better the model is.
‘M2segments’ seems to be the most parsimonious model relative to the other models, also when penalized for the number of parameters. It also has the most explanatory power for clustering sessions into distinct segments compared to the other models.
4.2.5. Independence of Irrelevant Alternatives
28
online purchasing. The IIA will be tested by means of the Hausman-McFadden (hmf) test. Four tests have been conducted, each version leaving out a different segment. The hmf-test will compare these restricted models to the full model. If the hmf-test returns a significant result, the multinomial logit is considered inappropriate (Cheng & Long, 2007). The outcome of the hmf test can be found in table 7.
Table 7: Independence of Irrelevant Alternatives Assumption
p Reject/Accept
IIA
Model without Showrooming <.001*** Reject
Model without Webrooming <.001*** Reject
Model without Onlinepurchasing <.001*** Reject
Model without Other <.001*** Reject
All models returned significant p-values, meaning that the IIA assumption has to be rejected. The availability of the alternative segments, thus affects the likelihood to be part of a specific segment. A potential next step is conducting a nested logit, which is a generalization of the multinomial logit (Hausman & McFadden, 1981). However, first the underlying reason for violation of the IIA will be further investigated, since in some cases adhering to the IIA seems irrational (Benson, Kumar, & Tomkins, 2016). When testing the different alternative subsets the reference level has been set at the “Other” segment. The same process has been applied with different reference levels, after which the IIA still gets violated. As could be seen in table 2, “Other” is the largest of the four segments. To check whether the segment size influenced the results, the IIA has been tested on a subset consisting of just the three other segments. Once again the IIA got violated. Also samples of the two largest segments have been made to be of similar size as the other segments. Still, the IIA is violated. The proper way to continue the analysis is by conducting a nested multinomial logistic regression, which alleviates the IIA assumption (Leeflang et al., 2016).
4.3. Nested multinomial logistic regression
29
and the other way around. The same story holds for the second nest. Between these nests however, the assumption is violated, e.g. the possibility to webroom does affect the likelihood to be showrooming. To ensure that the IIA is satisfied within the nests, the most similar segments are included in the same nest. The most important distinction between these two nests is that the first one requires that an online purchase has been made, whereas the other not.
Figure 2. Nests of nested logit.
4.3.1. Model fit & model selection
The nested logit model will be compared to the multinomial model by means of the AIC, BIC, likelihood-ratio scores, McFadden R² and log likelihood. A model is considered better than the other when it has a lower AIC, BIC and log likelihood score. On the other hand, a model is considered better when it has a higher likelihood ratio score and McFadden R² than the other. For both types of regressions, the model consists of the same explanatory variables. The reference level of both models is set at “Other”. The difference between the multinomial logit and nested logit is that the nested logit model consists of the 2 nests (“Showrooming” & “Onlinepurchasing”) and (“Webrooming” & “Other”).
The scores reveal that the nested logit model is the better model of the two, as could be seen in table 8. The nested model has lower AIC and BIC scores, and a higher LR score and higher McFadden R² score as well. Also the log likelihood score is closer to zero for the nested model. These values all indicate that the nested logit model outperforms the multinomial
Nested Model
Nest 1
Showrooming Onlinepurchasing
Nest 2
30
model. Since the IIA is violated for the multinomial logit, and the nested model is the better performing model, the nested logit model will be interpreted.
Additionally, a t-test has been conducted to test whether the true model is a standard logit model or not. The inclusive value of the nested model is used for this. The t-test returned a value of 5.73285. Since this value is above the threshold of 1.96, it can be concluded that the true model is not a standard logit model. Everything taken together, a multinomial logistic regression or a standard logistic regression are less appropriate than a nested logistic regression. It can be concluded that the nested logit model is the right model for this analysis. Table 8: Model selection criteria MNL & Nested Logit
AIC BIC LR McFadden R² Log
Likelihood
Nested logit 165,676.1 165,729.4 187,650 0.53128 (P<.001***) -82,774
Multinomial logit 165,723.6 165,776.1 187,600 0.53114 (p<.001***) -82,799
The estimated model is as follows:
With: j = segment 1, …, 4 Where: = Segments = intercept for segment j
= dummy display advertisement for segment j = dummy search advertisement for segment j
31 = number of basket pages viewed relative to total page views for segment j = number of product category pages viewed relative to total page views for segment j
= number of product overview pages viewed relative to total page views for segment j
= number of product detail pages viewed relative to total page views for segment j = number of total page views for segment j
= total time on page for segment j
= total time on homepage for segment j = total time on category page for segment j = total time on basket page for segment j = total time on product category page for segment j = total time on product overview page for segment j = total time on product detail page for segment j = dummy store page for segment j
= interaction effect of total time on page and total number of pages viewed for segment j
= interaction effect of total time on page and total time on product overview page for segment j
= interaction effect of total time on page and total time on product detail page for segment j
= error term for segment j
This model will be interpreted in the next section.
4.3.2. Model interpretation
32
independent variables and the segments. This will suffice for either accepting or rejecting the hypotheses. However, the odds ratios have been included in the model as well to interpret the actual effect sizes. These effect sizes are calculated by taking the exponent of the coefficients. The results will be interpreted relative to the “Other” segment. The formulated hypotheses and the results will be covered per segment. Findings not directly related to the hypotheses, and interaction effects will be interpreted in the discussion section. The numbers mentioned in the following section are all retrieved from table 9. The β explains the sign of the relationship between an independent and dependent variable. The odds ratio explains how strong this relationship is. The p-value explains whether the relationship is significant. An overview of the accepted and rejected hypotheses can be found at the end of this section in table 10.
4.3.2.1. Showrooming results
In this section findings regarding the determinants of showrooming behavior will be mentioned. The beta, odds ratios and significance levels will be used for this.
‘TotPageViews’ (β = .1917, OR = 1.2112, p < .05) is the total number of pages viewed per session. The coefficient shows that it is a positive relationship. The odds of choosing Showrooming over Other changes by a factor of 1.2112 if the total number of pages viewed in a session increases by 1 unit, everything else kept constant. In other words, the more pages viewed, the more likely the person will be showrooming in this session, compared to Other. A similar positive relationship for ‘sessionDuration’ (β = .1177, OR = 1.1249, p < .05) is found. These findings both seem to be contradictory to hypotheses H1a and H2a. Also, the main effects cannot be interpreted like this, because these variables were correlated. To account for this correlation, the interaction effect ‘sessionDuration:TotPageViews’ (β = -.0042, OR = .9958, p <.05) should be interpreted. The more time spent on the website, the less positive the effect of the number of pages viewed on showrooming becomes. Since the separate effects of these variables cannot be interpreted, H1a and H2a remain inconclusive.
33
‘TotTimePDP’ (β = .1158, OR = 1.1228, p < .05) is the total time spent on product detail pages. Since this variable interacts with session duration, the main effect could not be interpreted, ‘sessionDuration:TotTimePDP’ (β = -.0023, OR = .9977, p < .05). The more time spent on the website, the less positive the effect of the time spent on product detail pages on showrooming becomes. Because there is no direct effect, H4a remains inconclusive.
‘StorePage’ (β = -1.6878, OR = .1849, p < .05) is a dummy variable indicating whether a storepage has been visited or not. The odds of choosing Showrooming over Other changes by a factor of .1849 if a storepage is visited. This finding implies that when a storepage is visited in a specific session, the customer is less likely to be showrooming, everything else kept constant. The negative relationship between visiting a store page and showrooming is in line with H5a. Hence, H5a is accepted.
‘SearchAd’ (β = .8757, OR = 2.4006, p <.05) explains whether the customer entered the website via a search advertisement or not. The odds of choosing Showrooming over Other changes by a factor of 2.4006 if the website is entered via a search ad, everything else kept constant. The positive relationship means that someone entering the website via a search ad is more likely to be showrooming. This corresponds to H6a, therefore H6a is accepted.
4.3.2.2. Webrooming results
By means of the betas, odds ratios and significance levels, the factors influencing webrooming behavior will be interpreted.
Just as for the showrooming segment, the relationship between page views and webrooming as well as the relationship between session duration and webrooming cannot be interpreted because of the interaction effect between, ‘sessionDuration:TotPageViews’ (β = -.0043, OR = .9957, p < .05). The effect of the number of pages viewed on showrooming depends on the length of the session duration. The more time spent on the website, the less positive the effect of the number of pages viewed on webrooming becomes. Nonetheless, some of the more specific pages viewed, and the time spent on these pages could be interpreted. H1b and H2b remain inconclusive.
34
likely that a person is to be webrooming. This positive relationship confirms H3b, and will as such be accepted.
The time spent on these product category pages ‘TotTimePCP’ (β = -.1392, OR = .8755, p <.05) is negatively related to webrooming. The odds of choosing Webrooming over Other changes by a factor of .8755 if the time spent on product category pages increases by 1 minute, everything else kept constant. Hence, the more time spent on PCP’s, the less likely that a person is webrooming in that session, meaning that hypothesis H4b is rejected.
‘POPTotPer’ (β = .0571, OR = 1.0587, p < .05) is the number of product overview pages viewed relative to the total number of pages viewed in a session. The odds of choosing Webrooming over Other changes by a factor of 1.0587 if the number of POP’s viewed relative to the total number of pages viewed increases by 1%, everything else kept constant. The more POP’s viewed relative to the total page views in a session, the more likely that the person will be webrooming. The relationship is thus positive and in line with H5b. therefore, H5b will be accepted.
The main effect of time spent on these POP’s ‘TotTimePOP’ (β = -.0783, OR = .9246, p < .05) cannot be interpreted, because it interacts with session duration, ‘sessionDuration:TotTimePOP’ (β = .0015, OR = 1.0015, p < .05). The same holds for the time spent on product detail pages, since it interacts with session duration as well ‘sessionDuration:TotTimePDP’ (β = .0011, OR = 1.0011, p < .05). The relationship between TotTimePOP and TotTimePDP with webrooming thus stays indecisive. H6b and H8b remain inconclusive.
The effect of the number of PDP’s viewed ‘PDPTotPer’ (β = .0464, OR = 1.0419, p < .05) on webrooming, can be interpreted. The odds of choosing Webrooming over Other changes by a factor of 1.0419 if the number of PDP’s viewed relative to the total number of pages viewed in a session increases by 1%, everything else kept constant. So the more relative PDP’s viewed, the more likely that a person will be webrooming. This confirms the hypothesized positive effect between PDP’s viewed and webrooming, accepting H7b.
Whether a storepage is viewed in a session or not, ‘StorePage’ (β = -1.6243, OR = .1970, p < .05) is negatively related to webrooming behavior. The odds of choosing Webrooming over Other changes by a factor of .1970 if a storepage is visited, everything else kept constant. Hence, visiting a storepage decreases the likelihood of a person to be webrooming. This contradicts with the proposed hypothesis, as such H9b will rejected.
35
choosing Webrooming over Other changes by a factor of .0139 if the website is accessed via a display ad, everything else kept constant. So when a person accesses the website via a display advertisement, he or she is less likely to be webrooming. This negative relationship is in contrast with H10b, which will accordingly be rejected.
4.3.2.3. Online purchasing results
Lastly, the effects of the independent variables on the online purchasing segment will be interpreted.
Just as for the Showrooming and Webrooming segments, no conclusions could be drawn for the first 2 hypotheses (H1c and H2c) because of the interaction effects of sessionDuration with TotPageViews, ‘sessionDuration:TotPageViews’ (β = -.0032, OR = .9968, p < .05). H1c and H2c remain inconclusive.
‘PCPTotPer’ (β = .0625, OR = 1.0645, p < .05) illustrates that the number of product category pages viewed is positively related to online purchase behavior. The odds of choosing ‘Onlinepurchasing’ over Other changes by a factor of 1.0645 if the number of PCP’s viewed relative to the total number of pages viewed increases by 1%, everything else kept constant. So the more PCP’s viewed relative to the total number of pages viewed, the higher the likelihood that the person will be an online purchaser, conforming to H3c. H3c will be accepted.
‘TotTimePCP’ (β = -.2237, OR = .7996, p = .3896) indicates the total time spent on product category pages relative to the total number of pages viewed. Because this coefficient is insignificant, the result is not accurate and should not be interpreted. There is no support for H4c thus, H4c will be rejected.
The number of product overview pages viewed relative to the total number of pages viewed is illustrated by ‘POPTotPer’ (β = -.0255, OR = .9749, p < .05). The odds of choosing ‘Onlinepurchasing’ over Other changes by a factor of .9749 if the number of POP’s viewed relative to the total number of pages viewed increases by 1%, everything else kept constant. Thus, relatively seen, the more POP’s viewed in a session, the less likely that the person will be an online purchaser. This negative effect between number of POP’s viewed and online purchasing is in contrast with H5c, as such H5c will be rejected.
36
‘sessionDuration:TotTimePDP’ (β = .0007, OR = 1.0008, p <.05). H6c and H8c remain inconclusive.
The effect of the number of product detail pages viewed, on the other hand, can be interpreted ‘PDPTotPer’ (β = -.0233, OR = .9683, p < .05). The odds of choosing ‘Onlinepurchasing’ over Other changes by a factor of .9683 if the number of PDP’s viewed relative to the total number of pages viewed increases by 1%, everything else kept constant. In other words, relatively seen, the more PDP’s viewed in a session, the less likely that the person will be an online purchaser. This contradicts to H7c. Therefore H7c will be rejected.
Visiting a storepage, ‘StorePage’ (β = -1.3533, OR = .2584, p < .05) is negatively related to online purchasing. The odds of choosing ‘Onlinepurchasing’ over Other changes by a factor of .2584 if a storepage is visited, everything else kept constant. So if a storepage is visited in a session, the more likely it is that the person is an online purchaser. This is in line with the proposed hypothesis, therefore H9c will be accepted.
‘DisplayAd’ (β = -2.3240, OR = .0979, p < .05) explains the effect of accessing the website via a display ad on online purchasing behavior. The odds of choosing ‘Onlinepurchasing’ over Other changes by a factor of .0979 if ‘the website is accesses via a display ad, everything else kept constant. If a person enters the website via a display ad, it is less likely that the person is an online purchaser. This is in contrast with the formulated hypothesis. As such, H10c will be rejected.
37
Table 9: Output Nested Logit & Odds Ratio
Showrooming Webrooming Online purchasing
38
Table 10: Overview findings hypotheses
Segment Hypotheses Findings
Showrooming H1a: Page views is negatively related to showrooming behavior n.f.
H2a: Time on page is negatively related to showrooming behavior. n.f.
H3a: Number of PDPs viewed is negatively related to showrooming behavior. Accepted
H4a: Time on PDPs is negatively related to showrooming behavior. n.f.
H5a: Visiting a Store page is negatively related to showrooming. Accepted
H6a: There is a positive relationship between search ads and showrooming behavior. Accepted
Webrooming H1b: Page views is positively related to webrooming behavior. n.f.
H2b: Session duration is positively related to webrooming behavior. n.f.
H3b: Number of PCPs viewed is positively related to webrooming behavior. Accepted
H4b: Time on PCPs is positively related to webrooming behavior. Rejected
H5b: Number of POPs viewed is positively related to webrooming behavior. Accepted
H6b: Time on POPs is positively related to webrooming behavior. n.f.
H7b: Number of PDPs viewed is positively related to webrooming behavior. Accepted
H8b: Time on PDPs is positively related to webrooming behavior. n.f.
H9b: Visiting a Store page is positively related to webrooming. Rejected
H10b: There is a positive relationship between display ads and webrooming behavior. Rejected
Online purchasing
H1c: Page views is positively related to online purchase behavior. n.f.
H2c: session duration is positively related to online purchase behavior. n.f.
39
H4c: Time on PCPs is positively related to online purchase behavior. Rejected
H5c: Number of POPs viewed is positively related to online purchase behavior. Rejected
H6c: Time on POPs is positively related to online purchase behavior. n.f.
H7c: Number of PDPs viewed is positively related to online purchase behavior. Rejected
H8c: Time on PDPs is positively related to online purchase behavior. n.f.
H9c: Visiting a Store page is negatively related to online purchasing. Accepted
H10c: There is a positive relationship between display ads and online purchasing. Rejected
H11c: There is a positive relationship between search ads and online purchasing. Accepted
40
4.4. Discussion
Before discussing the results, the representativeness of the data will shortly be explained. Since the data used for this analysis is a random subset of the actual data available, it can be considered quite representative for the whole dataset. As such, when the same analysis would be run on the whole dataset, similar results are expected. Therefore, the segments and the findings are representative for the data at hand. Due to computational constraints, running the analysis on the whole dataset unfortunately was not possible.
In this section, the expected and unexpected findings will be discussed. The characteristics of the segments as discovered in the analysis will be explained and the segments will be compared to each other. The interpretation is based upon the results in table 9.
4.4.1. Showrooming
The results of this research reveal that not all hypotheses are supported. Take for example page views and visit duration. As previously recommended by Bucklin & Sismeiro (2003), page views and visit duration are often jointly modeled, since they provide a parsimonious representation of the browsing behavior. The findings of session duration and the amount of pages viewed indeed emphasize that these variables represent browsing behavior, albeit not separately. The more time spent on the website, the less positive the effect of the number of pages viewed on showrooming becomes.
Daunt & Harris (2017) argued that showroomers engage with products in-store, where they already assess all its aspects and acquire relevant information before going to the web. Therefore one would expect that less product detail pages need to be visited, which is exactly what is found in this research. Because of the interaction effect with session duration, it could not be concluded that time spent on these PDP’s negatively affects showrooming behavior. Identifying this effect is something to be identified by future research. If there actually turns out to be a positive effect, it might imply that although showroomers are well aware of the product they are willing to buy, they still carefully examine the product before making the purchase. They might still conduct some last checks about the size or color of the product. How showroomers spend their time on product detail pages might provide interesting additional insights and further clarify showrooming behavior throughout the customer journey.
