1
Evaluating the effect of device usage across the purchase funnel in online shopping
3 Evaluating the effect of device usage across the purchase funnel in online shopping
Annemijn van der Vaart University of Groningen Faculty of Economics and Business
MSc Marketing Intelligence & MSc Marketing Management
5
Preface
This master thesis is my final project before ending my academic career and completing the masters ‘Marketing Management’ and ‘Marketing Intelligence’ at the University of Groningen. It has been a long and formative process in which I have learnt a lot about myself, but especially extended my knowledge in the field of Marketing.
I would like to make use of this opportunity to express my gratitude to some people who have been very important during my studies, as well as during the last months when writing this master thesis. First of all, I want to thank my friends and family for their support and patience with me. Special thank goes out to the organization who provided me the data to conduct this research.
I took great pleasure in writing my thesis the past half year because it offered me the chance to obtain practical experience in the field of data analysis. I got the opportunity to choose a topic of my own interest and learn a lot in the area of marketing intelligence of which I am very grateful. As I have always been interested in consumer behavior, the topic of my thesis perfectly suits my personal interest. The Master thesis has made me realize that I made the right decision when I decided to combine Marketing Management with Marketing Intelligence and I definitely want to pursue my career in these fields.
I hope you will enjoy reading this thesis.
7
Management summary
Using multiple devices across the online customer journey is becoming more and more popular. Mobile devices enable customers to access a retailer’s website ‘anytime’ and ‘anywhere’, which has led to an enormous increase of mobile commerce over the last years. However, retailers are skeptical about whether these additional devices lead to benefits, as conversion rates and order values of orders originating from a PC are higher than those originating from mobile devices.
Even though research on device usage is growing, the effect of the devices used across the purchase funnel is underexplored. Also retailers only consider the device that is used to make the purchase. This evaluation process is also known as the last-click metric. By integrating the intermediate steps of the customer journey this study evaluates the differential effect of devices across the purchase funnel.
Results show that device usage has a significant effect on conversion and this effect differs between stages and devices. Retailers have no right to be longer skeptical about the effectiveness of mobile devices. This study shows that in the beginning of the purchase funnel the smartphone has the largest effect on conversion, towards the end of the funnel the PC becomes the device with the highest impact. Furthermore, the affective stage, the part of the customer journey where a customer creates a feeling towards the product, is found to significantly contribute to order value. This research shows that het last-click metric is outdated as devices used before a purchase is made also influence conversion and order value.
Tangible recommendations are given to online retailers based on these outcomes. Since smartphones have the largest influence at the start of the customer journeys, retailers should invest more in mobile advertisements. Additionally, retailers should focus on retargeting customers that visit the customer using a PC in the beginning of the purchase funnel in order to retain them. To increase order value, online retailers should educate smartphone and tablet users that payments made on mobile devices are safe. Moreover, online retailers could also consider to implement the option ‘buy now and pay later’ for mobile device users.
9
Abstract
Given the ongoing increase in electronic retailing and the use of novel sorts of computing devices, a better understanding of online purchasing behavior has become an important issue for online retailers. This study examines how shopping behavior varies between different internet-connected devices.
This study investigates the effect of three distinct devices on online shopping behavior. The devices included in this study are smartphone, tablet and PC. These devices are examined across three stages of the purchase funnel, the cognitive, affective and conative stage. All nine possible combinations of device used and purchase funnel stage are examined to evaluate their individual effect on conversion and order value.
The examined dataset includes all the identified visits originating from the three devices of customers during their customer journey. Using this unique dataset, a Type-II Tobit model was performed in order to create a parsimonious yet complete view that captures the decisions visitors face during a customer journey. The first decision is whether the customer makes a purchase or not. The succeeding decision is how much to spend, given that he converted.
The results show significant differences between the devices in their effect on conversion. The smartphone has the largest influence in the cognitive and the affective stage and the PC has the largest effect in the conative stage. Additionally, the tablet and PC are found to have a positive effect on order value. Overall, this research offers actionable implications for online retailers, as well as contributions to the current literature on the effect of device usage.
Keywords: device usage, smartphone, tablet, PC, purchase funnel, hierarchy of effects, Type-II
11
Content
Preface ... 5 Management summary ... 7 Abstract ... 9 1. Introduction ... 13 2. Theoretical Framework ... 17 2.1 Internet-connected devices... 172.2 The Purchase Funnel ... 18
2.3 Device-usage across the purchase funnel ... 19
3. Hypotheses and conceptual framework ... 21
3.1 The effect of device choice across the purchase funnel on conversion... 21
3.2 The effect of device choice across the purchase funnel on order value ... 24
3.3 Control variables... 25
3.4 Conceptual framework ... 28
4. Methodology ... 29
4.1 The model ... 29
4.2 Data sources ... 30
4.3 Operationalization of key variables ... 31
5. Results ... 34 5.1 Parameter estimates ... 34 5.2 Validity ... 38 5.3 Robustness check ... 39 5.4 Hypotheses testing ... 40 6. Discussion ... 42 6.1 Theoretical Implications ... 42 6.2 Managerial Implications ... 43
6.3 Limitations and future research ... 45
6.4 Conclusion ... 46
7. References ... 47
Appendices... 51
Appendix I: Seasonality effect ... 51
Appendix II: Brand effect ... 52
Appendix III: Conative stage; carting and buying sessions divided ... 53
13
1. Introduction
Novel sorts of computing devices with innovative interface types and standards have risen in prominence over the last several years. It is expected that by the end of 2016 there will be 6,4 billion devices connected to the internet worldwide. This number is expected to grow to almost 21 billion by 2020 (Gartner, 2015). Especially the smartphone has become ubiquitous over the last few years. It is no secret that consumers are totally attached to their smartphones. It has become an indispensable device for communication, entertainment and e-commerce. The IDC has predicted that by 2017 more than 70 percent of all connected devices in the world will be smartphones, whereas tablets account for 17 percent and personal computers (PCs) for 13 percent (Columbus, 2013). The prominence of these devices leads to more diverse and thorough use and a larger control over the purchase funnel (van der Veen & van Ossenbruggen, 2015). With numerous touch points and easy access to information, the customer journey is no longer a straightforward path.
Biased attribution in online shopping
The Internet has enabled consumers to interact with retailers on their own terms (Shankar & Malthouse, 2007), the arrival of mobile devices enabled consumers to browse and purchase products from retailers anytime and anywhere (Blázquez, 2014). The technological development in the computing devices resulted in significant changes in online shopping. Not surprisingly, online shops are increasingly explored by various connected devices (Mosteller, et al., 2014). Mobile devices, comprising both smartphones and tablets, are used more and more for online shopping compared to the more traditional PCs. On several websites, over one-half of the views originate from a mobile device. The share of mobile transactions in e-commerce has also grown extensively. In 2015, mobile commerce accounted for 35% of total online spending worldwide (Criteo, 2015).
Despite this good news, not all retailers perceive mobile devices as a blessing. Mobile devices cause an uplift in website traffic, but these visits result in substantially lower conversion rates than PCs (Bosomworth, 2015). There is a clear disconnect between traffic and the actual purchase of goods originating from mobile devices (Kumar & Mukherjee, 2013). A study by Barrilliance (2014) revealed that one out of three customers that start their journey on a mobile device, convert on a different device, giving reasons for online retailers to doubt the return on their mobile investments. Visitors on a smartphone are more likely to complete their purchase on a different device, while visitors using a PC during the journey are more likely to remain on that device and complete the purchase on it.
14 Kannan, 2014), an aggregate measure to evaluate the marketing investment. Hereby, the conversion is attributed to the last device used. Other devices used in the funnel are ignored which leads to biased estimations of conversion (Martin, 2009).
Understanding the true role each device plays in the customer journey is crucial in enhancing the customer experiences and improving business results. Therefore, devices should not be isolated and evaluated on conversion rates by means of the last-click metric solely (Li & Kannan, 2014). Rather, retailers have to scrutinize how the various devices complement each other in the purchase funnel. A clear image of the influence of the different devices across the funnel is needed for retailers to obtain a better understanding of consumers’ online shopping behavior.
Research Direction
Given the enormous growth of the use of various connected devices, a wholesome understanding of online shopping behavior has become an important topic for online retailers. Treating customers according to their preferred device, whether they browse and shop on a smartphone, tablet or PC has become a major challenge of marketing strategy (McKinsey, 2014).
The purpose of this study is to investigate the differential effect of the devices used across the purchase funnel on conversion and order value. Conversions made on smartphones for example, occur at a lower frequency and are of lower value than conversions made on other devices (IBM, 2014). This does not mean that customers who browse on their smartphone are less likely to convert, instead they might change their device and complete their shopping journey after the shift. One device might fit better at a specific stage than another device, since every device has its advantages and disadvantages (de Haan, 2015) and every step in the purchase funnel has its own purpose (Larivière, et al., 2013).
A model that integrates the intermediate devices used during sessions at the different stages of the purchase funnel is necessary to correctly measure the contribution of the devices to the customer journey. Such insights help the retailer to obtain a better understanding of the true value of the different devices. In short the research aim of this paper is
‘To what extent does device choice across the purchase funnel influence the customer purchase behavior’
For this main research question, the following two sub questions can be derived:
1. What is the effect of the different devices used across the purchase funnel on conversion probability?
2. When there is a conversion, what is the effect of the devices used on order value?
15 customer journeys at a large online retailer, a Type II Tobit model is proposed to estimate the probability of conversion and the order value, given the devices used.
Contribution to the literature
Although literature on cross-device shopping is growing, the joint view on devices and the purchase funnel is underexplored. This research however, examines the effect of smartphone, tablet and PC usage across the purchase funnel on conversion.
The contribution to the literature is thereby twofold. First, this study considers smartphones, tablets and PCs as devices used for online shopping. Prior research primarily focused on the differences in usage behavior between smartphones and PCs (Ghose, et al., 2012), not considering tablets, or clustering smartphones and tablets as mobile devices (Shankar, et al., 2010). Research on the effect of tablet use is still very scarce. Tablets are shown to have similar characteristics as smartphones, as they are portable, but the usage behavior mimics the behavior on PC, since the conversion rates for tablets more closely resemble those of PCs. It is therefore important to consider the tablet as a separate device (Blum, 2012). Second, while previous studies gave a lot of interesting insights in search activities (Wang, et al., 2013) (Montañez, et al., 2014) and purchase behavior (Google, 2012) (IBM, 2014) using various devices, the intermediate steps of the purchase funnel are hardly investigated. As all the stages in the purchase funnel serve different goals (Lavidge & Steiner, 1961) it is logical to assume that this might lead to varying effects on customer journey outcomes. The study of De Haan et al. (2015) is the first to examine the use of multiple devices across the path to purchase, yet mainly focusing on device shifts.
Overall, this paper provides an understanding of how purchase funnels are different in their effect on conversion and order value, because of the devices used. To the author’s best knowledge this is the first study to examine the three different devices across the entire purchase funnel. Furthermore, this research not only examines the probability of conversion as most of the studies do, by implementing the Type II Tobit model, it also investigates the effect of devices used on order value when a conversion occurs.
Contribution to the field
As McKinsey (2014) devoted the ‘device shift’ as one of the six major consumer trends, it can be concluded that insights about this topic are extremely valuable to the field. With the various connected devices becoming de facto for online activities, it has become even more important for online retailers to understand the dynamics behind it. This study provides insights that create a better understanding of the conversion probability and order value of customers, which is the key to personalized targeting.
16 on marketing spending is of influence on the marketing budgets, an unbiased attribution model should be used to determine the optimal allocation of budget.
Using a unique dataset from a large European retailer, this research evaluates the changes in customers’ conversion behavior and spending upon using different devices. This research provides insights into the added value of different devices to the customer journey, which is unknown so far. The main purpose of this analysis is to help the business better understand its customer’s behavior and therefore conduct customer-centric marketing more effectively. Concluding, this research provides the retailer with valuable insights to revise its cross-device strategies.
17
2. Theoretical Framework
The purpose of this research is to discover to what extent devices used during the customer journey affect conversion and order value. With the help of existing literature the following elements will be explored in this literature review. First, the different devices used in online shopping will be elaborated. Second, the different stages of the purchase funnel are discussed. Third, the use of the various devices across the purchase funnel will be expounded. After this, the possible relation between the use of the different devices on conversion and order value are discussed and hypotheses will be composed.
2.1
Internet-connected devices
Devices can be broadly divided into mobile devices and PCs. Mobile devices are smartphones and tablets, desktops and laptops represent PCs. Technology keeps developing and more and more connected devices enter the market. However, this study solely focuses on smartphones, tablets and PCs, which are the most common devices used in online shopping at the moment.
Even though all Internet accessible devices provide instant access to the same Internet sources (Ghose, et al., 2012), various differences exist. The main differences; product appearance, accessibility and trust will be discussed in this section.
Product appearance
Smartphones and tablets are pocket- or purse size devices, which rely on touch-based control (Wang, et al., 2015). Moreover, mobile devices are known for their relatively smaller screen, low display resolution, inconvenient input facilities, limited computational power and memory capacity (Yang & Kim, 2012) (Coursaris & Kim, 2011). Generally, tablets have a larger screen than smartphones. However, as technologies keep improving the usability of mobile devices, some larger smartphones have already surpassed smaller tablets in screen size. Inconveniences from the smaller screen do not entirely vanish by these improvements and might still lead to a lower perceived ease of use for these devices in online shopping (Jayasingh & Eze, 2012). Mobile devices have constraints which are related to the smaller screen with resulting typing inconveniences (Chandra, et al., 2010).
The largest internet accessible devices are the desktops and laptops. PCs have a relatively large screen, are mostly accompanied by a keyboard, track pad or mouse and offer higher functionalities such as more processing power (Wang, et al., 2015). This leads to a more fluent and convenient usage of these devices. Desktops require a connection with electricity, while laptops are the portable versions of desktops, but with equal functionalities.
Accessibility
18 connection and are used as communication device. Most of these characteristics also hold for the tablet, however the larger size makes it less attractive for consumers to always carry it with them.
The desktop is a fixed device, following that the location of its use is immobile. A laptop on the other hand is not fixed to one location. However, due to its size, it is less portable than a mobile device. The required electricity and the larger product size lead to lower flexibility to the user.
Trust
Consumers perceive mobile devices as less secure for online shopping and transferring money (Chin, et al., 2012). Consumers perceive a lack of privacy and security when using mobile devices. A PC on the other hand is often used in the security of the consumer’s home or workplace with a fixed internet connection which increases the perceived secureness.
2.2
The Purchase Funnel
While most studies in online retail focus on the purchasing stage, the online shopping behavior is not monolithic. Customers move towards a purchase in a series of stages, which is called the purchase funnel (Alba, et al., 1997). Consumers commence with a broad idea of a need they have, the wide end of the funnel. The several touch points of interaction the customer has with the retailer in the proceedings of its search are intended to narrow down the funnel and direct the customer to a purchase (McKinsey, 2009). A consumer engages in receiving product information before it places an order. The Internet offers many possibilities to increase the productivity of online shopping behavior by refining the availability of product information, enable product comparisons and reduce the search cost, because of its interactivity (Alba, et al., 1997). The consumer does not make a single decision, but the entire purchase funnel consists of numerous decisions in different stages.
Hierarchy of effects
The sequential stages of the purchase funnel, a series of stages between unawareness and the final purchase, have been captured in a model. This model, which is called the hierarchy of effects model indicates that the consumer moves from one stage to the next.
One of the most well-known hierarchy of effects models is the model of Lavidge and Steiner (1961), distinguishing three different stages leading to a purchase. The entire process of consumer behavior can be divided into six phases. These phases are awareness, knowledge, liking, preference, conviction and purchase. The six phases can be broadly distributed into three stages, which is necessary in order to conduct empirical analyses.
Stages of the Purchase Funnel
19 transfer of information of the retailer’s website to the consumer through browsing (Gefen, 2002). The importance of information search has been broadly recognized. In order to gain sufficient knowledge to make well-informed decisions it is critical to learn the product specifications and requirements (Choudhury & Karahanna, 2008). The second stage is the affective stage, where the customer starts to create attitudes or feelings toward the product by comparing different products. This stage is about ensuring that the consumer likes the product and that he creates a preference towards it. Finally, the conative stage sets in when the customer turns into action and makes a purchase (Bruce, et al., 2012). In this behavioral stage there is an exchange between monetary value, customer information and the focal product (Pavlou & Fygenson, 2006).
2.3
Device-usage across the purchase funnel
The different characteristics of the devices are likely to have consequences on the online customer journey in terms of switching behavior and the impact on conversion and order value.
Perceived ease of use and perceived usefulness
Device acceptation is explained by the technology acceptance model (TAM) in literature. This framework is found to be the simplest, most powerful and easiest framework to predict technology acceptance (Igbaria, et al., 1995). The framework examines two dimensions, the perceived ease of use and the perceived usefulness, to be fundamental elements of usage behavior (Davis, 1989). The perceived ease of use is defined as ‘the degree to which a person believes that using a particular system would be free from effort’. Perceived usefulness is ‘the degree to which a person believes that the use of a particular system would enhance his or her job performance’ (Davis, 1989). Several prior studies have utilized this framework to analyze numerous mobile-based innovations. These studies focused on the ease of use of a smartphone (Galletta & Dunn, 2014), comparisons between the ease of use of tablets and laptops (Wetzlinger, et al., 2014) and the effect of screen size on smartphone adoption (Kim & Sundar, 2014).
The current study associates the perceived usefulness of a device with the device choice per stage in the purchase funnel. Even though this research does not investigate the device adoption in itself, it is believed that the TAM-framework is a good indicator for device choice. The choice for a specific device will be induced by the perceived value of the transaction, or in other words, the perception of the net benefits (Chang, et al., 2009). Among these benefits are the perceived ease of use and the perceived usefulness, the primary determinants of intentions to choose for a device (Trivedi & Kumar, 2014). Hence, it can be deduced that the perceived ease of use and usefulness of the different devices will impact consumers’ intention to use these devices for online shopping activities.
20 the Internet accessible devices have influence on the perceived benefits and risks of the smartphone, tablet and PC and will play a significant role in online shopping behavior (Malik, et al., 2013). It affects the manner the customer moves through the purchase funnel (de Haan, et al., 2015). Customers want to maximize their benefits and minimize their risks while performing activities as searching, browsing and ordering on the website.
Device usage
Previous research has found that customers often start their purchase funnel on a useful device for search activities but use another device to continue their journey. This switching behavior is also called the ‘research phenomenon’ (Verhoef, et al., 2007).
Online sessions on a mobile device are usually shorter and of a higher frequency then sessions on a PC (Cui & Roto, 2008). Mobile devices, and especially smartphones, are often used on the go, in public locations and even in physical shops (Forrester Research, 2012) but as well just at home (GfK , 2013). Because of their portability and accessibility, most people always have their smartphone with them. Mobile devices are ideal for quick information search and task oriented activities.
Most consumers are more immersed in the use of a smartphone than any other device. The perceived ease of use for search activities is probably the reason that a lot of online traffic originates from smartphones. However, the conversion rates on mobile devices, especially on smartphones, are substantially lower than on PCs (Google, 2012). This is also partly caused by the smaller screens of mobile devices. Mobile device users, and especially smartphone users, are less capable of engaging in a rigorous and deep information seeking process to fulfill all their needs. Mobile device users on the other hand engage in extractive behavior, indicating short purposeful engagements (Humphreys, et al., 2013).
A tablet is used as a content consumption device instead of a communication tool by most consumers. Consumers prefer to use a tablet for entertainment and browsing activities, preferably at home (Google, 2012). Most activities performed on a tablet show deeper engagement levels, resulting in more pages viewed and longer sessions than on a smartphone. The smaller screen of mobile devices limits the ability of the consumer to immediately accomplish what they want. For example, filling out payment forms to finish an order is harder on a small touchscreen, compared to a PC with a keyboard (Shankar, et al., 2010).
21
3. Hypotheses and conceptual framework
The journey that customers take through the purchase funnel is highly dependent on the different characteristics of the devices (de Haan, et al., 2015). In every step of the purchase funnel a consumer makes a tradeoff between the benefits received by the device used and its costs. In this part a conceptual framework will be established to examine the effect of the devices used on shopping behavior across the path to purchase.
3.1
The effect of device choice across the purchase funnel on conversion
The conversion rate of an online retailer is the percentage of website visitors who make a purchase on that website. The conversion rate is one of the most important metrics in e-commerce to measure the success of a retailer (Ayanso & Yoogalingam, 2009). A higher conversion rate is likely to generate more revenues to the firm.
The effect of device choice in the cognitive stage on conversion
During the cognitive stage, customers become aware of a product need and start to gather information (Lavidge & Steiner, 1961). A device will be perceived as useful when a consumer believes that the use of that device would enhance his task performance.
Because of its portability (Yang & Kim, 2012), a smartphone user has access to timely information (Ghose, et al., 2012) and is able to interact with a firm ‘anytime’ and ‘anywhere’ (Shankar, et al., 2010). Search costs for fast information are lower on a smartphone because of its ubiquitous access (Yang & Kim, 2012). Smartphones are convenient when customers are reluctant to devote a lot of time to search information and simply want to fulfill their needs fast (Ghose, et al., 2012). However, searching on small screens requires numerous manoeuvres of scrolling, which limits the type and amount of information customers obtain (Shankar et al 2010), making it hard to retrieve target information (Ghose, et al., 2012), and which might be annoying (Adipat, et al., 2011). Additionally, Ghose, et al. (2012) find that ranking effects are higher for smartphones, indicating that searchers on a smartphone usually click on a higher ranked piece of content than PC users, not scrolling down for more information.
22 a convenient device for extensive search tasks, it will have a high perceived ease of use to gather information on more difficult topics.
The perceived ease of use of a smartphone will be high in the cognitive stage for fast information search that requires low cognitive efforts. Yet, this does not necessarily lead to conversion. The limited cognitive abilities of a smartphone (Ghose, et al., 2012) make it a less suitable device to proceed in the purchase funnel on. The usage behavior on a smartphone is characterized as ‘extractive’, indicating that the user does not dive deep into the information. The level of involvement on a smartphone is clearly lower than on tablet and PC, on which consumers show deeper levels of involvement (O'Malley, 2013). As the involvement of customers is of significant influence on their buying behavior and leads to higher frequencies of purchases (Wang, et al., 2006), it is expected that the conversion probabilities for smartphones are lower due to lower levels of involvement.
In this research, it is expected that the smartphone is widely used during the cognitive stage but that consumers initiating their session on a smartphone in the cognitive stage are the least likely to proceed to a conversion. Thus, the following hypothesis will be tested:
H1: Sessions in the cognitive stage originating from a smartphone are least likely to lead to conversion compared to sessions originating from other devices.
The effect of device choice in the affective stage on conversion
In the intermediate stage of the purchase funnel, the affective stage, a customer has the desire to create a feeling about the product. To do so, the customer evaluates the product characteristics by exploring content, reading reviews and considering different opportunities. The affective quality of a device is of particular interest in this stage to the consumers’ device choice due to its ability to provoke a positive feeling (Kim & Sundar, 2013).
As mentioned previously, a smartphone is used for short purposeful engagements. The usefulness of a smartphone in the affective stage might therefore be questioned. The smaller screen makes it challenging to engage in a deeper level with the content provided, which is required in the affective stage. However, tablets appear to be easier to use in the affective stage. Tablet users are found to view 70 percent more pages per website than smartphone users (O'Malley, 2013), supporting the statement that tablet usage behavior closely resembles PC usage. Even though customers are taking advantage of the convenience provided by the smartphone, the increased search costs in the affective stage can negatively affect their cognitive abilities in locating or recalling web information (Ghose, et al., 2012), indicating that devices with larger screens are easier to use in the affective stage.
23 2013). Especially tablets are used for entertainment and exploratory tasks (Ghose, et al., 2012), indicating higher playfulness and enjoyment. Therefore, in this study we expect the perceived ease of use in the affective stage to be the highest for the tablet.
Since the tablet is equally used across all the stages of the purchase funnel (comScore, 2011) it is not only expected that customers use a tablet in the affective stage, but customers doing so are likely to progress further in the purchase funnel and convert. A higher enjoyment and playfulness result in behavioral intentions as purchasing (Fang, et al., 2009) (Koufaris, 2002). Therefore, in this research we expect that the higher enjoyment and involvement on a tablet will have positive influence on the conversion.
The hypothesis is as follows:
H2: Sessions in the affective stage originating from a tablet are most likely to lead to conversion compared to sessions originating from other devices.
The effect of device choice in the conative stage on conversion
Since the conative stage is the last stage in the purchase funnel, the device with the largest perceived ease of use in this stage will also be having the largest effect on conversion. Conversion rates increase by screen size of the device (Criteo, 2015). This is in line with Chaffey (2015) who found that conversion rates are highest for fixed devices, followed by tablets and then smartphones.
The PC is often perceived as the most secure device, as it is usely used in the safe environment of the consumers’ home and are therefore providing more convenience when purchases are made and money has to be transferred (Chin, et al., 2012). Consumers wait until they are home to actually make a purchase on a more secure device (Cui & Roto, 2008). The PC is the preferred device to make purchases with (Lee, et al., 2005), which is likely caused by the fact that activities in the conative stage require focus. These tasks are easier performed on a PC.
Using a mobile device to make a purchase continues to be riddled by external and internal barriers. The external barriers include elements as the smaller screen size, lack of steady Internet access, and non-responsive webpages, which do not anticipate to the device used. As technologies improve, these external barriers might be overcome and such statements will be harder to support, yet the internal or psychological barriers of lower perceived safety remain (Malik, et al., 2013). These reasons combined build larger uncertainties and risks, which hamper consumers to use mobile devices in the conative stage.
24 H3: Sessions in the conative stage originating from a PC are most likely to lead to conversion
compared to sessions originating from other devices.
3.2
The effect of device choice across the purchase funnel on order value
The way in which a decision is made should in itself not affect the decision on the monetary value as is explained by the economic theory of preferences (Avnet & Higgins, 2006). In the context of this study, a customer chooses to purchase a product in the end of a purchase funnel which can consist of several devices. The money that a customer then offers to spend should not depend on the device used, as long as the product remains equal. Yet, the idea that customers assign different values to the same product in different scenarios is no longer surprising. Reports by Google (2012) and Deloitte (2013) show that PCs have the highest order value, followed by tablet and then smartphone. Additional research has shown that perceived value can be transposed into monetary value. The monetary value customers assign to a product can depend on the strategy they have chosen to obtain the product (Avnet & Higgins, 2003), in the light of this research, the devices he has used across the purchase funnel.
Perceived value
The way in which a customer moves through the purchase funnel can influence the customers’ perceived value of the product bought (Avnet & Higgins, 2003). The utility a consumer attaches to a product is a function of the consumers’ risks and benefits during the process, his current goal orientation and the way in which he makes a choice. This perceived value can subsequently convert into a higher or lower monetary value of the order depending on the devices chosen in the purchase funnel.
As the perceived value is influenced by the ease of use, which is also an important indicator for conversion, it is expected to find similar effects for the devices used during the purchase funnel on conversion as on order value. The proposed effects will only be shortly discussed in this part as the perceived values are extensively discussed in the previous sections.
A study by Lynch & Ariely (2000) found that consumers are less price sensitive when information is easier to navigate. As smartphones suffer from ranking effects and high search costs it is expected that customers using a smartphone to browse the retailer’s webpage are more price sensitive. In a similar vein as the effect on conversion the hypothesis is as follows:
25 In addition to the ease of navigation, other aspects play an important role in determining the perceived overall benefits, costs and value in an exchange as well. Customers with a stronger commitment are found to spend more on their purchases (Venkatesan, et al., 2007). As tablet users show high engagement levels and equally use this device across the purchase funnel it is very likely that this higher commitment results in higher spendings. It is expected that customers using a tablet in the affective stage spend more. The hypothesis that will be tested is:
H5: Sessions in the affective stage originating from a tablet are likely to result in a higher order value compared to sessions originating from other devices.
Because of its longer history, the PC offers higher levels of familiarity, safety and confidence than smartphones and tablets. Negative outcomes will be minimized because of the higher levels of trust and the higher ease of use in the conative stage. The relatively newer devices, smartphone and tablet, evoke higher uncertainty since customers are less experienced in using them and they are characterized as less secure (Chandra, et al., 2010). It is expected that consumers are less price sensitive when using a PC during the final stage of the purchase funnel because of the higher perceived ease of use. Together with the previous reasoning that consumers are less price sensitive when information is easier to navigate it is expected that a PC will evoke higher spendings. The hypothesis that will be tested is:
H6: Sessions in the conative stage originating from a PC are likely to result in a higher order value compared to sessions originating from other devices.
3.3
Control variables
The aim of this research is to examine the effect of the different devices used across the purchase funnel on conversion and order value. The hypotheses for these effects are established in the previous chapter. Prior research on conversion and order value in retail settings refer to various other aspects that affect consumer behavior. To be able to measure a purer effect of the device choice across the purchase funnel on conversion and order value, various control variables are included. These variables are not the focus of this research but their existence has impact on the dependent variables, which cannot be ignored.
Multi-device usage
26 wireless connection in all kinds of situations (Shankar, et al., 2010). The customer has multiple opportunities to make an order and increase its spending. Customers can shop and modify their shopping cart during the additional sessions made available through wireless devices and Internet connections (de Haan, et al., 2015). Multiple devices provide greater convenience to the customers, increasing their purchase frequency. The customer can combine the benefits he received from the different devices and realize a larger value and increase its spending, creating synergies between devices. Since multi-device customers are likely to show different shopping behavior this study will control for it.
Experience Effects
Measures of customers’ prior shopping behavior are fundamental predictors of their future behavior. Prior research of Bult & Wansbeek (1995), Fader, et al. (2005) and Cheng & Chen (2009) summarizes this prior behavior in recency, frequency and monetary value, which is known as the RFM framework. The recency of the last purchase indicates how recently the customer bought from the retailer. Frequency indicates the number of prior purchases and the monetary value indicates the amount of money spent in previous transactions (Bult & Wansbeek, 1995). A shorter period since the last purchase, a higher frequency (Wu & Lin, 2005) and a higher monetary value (Cheng & Chen, 2009) increase the likelihood of a customer to buy again from the retailer, thereby increasing its probability of conversion.
Customers who are more loyal to the firm have less concern about their security as they trust the retailer. Habitual behavior is reinforced by prior behavior (Neal, et al., 2012), which indicates that experience in purchasing with a certain retailer leads to higher conversion probabilities for that retailer. Not only the conversion probability increases as the customer is more experienced, so does it increase order value (Wang, et al., 2015). It is thus expected that the more experienced the customer is by placing orders at a retailer, the higher the conversion rates and order value will be. Since these factors are commonly used to predict customer behavior they will be included as control variables in this study.
Customer Characteristics
De Haan et al. (2015) found different usage behavior for older customers. It is argued that older customers are less likely to have experience with the latest technology and are more inclined to use PCs. PewResearch (2012) finds that male customers are more active across devices, especially on tablets, compared to female customers. As these differences are likely to influence conversion and order value at an online retailer, gender and age are included as control variables.
Seasonality
27 after they received their ‘vacation loan’. Seasonality is likely to have an impact on the conversion and order value dependent upon the calendar. To account for these effects this study will control for seasonality effects.
Order content
28 3.4
Conceptual framework
The theories elaborated on in the literature review and the subsequent hypotheses that are formed are visualized in the following conceptual model. This model exhibits the relation between the device choice across the purchase funnel as the independent variables and the conversion and order value as the dependent variables. Moreover, the control variables are displayed.
Control variables Multi-device usage Experience effects Customer characteristics Seasonality Order content Conversion Order value Smartphone Tablet Personal Computer DEVICE USE
PURCHASE FUNNEL STAGE
Conative Affective Cognitive
29
4. Methodology
In this section, the research methodology is discussed. First, the choice for the Type-II Tobit model will be explained. Subsequently, the data collection procedure and measurement scales are described. A unique feature of this study is the development of a dataset including all sessions during the customer-journey in the product category originating from three Internet accessible devices, coupled with the purchase activity for these customers to create a single-source dataset.
4.1
The model
The purpose of the established model is to capture shopping behavior and to do so at the aggregate level of the customer journey. A parsimonious yet complete view of the decisions visitors face during a customer journey is given by two aspects of shopping behavior. The first aspect is the visitor’s decision to make a purchase or not. The second aspect is the same visitor’s decision on how much to spend, given that he converted. A natural model for this two-step process is the Type-II Tobit model (Fox, et al., 2004) (van Heerde, et al., 2008), which is also known as the two-step Heckman selection model. In this Type-II Tobit model, the dependent variable, order value, is only observable for a selection of the data.
A Type-II Tobit model provides two appealing features. First, it is well suited for the type of censoring observed in shopping data. When modeling conversion of visitors, a significant proportion of observations indicate zero use, that is, the visitor did not buy anything. The Type-II Tobit model accounts for these censored observations. The second appealing feature is that it integrates conversion and order value while allowing for either a positive or negative correlation of both elements.
The Type-II Tobit assumes a Probit regression for the binary variable conversion, the selection part, and a least square regression for the response equation of order value, which is observed only for those customers that make a purchase. By jointly modeling the two aspects, this model enables to test whether conversion decisions correlate with order value, and vice versa.
The Probit model for conversion is defined as
𝑦𝑗 = { 1 𝑖𝑓 𝑦𝑗∗> 0
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 Where 𝑦𝑗∗ is a latent variable constructed from a linear model as
𝑦𝑗∗= 𝑥
𝑗′𝛼 + 𝜀𝑗
Here 𝑦𝑗 indicates whether a conversion occurs in customer journey j, the first dependent variable.
The independent variables for the customer journey j are displayed by the vector 𝑥𝑗′, which potentially
30 The entire Type-II Tobit model can be formulated as:
𝑌𝑗 = 0 𝑖𝑓 𝑦𝑗∗= 𝑥
𝑗′𝛼 + 𝜀1𝑗≤ 0
𝑌𝑗= 𝑥𝑗′𝛽 + 𝜀
2𝑗 𝑖𝑓 𝑦𝑗∗= 𝑥𝑗′𝛼 + 𝜀1𝑗 > 0
The second dependent variable, order value, is given by 𝑌𝑗. The elements in α describe the effect
of the independent variables on conversion, whereas the elements in β indicate the effect of the independent variables on the amount spent in the customer journey. In principle α≠β, accounting for the fact that the explanatory variables may have other effects on conversion than on order value.
4.2
Data sources
For this study, a dataset based on individual-level clickstream data from a large European retailer is used. In concentrating on a visitor’s decision of converting or not and how much to spend in a session, this study is limited to those Internet users who already chose to go to the website and start to navigate it. Therefore this study perfectly mirrors the on-site effect of device choice.
Data preparation.
To be able to identify purchase funnels, this study focuses on a single product category. Therefore, the final dataset solely contains website visits that touched the chosen product category.
The data set is constructed by selecting all the customers that visited the category on the retailer’s website between December 1, 2014 and May 1, 2015. Subsequently, the entire customer journey of these customers is tracked. All sessions of these selected customers that touched the category in this period and three months after are included in the dataset. Finally, all customers that visited the category three months before or after the funnel observation period are withdrawn from the data to make sure the data contains only entire customer journeys. This approach assures that the start of a customer journey lies between December 1, 2014 and May 1, 2015 and that the journey ends before August 1, 2015. Customer journeys that last longer than 90 days are deleted as it is not likely that all these sessions belong to a single customer journey.
2014 2015 Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct
Figure 2 Timeline of the dataset
31
The final data sample comprises 67.555 identified customers who engaged in an equal amount of customer journeys containing 119.503 usable sessions. In this study, only data from registered customers is used, since it was not possible to capture multi-device usage of unregistered users. By tracking cross-device use, the independent sessions on different devices become related actions. However, by solely focusing on sessions in which the customer is identified, this study might miss many sessions that do belong the the observed customer journeys as well.
For each session, detailed information on which device a customer uses to visit the website, pages viewed, products bought and order value is available next to individual customer data. The data used in this study spans eight months, from December 1, 2014 to August 1, 2015. The session-level data is aggregated to the journey level using measures that are described in section 4.3. These measures are designed to describe the nature of the sessions within each customer journey.
4.3
Operationalization of key variables
Dependent Variables
This research contains two dependent variables. The first dependent variable is conversion, whether the purchase funnel ends with a purchase or not. This dependent variable is a dummy variable, indicating 1 when a conversion occurred and 0 otherwise.
The monetary value of the conversion is the second dependent variable, the order value. Consistent with other studies using the Type-II Tobit method to model customer spending (Fox, et al., 2004) (Singh, et al., 2006), the logarithm of the order value is taken. The logarithm transformation creates a distribution which is closer to normal than the distribution of order value, because of the skewed data of customer spending.
Categorizing Devices
32
Categorizing shopping sessions
The distinct sessions of the customer journeys will be assigned to the different stages of the purchase funnel, which are determined by Lavidge & Steiner (1961). These stages are also used in Wiesel, et al. (2011), where the actions of the customer on the retailer’s webpage are indicators of the stages of the purchase funnel. In this study a similar approach is followed in categorizing the shopping sessions.
The stages are determined by the pages a customer touches during a session. Sessions consisting of visits to a category page, shopping page and or a maximum of three product pages are considered to be of searching intent, and therefore considered cognitive. Sessions consisting of more than three product pages are considered to be in the affective stage. The customer is assumed to be reviewing alternatives when visiting more than three product pages. And finally a session is considered to be in the conative stage when during a session a product is carted or bought. In this study, sessions can only belong to one stage in order to conduct an empirical analysis. If a session fulfills the requirements of multiple stages, the latest stage in the customer journey counts.
Nine dummy variables indicating all the possible combinations of the device used and customer journey stage are included in the model to empirically analyse the effect of the devices across the purchase funnel.
Seasonality
In order to check whether a seasonality effect exists, a one-way ANOVA test is conducted to investigate whether a significant difference between the months on conversion and order value exists. The test on conversion shows significant differences between the months (F=2,259 , p<.01) and the test on the logarithm of order value indicates the same difference (F=2,997 , p<.05). The entire output of the ANOVA test can be found in Appendix I. In order to control for this seasonality effect, all months are included as dummy variables, where December is taken as benchmark.
Order content
The retailer examined in this study sells products of different brands. As there are large differences between the prices for the different brands, dummies are included to indicate the brand bought. An empirical test on the differences between the brands on logarithm of order value confirms that there is a significant difference between the brands (F=1040.391, p<.01). The output of the ANOVA test can be found in Appendix II.
33 The control variables in both parts of the model do not contain exactly the same variables. As the brand and the amount of products bought is only known after a conversion occurs these variables cannot be used to estimate the probability of conversion. The retailer cannot use this information to predict conversion. Therefore these variables only occur in the response equation. All other variables discussed before are equal for both parts.
A complete overview of the variables in the model can be found in table 1.
Variables
Dependent Variables
Conversion Dummy variable indicating an order made by a customer (conversion = 1) Ln Order value Logarithm of total euros spent by the customer on the order
Independent Variables
Mobile Cognitive Dummy variable indicating the use of a smartphone at the cognitive stage Affective Dummy variable indicating the use of a smartphone at the affective stage Conative Dummy variable indicating the use of a smartphone at the conative stage Tablet Cognitive Dummy variable indicating the use of a tablet at the cognitive stage
Affective Dummy variable indicating the use of a tablet at the affective stage Conative Dummy variable indicating the use of a tablet at the conative stage PC Cognitive Dummy variable indicating the use of a PC at the cognitive stage
Affective Dummy variable indicating the use of a PC at the affective stage Conative Dummy variable indicating the use of a PC at the conative stage Control Variables
Multi-device Dummy variable indicating whether a customer uses one or more devices (Multi-device = 1) Recency Number of days between the customers’ last conversion and the start of the purchase funnel Frequency Number of orders of the customer in the last year
Monetary value Total amount of euros spent by the customer over the last year Age Age in years at the start of the purchase funnel
Gender Dummy variable indicating the gender (male = 1)
Months December, January, February, March, April, May, June, July Brand* Brand 1-8
Multiple products* Number indicating the amount of products sold minus 1
*Only occurs in the response equation
34
5. Results
Initially some descriptive statistics are discussed to show the impact of the devices used on the purchase outcome. The majority of sessions originate from a PC. With the highest conversion rate and on average the highest order value, the PC outperforms the tablet and smartphone on these performance indicators. The smartphone and tablet occupy a more or less equivalent part of the sessions, however sessions originating from a tablet more often lead to a conversion and higher order values, which is in line with Chaffey (2015) who found that the conversion rate for fixed devices was the highest, followed by tablets and smartphones showing the lowest conversion rates.
When splitting up the sessions by device and the stage of the customer journey, it is shown that most sessions take place in the beginning of the customer journey, the cognitive stage. Sessions in the affective stage are the least common. A reason for this might be that customers use other sources as comparison websites to review alternative products. Again, the PC is the most commonly used device. In every stage, sessions originating from a PC occur two to four times as often as smartphones and tablets. As explained in the literature part, smartphones are often used for search activities and therefore create a lot of traffic to the website. However, in this study, the smartphone is used the least in the cognitive stage. This can be explained by the fact that the dataset only contains purchase funnels for a product category that might require extensive research, for which the smartphone is less convenient (Wang, et al., 2015). In the conative stage the difference between occurrence of mobile devices and PCs is the largest; this will probably be because of the perceived inconveniences of making payments on smaller devices.
The conversion rates on the journey level, represent the share of purchase funnels that result in an order. As journeys can consist of multiple sessions, the conversion rates for customer journeys are higher than for single sessions.
The large majority of the customer journeys in the dataset consist of a single device. Multi-device journeys do clearly show higher conversion rates and order values. The most successful journey contains all the three devices, resulting in the highest conversion rate and order value. Journeys containing a tablet have a higher conversion rate and order value than journeys containing a smartphone. Across funnels consisting of multiple devices, the tablet seems to have a stronger impact than the smartphone (Conversion rate & Order value Smartphone+PC < Conversion rate & Order value Tablet+PC) and the PC seems to have a stronger impact than the tablet (Conversion rate & Order value Smartphone+Tablet < Conversion rate & Order value Smartphone + PC).
5.1
Parameter estimates
35 whether a customer purchases a product in the product category. The order value column indicates the effect of each variable on order value.
Predictors of conversion
The positive and significant signs for all the dummies that indicate the device used and the phase of the customer journey indicate that all possible combinations have a positive effect on conversion.
When looking at the effects per purchase funnel stage it can be concluded that during the cognitive stage the smartphone has the strongest effect on conversion. This is contradicting H1, in which it was expected that the smartphone has the weakest effect on conversion. Online shoppers intent to purchase increases with the use of a smartphone in the cognitive stage compared to the tablet and PC (α2 = 0.506 > α5 = 0.223 > α8 = 0.155).
During the affective stage, the smartphone has the largest influence on conversion as well, compared to the other devices tested (α3 = 0.660 > α6 = 0.438 > α9 = 0.318). This is quite surprising as it was expected that the smaller screen of a smartphone is less convenient in evoking emotional responses. H2 can therefore not be supported. Instead, the tablet is second best in evoking conversion in the affective stage. In this stage the strength of all devices are quite close to each other, indicating that the device used does not make a large difference. In the conative stage the PC has the strongest effect on conversion (α10 = 2.694 > α7 = 2.445 > α4 = 1.860), which is in line with H3. In turn, the smartphone has the weakest influence on conversion in the conative stage. As it is explained in the literature, the larger screen size and convenience of the PC are most likely the reason for the highest parameters.
Conversion Order value
36 February α18 -0.127*** β18 -0.029*** March α19 -0.129*** β19 -0.011*** April α20 0.028*** β20 -0.017*** May α21 -0.550*** β21 0.051*** June α22 -0.418*** β22 -0.035*** July α23 -0.095*** β23 -0.072*** Brand 1 β24 -0.058*** Brand 2 β25 0.931*** Brand 3 β26 0.129*** Brand 4 β27 -0.197*** Brand 5 β28 -0.200*** Brand 6 β29 1.054*** Brand 7 β30 -0.396*** Brand 8 β31 -0.075*** # of products β32 0.447***
*Significant at the 10% level **Significant at the 5% level *** Significant at the 1% level
Table 2 Parameter estimates for the Type-II Tobit model
Even though it is not hypothesized it is also interesting to explore the effects on conversion per device. All the devices show a stronger effect on conversion towards the end of the journey, which is reasonable. However, the difference between the beginning and the end of the journey is especially large for the PC and to a lesser extent for the tablet as well. When using a PC or tablet it is very easy to go to another website. In online environments switching costs in the beginning of the purchase funnel are quite low; competitors are just one click away from the retailer’s own website.
Predictors of order value
The second panel of table 2 shows coefficients for the continuous component of the model, that is, how much the consumer will spend, given that he purchases a product. As can be seen in table 6, significant estimates can only be found in the affective stage. It can therefore be concluded that the tablet has the largest influence on order value, followed by PC in the affective stage. All the other possible combinations of devices and stages do not significantly contribute to the order value. This finding partially supports H5. Due to insignificant estimations H4 and H6 cannot be supported.
37
Differences between conversion and order value
All dummies are found positive in the conversion part of the model. However, only the tablet and PC in the affective stage show significant effects on order value. This study shows that the devices used across the different stages of the purchase funnel do influence conversion, however this effect is not found for order value. This is in line with Avent & Higgins (2006) who state that the way in which a decision is made should not affect the decision on monetary value.
Control variables
The negative sign for multi-device usage seems counterintuitive as literature explains multi-device users to have a higher purchase frequency. However, the negative sign of multi-device users indicates that the device user has a lower initial conversion probability than a single device user. The multi-device user touches multiple multi-devices in its purchase funnel and thereby increases its conversion probability by every device he uses, as all the dummies are positive and significant. Additional proof for this reasoning can be found in Appendix IV, in which all the different types of journeys are taken as independent variables. It shows that the multi-device journeys have higher conversion probabilities than single-device journeys (α2, α3 < α4 – α7).
Recency, frequency and monetary value all have positive significant effects on conversion. As was explained by literature, more loyal customers have higher conversion probabilities. Surprisingly, recency shows a positive sign; even though it is small, it says that the longer time ago since a purchase, the higher probability of conversion. The positive sign for recency can be explained by the fact that the retailer primarily sells products which are not bought on a regular basis. Therefore a higher recency might indeed positively effect conversion. The RFM-framework only shows significant effects on order value for frequency and monetary value. Frequency shows a negative effect on order value, indicating that customers who buy more often at the retailer spend less money. This is surprising as more loyal customers are expected to spend more money. However, customers who buy more often from the retailer might also be more familiar with the retailer’s promotional activities and understand how to take the benefits of promotions and discounts.
Interesting effects are found for age and gender. Age is found to have a significant effect on conversion but not on order value. On the other hand, gender solely shows a significant effect for order value. It can therefore be concluded that men are not more likely to buy than women, but when they do, they spend more. Additionally, older customers are more likely to make a purchase but do not spend more than younger customers.
38 As is expected, all brands show significant effects on order value. The size and rank of the parameters is in line with the average prices of the different brands.
5.2
Validity
The performance of the selection part of the Type II Tobit model is evaluated by computing a hit rate. The hit rate indicates the percentage of observations for which the predicted value of the model is in line with the true value. In the estimation sample, a hit rate of 89,26% is achieved. The selection equation corrected 13,49% of the incorrect predictions by the default specification which holds constant probability. The highest amount of correctly predicted journeys, 95.75%, comes from journeys which did not convert, which is extremely high. Conversion journeys are well predicted in 43.49% of the cases. Even though the model predicts a relatively small part of conversions correct, a hit rate of 89,26% indicates good performance. Table 3 shows the predicted and observed conversions. Overall, it can be concluded that the model has a fairly good predictive power.
Hitrate
Observed
Predicted Conversion = 0 Conversion = 1 Total Conversion = 0 50993 4264 55257
Conversion = 1 2263 3281 5544
Total 53256 7545 60801
%Correct 95.75 43.49 89.26
%Incorrect 4.25 56.51 10.74
Table 3 Hitrate of the selection equation
Additionally, the McFadden R-squared of the chosen model is evaluated to estimate whether the predictors are effective. Or in other words, whether the dummies that indicate the device used and the phase of the purchase funnel contribute positively to the model. The estimated model will be compared to two non-nested models; one that does not contain independent variables at all and a model that only includes the devices used and not the stages of the purchase funnel. Table 4 shows that the estimated model performs extensively better than the two other models. Not only the McFadden R-squared is considerably higher, the AIC and BIC, which penalize for more variables, decrease as well. As the model with the lowest values for AIC and BIC is the preferred model it can be concluded that the estimated model outperforms the non-nested models. Adding the dummies that indicate the device used and the stage of the purchase funnel improve the model.
Validity measures
Chosen model Devices solely as independent variable No independent variables
McFadden R-squared 0.477 0.024 0.014
AIC 0.393 0.733 0.740
BIC 0.396 0.735 0.742
39 The validity of the second part of the model is tested by a random holdout sample of 10%, including 6755 customer journeys. Using the parameter estimates from the second panel of table 2, the logarithm of order value is predicted for the customers of the holdout sample. The predictive validity measure that is used to make estimate the forecast accuracy is the Mean Absolute Percentage Error (MAPE). This validity measures computes the relative error to the actual value (Leeflang, et al., 2014). The MAPE for this model is 5.15%, indicating that the predictive values in the holdout sample on average have an error of slightly more than five percent. This is a very low error rate and therefore it can be concluded that the model gives accurate predictions.
The overall fit of the entire model is also significant (Wald 7060.12, p<.01), indicating that the measurement model fits the data well.
5.3
Robustness check
To ensure that the results are robust in the presence of potential errors in the data a sampling robustness check is conducted. As the data only consists of identified customers a potential concern is that the entire dataset may be subject to macrosampling errors such as oversampling the single device users. This, in turn, can bias the coefficient estimates. To that end, a subsample of the data is constructed which only contains customers who visited the retailer by two or more devices. Subsequently, this dataset is analysed in the same manner as the entire dataset.
Conversion Order value
40 June α21 -0.531*** β21 -0.060*** July α22 -0.045*** β22 -0.008*** Brand 1 β23 -0.077*** Brand 2 β24 0.888*** Brand 3 β25 0.103*** Brand 4 β26 -0.178*** Brand 5 β27 -0.190*** Brand 6 β28 0.821*** Brand 7 β29 -0.382*** Brand 8 β30 -0.068*** # of products β31 0.375***
*Significant at the 10% level **Significant at the 5% level *** Significant at the 1% level
Table 5 Parameter estimates for robustness check
Table 5 shows that the results are robust to the subsample. Overall, the parameters for the independent variables in the selection part remain qualitatively the same in terms of sign and significance. The dummies exhibited the same patterns as those derived from the full data set. An interesting observation from table 5 is that as we move from the original sample to the multi-device sample, the magnitude of coefficient estimates for conversion decreases. This is not very surpising as the negative sign of multi-device usage in table 2 is spread out over all the dummy variables. Additionally, it might suggest that multi-device users are less device sensitive as the conversion probabilities are lower.
The significant effects on order value that were found in table 2 are not shown in table 5. The positive effect of the tablet and PC in the affective stage are likely to be solely for the single device users. However caution is needed to draw conclusions about the difference between multi-device and single-device users, as this test does not significantly test the differences. More research is needed to make these statements. Overall, the control variables remain equal, solely for recency, frequency, February, July and Brand 8, parameters become insignificant. This might be a result of a smaller data set. It can be concluded that the results found in table 2 are robust to the subsample. Underlined parameters in table 5 show parameters that changed became insignificant.
5.4
