How device switching aﬀects customer purchase behavior: An investigation in multichannel context

(1)

purchase behavior: An investigation in

multichannel context

Author: N.F. Lagerweij

Student number: s2527383

(2)

University of Groningen

Supervisor: dr. H. Risselada

(3)

behavior: An investigation in multichannel

context

Norbert F. Lagerweij

Januari 2020

Abstract

Customers frequently switch between multiple devices in the online journey prior to purchasing. Besides the growth of cross-device usage, customers also utilize both the online platform and the brick-and-mortar store on the path to purchase. In this study, we examine the relationship between device switching and online and offline conversion rates. We find that switching from a smartphone to a tablet or PC leads to significantly higher online conversion rates, whereas switching from a smartphone to a PC also results in lower offline purchase probabilities. We show that these switching effects differ with the extent in which customers have previously purchased in the online and offline channel. These findings illustrate the importance of focusing on the combination of the online and offline channel when analysing device switching effects.

(4)

1 Introduction 4

2 Conceptual framework 6

2.1 Device switching effects . . . 8

2.2 Moderating effects . . . 10 2.3 Control variables . . . 12 3 Data description 13 3.1 Online data . . . 13 3.2 Offline data . . . 14 3.3 Combined data . . . 14 4 Methodology 16 4.1 Propensity score weighting . . . 17

4.2 Balancing results . . . 20

4.3 Model specification . . . 22

5 Results 23 5.1 Online conversion models . . . 24

5.2 Offline conversion models . . . 26

5.3 Robustness check . . . 30 6 Conclusion 30 References 33 Appendices 37 A Descriptive statistics . . . 37 B Balance statistics . . . 38

C Overlap of propensity scores . . . 39

D Estimation results (continued) . . . 41

(5)

(6)

1 Introduction

The growth of the Internet has influenced customer purchase behavior and the advertising landscape at a rapid pace. Ever since the first online banner advertisement in 1994 (Kaye and Medoff 2001), the digital advertisement spend in the United States has grown to $88 billion in 2017 (Interactive Advertising Bureau 2018). This substantial market could only have developed due to the billions of individuals who adopted the Internet over the past decades. Due to the endorsement of the Internet, customers are now offered a choice of whether to buy their products online or offline. Nevertheless, it has appeared that many customers use both channels on their path to purchase. The “research shopper” phenomenon, proposed byVerhoef et al.(2007), describes customers who do their research in one channel (e.g., an e-commerce website), but purchase the product through another channel (e.g., a brick-and-mortar store). The foundation for this phenomenon has its roots in the perceived risks and benefits associated with the online and offline channel. In numerous of these “online vs offline” studies (e.g.,Forman et al. 2009;Herhausen et al. 2015), the risk of the online channel is treated in a singular context without differentiating between the devices that are used to access the online channel. Research has shown, however, that the devices used for online shopping differ in terms of perceived security risk (Chin et al. 2012). The decision whether to purchase online or offline therefore also depends on the device that is being used by the customer. As a result, it is essential to include device type when analysing customer channel choice and purchase behavior (Singh and Swait 2017).

(7)

(Xu et al. 2016). The customers that use a tablet or PC as a subsequent device, however, show different behavior than customers using a tablet or PC from the outset. The former having engaged in a larger number of sessions prior to buying (Wang et al. 2015) and a higher conversion rate (de Haan et al. 2018) than the latter. Hence, in addition to the type of device, also the order in which devices are used affects customer behavior.

Although cross-device usage has been proven to influence customer actions, academic literature has yet to develop a broad theory of how this mechanism affects purchase be-havior. While there is limited research regarding device switching effects in an online context (e.g.,de Haan et al. 2018;Xu et al. 2016), there exists, to our knowledge, no em-pirical research which also includes the offline channel in its research design. Therefore, we do not yet know the full extent of the effects of device switching on total purchase behavior. Thereby leaving retailers and academics in the dark about the complete mag-nitude of device switching effects on online and offline sales.

The aim of this thesis is therefore threefold. First, building on the findings ofde Haan et al. (2018), we investigate the magnitude of several specific device switching effects on the online conversion rate. That is, we examine to what extent switching from a mo-bile device to a tablet or PC affects the online conversion rate. Second, we examine to what degree cross-device usage has an impact on offline purchase behavior, i.e., do offline conversion rates change if consumers switch between devices during their online journey prior to buying offline. Third, we investigate to what extent prior channel experience influences device switching effects on purchase behavior. That is, we investigate whether device switching effects change with the extent in which customers have previously pur-chased in the online or offline channel.

(8)

becomes stronger if the customer is more experienced with the offline channel, whereas the effect reduces if the customer has more online purchase experience with the retailer. Our results also imply that the switching effects on offline purchase probabilities dimin-ish as online or offline experience grows. From a managerial perspective, this thesis can grant practical insights about budget allocation decisions across multiple device types in consumers’ path to purchase. The findings can thereby increase understanding of the role of customer channel experience in these budget allocation decisions. As a result, this study enhances our general understanding of cross-device usage in a multichannel purchase context.

The remainder of this study is structured as follows. In Section 2 we describe the conceptual framework and we state our hypotheses. Section 3 gives an overview of the online and offline data used in our model estimation. Section 4 describes and evaluates our propensity score methodology and specifies our empirical model. In Section 5 we interpret the results and Section 6 concludes.

2 Conceptual framework

The online-offline competition in marketing and retail has been the subject of numerous studies in the past quarter-century (e.g., Avery et al. 2012; Brynjolfsson et al. 2009;

(9)

brick-and-mortar stores (Brynjolfsson and Smith 2000;Ghose et al. 2006).

Despite these advantages, the online channel also comes with a number of drawbacks compared to the traditional offline store. Regardless of the greater selection of products, the online channel faces the issue of product inspection. Offline stores allow customers to feel and observe the product in reality, while e-commerce websites only provide a digital inspection. Another drawback of the online channel is the involved transportation time (Forman et al. 2009). Consumers that buy products in an offline store have immediate access to their purchase, whereas products ordered through the online channel include a shipping period.

Due to these different characteristics, the online and offline channel used to be seen as vastly different. This distinction in scope resulted in businesses being active in either the online channel or the offline channel. Nowadays, the channels are seen as more or less integrated. Studies have also shown that using both channels in synergy is more effective than using the channels in isolation (e.g., Herhausen et al. 2015; Weinberg et al. 2007). That is, customers that use multiple channels spend on average more than single-channel shoppers (Weinberg et al. 2007) and firms that are active in both channels have greater perceived service quality (Herhausen et al. 2015). Important arguments for this online-offline synergy are the “showrooming” and “research shopper” phenomena. That is, doing research in the offline channel and buying online (showrooming), or doing research online and buying offline (research shopper). One of the key factors affecting customer channel choice is perceived risk (Andrade 2000). The concept of risk was introduced to the marketing field by Bauer(1960) as: “Consumer behavior involves risk in the sense that any action of a consumer will produce consequences which he cannot anticipate with anything approximating certainty, and some of which at least are likely to be unpleasant.” The online channel and all devices used to access it, are perceived as more risky in comparison to the offline channel. Consumers are therefore more likely to shop online for items that correspond to low purchase risk or items with a well-known brand (Lee and Tan 2003).

(10)

Online conversion Offline conversion Device Switching (from mobile device to tablet/PC) Offline experience with retailer Online experience with retailer Control variables

Figure 1. Conceptual model. −∗indicates negative moderating effect found byde Haan et al. (2018).

et al. 2012) or the risk of battery failure during a transaction (Cao et al. 2015). As a result, customers often use a mobile device only for habitual purchases, but switch to a different device (e.g., a tablet or PC) as the purchase becomes more complex (Wang et al. 2015). Extant literature has also shown that the tablet acts as a complements to mobile devices (Xu et al. 2016). This pattern of cross-device usage results in enhanced sales outcomes (de Haan et al. 2018;Xu et al. 2016).

Using the above literature as building blocks, we have developed the conceptual model depicted in Figure1. In the next subsections we elaborate on the proposed relationships defined in the model and define our hypotheses. These include the main effects of device switching on online and offline conversions. Furthermore, we also discuss the hypothesized moderating effects of online and offline retailer experience. We conclude this Section with a discussion of the controls used in our model.

2.1 Device switching effects

(11)

websites are still made with either a PC, tablet or mobile device. While all these devices can be used to access the online channel, they have distinctive features. Mobile devices have small dimensions, which allows consumers to bring it along to wherever they go. The small size combined with the wide accessibility of the Internet allows the consumer to be online “anywhere, anytime” (Shankar et al. 2010). This mobile convenience leads to purchase intentions as long as customers want to satisfy habitual needs which do not require much search effort (Wang et al. 2015).

Less mobile devices, such as tablets, are less convenient in terms of mobility due to their greater size. Unlike mobile devices, tablets are often not equipped with all-day telecom connections and are often dependent on Wi-Fi. As a result, tablets are often used at home or at places with a stable Wi-Fi connection. Also PCs are considered less convenient than mobile devices due to their greater size and limited access to the Internet. However, tablets and PCs do have larger screens, which provides more comfort in terms of comparing alternatives when shopping in an online environment. Furthermore, as argued byChin et al.(2012), PCs are perceived as less risky in terms of purchase risk, compared to mobile devices. That is, PCs often provide a wider bandwidth, better battery power and connection stability compared to mobile devices (Cao et al. 2015). Also, as tablets are often used at home, the internet connection and stability may provide a more secure environment to purchase than on a mobile device, which is often used on the go.

(12)

H1: The online conversion rate is higher for customers that switch from a mobile

device to a (a) tablet or (b) PC, compared to customers that did not switch between devices.

Customers have appeared to differ in terms of risk perception of the online channel (Chin et al. 2012). That is, one customer may perceive the online channel as very risky, and may therefore prefer to buy in a brick-and-mortar store, while another customer has no issue with shopping online. Customers that start their online journey on a mobile device may therefore also have different strategies to reduce purchase risk. Whereas some customers switch to a tablet or PC as a means to reduce purchase risk (de Haan et al. 2018), others may move to the offline store to finalize their purchase. By switching from a mobile device to a tablet or PC, customers that usually shop offline may now feel comfortable enough to finalize the transaction in the online channel. That is, customers that usually purchase offline after browsing on their mobile device, may now purchase online due to the purchase risk reduction resulting from switching. As a result, offline purchase probabilities may drop for the customers that switch from a mobile device to a tablet or PC. Hence, we posit that:

H2: The offline conversion rate is lower for customers that switch from a mobile

device to a (a) tablet or (b) PC, compared to customers that did not switch between devices.

2.2 Moderating effects

(13)

has perceived high risks in the past and may therefore prefer to shop through the offline channel as a means to reduce purchase risk. These customers with higher risk perceptions of the online channel, as argued byChin et al. (2012), also have a greater need to reduce risk. That is, these customers are used to purchase their goods in the offline channel, which corresponds to low purchase risk. Hence, in case of online shopping, the need to switch from a mobile device to a tablet or PC in order to reduce purchase risk may also be greater. Therefore, the customers that have more experience with the offline channel of the retailer may have a greater advantage of switching in order to finalize a purchase. The effect of switching from a mobile device to a tablet or PC on online conversions may therefore be stronger if the customer is more experienced with the offline channel. Hence we posit that:

H3: The observed higher online conversion rate from switching from a mobile device

to (a) a tablet and (b) a PC is even larger if the customer has more offline experience with the retailer.

Followed by the argument above, offline experience should therefore also increase the negative effects of device switching on offline conversions. That is, customers may purchase even less in the offline channel after switching from a mobile device to a tablet or PC as offline experience increases. Due to the greater level of expertise resulting from more offline experience (Alba and Hutchinson 1987), the gap between the online and offline channel in terms of purchase risk reduces. Hence, the level of purchase risk that a customer is willing to accept before making an online purchase, may be reached at an earlier stage in the customer journey. This may lead to more customers purchasing online, who otherwise would have purchased in the offline channel. The negative effect of device switching on the offline conversion rate may therefore also become more negative as offline experience grows. Hence, we hypothesize that:

H4: The observed lower offline conversion rate from switching from a mobile device to

a (a) tablet or (b) PC is even larger if the customer has more offline experience with the retailer.

(14)

as online experience increases, the difference in terms of perceived purchase risk between the online and offline channel also decreases. Whereas inexperienced customers have a need to reduce purchase risk by switching to a less mobile device, this need becomes smaller as the customer becomes more experienced with the online channel. Therefore, the customers that were first convinced to purchase online due to the switching effects, the more online experienced customer may perceive the online channel as more or less equal to the offline channel in terms of purchase risk. As a result, the effect of switching from a mobile device to a tablet or PC on offline purchase rates may become less substantial as online experience grows. Hence we posit that:

H5: The observed lower offline conversion rate from switching from a mobile device

to a (a) tablet or (b) PC is smaller if the customer has more online experience with the retailer.

2.3 Control variables

To obtain reliable estimates for device switching effects, we also include several control variables in our model. Specifically, we control for the time between subsequent sessions. Previous sessions on a mobile device will most likely result in an information growth by the customer. However, as time passes, the information gain will most likely decrease as the human brain will slowly forget the previously collected data. By forgetting informa-tion, the advantage from switching from a mobile device to a tablet or PC also decreases (de Haan et al. 2018). Besides the subsequent time between two sessions, we also account for the duration of the previous session. The duration of the previous session increases the amount of information the customer has at the start of the next session. Thereby also influencing the experience and therefore expertise of the customer, which may influence purchase behavior in the subsequent session. We also control for the number of pages viewed in the current session, as that leads to higher online conversion probabilities (Moe 2003). Furthermore, we also account for the day of the week. Several studies have shown that the proportion of device usage differs throughout the week (e.g.,de Haan et al. 2018;

(15)

a different age, since age may influence technology adoption (Bigne et al. 2005). Also gender is included as a control, because males and females have appeared to differ in their attitudes towards certain devices (Konu¸s et al. 2008).

3 Data description

To examine device switching effects on purchase behavior, we use data from a large international firm in the personal care and cosmetics sector. Given the multichannel context of our research, we use two types of data. We use clickstream data for the online component and brick-and-mortar store purchase data for the offline part. Both datasets span the period from April 1, 2019 to October 31, 2019. Besides the firm’s e-commerce website, the company has brick-and-mortar stores across the whole Western European country of which the data was used in this study. This is essential in our research design, as customers should have the possibility to purchase in either channel.

3.1 Online data

(16)

observe if a customers has switched between devices in subsequent sessions.

While this use of data is in line with the approach of de Haan et al. (2018), we also briefly note the shortcoming of our data. While it is mandatory to be logged-in to finalize a purchase, it is likely that in some cases our data does not capture the complete path to purchase. That is, customers may not be logged-in during all the sessions prior to a transaction. For example, suppose that a customer creates an account in the third session. We can then trace back each subsequent session in which the customer was logged-in to that specific individual. The first and second session, however, cannot be traced back to that particular customer. Hence, if the customer switched between devices from the second to the third session, we are not able to detect this. While we are aware of this pitfall in our data, there is, to our knowledge, no extant research regarding the impact of this identification issue or an available solution for that matter.

3.2 Offline data

For the offline component of our study, we use transaction data of brick-and-mortar stores. We only use transactions in which a membership card of the firm was scanned at payment, as this allows us to identify customers across the online and offline channel. That is, the membership card is linked to the same email address as the customer’s online account. This allows us to identify which devices were used to browse the retailer’s website prior to purchasing offline.

This procedure of offline data selection, however, also comes with certain limitations. That is, only transactions of loyalty program members can be traced back to online sessions. Therefore, the offline data could be an unrepresentative sample of the true underlying population, as members of a loyalty program have been shown to have in-creased expenditures and purchase-frequency compared to non-members (Dorotic et al. 2012;Meyer-Waarden 2008). Therefore, this should be taken into account when drawing general conclusions from the results regarding offline purchase behavior.

3.3 Combined data

(17)

session. Hence, in the final data set we only include offline transactions of those consumers who have also browsed the retailer’s website prior to buying in the offline store. This allows us to have a complete overview of the devices used on the path to purchase.

The final data set then contains 200,972 usable sessions of 49,441 unique customers. As can be observed in Table1, mobile devices are the most frequently used device (57.2%), followed by the PC (33.2%) and the tablet (9.6%) accounts for the smallest number of sessions. When we look at the activity per device, we see that the duration, which denotes the time from the beginning of the session until termination, and the pages viewed per session are the lowest for sessions on a mobile device followed by tablets and then PCs. It follows that customers spend more time browsing on less mobile devices. This distribution of activity per device in a purchase context is in line with other studies (e.g.,de Haan et al. 2018;Raphaeli et al. 2017). When we look at the device used in the previous session, we see that on average most people do not switch between devices in subsequent sessions. More specifically, customers on a mobile device are the most likely to also use a mobile device in the next session. Tablet users most frequently switch to a different type of device, whereas still 82.5% of the previous sessions were also on a tablet.

When we analyse the online conversion rate per device category, we find substantial differences. More specifically, the online conversion rate is the highest on PCs, followed by the tablet and then mobile devices. This is also in line with data from other studies (e.g.,de Haan et al. 2018). When we analyse the conversion rate taking into account the device used in the previous session, we find that conversion rates increase if the customer switches between devices. Especially the conversion rates for sessions on tablets differ if one assesses the device used in the previous session. This indicates that device switching may effect purchase behavior, in line with H1.

(18)

Table 1. Data description per device category

Device category

Mobile Tablet PC

Sessions

Number of sessions 115,077 19,272 66,623

Previous session on mobile 90.8% 10.8% 11.6%

Previous session on tablet 1.9% 82.5% 2.3%

Previous session on PC 7.3% 6.7% 86.1%

Page views per session 7.4 9.2 9.8

Duration per session (sec.) 334 463 476

Conversions

Online conversion rate 7.4% 13.2% 17.8%

Previous session on mobile 7.2% 22.2% 25.1%

Offline conversion rate 12.3% 11.1% 9.3%

Previous session on mobile 12.5% 11.2% 8.5 %

PCs are the highest if the same device was used in the previous session. This indicates that switching between devices is negatively related to offline purchase rates.

4 Methodology

(19)

self-select the order and the type of device when browsing the Internet. Consequently, the customers that switch between devices and those who do not, may systematically differ from each other. This difference could result in biased estimates of device switching, since the sessions of both groups are not comparable. In order to increase comparability of the sessions, we apply the inverse probability of treatment weighting (IPTW) method ofMcCaffrey et al. (2013). While there are several alternative methods that address the issues of possible confounding in observational studies, such as regression adjustment or matching on the propensity score (Rosenbaum and Rubin 1983), the method set out by

McCaffrey et al.(2013) provides tools for dealing with multiple treatment groups simul-taneously. Furthermore, most alternative methods for estimating propensity scores are parametric-based and their performance is therefore strongly affected by proper variable selection (McCaffrey et al. 2013). The method ofMcCaffrey et al. (2013) overcomes this issue, as it yields propensity scores in a non-parametric framework. In the next subsec-tions we will first discuss this balancing procedure and its results and then specify the model that willl allow us to test our hypotheses.

4.1 Propensity score weighting

The treatments in the present study, which we denote by Ti,s, are the possible device

switching combinations as described in Table2. That is, Ti,s= 1 indicates that customer

i used a mobile device in session s − 1 and a tablet in session s, Ti,s = 2 indicates that

customer i used a mobile device in session s − 1 and a PC in session s and so on. We use these treatment definitions, as it allows us to measure the effect of switching from a mobile device to a tablet or PC compared to not switching, while controlling for other potential switching effects (i.e., the treatment with Ti,s= 3).

Table 2. Treatment characteristics

Ti,s Treatment name Description

1 MobileToTablet From mobile device in s − 1 to tablet in s

2 MobileToPC From mobile device in s − 1 to PC in s

3 OtherSwitch Other combinations of two different devices in

s − 1 and s

(20)

The goal of the IPTW method is to make the sessions with these different treatments as comparable as possible. Procedurally, we first estimate the generalized propensity scores for each session s, based on a set of pre-treatment covariates. These generalized propensity scores are conditional probabilities that a customer will use a device sequence according to treatment Ti,s = t, where t ∈ T = {1, 2, 3, 4} and T denotes the set of

possible outcomes of treatment Ti,s. More formally, let the generalized propensity scores

of individual i in session s be defined by Pi,s(xi,s) =

p1,i,s(1, xi,s), p2,i,s(2, xi,s), p3,i,s(3, xi,s), p4,i,s(4, xi,s)

, (1)

where

pt,i,s(t, xi,s) = Pr Ti,s= t|xi,s

(2) denotes the propensity score of treatment t of individual i in session s and xi,s contains

the observable covariates: online and offline purchase experience with the retailer (ex-perienceOnline and experienceOffline), demographics of the customer (age and gender), the average price of products viewed in previous sessions (productPrice), days since the previous session (recency), duration of the previous session (durationLag), pages viewed in the current session (pages) and the day of the week (day) as described more exten-sively in Appendix A. We use these variables as they contribute to explain differences between customers and corresponding sessions and indicate something about the stage a customer is in during their path to purchase. Furthermore, these covariates were also used byde Haan et al. (2018) to estimate the propensity scores for switching to a more or less mobile device, which is similar to our goal.

We estimate the generalized propensity scores given in (1) as follows. First we create an indicator variable zt,i,s for each treatment, where zt,i,s is equal to 1 if person i has

shown device switch t in session s and 0 otherwise (e.g., z1,i,s = 1 if person i switched

from a mobile device in session s − 1 to a tablet in session s and 0 otherwise). Then, we fit a generalized boosted model (GBM) to each treatment indicator, using the covariates in xi,s as input, to obtain the estimated propensity scores for each of our treatments.

To improve the readability, we explain the following estimation procedure in terms of a single treatment that is indicated by zi,s, where zi,s= 1 if customer i received treatment

(21)

first replace zi,sby z1,i,sto obtain the estimates for p1,i,s(1, xi,s), then we repeat the same

procedure but we use z2,i,sas the treatment indicator to get the estimates for p2,i,s(2, xi,s)

and so on. We then combine all the propensity scores estimates to obtain an estimate of the generalized propensity scores for each session.

Using GBM to estimate propensity scores, as extensively described by McCaffrey et al.(2004), involves an iterative process to estimate the possibly complex and nonlinear relationship between treatment assignment and a large set of pre-treatment covariates. That is, GBM models the log odds of a specific treatment assignment, denoted by g(xi,s),

which is defined by g(xi,s) = log p(xi,s) 1 − p(xi,s) , (3)

where p(xi,s) denotes the propensity score of the treatment. We use log ¯z/(1 − ¯z) as a

na¨ıve initial value for the log odds, where ¯z ∈ [0, 1] denotes the proportion of sessions in the entire sample that received treatment. Then, the algorithm searches in an iterative fashion for a small alteration hm(xi,s) to add and thereby improve the fit of the model to

our data. The iterative adjustment hm(xi,s) is a regression tree that models the residuals

from the current fit. The basic idea behind a regression tree is that it breaks down a dataset in subsets based on the values of the covariates in the data (seeBreiman et al.

(1984) for a detailed explanation of classification and regression trees). Hence, at every iteration m, we model the residuals hm(xi,s) = zi,s− ˆpm−1(xi,s), where

ˆ

pm−1(xi,s) =

1

1 + exp(−ˆgm−1 xi,s)

(4)

is the estimate of the propensity score of the most recent fit (i.e., at iteration m − 1) and ˆgm−1 xi,s) denotes the estimate of the log odds at m − 1. Then, for all the sessions

belonging to node k = 1, . . . , K of the regression tree, denoted by Ak, we calculate

hk,m(xi,s) =

P

xi,s∈Akzi,s− ˆpm−1(xi,s)

P

xi,s∈Akpˆm−1(xi,s) 1 − ˆpm−1(xi,s)

, (5)

where hk,m(xi,s) denotes the adjustment of node k at iteration m. Then, we calculate

the fit of ˆgk,m(xi,s) = ˆgm−1(xi,s) + hk,m(xi,s) at each node, using the following Bernoulli

log-likelihood L ˆgk,m(xi,s) = N X i=1 S X s=1

zi,sgˆk,m(xi,s) − log

1 − exp ˆgk,m(xi,s)

(22)

where large values of the log-likelihood imply a better fit. The model then determines which of the K adjustment hk,m(xi,s) provides the largest increase in fit. The log odds

ˆ

gm−1(xi,s) is then updated with the partitioning of node k that offers the largest

in-crease in fit and the model iterates. The model will stop when a pre-specified number of iterations is reached. It values the balance between the covariates in xi,s across the

treatment and control (non-treatment) group at each iteration. It then selects the model of the iteration that results in the best balance across both groups and computes the corresponding propensity scores of that model. We use two separate measures to define balance at each iteration: the average difference (AD) and maximum difference (MD) of the standardized absolute mean of covariates. That is, for each covariate we calculate the absolute value of the difference in mean between the treatment and the control group divided by the standard deviation of the treatment group. Then, the average and maxi-mum of this measure across all the covariates of that iteration is computed. The model then selects the iteration that minimizes the AD or MD.

The final GBM model is then a sum of many regression trees. In our computations we use 17,500 regression trees for each of the four propensity score estimation procedures. After estimating the generalized propensity scores for each session, we compute the weight of each observation by taking the inverse of the propensity score of the actual treatment received by the customer. That is, let the weight of session s of customer i be defined by

wi,s=

1 P4

t=1zt,i,spˆt,i,s(t, xi,s)

, (7)

where ˆpt,i,s(t, xi,s) denotes the estimated propensity score of treatment t and, evidently,

P4

t=1zt,i,s = 1, since each session can only have one treatment. As a result, we obtain

two vectors with weights: one based on the AD balance criterion and one based on the MD criterion. Both weights should balance the group receiving treatment t and the entire sample (McCaffrey et al. 2013).

4.2 Balancing results

In order to assess whether our balancing procedure was successful, we compare the co-variate balance before and after weighting, and evaluate the overlap of propensity scores. We do so for both sets of weights, as this will allow us to analyse to what extent the two balancing methods differ.

(23)

defined by ¯ xc,t= PN i=1 PS

s=1xc,i,szt,i,swi,s

nt

, (8)

where nt denotes the total number of sessions with treatment t. For each covariate we

compute the pairwise standardized mean differences (SD) between all treatment groups as follows SDc,t,r= ¯ xc,t− ¯xc,r δc , (9)

where ¯xc,t denotes the mean of covariate c of the sessions with treatment t, r = 2, 3, 4

denotes the comparison treatment group with r > t and δcdenotes the standard deviaton

of covariate c in the entire sample. This measure allows us to compare differences between treatment groups relative to the entire population. Then, in order to assess balance, we use the maximum of the absolute standardized mean difference per covariate (MASDc),

as described byLopez and Gutman(2017), which is defined by

MASDc= max

|SDc,1,2|, |SDc,1,3|, |SDc,1,4|, |SDc,2,3|, |SDc,2,4|, |SDc,3,4|

. (10)

Using MASDcallows us to analyse the largest discrepancies in covariates across all

treat-ment groups. That is, the value for MASDc measures the largest mean difference in

covariate c between all treatment groups (i.e., the ”worst-case” scenario in terms of im-balance (Lopez and Gutman 2017)).

In order to investigate whether our propensity score methodology was successful in balancing the data, we first compare the MASDc of each covariate before and after

weighting by the inverse of the propensity score. McCaffrey et al. (2013) indicate that one should be wary of imbalanced data for MASDc values greater than 0.2. As can be

observed in detail in Appendix B, MASDc drops for all covariates below the cut-off of

0.2 for both weighting procedures. That is, we find that the highest value of MASDc

is 0.588 for the unweighted data, whereas it is 0.071 and 0.058 after weighting by the propensity scores resulting from AD and MD balancing method, respectively. Hence, for both balancing methods, the propensity score weighting procedure results in a data set which is balanced according to the criteria ofMcCaffrey et al.(2013), whereas it was not properly balanced prior to the propensity score procedure.

(24)

As can be observed in Appendix C, we find that in general this assumption appears to hold for both the AD-based and MD-based propensity scores. That is, each treatment group appears to have a solid overlap in the distribution of the propensity scores. This implies that, independent of the treatment that was actually received, the sessions are comparable in terms of probabilities of treatment. Hence, also this criterion appears to be met.

Based on the MASDc scores and the overlap in propensity scores, we can conclude

that the weights resulting from the AD and MD criteria successfully balance the data. Therefore, by making the sessions comparable, we can proceed to testing our hypotheses of device switching effects.

4.3 Model specification

To analyse to what extent device switching affects purchase behavior, we use a multino-mial logistic regression model. That is, we test our hypothesis regarding device switching effects using the following model

Φi,s= αi+ 4 X q=1 βqkq,i,s+ 3 X t=1 γtzt,i,s+ 2 X r=1 δrmr,i,s+ 3 X t=1 2 X r=1 ζt,rzt,i,s,mr,i,s + 15 X c=1 φcvc,i,s+ i,s, (11) where Φi,s= log Pr Ui,s= j Pr Ui,s= j∗ ! (12)

is the log-odds with Ui,s indicating the type of conversion, where j = 1 indicates an

online conversion, j = 2 indicates an offline conversion and j = 3 is no conversion (where j∗ ∈ {1, 2, 3} is the reference level with j∗ _{6= j) by customer i during session s, k}

q,i,s are

device related variables (i.e., device type and operating system), zt,i,s are device

switch-ing dummy indicators for treatment t, mr,i,s are the hypothesized moderating variables

(i.e., online experience and offline experience ), αi is a random intercept accounting for

customer heterogeneity, βq, γt, δr, ζt,r, φc are coefficients and i,s is an error term. A

more extensive description of the variables kq,i,s, mr,i,s and vc,i,s are provided by

(25)

estimate the model with no conversion as the reference category (i.e., j∗ = 3). This al-lows us to analyse how the log-odds of purchasing change if one switches between devices. Finally, we also estimate the model using online conversions as the reference category (i.e., j∗ = 1), to see how device switching impacts the log-odds of offline conversions to online conversions.

We estimate model (11) in two different ways. First we estimate the model without weighting the sessions (i.e., each session is equal in weight). Then, we also estimate the model using the weights obtained from our propensity score weighting procedure. We use the weights resulting from the AD approach as advised byMcCaffrey et al. (2013). Furthermore, we mean-center the variables in mr,i,sand vc,i,sbefore estimating the model.

By mean-centering these variables, we can interpret the estimates of γ1 and γ2 (i.e., the

effects of switching from a mobile device to a tablet and PC on purchase rates) relative to all other variables having a mean value. This allows us to test our hypotheses in H1

and H2, where we thus expect that the the estimates of γ1 and γ2 will be positive and

significant for online conversions (H1) and negative for offline conversions (H2). For H3,

H4 and H5, we will use the estimation results for ζt,r, since these coefficients test the

hypothesized moderating effects of online and offline experience.

5 Results

(26)

5.1 Online conversion models

The estimates of the multinomial logistic regression model regarding online conversions are given in column (1) and (2) of Table 3 (and Appendix D.1 for the control vari-ables). The weighted and unweighted model estimates are similar in terms of sign and significance. For H1a, we find a significant and positive estimate for MobileToTablet

(ˆγ1= 0.400, p < 0.01). This is in line with our expectation that switching from a mobile

device to a tablet is related to a higher conversion rate, compared to not switching. In economic terms, this implies that for a customer with mean levels for variables mr,i,sand

vc,i,s, who switched from a mobile device to a tablet, has odds of 1.49 (= e0.4) of

purchas-ing online, compared to a similar individual who did not switch. With this findpurchas-ing, H1a

is supported. Also for H1b, we find a positive and significant estimate for MobileToPC

(ˆγ2= 0.130, p < 0.01). Hence, switching from a mobile device to a PC results in odds of

purchasing online of 1.39, compared to not switching. Therefore, also H1b is supported.

For H3a, the moderating effect of offline purchase experience on the effect of switching

from a mobile device to a tablet, we find a significant and positive estimate (ˆζ1,1 =

0.315, p < 0.01). This implies that the positive switching effect becomes stronger as the customer is more experienced with the offline channel of the retailer. Therefore, we find support for H3a. For H3b, we also find a significant positive estimate ( ˆζ2,1 = 0.193,

p < 0.01) for the second hypothesized moderating effect on online conversions. Hence, the effect of switching from a mobile device to a PC is positively moderated by the offline experience of the customer.

In addition to the hypothesized effects, we find a negative and significant estimate for the main effect of the mobile device ( ˆβ1= −0.504, p < 0.01) and the tablet ( ˆβ2= −0.324,

(27)

Table 3. Estimation results of online conversion (baseline: no conversion)

Dependent variable: Online conversion vs. no purchase

(28)

5.2 Offline conversion models

The estimation results of the model regarding offline conversions are presented in Table4

(and Appendix D.2 for the controls). For H2a, we find a positive and significant estimate

for MobileToTablet (ˆγ1 = 0.147, p < 0.01). Hence, we find no support for H2a, but we

find evidence for the opposite effect. That is, switching from a mobile device to a tablet has a positive effect on offline conversions, in contrast with our hypothesized negative effect. For H2b, we find a negative and significant estimate for MobileToPC (ˆγ2= −0.121,

p < 0.05). This implies that the odds of making an offline purchase is 0.886 (= e−0.121) for a customer with mean levels that switched from a mobile device to a PC, compared to a similar customer that did not switch. Hence, we find support for H2b.

When we examine the estimates of the interaction effects of device switching and offline experience we find significant and insignificant results. For H4a, we find a negative

and significant estimate (ˆζ1,1 = −0.274, p < 0.05). Hence, the positive effect of switching

from a mobile device to a tablet on the offline conversion rate reduces as the customer is more experienced with the offline channel. While the direction (i.e., sign) of the moderating effect is in line with our expectation, we find no support for H4a, since the

main effect of switching from a mobile device to a tablet differs from our hypothesis. The estimate of the interaction effect of offline experience and switching from a mobile device to a PC, on the other hand, is insignificant (ˆζ2,1 = −0.035, p > 0.10). This implies

that we find no support for the moderating effect as stated in H4b. That is, the negative

effect of switching from a mobile device to a PC on offline purchase probabilities does not appear to be moderated by the customer’s offline purchase experience.

Furthermore, when we examine the estimates of online experience, we find that the positive effect of switching from a mobile device to a tablet on offline conversions is lower when the customer is more experienced with the online channel (ˆζ1,2 = −0.143,

p < 0.01). Hence, comparable to the moderating effect of offline experience, the direction of the moderating effect is in line with our hypothesis, but the main effect of switching differs. For H5b, we find a positive significant effect ( ˆζ2,2 = 0.194, p < 0.01) for the

(29)

Table 4. Estimation results of offline conversion (baseline: no conversion)

Dependent variable: Offline conversion vs. no purchase

(30)

Finally, we examine the estimates of multinomial logistic regression model with online conversion as the reference level (i.e., j∗ = 1). The estimation results of this model are given in Table 5 (and the estimation results of the control variables in Appendix D). It follows that we find negative and significant estimates for MobileToTablet (ˆγ1= −0.238,

p < 0.01) and MobileToPC (ˆγ2 = −0.245, p < 0.05). This implies that the odds of

purchasing in a brick-and-mortar store relative to purchasing online are 0.79 and 0.78 for customers that switch from a mobile device to a tablet and PC, respectively, compared to a customer that did not switch. Customers that switch to a less mobile device are therefore less likely to purchase offline, compared to customers that used the same device as in the previous session.

Furthermore, when we examine the interaction estimates, we find that the negative effect of switching decreases even further as offline experience grows. That is, the estimate of the moderating effect of offline experience on the effect of switching from a mobile device to a tablet is negative and significant ( ˆζ1,1= −0.563, p < 0.01) and also the effect

of switching to a PC is negative and significant ( ˆζ2,1 = −0.233, p < 0.01). This implies

that the negative effect of switching on the odds of purchasing offline relative to online, increases if the customer is more experienced with the offline channel of the retailer.

For online experience, we find a moderating effect in the opposite direction. That is, the estimates of the interaction effects of device switching and online experience are pos-itive and significant (i.e., ˆζ1,2= 0.234, p < 0.01 and ˆζ2,2 = 0.473, p < 0.01, respectively).

This implies that the negative device switching effects on the likelihood of purchasing offline relative to online diminishes if the customer is more experienced with the online channel.

When we examine the remaining estimates in Appendix D.3, we find a negative and significant estimate for male customers ( ˆφ13= −0.240, p < 0.01). This implies that male

customers are less likely to purchase offline compared to female customers. Also, looking at the estimates with respect to the online activity of the customer (i.e., DurationLag ( ˆφ3 = −0.062, p < 0.01), and Pages ( ˆφ4= −0.042, p < 0.01)) we find that customers who

(31)

Table 5. Estimation results of offline conversion (baseline: online conversion)

Dependent variable:

Offline conversion vs. online conversion

(32)

5.3 Robustness check

In order to check the robustness of our results, we also estimate the model using the weights resulting from the MD balance criterion. This collection of weights also balances our data as argued in Section 4. It is therefore appropriate to estimate device switching effects using these weights as it also removes potential selection bias. The estimation results of the multinomial regression model with no purchase as the reference level are given in Appendix E.

If we compare the estimation results of the online conversion models (i.e., column (2) of Table 3 and column (1) of Appendix E), we find no substantial differences in the estimation results in terms of sign, magnitude and significance. The only small magnitudinal difference occurs in the interaction effect of online purchase experience and switching from a mobile device to a PC. In the initial model using the AD-based weights, the estimate was −0.204, whereas the estimate using the MD-based weights becomes −0.367. Hence, the negative moderating effect in H3b is stronger in the model using the

weights resulting from the maximum standardized mean difference approach.

Next, we compare the offline conversion model estimation results (i.e., column (2) of Table 4 and column (2) of Appendix E). The estimates of the main effects of device switching on offline conversions are similar in terms of sign, magnitude and significance in both models. The estimates of the interaction effects of device switching and purchase experience are to a great extent also similar, with the exception of the moderating effect of offline purchase experience on the effect of switching from a mobile device to a PC. In the previous subsection, we found that the corresponding estimate was insignificant. In the model using the MD-based weights, this estimate is significant at the 10%-level. This finding would imply weak support for H4b. However, if we would base our results on a

5%-level of significance, the conclusions regarding our hypotheses would be the same for both models. Hence, we can conclude that the our results are relative robust to different weighting procedures.

6 Conclusion

(33)

a PC leads to higher online conversions but lower offline sales. Other studies investigating the mechanism of device switching on purchase behavior (e.g., de Haan et al. 2018; Xu et al. 2016), have focused solely on the online channel. These studies show that switching to a less mobile device is related to higher online conversions. While our results confirm these findings, we also show that device switching impacts offline purchase behavior. By also exploiting the causal impact of cross-device usage on offline sales, this study contributes to the existing literature on device switching effects in a purchase context as well as cross-channel marketing.

Besides the main effects of device switching on purchase behavior, we focus on the customer’s purchase experience with the retailer as a theoretical moderator underlying this mechanism. At the consumer level, we show that the switching effects on online conversions are weaker if the customer is more experienced with the online channel of the retailer, but stronger as offline purchase experience increases. For device switching effects on offline conversions, we find that both online and offline purchase experience reduces the positive effect of switching from a mobile device to a tablet. The negative effect of switching from a mobile device to a PC on offline conversions, however, is only negatively affected by online experience. These findings imply that the effects of device switching diminish or increase if the customer is a recurring shopper of the retailer. Our results therefore confirm the important role of purchase experience in consumer behavior (Alba and Hutchinson 1987;de Haan et al. 2018).

(34)

aim to maximize online revenue could lead to the opposite effect in the offline channel. Our study has a few limitations. First, we may not have correctly identified every session to the right customer. As we are only able to link sessions to each other, in which the customer was logged-in, we might be missing certain sessions in which the customer was logged-out. Also, the offline data might not be complete, as customers could have forgotten their membership card while purchasing offline. Therefore, our data may not have captured every complete path to purchase. Second, our study only uses data from a single retailer in the cosmetics sector. The results can therefore be different if one uses data of other retailers in the cosmetics sector or uses data from another branch. Third, our data does not distinguish laptops from desktops and these sessions are all labeled as PC sessions. Laptops are more convenient in terms of mobility and may therefore also share some of the advantages of other devices such as tablets or mobile devices (de Haan et al. 2018). Therefore, our results regarding PCs can be different for laptops and desktops.

(35)

References

Alba, J.W. and J.W. Hutchinson (1987). Dimensions of consumer expertise. Journal of consumer research 13(4), 411–454.

Andrade, E.B. (2000). Identifying discriminating variables of online and offline buyers: A perceived-risk approach. AMCIS 2000 Proceedings, 184.

Andrews, M., X. Luo, Z. Fang, and A. Ghose (2016). Mobile ad effectiveness: Hyper-contextual targeting with crowdedness. Marketing Science 35(2), 218–233.

Avery, Jill, Thomas J Steenburgh, John Deighton, and Mary Caravella (2012). Adding bricks to clicks: Predicting the patterns of cross-channel elasticities over time. Journal of Marketing 76(3), 96–111.

Bauer, R.A. (1960). Consumer behavior as risk taking. Chicago, IL, 384–398.

Beldona, Srikanth, Alastair M Morrison, and Joseph O’Leary (2005). Online shopping motivations and pleasure travel products: a correspondence analysis. Tourism Man-agement 26(4), 561–570.

Bigne, E., C. Ruiz, and S. Sanz (2005). The impact of internet user shopping patterns and demographics on consumer mobile buying behaviour. Journal of Electronic Commerce Research 6(3), 193–209.

Breiman, L., J.H. Friedman, R.A. Olshen, and C.J. Stone (1984). Classification and regression trees. Chapman Hall (Wadsworth, Inc.), New York,.

Brynjolfsson, E., Y. Hu, and M.S. Rahman (2009). Battle of the retail channels: How product selection and geography drive cross-channel competition. Management Science 55(11), 1755–1765.

Brynjolfsson, E. and M.D. Smith (2000). Frictionless commerce? a comparison of internet and conventional retailers. Management science 46(4), 563–585.

(36)

Chin, E., A.P. Felt, V. Sekar, and D. Wagner (2012). Measuring user confidence in smartphone security and privacy. In Proceedings of the eighth symposium on usable privacy and security (SOUPS) 2012. July 11-13, Washington, DC.

Cozzarin, B.P. and S. Dimitrov (2016). Mobile commerce and device specific perceived risk. Electronic Commerce Research 16(3), 335–354.

Dorotic, M., T.H.A. Bijmolt, and P.C. Verhoef (2012). Loyalty programmes: Current knowledge and research directions. International Journal of Management Reviews 14(3), 217–237.

Forman, C., A. Ghose, and A. Goldfarb (2009). Competition between local and electronic markets: How the benefit of buying online depends on where you live. Management science 55(1), 47–57.

Gensler, S., S.A. Neslin, and P.C. Verhoef (2017). The showrooming phenomenon: it’s more than just about price. Journal of Interactive Marketing 38, 29–43.

Ghose, A., M.D. Smith, and R. Telang (2006). Internet exchanges for used books: An empirical analysis of product cannibalization and welfare impact. Information systems research 17(1), 3–19.

Google (2019). How a web session is defined in analytics. url: https://support.google. com/analytics/answer/2731565?hl=en, last accessed on 30/10/19.

de Haan, E., P.K. Kannan, P.C. Verhoef, and T. Wiesel (2018). Device switching in online purchasing: examining the strategic contingencies. Journal of Marketing 82(5), 1–19.

Herhausen, D., J. Binder, M. Schoegel, and A. Herrmann (2015). Integrating bricks with clicks: retailer-level and channel-level outcomes of online–offline channel integration. Journal of retailing 91(2), 309–325.

(37)

Interactive Advertising Bureau (2018). Digital ad spend reaches an all-time high of $88 billion in 2017, with mobile upswing unabated, accounting for 57% of revenue.

url: https://www.iab.com/news/digital-ad-spend-reaches-all-time-high-88-billion-2017-mobile-upswing-unabated-accounting-57-revenue/, last accessed on 17/09/19.

Kaye, B.K. and N.J. Medoff (2001). Just a click away: Advertising on the Internet. Allyn & Bacon, Inc.

Konu¸s, U., P.C. Verhoef, and S.A. Neslin (2008). Multichannel shopper segments and their covariates. Journal of Retailing 84(4), 398–413.

Kowatsch, T. and W. Maass (2010). In-store consumer behavior: How mobile recom-mendation agents influence usage intentions, product purchases, and store preferences. Computers in Human Behavior 26(4), 697–704.

Lee, K.S. and S.J. Tan (2003). E-retailing versus physical retailing: A theoretical model and empirical test of consumer choice. Journal of Business Research 56(11), 877–885. Lopez, M.J. and R. Gutman (2017). Estimation of causal effects with multiple treatments:

a review and new ideas. Statistical Science 32(3), 432–454.

Luo, X., M. Andrews, Z. Fang, and C.W. Phang (2013). Mobile targeting. Management Science 60(7), 1738–1756.

McCaffrey, D.F., B.A. Griffin, D. Almirall, M.E. Slaughter, R. Ramchand, and L.F. Burgette (2013). A tutorial on propensity score estimation for multiple treatments using generalized boosted models. Statistics in medicine 32(19), 3388–3414.

McCaffrey, D.F., G. Ridgeway, and A.R. Morral (2004). Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychological methods 9(4), 403.

Meyer-Waarden, L. (2008). The influence of loyalty programme membership on customer purchase behaviour. European Journal of marketing 42(1/2), 87–114.

(38)

Narang, U. and V. Shankar (2019). Mobile app introduction and online and offline purchases and product returns. Marketing Science.

Raphaeli, Orit, Anat Goldstein, and Lior Fink (2017). Analyzing online consumer be-havior in mobile and pc devices: A novel web usage mining approach. Electronic Commerce Research and Applications 26, 1–12.

Rosenbaum, P.R. and D.B. Rubin (1983). Constructing a control group using multivari-ate matched sampling methods that incorpormultivari-ate the propensity score. The American Statistician 39(1), 33–38.

Shankar, V., M. Kleijnen, S. Ramanathan, R. Rizley, S. Holland, and S. Morrissey (2016). Mobile shopper marketing: Key issues, current insights, and future research avenues. Journal of Interactive Marketing 34, 37–48.

Shankar, V., A. Venkatesh, C. Hofacker, and P. Naik (2010). Mobile marketing in the retailing environment: current insights and future research avenues. Journal of inter-active marketing 24(2), 111–120.

Singh, S. and J. Swait (2017). Channels for search and purchase: Does mobile internet matter? Journal of Retailing and Consumer Services 39, 123–134.

StatCounter (2019). Desktop vs mobile vs tablet market share worldwide - august 2019. url: https://gs.statcounter.com/platform-market-share/desktop-mobile-tablet/worldwide#monthly-200908-201908, last accessed on 19/09/19.

Verhoef, P.C., S.A. Neslin, and B. Vroomen (2007). Multichannel customer management: Understanding the research-shopper phenomenon. International journal of research in marketing 24(2), 129–148.

Wang, R.J.H., E.C. Malthouse, and L. Krishnamurthi (2015). On the go: How mobile shopping affects customer purchase behavior. Journal of Retailing 91(2), 217–234. Weinberg, B.D., S. Parise, and P.J. Guinan (2007). Multichannel marketing: Mindset

and program development. Business Horizons 50(5), 385–394.

(39)

Appendices

A Descriptive statistics

Sessions=200,972

Variable Operationalization Mean St. dev. Min Max

Device variables (k)

Device Two dummies indicating the

device used during session

OperatingSystem Two dummies indicating the

operating system of the device

Moderating variables (m)

ExperienceOnline Number of previous online purchases

0.77 1.39 0 53

ExperienceOffline Number of previous offline purchases

0.31 0.81 0 25

Control variables (v)

Recency Number of days since the

shopper’s previous session

11.07 23.51 0 182

ProductPrice Mean price of products viewed

in previous session (in euro)

22.40 10.63 0.35 179.90

DurationLag Duration of shopper’s previous

session (in seconds)

393.46 594.84 1 24,030

Pages Total number of pages viewed

during session

8.35 8.35 1 122

Day Six dummy variables indicating

the day of the week

Age Age of the shopper at the start

of data period

38.13 11.27 9 102

Gender Two dummy variables indicating

(40)

B Balance statistics MASDc Weighted by PS covariate Unweighted ADa _MDb Recency 0.287 0.004 0.007 ProductPrice 0.089 0.014 0.011 DurationLag 0.426 0.016 0.021 Pages 0.588 0.042 0.055 GenderFemale 0.092 0.007 0.009 GenderMale 0.182 0.021 0.023 GenderUnknown 0.096 0.019 0.018 ExperienceOnline 0.278 0.071 0.064 ExperienceOffline 0.107 0.048 0.058 DayMonday 0.107 0.031 0.027 DayTuesday 0.077 0.016 0.019 DayWednesday 0.064 0.047 0.044 DayThursday 0.040 0.022 0.023 DayFriday 0.054 0.018 0.020 DaySaturday 0.163 0.028 0.032 DaySunday 0.091 0.015 0.015 Age 0.469 0.044 0.059

(41)

C Overlap of propensity scores

C.1 Overlap of AD-based propensity scores

MobileToPC propensity scores by Tx group

Treatment Propensity scores 0.0 0.2 0.4 0.6 0.8 1.0

MobileToPC MobileToTablet NoSwitch OtherSwitch

MobileToTablet propensity scores by Tx group

NoSwitch propensity scores by Tx group

OtherSwitch propensity scores by Tx group

(42)

C.2 Overlap of MD-based propensity scores

MobileToPC propensity scores by Tx group

MobileToTablet propensity scores by Tx group

NoSwitch propensity scores by Tx group

OtherSwitch propensity scores by Tx group

(43)

D Estimation results (continued)

D.1 Table 3. (continued)

(44)

(45)

Dependent variable:

(46)

E Model estimation - Robustness check

Dependent variable: Conversion vs. no conversion

(47)

Dependent variable: Conversion vs. no conversion