Creating Customer Engagement Via Mobile Apps-How Apps Usage Drive
Creating Customer Engagement Via Mobile Apps-How Apps Usage Drive
Isaac Dinner
University of North Carolina
Scott A. Neslin
Dartmouth College
October 2015
ABSTRACT
The proliferation of smartphones has spawned a new industry – mobile apps. Managers
increasingly recognize the potential for mobile commerce apps to “engage” customers and
thereby grow sales. To measure this potential, this paper examines what drives customer usage of
apps and whether app usage drives purchases in the online and offline channels. The paper
proposes three key drivers of app usage – marketing efforts, app-related consumer factors and
app design. The drivers are empirically tested on a unique database provided by an upscale
retailer on 1286 customers who have downloaded the retailer’s mobile app. The data record app
usage and online and offline purchase behavior. The authors estimate models for using or
“accessing” the app and the translation of access into both online and offline purchase. The
results demonstrate that app access is driven by online advertising, social media messaging, app
upgrades, state dependence and purchase history. Importantly, accessing the mobile app
translates into both online and offline purchases, although as expected, more so online than
that increase the functionality of smartphones beyond mere communication. The two dominant
“stores” from which consumers can download apps are Google Play and the Apple App Store.
Google Play currently has 1.5 million apps ready for download; the App Store has 1.4 million
(Android and other sources, 2015). Consumers have downloaded a total of 50 billion apps from
Google Play, and 100 billion from the App Store (Apple & TechCrunch 2015; Mashable &
Google 2015).
While gaming apps account for a significant 21% of these downloads (PocketGamer.biz
2015), mobile commerce apps are finding their way onto the screens of many mobile device
users. For example, a 2013 Arbitron report found that 43.9% of Android smartphone users and
67.5% of iOS users accessed mobile commerce apps in a given month (Mobile Minute 2015).
Managers increasingly recognize the potential for apps to “engage” the customer and thereby
grow sales. A Forrester report trumpets, “Mobile apps are the face of new systems of
including health services, travel services, financial services, food and health and fitness,
restaurants, and retail now offer “branded apps” (Bellman et al. 2011) to enhance customer
relationships and grow their businesses. An Adobe study states that 38% of tablet shoppers and
42% of smartphone shoppers say apps strengthen “brand connection” (Adobe 2015). Forrester
notes that 25 of the top 30 online US retailers have built native iPhone apps (Schadler and
McCarthy 2012). The list of top mobile commerce apps includes apps for Amazon, Groupon,
Walgreen, Target, Etsy, Best Buy, Home Depot, Macy’s, and Gilt Groupe (Siwicki 2014).
Figure 1 shows examples of apps available on the App Store from Capital One (financial
(retail). The particular features of these apps vary, but they are all geared toward providing the
customer information or entertainment. The link to sales may be explicit (e.g., Macy’s and
McDonald’s) or implicit (Tide, Capital One), but the overall strategy clearly is to grow sales or
The critical problems this paper examines are what determines customer usage of these
apps, and whether usage translates into sales. There has been preliminary work, notably in the
computer science literature, on the determinants of app usage (e.g., Bellman et al 2011, Xu et al.
2013, Taylor and Levin 2014). Bellman et al (2011) employ a lab experiment to demonstrate
that app usage can enhance brand attitudes and purchase intentions. Taylor and Levin (2014) use
a survey to show that the design of an app can influence interest in that app as well as intention
time series field data to measure the drivers of app usage and the translation of app usage into
sales. In addition, while there is the recognition that app usage has the capability to influence
online as well as offline sales (Schadler and McCarthy 2012), no previous research has examined
this empirically.
These are the issues this paper attempts to address. In particular, we examine:
We study the performance of a retailer app. We develop a framework that categorizes the
potential drivers of app usage, what we call “accessing” the app. We utilize a unique database
that records when customers access the app and when they purchase, either offline and online.
4
We formulate and estimate a multi-equation panel model consisting of an Access model and a
Purchase model. The model controls for customer heterogeneity and for endogeneity.
Our framework consists of three classes of app access drivers: app marketing efforts,
app-related consumer factors (e.g., state dependence) and app design. Our results suggest that
each of these drivers influences access. We then find that app access indeed is associated with
increased purchasing. The increase is exerted both online and offline, although to a greater
extent online. We also allow for the firm’s non-app specific marketing communications (e.g.,
influences access but has no direct effect of either online or offline purchase. Therefore, the app
essentially serves as a portal translating firm advertising into sales. We elaborate on these and
other results, and discuss their implications for app design and management.
We proceed with a review of the literature. We draw on that literature in developing our
framework, which we present next. Then we discuss our data, specify the model, and describe
how we estimate it. We next present results and conclude with a discussion of implications for
LITERATURE
As noted earlier, the computer science literature has examined factors that influence app
usage. These studies often use data collected directly from the users’ phone, such as location.
The focus of these studies is often on which of several available apps is downloaded (e.g. Garg
and Telang 2013, Ghose and Han 2014), rather than focusing on usage of a particular app. From
5
the perspective of marketing models, the focus of these papers is on app choice rather than
incidence for a specific app, which is the focus of our research. Nevertheless, these studies
provide important evidence on the types of variables that influence app usage.
A number of studies have investigated how consumer characteristics impact app usage. In
particular, Böhmer et al (2011), Shin et al (2012), and Huang et al (2013) find significant
evidence of state dependence. The study by Böhmer et al (2011) investigates a large database of
mobile subjects (4100 users) and finds that previous app usage results in a high correlation with
subsequent app usage. This study also finds that time of the day and location of the customer
influence the particular app used. Verkasalo (2009) finds more evidence that location of the user
influences which app will be accessed. Huang et al (2012) find still more evidence of state
dependence – app used last time predicts app used this time – along with the importance of
location and user profile (silent mode, general mode, etc.). Xu et al (2013) use collaborative
filtering to demonstrate that a focal user’s app usage is related to app usage of similar users in
Recent studies investigate how the design and characteristics of the app lead to increased
or decreased usage. Since the strategy behind mobile commerce apps usually centers on
engagement, it is no wonder that papers have begun to examine the design elements that
comprise a successful app. Starting from the choice of platform, Taylor and Levin (2014) find
that the users of Apple’s iOS are more receptive to retail apps than users of Android or other
mobile operating systems. Kim, Lin, and Sung (2013) distill the design elements of 106 apps
offered by global brands. They consider both informational apps (which they define to include
information such as coupons, sweepstakes/prizes, and job placement) and experiential apps
(operationalized as whether the app provided game functions). They find seven fundamental
6
design elements they interpret as related to engagement: vividness, novelty, motivation, control,
customization, feedback, and multi-platforming. These factors are quite interesting and an
important follow-up would be to examine how these factors relate to customer interest and usage
of the app. Zhao and Balagué (2015) advocate that app design should reflect the business
objective of the app (e.g., mobile commerce). They argue that personalization is critical for
mobile commerce apps, consistent with Kim, Lin, and Sung’s (2013) finding that many apps
offered by global brands include customization. These studies identify very interesting elements
of what might constitute successful app design. None of them explicitly relates design to app
performance, although that assumption is implicit. Unfortunately, we do not have the data
required to test specific design elements, but we will examine the impact of certain design
Firms hope that users will not only download their app, but also continue to use it well
past the time of initial download. However, evidence suggests that usage of most mobile apps is
rather fleeting (eMarketer 2011). Therefore, firms initiate activities to enhance engagement with
In summary, there is evidence for two classes of usage drivers: (1) consumer factors such
as location, time, previous app usage, previous retail behavior, and the behavior of similar users
and (2) design factors, including the operating system and app quality. We will consider these as
Marketing research has only begun to uncover the cross channel impact of mobile
applications: the effect of mobile on online and offline purchase intentions and behavior. For
7
example, Xu et al (2014) show that introduction of a mobile news app leads to an increased
likelihood of visiting the website of the same organization. While this study only refers to page
views, not sales, this is important because it provides evidence that a firm’s introduction of a
mobile app can have positive consequences in other channels. Bellman et al. (2011) use a lab
experiment to examine the impact of app usage on purchase intent and attitude toward the brand.
They utilize a before/after experimental design. One to two weeks before the lab test, they collect
survey data on attitudes from 228 consumers recruited from online panels. Respondents were
then provided with an iPod touch, pre-loaded with apps from eight firms – Target, Kraft, Gap,
Lancôme, Best Buy, Weber, BMW, and Gillette – and given the opportunity to use these apps in
a one-and-a-half-hour session. At the conclusion of the session, the authors found app usage
translated into more favorable attitudes and higher purchase intentions for the app’s brand.
While the lab-based nature of this study does not guarantee the enhanced attitudes will translate
into actual purchase, much less online vs. offline purchase, this study is important because it
Taylor and Levin’s (2014) survey of 345 customers of a major U.S.-based retailer
provides further evidence of the link between a retailers mobile app and a customer’s intention to
purchase. This survey asks respondents whether they would be interested in the retailer’s app,
and their intention to use it for purchase or for “information-sharing”, i.e., to learn about the
firm’s products through videos, interviews, and photoshoots. The authors find a positive
relationship between app interest and both purchase intention and information-sharing usage.
These relationships are amplified if the customer has recently visited the retailer’s physical store,
These studies are important because they provide evidence that app usage may translate
into sales. However, they do not distinguish between online and offline sales impact, and they
are based on either a lab test or a survey, and therefore do not examine the relationship between
actual app access and actual purchase behavior over time. This is what we do in our paper.
FRAMEWORK
We draw on the literature to articulate a framework by which we can study the drivers of
app access and the translation of app access into sales. Figure 2 shows the framework. We
categorize app access drivers into three groups: (1) app marketing, (2) app-related consumer
factors, and (3) app design. App design follows from the work cited earlier that finds that the
operating system influences usage, and the literature that proposes app success is related to the
design of the app. Consumer factors include contextual variables such as location, time, and
current user profile, as well as user preference. The design and customer factors follow directly
from the literature. We add a third – app marketing – which refers to marketing intended to
increase usage of the mobile app. An example would be social media posts that promote the app
advertising and promotions) to guide consumers directly to purchase. We consider how these
marketing communications directly determine purchases of the firm. However, there are
possible spillover effects from these marketing communications that also might drive app access.
For example, online advertising might encourage the customer to access the app to learn more
9
about the advertised product, while also convincing the customer to purchase directly.
Figure 2 also includes feedback effects. For example, we allow previous purchase to
influence app access. We also allow for state dependence: previous access influences current
access, and previous purchasing influences current purchasing. These state dependence effects
will be included alongside controls for consumer access and purchase rates, reflecting static
preference for the app and the retailer. These preferences are part of the consumer factors and
A key element of our framework is that mobile app access can influence both online and
offline purchases. As noted earlier, the need to consider the omni-channel environment suggests
consideration of both online and offline purchase effects. The obvious impact of app access is
on online sales. But, as Dinner, van Heerde, and Neslin (2014) show, online media can have
cross-effects on offline sales. Dinner et al. (2014) do not investigate app usage, but the same
mechanisms that create online-to-offline cross effects (e.g., research shopping) could operate
with respect to apps: the app drives the customer to the website where the customer gathers
DATA DESCRIPTION
The data in this study come from a well-known American clothing and apparel retailer
that wishes to remain anonymous. This is a high-end retailer that has a significant online
10
presence, and operates a single brick-and-mortar flagship store in a major metropolitan area. In
February of 2010 the retailer released a mobile app. The mobile app highlights a “product of the
day,” a single new product from a consistent product category each weekday, and provides the
mobile user an opportunity to purchase this product. Importantly, this particular product is not
the subject of a price discount, so the setting is quite different from a “deal of the day” situation,
where substantial discounts are offered. The app also provides a link to the store’s website for
Any iPhone user can download this app and view the product of the day. Many of these
users “register” with the firm by providing an email address. Registration allows the firm to
create a list of users identified by email address. In addition, the firm compiled an email list from
online and offline purchases. We match these email addresses with those of the registered app
users to determine individuals who have a) downloaded the mobile app and b) are a customer of
the firm in either the online or offline channel. In total, we are able to match 1286 individuals
who use both the mobile app and purchase in either the online or offline store channels.1
We study these customers over a period of 77 weeks following the introduction of the
mobile app. Over this time period, these customers access the app an average of 1.31 times per
week, and purchase in 5.09% of the weeks. They are slightly more likely to make a purchase in
the offline channel (2.86%) than in the online channel (1.99%). Geographically, 27.4% of these
individuals are located in the same county as the flagship store, or in a county adjacent to the
location of the flagship store. We do not have complete demographic data on the users, but the
nature of the retailer implies that these likely are high net worth individuals. In addition, based
1
We are also able to continue tracking customers even if they switch mobile phones, as long as they register both
devices. In addition, using the iPhone’s international mobile station equipment identity number (IMEI), we are able
to account for individuals that use an app before they officially register with the firm.
11
customers. As a result, they may be more familiar and comfortable with mobile technology than
the average customer of the retailer. Table 1 describes the variables used in the analysis. Table 2
App-related marketing: The retailer is also active in communicating to its customers via
social media, in particular with Facebook and Twitter. We are able to quantify the number of
Facebook or Twitter posts that refer specifically to the mobile app (SocialPostst). These posts
encourage app usage, and averaged .391 posts per week. We should note that while the firm’s
Instagram account is now very popular, Instagram was not released until October 2010, and was
not used to promote products on the mobile application during the time period of our analysis.
Consumer Factors: Based on prior research on app usage (e.g., Shin et al 2012), we
control for consumer factors, most prominently the number of accesses in the previous time
period, as well as if the customer made a purchase in the previous time period. In addition, there
is also evidence that an individual’s app usage may decrease over time. For example, an
eMarketer (2011) survey with MTV found that only 22% of consumers still used a favorite TV
or Movie app after six months, and only 11% of consumers still used that app after a year. We
account for this effect by including the log of the time since the app was downloaded by the
consumer.
App design: The firm’s app is only available on Apple’s App Store. This was a deliberate
decision by the firm to focus resources on the platform that more closely mirrors the
characteristics of their shoppers. The decision to choose Apple over Android is not surprising
because the firm primarily sells expensive products and iPhone users are known to spend more
12
on mobile applications (Fortune 2014), and also have a higher average household income than
Android users (Forrester Research 2011). Interestingly, during our period of analysis, Apple’s
mobile operating system, iOS, was upgraded twice. We capture the effect of these exogenous
During the sample period, the firm also upgraded the functionality of the mobile
application four times. Some of these upgrades were executed simply to fix bugs and make the
application more reliable, while others were devoted to adding functionality. We capture the
advertising aimed at increasing purchase; not directly aimed at increasing app usage. The bulk of
this budget is on traditional advertising (91.5%), a mix of radio, print media, television and
billboard campaigns. Most of these campaigns are very short and unique, which makes analysis
of any particular type extremely difficult. Therefore, we aggregate the firm’s advertising
expenditures across these media. The remaining 8.5% of the retailer’s advertising spend is on
online display advertising. Most of this advertising is general in nature, and does not involve
retargeting. Both online and traditional advertising data are available at the national level.
The retailer often runs joint online and offline promotions to spur sales of a particular
brand, segment or category. Discussions with the retailer revealed that most of these promotions
are not price-related, but typically involve a gift-with-purchase or a new product. In addition,
there are biannual clearance sales when much of the promoted merchandise is discounted. We
operationalize Promotionst to be the sum of the promotion-days in a given week t. For example,
if men’s dress shirts are promoted 4 days and women’s sunglasses are promoted for 5 days, then
this week would have 9 promotion-days. The average week has a total of 4.62 promotion days.
13
We note that the firm’s traditional (offline) and display (online) advertising occasionally
announces a promotion.
Other factors: We account for seasonality and trend. Namely, we use a Christmas
dummy variable to account for any increase in purchases related to the holiday period. We also
account for any systematic changes within the study time period with a trend variable.
MODEL
decision to access the mobile app and to purchase online and offline. We specify these models
Access Model
Accessing the mobile app ( ) is measured by the number of times customer i
uses the app in week t. We use a fixed-effect model for as a function of its
determinants:2
2
To address the count nature of the dependent variable we also tried alternative models, including Poisson and
Negative Binomial. Both of these models provided similar results to the fixed-effects model. However, count
models with autoregressive regressors required to capture state dependence can be unstable and there is no
established solution (Cameron and Trivedi 2013, p. 281). Since model (1) is a panel data model with a lagged
dependent variable, we also considered the Arellano-Bond estimator. It again yields very similar results, in line with
the notion that for panels with a large T (e.g., T>30) there is hardly any difference between Arellano-Bond estimates
and fixed-effect estimates (Roodman 2006).
14
where is a fixed effect (dummy indicates customer i) that captures any unobserved
between-customer differences in accessing the app. The three stock variables in (1) allow for the
where ;!( is the weekly spend on online advertising (display ads) and
;!( are the expenditures on offline advertising (aggregated across TV, radio,
print billboard). The carryover parameters are 34 for app access; 39: for online advertising and
39<< for offline advertising. They are contained in the interval [0,1), where 0 means no
carryover and a high value (such as .9) means a long carryover effect.
In equation (1), !" measures promotions in week t (as explained before) and
$ is the weekly number of Tweets that the company issues about the app. Thus, the
access model includes one app-related marketing effort (SocialPosts) and three general
The model’s app design variables include '(!$&, , which is a step dummy for
upgrade j for the app (for j=1,..,4) upgrades, and '(!$*, , a step dummy (for k=1,2) for
15
upgrades to the operating system of iPhones and iPads. These upgrades may help app usage (e.g.,
through fixing bugs) or hurt (e.g., by adding features that makes the app less familiar to users).
whether the customer made a purchase (online or offline) in the past week and thus captures
feedback effects: a purchase could make the customer more interested in using the app
(indicating higher levels of engagement) or it could make the customer less interested (e.g., the
purchase reduces the urge to check out the app). The model also controls for
-(_/" 0$ , the log of the time (in weeks) since customer i downloaded the
app, capturing any general increase or decrease in interest in accessing the app over time.
Finally, the model also controls for two time-related aspects: /! is a counter (1 in
the first week of the data, 2 in the second week, etc) that represents any systematic trend
common across customers. 1ℎ!"$ is the sum of the number of days in a particular week
when the week is between Thanksgiving and Christmas, and it captures the fact this period may
Purchase Model
The customer might purchase online, offline, or both in a given week. We therefore use a bivariate
probit model for customer i’s purchase decision in week t. Underlying the purchase decision is the
∗ ∗
unobserved utility to purchase online ('9: ) and offline ('9<< ). A consumer decides to purchase online if
∗ ∗
'9: > 0 and to purchase offline if '9<< > 0. The utilities are modelled as:
16
∗
(5) '9: = A + A + A +A + A +
∗
(6) '9<< = F + F + F +F + F +
where A and F are random intercepts for customer i: A ~H(0, I ) and F ~H(0, I ).
-$ is a binary variable that is 1 if the individual resides in close proximity to the store.3
We operationalize this as holding a residence in the same county as the store, or a county
adjacent to the county of the store. The other variables have been defined previously. The
independent variable of central interest is : does accessing the app (and the
associated customer engagement) increase the online and offline purchase likelihood?
The errors terms come from a bivariate normal distribution with a variance of 1 (for
1 J
LE , E M~NOH(0, Σ) Σ=P Q
J 1
(7)
Model Estimation
We estimate model (1) with a fixed effects within-estimator and estimate (5)-(6) with
simulated maximum likelihood. Since the stock variables defined in (2)-(4) involve the unknown
carryover parameters 34 for app access; 39: for online advertising and 39<< for offline
3
Note Location is not included in the Access equation because it is constant we have fixed effects in that equation.
With the random effects in the bivariate probit, we can include location.
17
advertising, we use a grid search to estimate them. We vary each of the three carryover
next select the set that maximizes the joint likelihood of purchase equations (5)-(6). We find that
34 = 39: = 39<< = 0 yields the highest likelihood. Figure A1 in the Online Appendix plots the
likelihood function. That all carryover coefficients equaled zero is a bit surprising, but there are
two reasons this may occur: 1) the dynamic nature of lagged purchase effects may already
account for the long term effects of the stock variables, and 2) we are modeling a very specific
set of customers who download the retailer’s app. It is plausible that these customers are
transactional in the sense that they attend to immediate stimuli and if they act on it, act
immediately.
endogeneity concern is app access in the purchase equation because there may be unobserved
factors that drive both app access and the propensity to purchase. For example, a consumer may
be planning to make a purchase and will open the mobile app to see what product is available. In
this scenario, app usage correlates with purchase but did not cause the purchase occasion. To
account for this possible endogeneity problem, we use a control function approach (Petrin and
Train 2010). In this setting, the control function approach is more suitable than 2SLS because the
purchase model is not a linear model but a bivariate probit (Petrin and Train 2010). The access
equation (1) can be considered a first-stage regression that includes exogenous explanatory
variables, notably the six upgrade dummies and the time since download, that are likely to
influence app usage but are excluded from the purchase behavior model (5) and (6). As such,
these exogenous variables serve as instrumental variables (IVs). An incremental F-test shows
To implement the control function for app access in equations (5)-(6), we first take the
residuals from the access equation (1). For a given value of the carryover parameter 34 for app
access, we calculate the stock version of these residuals using 34 , because access stock is the
variable that is potentially endogenous (rather than just the current-week access), and include the
residual stock as a control function term in the equations for purchase utility (5) and (6). We do
Once the grid search is executed, the control function approach also allows us to test
whether endogeneity is in fact an issue (Wooldridge 2002, p. 121). In the selected model (with
34 = 39: = 39<< = 0), we find that the control function term for Access is significant (p<.1) in
the online purchase equation (5), indicating the need for endogeneity correction in this instance.
We do not find evidence for endogeneity in the offline purchase equation (p>.1). Hence, we
retain the control function term only in the online purchase equation but not in the offline
purchase equation. As the control function term is an estimated rather than observed variable, we
need to account for this in calculating the standard errors of the coefficients. We follow Petrin
and Train (2002) by bootstrapping on the access equation and then inserting the estimated
control functions into the purchase equations. The variance in the coefficients of these purchase
equations is then added to original variance. As shown by Karaca-Mandic and Train (2003), this
We also address the potential endogeneity of online and offline advertising in the access
and purchase equations. We specify separate first-stage regressions for each type of advertising.
We include two sets of IVs. The first are the quarterly advertising costs for TV advertising, radio
advertising, print advertising, and internet advertising (source: Bureau of Labor Statistics). These
IVs are likely to influence advertising decisions of the focal retailer but unlikely to be related to
19
mobile app access or consumer purchase behavior. The second set is the combined advertising
spend by other retail companies that are not direct competitors of the focal retailer, separated by
These variables may be correlated with the focal retailer’s advertising patterns due to industry
practice but again will have little to do with the dependent variables of interest (app access and
purchasing). Again, incremental F-tests (p<.05) confirm the strength of the IVs for both online
and offline advertising. Table A1 in the Online Appendix shows the first-stage regression results.
We next proceed as we did for access: we calculate the stock version of the first-stage
regression and include its residuals as a control function term in the access equation (1) and the
purchase utility equations (5) and (6). We find that in none of the equations (1), (5), (6) is there
significant evidence of endogeneity of online or offline advertising (p>.1 for all control function
terms) and therefore do not retain these terms. As a robustness check, Table A2 in the Online
Appendix shows the access model that allows for the endogeneity of online and offline
advertising and Table A3 shows the purchase models with all control function terms included.
The models include other variables that are under management control: promotions and
app-related social posts. We argue that these are unlikely to be endogenous. We discussed the
way promotions are scheduled with company management, who were clear that the promotional
calendar is determined well in advance (between a quarter and a year ahead). In addition, the
promotions are extremely predictable from year to year. In fact, the year-over-year correlation in
promotions is .69. This makes it unlikely that the firm is organized to capitalize on week-to-week
unobserved demand shocks, mitigating endogeneity concerns. As for social posts: the company
uses Twitter to post messages that relate to the app. These Tweets clearly focus on the app and
not directly on promoting purchase. The Tweets go to all Twitter followers of the company, not
20
just the app users. Because of the nature of this communication, it is unlikely that they are based
on unobserved shocks in app access or offline and online demand shocks for the app users,
RESULTS
We now discuss the results for the mobile app usage and purchase models. Throughout
we use p<.05 for significance unless indicated otherwise. The key findings are that mobile access
has a significant positive effect on purchase (both online and offline) and that several marketing
efforts exert a positive influence on app usage. We first discuss the parameter estimates and then
report simulations to capture the dynamic impact of shocks to variables in the system.
The model for app access (equation (1)) shows a reasonable fit with an R2 of .74 (Table
4). In term of app marketing, we find that social posts (b =.0143), promotions (b=.0030), and
online advertising (b = .0044) have positive effects on the app usage. Hence these represent
levers a manager can use to drive app usage. Offline advertising does not significantly affect app
access, in line with the relative weak effects of traditional advertising in general (Sethuraman,
Tellis, and Briesch 2011). Another consideration is that the customers in our sample may be
deeply tuned into the online world, and perhaps pay less attention to offline advertising.
App access is also driven significantly by consumer factors. Lagged access exerts a
positive influence on current access (b = .6326), indicating positive state dependence. A purchase
made in the previous period leads to a decrease in access (b = -.0373), in line with the notion that
21
a purchase may alleviate the need to look for what is offered via the app. Over time, using the
app gets less and less frequent, as indicated by the negative effect of the time since download (b
= -.1285). This suggests a trial period of high app usage immediately after the app is
Changes in app design also affect app usage. Upgrades to the app have mixed effects,
with some upgrades lifting access (e.g., Upgrade 2) but others hurting (e.g., Upgrade 4). These
findings shows that upgrades (which are typically meant to eliminate glitches or to add
functionality) may often not have the desired effect, perhaps by introducing new glitches or
alienating users. Similarly, upgrades to the iOS operating system have detrimental effects on app
The trend variable has a negative effect (b = -.0012) showing that app usage is decreasing
over time, even after controlling for time since download. The Christmas period is associated
with a drop in usage as well (b = -.0099). This is quite interesting as it suggests that during
Christmas season, customers may be more goal-directed and hence don’t use their apps as often
Determinants of Purchase
Table 5 displays results for the bivariate probit purchase model (equations (5) and (6)).
The models for online purchases (hit rate 97.4%) and offline purchase (hit rate 98.1%) fit the
data well. The correlation ρ between the error terms of the two equations equals is .156 and
significant (p<.01), suggesting there are unobserved factors that spur online and offline
purchases. Importantly, consumer purchase decisions are driven significantly by app usage: b =
.037 for offline purchases and b = .058 for online purchases. Figure 3 shows the impact visually.
22
These results suggest that the engagement generated by using the app enhances the inclination to
buy online and offline. This is even after controlling for the endogeneity of app access, which is
required in the online purchase equation only, as discussed before (b = .049 for the control
function term). As expected, the impact on online purchase is greater than that on offline
purchase (in Figure 3, the marginal response line for online is steeper), but importantly both are
statistically significant. This means there is a cross-channel effect from mobile app access (an
Promotions drive offline purchases (b = .011, p<.1) but not online purchases. This is
consistent with the fact that the promotions are offered mostly in the brick-and-mortar store and
much less so online. The offline environment may offer a more enticing environment for
customers to make impulse purchases driven by promotions than the online store. We do not find
a direct of effect of advertising on offline or online purchases. This result is surprising, but there
is ample evidence that not all advertising is effective. For example, Sethuraman et al.’s (2011)
meta-analysis finds that 41% of all short-run advertising elasticities are less than .05. Combined
with the positive impact online advertising has on app usage, the results suggests that, for this set
of customers, the app serves as a conduit between the firm’s advertising and the purchase
decision.
lagged purchase effect on offline purchasing (b = .161) and for online purchasing (b = .167).
Close proximity to the physical store enhances offline purchases (b = .770) but at the expense of
proximity to the store means the customer can shop there rather than online. There is no
23
significant trend in online or offline purchase propensities and Christmas has a significant
positive effect only on online purchases (b = .028, p<.1), again in line with an online-oriented
sample.
In brief, we find that promotions, social posts, online advertising are significant drivers of
app access. In turn, app access drives online and offline purchasing significantly. To gauge the
Dynamic Simulations
We use dynamic simulations to understand the magnitude and duration of the dynamic
effects of shocks in access and the variables that drive access. These simulations use the
estimated versions of model (1) and (5)-(6), taking into account how a change in one variable
affects other variables through its contemporaneous effects and through feedback and state
dependence effects (see Figure 2). We apply shocks to access, social media, online advertising
and promotion. Offline advertising does not have a significant effect on either app access or
purchasing, so we do not study its dynamic impact. Similar to impulse response functions, we
increase a variable by one standard deviation in period t and then study how this shock drives
access and purchases in periods t, t+1, t+2, etc. We then accumulate the changes in the
dependent variable of interest to obtain long-term effects. Finally, we express the change in the
focal dependent and independent variable as a percentage of their respective base levels,
Figure 4a shows that among the marketing efforts, the impact of online advertising on the
number of accesses is the strongest, followed by promotion and by app-related social media (i.e.
Tweets and Facebook posts about the app). The effect is not instantaneous, but lasts for several
weeks, until it dies out by about week ten. The reason why these variables affect access in the
24
current as well as subsequent periods is the strong state dependence effect for access (lagged
access variable in Table 4). Marketing efforts have an immediate positive impact on access,
which then stays high for a relatively long time due to state dependence in app access.
Figure 4b displays the impact of accessing the app on online and offline purchases. The
online purchase effect is stronger than the offline purchase effect, and the effects spike in the
first week and then decay to zero in about ten weeks’ time. Again, a shock to access in period t
generates purchases in period t, but also generates more accesses in subsequent periods due to
.34% increase in the offline purchase probability and a .49% increase in online purchase
probability, which are sizable elasticities. This suggests that if a firm can successfully stimulate
app usage, there are potential large gains in purchase propensities. There is also a strong impact
of shocking access on total app access, with a 1% shock in period 1 leading to a 2.77% increase
across periods t, t+1, t+2, etc. Figure 5 shows the dynamic simulation that produces this
elasticity. There are two forces at work: 1) access is state dependent (b = .63, Table 4) and 2)
there is a negative impact of previous purchase on access (b = -.0373). From Table 5, access in
period t increases purchase in period t. This feeds back to a small decrease in subsequent access.
However, access state dependence dominates this slight negative effect, as is evident from the
net positive impact in the access simulation for all periods in Figure 5.
While there is a strong payoff to access, Table 6 also shows that driving app usage is
difficult, with rather modest elasticities of .06 for online advertising, .03 for promotion and .01
25
for social media. Consequently, the elasticities of these marketing variables on purchases (online
or offline or total) are quite small (between .001 and .09 in absolute value). For instance, app-
related social media has a limited effect on access, which is further diluted because access only
partially drives purchasing. Promotions have the strongest elasticity (.09 for offline purchasing;
.04 for total purchasing), but this is largely due to its direct effect on purchase probability rather
In sum, these simulations show that access has relatively strong effects on online and
offline purchase probabilities. They also show that a positive shock in access lingers for quite a
long time due to state dependence. It is, on the other hand, hard to increase access through
marketing: while the effects are positive and statistically significant, the elasticities are not large
in magnitude.
Summary
This paper examines the factors that drive customer usage of mobile commerce apps, and
whether usage translates into online and offline sales. Many firms have introduced these apps
with the implicit or explicit goal of connecting with customers and building customer
the antecedents and consequences of app usage and propose econometric models to estimate
these effects, while using heterogeneity and endogeneity controls. Based on data tracking app
usage and purchase behavior for a sample of 1286 customers of a large upscale retailer, we find
strong effects of using the app on purchase behavior. The elasticity of online purchase
26
probability to mobile app access is .49, whereas for offline purchasing it is .34. These are sizable
elasticities, compared to other managerial levers, such as advertising, where elasticities often
hover around .1 or less (Sethuraman et al. 2011). Accessing the app feeds strongly on itself, as
evidenced by large state dependence effects in the access equation (b = .63). Hence, it appears
that apps do engage the customer, and engagement translates into purchase.
However, driving app usage is difficult. This may be due to the particular context of our
application. In our empirical setting, the retailer made no explicit attempts to drive app usage
with the exception of limited efforts through Twitter (.39 posts on average per week). Therefore,
while we do find significant effects of online advertising, promotions and these Tweets on app
usage, their magnitude is not very high, with the highest elasticity of app usage with respect to
online advertising (.06). Nevertheless, we are encouraged by these results, because they may
represent a lower bound on the effect sizes that can be expected as firms start driving app usage
more explicitly, as firms learn how to design apps more effectively, and as consumers become
Since many consumers are increasingly living their lives through smartphones and
tablets, engaging them on these platforms may be one of the few effective channels to reach
them going forward. For example, it is quite striking that in our empirical setting, online
advertising has no direct effect on purchasing and its only effect on purchasing happens through
driving usage of the mobile app. Hence, app usage serves as a portal between online advertising
and purchasing. Offline advertising is not significantly driving access or purchases, suggesting
that this instrument may have lost its effectiveness, at least for the relatively tech-savvy
It is also very interesting that all direct effects are instantaneous. While we use stock
variables to allow for possible carry-over effects for online advertising, offline advertising, and
access, we find that their effects are instantaneous (in the same week). This is in line with the
notion that in a world that is becoming increasingly online-focused, human attention span is
becoming shorter (The Guardian 2012). As noted earlier, it is quite possible that the customers
who download the app are transactional in nature. They attend to stimuli by accessing the app
immediately, and if that translates into purchase, the purchase is also made immediately. Despite
this, the dynamic simulations find that access has a longer-term impact on purchase, and the
marketing variables that influence access also have longer-term impacts. The key reason is
strong state dependence for access (autoregressive term b = .63, Table 5).
We also uncover some key new insights on the nature of app access. In our application,
the focal app very much tries to capitalize on the curiosity of its users for a featured product. The
strong levels of state dependence in app usage suggest there are periods with higher levels of
curiosity and perhaps an urge to buy something. Interestingly, once a purchase is made (which
may be the product on offer or something else altogether), app access decreases significantly, but
the decrease is not strong enough to decrease access back to zero, as shown by the simulations.
The consumer may be checking the app to avoid post-purchase cognitive dissonance. Future
We also find that app design drives app access, but not always in a positive way. Several
of the upgrades for the app or operating system lead to lower levels of app usage. These findings
show that care should be executed by the operator of the app both when issuing own upgrades
and when the operating system of the smartphone or tablet undergoes upgrades. These are the
moments with distinct changes in app usage; mostly negative, in this empirical setting.
28
Managerial Implications
This work has a number of new and important implications for firms using or considering
a mobile app. The key takeaway is that it is crucial to drive usage of the app, because app usage
strongly enhances the probability the consumer purchases from the firm. The question is how to
do this. The marketing efforts we study, even app-related social media, have limited success in
driving app usage. An alternative would be to emphasize the app and its usage much more in
online advertising, because tech-savvy consumers spend a lot of time online. For example,
Danaher et al (2015) show how text-message based mobile coupons can effectively drive sales
within a shopping mall. Similar within-app push notifications can also be used, but with caution,
We find that app design factors influence usage. Apps should be designed accordingly.
Consumers that access the app should easily be able to find what the firm offers and to purchase.
In a survey of millennials, 72% agreed that company apps should allow them to purchase
products or services via the company’s app (Oracle 2015). The same survey found that other
important features are the ability to pay bills (71%), to flag issues (65%) and to receive updates
on upcoming offers (62%). Consumer curiosity needs to be catered to on an ongoing basis, with
multiple updates and reasons to maintain engagement, similar to successful news apps (e.g.,
BBC News). Successful apps outside the realm of commercial apps often have game elements
(e.g., Candy Crush, Angry Birds, FIFA 15) or social elements (e.g., Minecraft, Facebook) that
hook users. We are not suggesting to create addictive commercial apps (as they may lead to other
issues for the firm, such as bad publicity), but some of the insights and expertise from successful
app builders could be utilized to improve app design. The app should keep customers engaged
29
for extended periods of time. In that sense, the app we study (with its declining usage over time
and unsuccessful updates) may present a lower-bound case of what is feasible in this domain.
A final important managerial implication is that apps, which are accessed on mobile
online devices, can be used to effectively drive offline sales. This is another form of marketing
cross-channel effect (Wiesel et al. 2011; Dinner et al. 2014). It says that the customer monitors
the app to see what the retailer has available, and goes to the store when s/he finds something of
interest. This means that the app needs to be designed to integrate well with what’s going on in
the store. E.g., if a “product of the day” is featured in the app, the store needs to have the
inventory required to serve the interest that will be generated by customers accessing the app.
In this study we look at consumers who have downloaded the app. The reason is that we
want to focus on how app usage affects purchase behavior, which cannot be done for non-
adopters of the app. A useful extension of the research would be to study the download decision
as well, and try to understand the drivers of this decision. This would allow for testing
differences between app adopters and non-adopters in purchase behavior over time.
Another obvious limitation is that we study only a single mobile app. Future research can
look at different apps and study the antecedents and consequences of their usage. This would
generate insight on which specific app design (Kim, Lin, and Sung 2013; Zhao and Balagué
2015) and consumer factors enhance usage and the impact of usage on purchase behavior.
Despite these limitations, we hope this paper will become a stepping stone in gathering
oriented world. We hope that future research will refine and extend this understanding.
30
REFERENCES
Adobe (2015)
https://www.adobe.com/aboutadobe/pressroom/pdfs/Mobile_Shopping_Retail_App_Usa
ge_on_the_Rise_Infographic.pdf
Android, & Apple, & Google, & Microsoft, & AppBrain, & BlackBerry, & Various sources
(WindowsCentral.com), & Amazon. (2015). Number of apps available in leading app
stores as of July 2015. In Statista - The Statistics Portal. Retrieved August 15, 2015, from
http://www.statista.com/statistics/276623/number-of-apps-available-in-leading-app-
stores/.
Apple, & TechCrunch. (2015). Cumulative number of apps downloaded from the Apple App
Store from July 2008 to June 2015 (in billions). In Statista - The Statistics Portal.
Retrieved August 15, 2015, from http://www.statista.com/statistics/263794/number-of-
downloads-from-the-apple-app-store/.
Bellman, Steven, Robert F. Potter, Shiree Treleaven-Hassard, Jennifer A. Robinson, and Duane
Varan (2011), “The Effectiveness of Branded Mobile Phone Apps,” Journal of
Interactive Marketing, 25(4), 191-200.
Böhmer, M., et al.: (2011), “Falling Asleep with Angry Birds, Facebook and Kindle – A Large
Scale Study on Mobile Application Usage,” in Proc. of MobileHCI, 47-56.
Cameron, A. Colin, and Praven K. Trivedi (2013), Regression Analysis of Count Data, 2nd
Edition, Cambridge, UK: Cambridge University Press.Danaher, Peter J., André Bonfrer,
and Sanjay Dhar (2008), “The Effect of Competitive Advertising Interference on Sales
for Packaged Goods,” Journal of Marketing Research, 45 (April), 211-25.
Danaher, Peter J., Michael S. Smith, Kulan Ranasinghe, and Tracey S. Danaher (2015) “Where,
When, and How Long: Factors That Influence the Redemption of Mobile Phone
Coupons,” Journal of Marketing Research, 52 (5), 710-725.
Dinner, Isaac M., Harald J. van Heerde, and Scott A. Neslin (2014), “Driving Online and Offline
Sales: The Cross-Channel Effects of Traditional, Online Display, and Paid Search
Advertising,” Journal of Marketing Research, 51(5), 527-625.
Forrester Research (2011) North American Technographics Telecom And Devices Online
Recontact Survey, Q3 2011 (US)
Garg Rajiv, Rahul Telang (2013) “Estimating app demand from publiclyavailable data,” MIS
Quarterly. 37(4), 1253–1264.
31
Greene, William H. (2000), Econometric Analysis, Prentice Hall: Upper Saddle River, NJ.
Ghose, Anindya, Sang Pil Han (2015) “Estimating Demand for Mobile Applications in the New
Economy,” Management Science, 60 (6), p 1470-1488.
Huang, Ke, Chunhui Zhang, Xiaoxiao Ma, Guanling Chen (2012), “Predicting Mobile
Application Usage Using Contextual Information,” UbiComp ’12, Septebmer 5-8,
Pittsburgh, USA.
Karaca-Mandic, Pinar and Kenneth Train (2003) “Standard Error Correction in Two-Stage
Estimation with Nested Samples,” The Econometrics Journal, 6 (2), 401-407.
Kim, Eunice, Jhih-Syuan Lin, and Yongjun Sung (2013), “To App or Not to App: Engaging
Consumers via Branded Mobile Apps,” Journal of Interactive Marketing, 13(1), 53-65.
Lambrect, Anja, Avi Goldfarb, Alessandro Bonatti, Anindya Ghose, Daniel G. Goldstein,
Randall Lewis, Anita Rao, Navdeep Sahni and Song Yao (2014) “How do firms make
money selling digital goods online?” Marketing Letters, 25, 331-341.
Mashable, & Google. (2015). Cumulative number of apps downloaded from the Google Play
Android app store as of July 2013 (in billions). In Statista - The Statistics Portal.
Retrieved August 15, 2015, from http://www.statista.com/statistics/281106/number-of-
android-app-downloads-from-google-play/.
Oracle (2015), “Millennials and Mobility: How Businesses Can Tap into the App Generation,”
http://www.oracle.com/us/dm/millennials-and-mobility-2508385.pdf, accessed on July
31, 2015.
Petrin, Amil and Kenneth Train (2002) “Omitted Product Attributes in Discrete Choice Models,”
Working Paper, Department of Economics, University of California, Berkeley.
Petrin, Amil and Kenneth Train (2010), “A Control Function Approach to Endogeneity in
Consumer Choice Models,” Journal of Marketing Research, 47(1), 3-13.
PocketGamer.biz. (2015). Most popular Apple App Store categories in June 2015, by share of
available apps. In Statista - The Statistics Portal. Retrieved August 15, 2015, from
http://www.statista.com/statistics/270291/popular-categories-in-the-app-store/.
Schadler, Ted and John C. McCarthy (2012), “Mobile is the New Face of Engagement,” Forester
Report, February 23, 2012.
32
Sethuraman, Raj, Gerard J. Tellis, and Richard A. Briesch (2011), “How Well Does Advertising
Work? Generalizations from Meta-analysis of Brand Advertising Elasticities,” Journal of
Marketing Research, 48 (June), 457-71.
Shin, Choonsung, Jin-Hyuk Hong, and Amind K. Dey (2012), “Understanding and Prediction of
Mobile Application Usage for Smart Phones, UbiComp’12, September 5-8, 2012,
Pittsburgh, USA.
Siwicki, Bill (2014), “Exclusive: The 25 Most Popular Mobile Commerce Apps,” Internet
Retailer, June 5, 2014.
Taylor, David G., and Michael Levin (2014), “Predicting Mobile App Usage for Purchasing and
Information –Sharing,” International Journal of retail and Distribution Management,
42(8), 759-774.
The Guardian (2012), “Say It Quick, Say It Well – The Attention Span of a Modern Internet
Consumer,” http://www.theguardian.com/media-network/media-network-
blog/2012/mar/19/attention-span-internet-consumer [Accessed September 10, 2015]
Verkasalo, H. (2009), “Contextual Patterns in Mobile Service Usage,” Personal and Ubiquitous
Computing, 13(5), 331-342.
Wiesel, Thorsten, Koen Pauwels, and Joep Arts (2011), “Marketing’s Profit Impact: Quantifying
Online and Offline Funnel Progression,” Marketing Science, 30 (4), 604-611.
Wooldridge, Jeffrey M. (2002), Econometric Analysis of Cross Section and Panel Data,
Cambridge: MIT Press
Xu, Jiao, Chris Forman, Jun B. Kim and Koert Van Ittersum (2014) “News Media Channels:
Complements or Substitutes? Evidence from Mobile Phone Usage,” Journal of
Marketing, 78(4), 97-112.
Xu, Ye, Mu Lin, Hong Lu, Giuseppe Cardone, Nicholas D. Lane, Zhenyu Chen, Andrew
Campbell, and Tanzeem Choudhury (2013), “Preference, Context and Communities: A
Multi-faceted Approach to Predicting Smartphone App Usage Patterns,” ISWC’13,
Zurich, Switzerland.
Zhao, Zhenzhen, and Christine Balagué (2015), “Designing Branded Mobile Apps:
Fundamentals and Recommendations,” Business Horizons, 58(), 305-315.
33
Figure 1
Examples of Mobile Commerce Apps
Figure 1a: Capital One Figure 1b: Tide Detergent
Figure 2
Framework for Examining the Drivers of App Access and
the Impact of Access on Purchase
Marketing
Communications: Purchases
• Online advertising • Online
• Offline advertising • Offline
• Promotions
Notes:
Figure 3
Direct Impact of Access on Purchase Probability
3.5%
3.0%
2.5%
Purchase Probability
2.0%
1.5%
1.0%
0.5%
0.0%
0 1 2 3 4 5 6 7
Number of Accesses
Figure 4
Impact of Marketing Efforts on Access and the Impact of Access on Purchasing
0.020
0.015
0.010
0.005
0.000
0 5 10 15
Week
0.004
0.003
0.002
0.001
0
0 2 4 6 8 10 12 14
Week
Simulated impact obtained by one standard deviation shock in the variable of interest in week 1
37
Figure 5
Simulated Impact of a Shock in Current Access on Future Access
1.4
1.2
Change in Number of Accesses
1
0.8
0.6
0.4
0.2
0
0 2 4 6 8 10 12 14
Week
Simulated impact obtained by one standard deviation shock in the variable of interest in week 1
38
Table 1
Variable Operationalizations
Table 2
Descriptive Statistics (per period)*
Variable Mean Std Dev
AppAccesses 1.306 .629
Purchase .0509 .0099
Purchase Online .0199 .0062
Purchase Offline .0286 .0067
Promotions 4.62 4.863
SocialPosts .391 1.037
OnlineAdvertising 8.503 9.650
OfflineAdvertising 91.497 139.154
Location .274 .018
Log_TimeSinceDownload 3.096 .905
Upgrade1 .783 .415
Upgrade2 .663 .475
Upgrade3 .641 .482
Upgrade4 .576 .497
iOSChange1 .696 .463
iOSChange2 .141 .35
Trend 46.5 26.703
Christmas .5652 1.836
*: N = 83,549. Data are comparatively re-scaled to prevent revealing the identity of the retailer.
40
Table 3
Correlation Matrix
Var 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
AppAccesses 1
Purchase 2 .04*
Purchase Online 3 .06* .64*
Purchase Offline 4 .01 .77* .04*
Promotions 5 .03* .01* .00 .02*
SocialPosts 6 .07* .00 .00 -.01 -.06*
Online Advertising 7 .15* .01* .01* .00 .03* .08*
Offline Advertising 8 -.01* .01* .00 .01* .30* .01* -.05*
Location 9 -.10* .09* -.02* .13* .00 .00 .01* .00
* * * * * * *
Log_TimeSinceDownload 10 -.26 -.03 -.04 -.02 -.05 -.28 -.42 .05* .02*
Upgrade1 11 -.19* -.01* -.01* .00 -.04* -.21* -.81* .00* -.01 .59*
Upgrade2 12 -.23* -.01* -.01* -.01* -.05* -.23* -.47* .01* -.02* .64* .59*
Upgrade3 13 -.22* -.01* -.01* -.01* .00 -.23* -.43* .09* -.02* .64* .55* .93*
Upgrade4 14 -.21* -.01* -.01* .00 .10* -.33* -.30* .33* -.02* .65* .45* .77* .82*
* * * * * * * * * * * *
iOSChange1 15 -.23 -.02 -.01 -.01 -.02 -.11 -.55 .06 -.02 .64 .66 .89 .83* .69*
iOSChange2 16 -.10* -.01* .00* -.01* -.14* -.13* .15* -.21* -.01 .34* .11* .18* .20* .24* .17*
Trend 17 -.24* -.02* -.01* -.01* -.07* -.31* -.31* .05* -.02* .77* .51* .73* .75* .81* .68* .54
* * * * * * * * * * * * * *
Christmas 18 -.03 .01 .01 .01 .04 -.09 -.02 .40 .00 .05 .07 .13 .14 .16 .11 -.09* .02*
*
: p-value < .05
41
Table 4
Model Results for Accessing the Mobile App
b S.E. t p-value
Table 5
Model Results for Bivariate Probit Model for Offline and Online Purchases
Purchase Offline Purchase Online
b S.E. t p-value b S.E. t p-value
Intercept -2.888** .128 -22.61 .00 -2.431** .094 -25.89 .00
Marketing
Communications
Other Controls
Access Equation
.049** .024 2.00 .05
Control Function
Standard Deviation in
.836 .032 26.13 .00 .598 .024 24.92 .00
random intercept
Number of
Observations 83,549
Log-Likelihood -14,703.95
Error correlation ρ .156 (.029)
** *
: two-sided p-value < .05; : .05 ≤ two-sided p-value < .10.
43
Table 6
Long-Run Elasticities for the Drivers of Access and Purchases
Elasticity of:
With Respect to 1% Offline Online Total
Change in: Access Purchasing Purchasing Purchasing
Access 2.774 .337 .492 .318
Social .010 .001 .001 .001
Promotion .033 .090 -.026 .044
Online Advertising .063 -.018 .061 .012
44
ONLINE APPENDIX
Table A1
First-Stage Regression Models for Online and Offline Advertising
Online Advertising Online Advertising Offline Advertising Offline Advertising
with Instrumental Variables without Instrumental Variables with Instrumental Variables without Instrumental Variables
b S.E. t p-value b S.E. t p-value b S.E. t p-value b S.E. t p-value
Intercept 265.915 391.931 .68 .50 12.904 1.734 7.44 <.01 -1988.118 2487.613 -.80 .43 49.821 14.168 3.52 .00
Lagged Aggregate Sales1 .000 .000 1.24 .22 .000 .000 1.06 .29 .000 .000 .11 .91 .000 .000 .01 .99
Promotions -.005 .086 -.05 .96 .090 .104 .87 .39 .039 .549 .07 .94 1.175 .846 1.39 .17
SocialPosts .020 .510 .04 .97 .323 .572 .57 .57 3.849 3.235 1.19 .24 8.944 4.675 1.91 .06
Upgrade1 -17.007 2.408 -7.06 <.01 -10.587 2.166 -4.89 <.01 -41.513 15.286 -2.72 .01 14.665 17.691 .83 .41
Upgrade2 .361 3.739 .10 .92 2.450 4.272 .57 .57 -82.513 23.733 -3.48 .00 -82.003 34.898 -2.35 .02
Upgrade3 2.318 3.346 .69 .49 -1.669 3.744 -.45 .66 12.009 21.236 .57 .57 -9.601 30.583 -.31 .75
Upgrade4 5.431 4.248 1.28 .21 .194 2.573 .08 .94 124.644 26.959 4.62 <.01 98.022 21.021 4.66 <.01
iOSChange1 -5.076 2.954 -1.72 .09 -3.227 3.254 -.99 .32 10.357 18.750 .55 .58 28.866 26.584 1.09 .28
iOSChange2 8.233 2.943 2.80 .01 .191 2.125 .09 .93 -7.510 18.679 -.40 .69 -50.701 17.359 -2.92 .00
Trend -.009 .099 -.09 .93 .054 .060 .89 .37 -1.678 .625 -2.68 .01 -.184 .491 -.37 .71
Christmas .018 .410 .04 .97 -.961 .317 -3.03 .00 1.787 2.602 .69 .49 5.824 2.591 2.25 .03
TV TNS2 .241 13.380 .02 .99 -357.857 84.926 -4.21 <.01
Magazine TNS -.233 .147 -1.59 .12 7.081 .931 7.60 <.01
Newspaper TNS -18.483 4.026 -4.59 <.01 -73.409 25.552 -2.87 .01
DISPLAY TNS 127.604 148.689 .86 .39 -346.672 943.740 -.37 .71
TV BLS3 .068 .220 .31 .76 2.718 1.397 1.95 .06
Internet BLS .392 .194 2.02 .05 3.293 1.232 2.67 .01
Radio BLS 1.576 .697 2.26 .03 .703 4.426 .16 .87
Print BLS -3.597 3.821 -.94 .35 13.021 24.251 .54 .59
N 91 91 91 91
R2 .723 .496 .830 .489
Adjusted R2 .648 .425 .785 .417
1
This refers to the lagged total sales by the retailer across all channels and customers
2
TNS refers to advertising spend by other (non-competing) retailers as provided by TNS
3
BLS refers to advertising cost as measured by the Bureau of Labor Statistics
Table A2
Model for Access of the Mobile App Estimated with 2SLS Allowing for Endogeneity of
Offline and Online Advertising
b S.E. t p-value
Table A3
Purchase Model with All Control Functions Included
Purchase Offline Purchase Online
b S.E. t p-value b S.E. t p-value
** **
Intercept -2.897 .134 -21.65 .00 -2.408 .096 -24.99 .00
Effect of Mobile App
AppAccessStockt .044** .017 2.61 .01 .059** .012 4.88 .00
Marketing Efforts
Promotions .011* .006 1.84 .07 -.003 .005 -.57 .57
OnlineAdStock -.003 .007 -.35 .73 .004 .005 .80 .42
OfflineAdStock .001 .001 .58 .56 -.001 .001 -1.10 .27
Other Controls
Purchaset-1 .160** .080 2.01 .05 .166** .062 2.70 .01
** **
Location .775 .066 11.80 .00 -.132 .053 -2.49 .01
Trend -.001 .002 -.83 .41 -.001 .001 -.89 .38
*
Christmas .005 .021 .22 .82 .027 .015 1.73 .08
1 *
Access Equation CF -.015 .034 -.44 .66 .047 .025 1.92 .06
OnlineAdvertising CF -.001 .013 -.05 .96 .001 .010 .08 .94
OfflineAdvertising CF .001 .002 .29 .77 .001 .001 .99 .32
Standard Deviation in
.837 .033 25.36 .00 .598 .024 24.92 .00
Random Intercept
Number of Observations 83,549
Log-Likelihood -14,701.286
Error correlation ρ .157 (.029)
**
: two-sided p-value < .05; *: .05 ≤ two-sided p-value < .10.
1
CF = Control Function
47
Figure A1
Likelihood Function over Values of Online Carryover (39: )
and Offline Advertising Carryover (39<< ) when Access Stock Carryover (34 ) = 0