Creating Customer Engagement via Mobile Apps:

How App Usage Drives Purchase Behavior

Isaac Dinner
University of North Carolina

Harald J. van Heerde

Massey University

Scott A. Neslin
Dartmouth College

October 2015

Electronic copy available at: http://ssrn.com/abstract=2669817

 1

Creating Customer Engagement via Mobile Apps:

How App Usage Drives Purchase Behavior

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

offline. The paper concludes with managerial implications.

Electronic copy available at: http://ssrn.com/abstract=2669817

 2

The proliferation of smartphones has spawned a new industry – applications or “apps”

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

engagement.” (Schadler and McCarthy 2012, p. 4). Companies in a variety of industries

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

services), Tide Detergent (housekeeping products), McDonald’s (restaurants), and Macy’s

 3

(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

enhance loyalty through customer engagement.

<Insert Figure 1 About Here>

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

to purchase. However, to our knowledge, no previous research has examined customer-level,

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:

• What are the drivers of app usage?

• Does app usage translate into sales?

• How is this translation allocated between online and offline sales?

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.,

advertising) to influence access as well as purchase. Interestingly, we find online advertising

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

researchers and managers.

LITERATURE

Determinants of App Usage

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

terms of their app preferences and location.

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

elements, namely operating system and app upgrades.

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

the mobile app and the firm as a whole (Oracle 2015).

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

well as a third: firm-initiated marketing efforts such as social media campaigns.

Impact of App Usage on Purchase

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

begins to show that link.

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,

a sign either of state dependence or static preference for that store.

 8

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

to the firm’s followers.

<Insert Figure 2 About Here>

Traditionally, firms use marketing communications (i.e., online advertising, offline

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.

Therefore, our framework allows for purchase-oriented marketing communications to drive

access (see the dashed arrow in Figure 2).

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

will be accounted for in our model.

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

information, then goes to the store to make the purchase.

DATA DESCRIPTION

Customer Access and Purchase Data

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

the purchase of other products besides this item of the day.

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

on their pursuit of mobile apps, we believe this is a strongly “online-oriented” sample of

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

offers descriptive statistics and Table 3 is a correlation matrix of key variables.

<Insert Tables 1-3 About Here>

Access and Purchase Predictors

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

events through separate dummy variables.

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

impact of these upgrades with separate dummy variables.

Marketing communications: The retailer undertakes online and traditional (offline)

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

We draw on our framework (Figure 2) to specify econometric models of the consumer

decision to access the mobile app and to purchase online and offline. We specify these models

below, after which we detail model estimation.

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

(1)

 = ∑   + 

, +  
 +

 
 +  !" + # $  + ∑& #%& '(!$&, +

∑* )%* '(!$*, +  +!ℎ$, +  -(_/" 0$ +

 /! + # 1ℎ!"$ ++

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

dynamic impact of past app access (

, ), online advertising

(
 ), and offline advertising (
 ) on access, operationalized as

(e.g., Danaher, Bonfrer, and Dhar 2008):

(2)

 = 34

, + (1 − 34 )

,

(3) 
 = 39: 
 + (1 − 39: )
;!(

(4) 
 = 39<< 
 + (1 − 3=<< )
;!(

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

marketing efforts (promotions and online/offline advertising).

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).

The model includes several consumer factors, including the aforementioned


, capturing state dependence. +!ℎ$, is an indicator variable for

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

influence (positively or negatively) with accessing the app.

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 
 +

A !" + A# +!ℎ$, + AB -$ + AC /! + AD 1ℎ!"$ + E

∗
(6) '9<< = F + F  + F

 +F 
 + F 
 +

F !" + F# +!ℎ$, + FB -$ + FC /! + FD 1ℎ!"$ + E

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

identification purposes) and a covariance of J:

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

parameters from 0 to .9 in increments of .1, for a total of 10×10×10 = 1000 combinations. We

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.

In estimating the model, we address potential endogeneity issues. One potential

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

that the IVs are sufficiently strong (p<.05).

 18

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

so for every value of the carryover parameter in the grid search.

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

bootstrapping technique gives very similar results to an asymptotic approach.

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

TV advertising, magazine advertising, newspaper advertising and online display advertising.

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,

alleviating endogeneity concerns for SocialPostsit.

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.

Determinants of App Access

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.

<Insert Table 4 About Here>

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

downloaded, followed by declining usage rates.

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

usage (b =-.2565 (p<.05) and -.0297 (p<.1), respectively).

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

and may just go directly to visit either the store or website.

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

online channel) to buying in the offline store.

<Insert Table 5 and Figure 3 About Here>

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.

There is evidence for significant state dependence in purchasing: there is a positive

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

online purchases (b=-.132), consistent with a cannibalization effect. This is understandable –

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

managerial significance of these effects, we use dynamic simulations.

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,

allowing us to calculate long-term elasticities.

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

state dependence, hence more purchases.

<Insert Figure 4 and Table 6 About Here>

Table 6 shows the corresponding long-term elasticities. A 1% shock in access leads to a

.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.

<Insert Figure 5 About Here>

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

than its indirect effect via access.

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 AND MANAGERIAL IMPLICATIONS

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

engagement and ultimately increasing sales. We develop a conceptual framework to understand

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

accustomed to mobile apps.

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

consumers that we study.

 27

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

research may try to test the validity of this proposed mechanism.

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,

as many consumers see these messages as intrusive (Oracle 2015).

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.

Limitations and Future Research

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

an understanding on how consumers can be engaged through apps in an increasingly online-

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.

eMarketer (2011) http://www.emarketer.com/Article/How-Encourage-Mobile-App-Discovery-

Usage/1008443

Forrester Research (2011) North American Technographics Telecom And Devices Online
Recontact Survey, Q3 2011 (US)

Fortune (2014) http://fortune.com/2014/06/27/apples-users-spend-4x-as-much-as-googles/

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/.

MobileMinute (2015) “Mobile Commerce App Usage iOS vs. Android”.

http://mobileminute.info/mobile-commerce-app-usage-ios-vs-android/

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/.

Roodman, David (2006), “How to Do xtabond2: An Introduction to “Difference” and “System”

GMM in Stata,” Working Paper Number 103, December, www.cgdev.org.

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 1c: McDonald’s Figure 1d: Macy’s

 34

Figure 2
Framework for Examining the Drivers of App Access and
the Impact of Access on Purchase

App Usage Drivers:

• App Marketing
• App-related Mobile App Access
Consumer Factors
• App Design

Marketing
Communications: Purchases
• Online advertising • Online
• Offline advertising • Offline
• Promotions

Notes:

• Straight arrows represent contemporaneous effects or stock variable effects

(contemporaneous plus lagged effects) for advertising and for mobile access.
• Curved arrows represent lagged effects
• Dashed arrow represents effect of advertising and promotion on mobile app access.
 35

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

Offline Purchase Probability Online Purchase Probability

 36

Figure 4
Impact of Marketing Efforts on Access and the Impact of Access on Purchasing

Figure 4a: Simulated Impact of a Shock in Marketing Efforts

on the Number of Mobile App Accesses
0.025
Change in Number of Accesses

0.020

0.015

0.010

0.005

0.000
0 5 10 15
Week

Social Media Impulse Online Advertising Impulse Promotion Impulse

Figure 4b: Simulated Impact of a Shock in App Access

on Online and Offline Purchase Probabilities
0.005
Change in Purchase Probability

0.004

0.003

0.002

0.001

0
0 2 4 6 8 10 12 14
Week

Offline Purchase Probability Online Purchase Probability Total Purchase Probability

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

Variable Operationalization Source

AppAccesses Total number of times the mobile app was accessed in a particular week Company
A binary variable that is 1 for weeks where a purchase was made in any
Purchase Company
channel (in-store, online or mobile) and zero otherwise
A binary variable that is 1 for weeks where a purchase was made through
Purchase Online Company
the online (but not mobile) channel, and zero otherwise.
A binary variable that is 1 for weeks where a purchase was made in the
Purchase Offline Company
store, and zero otherwise
The number of individual, minor, and major promotion sale days in a
Promotions particular week. Multi-day promotions are counted by the number of days Company
that the promotion runs.
The total number of social posts on Twitter or Facebook in a particular Twitter,
SocialPosts
week that specifically refer to the mobile app. Facebook
OnlineAdvertising Total dollars spent on internet display advertising. TNS
Total dollars spent on Newspaper, Magazine, Radio, Television and
OfflineAdvertising TNS
Billboard advertising.
A binary variable is 1 if the individual is within a county adjacent to the
Location Company
county of the store.
The logarithm of the time, in weeks, since the specific user has
Log_TimeSinceDownload Company
downloaded the app.
A step variable that is 0 before an upgrade of the mobile app, 1 after the
Upgrade(1-4) Apple iTunes
upgrade. There are 4 upgrades of the app during our sample period.
A step variable that is 0 before an upgrade of the iOS operating system,
iOSChange(1-2) and 1 after the upgrade. There are 2 upgrades of iOS during the sample Apple
period.
Time variable which increases linearly with the introduction of the mobile
Trend Company
app
A variable which is the sum of the number of days in a particular week
Christmas Company
when the week falls between Thanksgiving and Christmas.
 39

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

Intercept 1.0512** .0423 24.84 .00

App-Related Marketing
SocialPosts .0143** .0053 2.70 .01
Consumer Factors
AppAccessStockt-1 .6326** .0027 236.85 .00
Purchaset-1 -.0373** .0031 -12.03 .00
Log_TimeSinceDownload -.1285** .0157 -8.20 .00
App Design Factors
Upgrade1 -.0390 .0323 -1.21 .23
Upgrade2 .1033** .0373 2.77 .01
Upgrade3 .0246 .0303 .81 .42
Upgrade4 -.0479* .0256 -1.87 .06
**
iOSChange1 -.2565 .0279 -9.20 .00
iOSChange2 -.0297* .0179 -1.66 .10
Marketing Communications
OnlineAdStock .0044** .0014 3.14 .00
OfflineAdStock .0000 .0002 -.03 .98
**
Promotions .0030 .0009 3.48 .00
Other Controls
Trend -.0012* .0007 -1.75 .08
Christmas -.0099** .0029 -3.36 .00
Fixed Effects Not reported
Number of Observations 83,549
R2 .74
**
: two-sided p-value < .05; *: .05 ≤ two-sided p-value < .10.
 42

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

Effect of Mobile App

AppAccessStockt .037** .014 2.54 .01 .058** .012 4.84 .00

Marketing
Communications

Promotions .011* .006 1.83 .07 -.004 .005 -.79 .43

OnlineAdStock -.003 .006 -.45 .65 .004 .004 1.04 .30

OfflineAdStock .001 .001 .92 .36 .000 .001 -.71 .48

Other Controls

Purchaset-1 .161** .080 2.02 .04 .167** .062 2.71 .01

Location .770** .065 11.78 .00 -.132** .053 -2.50 .01

Trend -.001 .002 -.93 .36 -.001 .001 -.85 .40

Christmas .004 .020 .19 .85 .026* .015 1.75 .08

Access Equation
.049** .024 2.00 .05
Control Function

Standard Deviation in
.836 .032 26.13 .00 .598 .024 24.92 .00
random intercept

Hit rate 97.4% 98.1%

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

Intercept 1.1536** .0415 27.80 .00

App-Related Marketing
SocialPosts .0139** .0053 2.61 .01
Consumer Factors
AppAccessStockt-1 .6328** .0031 204.34 .00
**
Purchaset-1 -.0375 .0153 -2.46 .01
Log_TimeSinceDownload -.1294** .0158 -8.21 .00
App Design Factors
Upgrade1 -.1118** .0356 -3.14 .00
**
Upgrade2 .1059 .0381 2.78 .01
Upgrade3 .0195 .0304 .64 .52
Upgrade4 -.0385 .0256 -1.50 .13
iOSChange1 -.2641** .0290 -9.10 .00
iOSChange2 -.0113 .0189 -.60 .55
Marketing Communications
OnlineAdStock (predicted) -.0001 .0020 -.06 .95
OfflineAdStock (predicted) -.0001 .0002 -.33 .74
Promotions .0032** .0009 3.66 .00
Other Controls
Trend -.0012* .0007 -1.81 .07
**
Christmas -.0086 .0029 -2.92 .00
Fixed Effects Not reported
Number of Observations 83,549
R2 .74
**
: two-sided p-value < .05; *: .05 ≤ two-sided p-value < .10.
 46

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