Mobile App Analytics:
A Multiple Discrete-Continuous Choice Framework
Sang Pil Han
Assistant Professor in Department of Information Systems
W.P. Carey School of Business, Arizona State University
BA 301D, Information Systems, Tempe, AZ, USA, Email: [email protected]
Sungho Park
Assistant Professor in Department of Marketing
W.P. Carey School of Business, Arizona State University
BAC 421, Information Systems, Tempe, AZ, USA, Email: [email protected]
Wonseok Oh
Chair Professor in School of Business
Korea Advanced Institute of Science and Technology
85 Hoegiro Dongdaemoon-Gu, Seoul, Korea, Email: [email protected]
Mobile App Analytics:
A Multiple Discrete-Continuous Choice Framework
Abstract
The number of mobile apps is exponentially growing to over 2 million, but little is known about
how users choose and consume apps in numerous categories. This study develops a utility
theory-based structural model for mobile app analytics. Specifically, based on the theoretical
notions of utility and satiation along with the factor analytic approach, we model users’ multiple
discrete app choice and continuous consumption decisions simultaneously in order to uncover
the complex relationships among choice, consumption, and utility maximization for numerous
mobile apps. Using a unique panel data set detailing individual user-level mobile app time
consumption, we quantify the baseline utility and satiation levels of diverse mobile apps and
delineate how users’ app preferences and consumption patterns vary across demographic factors.
The findings suggest that users’ baseline utility diverges substantially across app categories and
their demographic characteristics explain a substantial amount of heterogeneity in baseline utility
and satiation. Furthermore, both positive and negative correlations exist in the baseline utility
and satiation levels of mobile web and app categories. Our modeling approaches and
computational methods could open new perspectives and opportunities for handling large-scale,
micro-level data, while serving as important resources for big data analytics and mobile app
analytics in particular.
Keywords: mobile analytics, mobile web and apps, time use modeling, satiation,
interdependence, structural econometrics
2
Mobile App Analytics:
A Multiple Discrete-Continuous Choice Framework
INTRODUCTION
The mobile revolution has fundamentally changed the ways consumers engage with media and
businesses. A recent study by eMarketer (2014), a global digital market research company,
revealed that US adults spent more time with digital media than with TV in 2014 for the second
consecutive year.1 In particular, the prevalence of digital content and the rapid adoption of
smartphone and tablet devices helped augment digital time spent; on average, users spend 2
hours 51 minutes with mobile devices daily, surpassing the 2 hours 12 minutes they spend on
PCs each day (eMarketer 2014). The fuel driving mobile’s huge growth is primarily app usage.
According to Flurry, a global leader in mobile analytics solutions, average US mobile consumers
spend approximately 86% of their mobile time on apps in particular – 2 hours and 19 minutes per
day (Khalaf 2014).
The predominance of apps in the emerging mobile paradigm has been empowered by their
colossal growth and expansion. As of July 2014, Google had 1.3 million apps available through
its Android platform, while its competitor, Apple, allowed users to choose from 1.2 million apps
through its iTunes app store.2 Despite the pervasiveness of mobile apps in everyday economic
and social exchanges, not all apps are created equal. App demand (i.e., download count) follows
a Pareto distribution (Carare 2012), indicating that there are a small number of popular apps
while there are countless unpopular apps. Even after app download, certain apps (e.g.,
1
Digital media include all online, mobile and other non-mobile connected-device activities, such as video streamed
through over-the-top services (eMarketer 2014).
2
http://www.statista.com/statistics/276623/number-of-apps-available-in-leading-app-stores/
3
communication messengers, games, news/stock feeds) are used with great frequency, while
many others are deployed only once or twice in their lifespans and then quickly disappear into
the void forever.
As mobile users are increasingly inundated with an overabundance of apps, they adopt a
variety of choice mechanisms to sift through the multiplicity of alternatives and manage their app
consumption effectively to maximize utility given time constraints. However, little is known
about the changing dynamics of the relationship between users’ app choice and consumption
decisions and their utility maximization. The plethora of apps and the relentless usage of apps
generate the mammoth amount of data in the form of social media exchanges, purchase
transactions, music downloads, car navigations, stock investments, location alerts, and search
queries. In an app-based, networked socioeconomic environment, every text, every transaction,
every digital process, and literally every touch and move through apps can become a data point.
Mobile app analytics tracks and measures such app usage data to study the behavior of
mobile app users. In the initial stages, the number of downloads was of key interest. However, as
mobile app analytics has gone through a maturing process, businesses have taken a great interest
in the direct, quantitative measurement of usage, habits and engagement.3 For example,
Google’s Mobile App Analytics solution measures user behaviors from first app discovery and
download to in-app purchases. It offers real-time reporting about where users take action, pause,
or disappear using event tracking and flow visualization, providing insights about app user
behaviors.4 However, methods and results from industry mobile app analytics solutions (e.g.,
Google’s Mobile App Analytics, Yahoo’s Flurry Analytics, and Countly’s Mobile Analytics) are
generally descriptive in nature and neglect interdependence in the usage of various mobile apps.
3
4
https://count.ly/resources/Countly_Mobile_Analytics_101_Whitepaper.pdf
http://www.google.com/analytics/mobile/
4
These analytics have therefore been difficult to relate to quantitative measures in utility theories
in order to understand users’ app usage within a utility maximization framework. Understanding
the underlying mechanisms of users’ app consumption has important implications for firms’
strategies and in particular for mobile app monetization, mobile consumer engagement and
mobile media planning. According to Flurry, between 2010 and 2012 approximately 80 percent
of apps were free, but by 2013, 90 percent of apps in use were free (Gordon 2013). As app
developers and publishers migrate from paid downloads to in-app purchase and ad-supported
business models, they have started shedding more light on how much time their users spend in
their apps. More time spent in an app is likely to generate additional in-app advertising revenues
and purchases. In addition, brands use mobile apps as a communication and transaction touchpoint to interact with their customers. Extra time spent in a brand’s app enhances consumer
engagement with the brand. Furthermore, as advertisers increasingly spend more on mobile
advertising, they aim to select optimal mobile media vehicles – mobile apps and websites – to
maximize their advertisement exposures and thereby to stimulate consumers’ economic
behaviors. According to Flurry, “in the world of advertising, time-spent by consumers is the
timeless currency (Khalaf 2014).” Thus, it is critical for businesses in the app-based networked
economy to understand how users determine their time use on apps.
Understanding users’ app time-use is complicated and constrained because of user
heterogeneity. In particular, user demographics affect intrinsic preferences for apps and the
marginal utility derived from app consumption. Since behavioral heterogeneities on IT usage
have been frequently observed across diverse demographic attributes (e.g., Taylor and Todd
1995), the choice and consumption patterns regarding apps are expected to vary across age,
gender, income, and educational levels. However, no comprehensive, systematic analytics has
5
yet to explore how demographic characteristics influence the consumption of numerous mobile
apps and their implications for individual utility trajectories. Furthermore, some apps are often
utilized in chorus in order to increase user experience and utility, while others replace each other,
competing for volume and frequency of use in the battle of the substitutes.
Although the number of apps is increasing at an unprecedented rate, the lack of validated
empirical schemes, robust computational frameworks and analytics, and actual usage data in
large quantity has plagued our understanding of app usage patterns and their interdependence
and competition. To systematically explore the patterns of consumption dependency for a wide
selection of apps, this study builds on the work of Bhat (2005) to develop a utility theory-based
structural model for multiple discrete/continuous choices in app use. The underlying mechanism
assumes that a consumer’s marginal utility diminishes as the level of consumption of any
particular app increases — a phenomenon known as satiation. To provide empirical insights into
our utility-based choice paradigm, we use a unique panel data set detailing individual user-level
mobile app consumption. Specific research questions include: What characterizes the baseline
marginal utility of a particular app category when no app is consumed? To what extent does
marginal return diminish as the consumption of a particular category of apps increases? How do
the baseline utility and satiation (i.e., diminished marginal utility) levels vary across different
user demographics? Finally, how do we estimate app category (dis)similarity in unobserved
attributes when the number of app categories is large (e.g., dimensionality issue)? Answering
these questions within a rigorous theoretical framework with empirical validation can enhance
our understanding of the baseline utility and satiation levels of numerous mobile apps in diverse
categories.
6
The frameworks and methods derived in this study may have several important contributions
to the emerging field of mobile analytics, which has yet to coalesce into a clearly defined
approach that simultaneously incorporates both consumer choice and utility. Unlike the typical
discrete choice models in which a single option is chosen, the multiple discrete-continuous
choice approach allows multiple alternatives to be selected in tandem. The diminishing marginal
utility provides a horizontal motivation for multiple discrete purchases. That is, if consumers
face diminishing marginal utility for each of several potential apps with comparable
attractiveness, then several alternatives will be chosen and evaluated simultaneously. We regard
the diminishing marginal utility as a manifestation of satiation.
Furthermore, our approach jointly models the app choice and consumption time decisions
within a single utility framework. For mobile app users, decisions regarding which apps to use
and how extensively to consume them are closely related since both questions are influenced by
the same factors (e.g., time constraints and the app developer’s marketing activity such as
offering a one-week free trial). Modeling the two decisions in isolation (i.e., using a multivariate
Logit model for the choice decision and regression models for the quantity decisions) may not be
accurate if the quantity decision is not statistically independent of the choice decision and vice
versa. The separate modeling approach is prone to resulting in biased and inconsistent regression
parameter estimates and inefficient Logit parameter estimates (Tellis 1988; Krishnamurthi and
Raj 1988). The bias occurs because the regression models omit a relevant variable, and the
inefficiency in the Logit model arises because the information provided by the data about
continuous use time is ignored. Therefore, these modeling challenges reinforce the need for a
new methodological paradigm for big data analytics in the context of mobile app markets where
consumers face numerous choices and need to make effective consumption decisions.
7
One of the key contributions our approach makes to mobile analytics is that it incorporates
factor analytic structures into the multiple discrete-continuous choice framework. The proposed
structural techniques and processes enable us to estimate correlations in both baseline utilities
and satiation levels of mobile app categories in a parsimonious and flexible way. Further, we
have extended our model from app category-level to app-level and found that the proposed
method remained robust. In addition, the framework can be applied to modeling individual userlevel time-use of various IT artifacts and their interdependence using large-scale, micro-level
data. Furthermore, this modeling approach can reduce the computational inefficiencies inherent
in estimating unobserved heterogeneity and interdependence when the number of app categories
is exceedingly overwhelming (e.g., dimensionality issue). Consequently, our methods and
processes could serve as important resources for big data analytics and open new perspectives for
mobile app analytics.
In addition to estimating the baseline utility and satiation related to app consumption, we
examine app time-use interdependence among apps in diverse categories by considering, for
example, whether consuming social network apps will increase or decrease the use of
communication apps or vice versa. Recently, Facebook (SNS category) acquired WhatsApp
(communication category) in a landmark deal for $19 billion to cement its position and create a
positive synergy. This corporate coupling further expands and diversifies Facebook’s existing
strategic app portfolio, which includes Instagram (photo category), Spaceport (game category),
Jibbigo (utility category) and others. We illustrate the strategic value of our quantitative
paradigm and mechanism to maximize the consumption of both apps through such portfolio
strategies. Finally, we offer managerial guidelines for establishing and implementing optimal
mobile media strategies in the app-based economy.
8
THEORETICAL BACKGROUND
User Behavior in Mobile Platforms and Apps
Research on mobile platforms and apps has proliferated commensurate with their increasing use.
Ghose and Han (2011) investigate user behavior on the mobile internet by mapping the
interdependence between the generation and usage of mobile content. They find that a negative
temporal interdependence exists: The more the mobile content is consumed in the previous
period, the less the content is generated in the current period or vice versa. Ghose et al. (2013)
report that the influence of ranking and geographical proximity are more pronounced on mobile
devices than PCs. Recently, Einav et al. (2014) find that adoption of the mobile shopping
application is associated with both an immediate and sustained increase in total platform
purchasing.
Our study also builds on an emerging stream of literature on mobile apps. For example,
using a reduced-form model and data from Apple’s App Store, Carare (2012) reveals that app
consumers are willing to pay an additional $4.50 for top ranked apps, but their preference
towards the apps with bestseller status declines sharply for top ranked products. Recently, Xu et
al. (2014) demonstrate that the introduction of a mobile app by a major national media company
leads to a significant increase in demand at the corresponding mobile news website. Using
publicly available data from Apple’s App Store, Garg and Telang (2013) discover that topranked paid apps available for iPhone elicit 150 times more downloads than other apps ranked in
the top 200. Ghose and Han (2014) find that app demand increases with the in-app purchase
option wherein a user can complete transactions within the app.
9
Based on a thorough review, we found that the current literature on mobile apps focuses
exclusively on demand in the form of downloads or paid purchases, but pays scant attention to
the actual choices and consumption patterns of mobile apps.
Multiple Discrete-Continuous Choice Models
Several analytics have been proposed to capture the dependence between choice and quantity
decisions: the single utility (or structural) approach (e.g., Hanemann 1984; Chiang 1991;
Chintagunta 1993; Kim et al. 2002) and the error-dependence (or reduced form) approach (e.g.,
Tellis 1988; Krishnamurthi and Raj 1988; Zhang and Krishnamurthi 2004). The former approach
specifies a utility function and assumes that the optimal choice and quantity are derived as an
equilibrium solution from the utility function. The dependence between choice and quantity
decision is captured in the utility function. The latter approach handles the dependence by
allowing correlations in error terms of the choice and quantity models. A major advantage of the
single utility approach is that it allows researchers to estimate structural parameters and metrics
of economic interest (e.g. compensating variation). The proposed model based on the multiple
discrete-continuous choice approach belongs to the single utility approach.
Mobile users' app usage decisions can be decomposed broadly into two elements - which
apps to select and how extensively to consume them. Since multiple discrete-continuous choice
models tackle both problems within a single utility maximization framework, they are
appropriate for analyzing our mobile app and web time-use data. Kim et al. (2002) propose a
translated nonlinear, addictive utility model in which a parsimonious specification provides both
corner and interior solutions in the context of the simultaneous purchase of multiple varieties.
The multiple discrete-continuous extreme value (MDCEV) model formulated by Bhat (2005,
10
2008) extends single discrete-continuous frameworks (e.g., Dubin and McFadden 1984;
Hanemann 1984; Chiang 1991; and Chintagunta 1993) to handle multiple discreteness in demand
and resolve the presence of heteroscedasticity and correlations that arise from unobserved
characteristics. Among various applications, MDCEV models have frequently been used for
analyzing time-use data. Bhat (2005) scrutinizes time-use allocation decisions among several
discretionary activities on weekends. Building on Bhat’s MDCEV time-use model, Spissu et al.
(2009) study within-subject variation over the 12-week sample period for six broad activity
categories along with another activity. Luo et al. (2013) incorporates dynamic components into a
MDCEV model to examine how consumers allocate time to a portfolio of leisure activities over
time.
In this study, we develop a unique structural model of app selection and time-use decision
by incorporating a factor analytic structure into a MDCEV framework. The vectors of individuallevel baseline utilities and satiation parameters are modeled as functions of observed mobile user
characteristics and a small number of unobservable user-specific factors. In literature, factor
analytic structures are combined with Probit or Logit models and the primary focus of these
models is to understand inter-brand competition by pictorially depicting locations of competing
brands in a perceptual map (Chintagunta 1994; Elrod and Keane 1995). Our approach offers a
methodological contribution to MDCEV models by allowing correlations in both baseline
utilities and satiation levels of various mobile app categories in a parsimonious manner using
factor analytic approaches.
11
EMPIRICAL BACKGROUND AND DATA DESCRIPTION
Mobile App and Web Time-Use Panel Data
We provide a brief overview of the empirical background for our data. We have gathered largescale panel data comprising mobile app and web time-use histories provided by Nielsen
KoreanClick, a market research company that specializes in consumers’ Internet and mobile
usage. Audience measurement gauges how many users are in an audience and how long they
remain, usually in relation to television viewership (e.g., Nielsen ratings), but also in relation to
increasingly traffic on websites and mobile apps. Nielsen KoreanClick maintains a panel of
mobile app users with Android operating system-based devices, aged 10 to 70, who are selected
based on stratified sampling techniques. Android is the dominant operating system of mobile
devices worldwide accounting for 67.5 percent of the global market. In Korea, nearly 93 percent
of smartphones are powered by the Android platform (Yonhap News 2014).
After individuals voluntarily agree to be panel members, they download and install a meter
application from Nielsen KoreanClick on their mobile devices. The participating users are
rewarded with points for installing the meter app and the incentive points can be accumulated
and redeemed for gift cards. This app runs in the background and collects data on panel
members’ use of mobile apps and the mobile web even while disconnected.5 The meter app
regularly transmits encrypted log files to a server via a secure cellular connection or Wi-Fi. In
addition, we acquire individual user-level information on user demographics such as age, gender,
monthly income, and education.
5
Mobile web refers to the collective term for websites accessed from mobile devices using browsers. So the mobile
web is often interchangeably used with mobile browser.
12
We collected the data between March 5 and April 30 2012 (eight weeks). These data include
1,425 panel members who used mobile apps and the web throughout the sampling period.
Moreover, our data incorporate individual-level, weekly information on the type, name, and
duration of mobile apps used and mobile websites visited. Nielsen KoreanClick classified the
mobile content into 14 categories; communication, game, map and navigation, entertainment,
lifestyle, personal finance, music and radio, photo, portal, schedule and memo, social
networking, utility, video, and combined mobile web activities.6 Table 1 demonstrates that panel
members in our sample used 11,063 apps and visited 7,944 websites.7 Apps in our sample
include not only top global mobile apps such as Facebook, YouTube, Twitter, and Skype, but
also top local mobile apps such as Kakao Talk, Naver, and Cyworld.
We have 11,400 (1,425 users  8 weeks) app choice occasions in our data. Column 1 of
Table 2 shows that users access the mobile web most frequently (99.7%), followed by
communication apps (99.3%), schedule/memo apps (98.7%), and utility apps (97.7%), but access
entertainment apps least frequently (38.0%). In addition, Column 2 of Table 2 reveals that
smartphone users in our sample spend an average of 1 hour and 47 minutes every day on
consuming content in their mobile devices. Daily time spent in mobile apps surpasses mobile
6
Google Play is a leading app market based on Android operating system. Notably when publishing a new app or a
new version of the existing app in Google Play, app developers self-select one or more appropriate app categories,
however, they are not required to go through the verification process. Hence there are some cases in which app
categories reported by app developers are incorrect or inconsistent. To address this issue, Nielsen KoreanClick
performed a thorough, manual re-classification task to ensure that a certain app is classified to the single, primary
app category. We use the Nielsen KoreanClick's categorization in our empirical analysis. If the categorization of
apps were arbitrary, it would be difficult to find meaningful results from the empirical analysis. However, our
empirical analysis based on the Nielsen KoreanClick's categorization show highly significant and meaningful results
(we will discuss the result in detail in Results section), adding empirical validity to the categorization. Of course, we
can adopt alternative categorizations in the proposed empirical framework. In Results section, we also perform an
empirical analysis using individual app level data.
7
The number of mobile websites visited was aggregated to the domain level. That is, multiple URLs with the same
domain name were counted as one. For example, two different URLs with the common domain name https://www.google.com/maps/preview and https://www.google.com/calendar/render- were treated as one website
when we counted the number of websites visited.
13
web consumption. To be specific, Figure 1 indicates that users spend the most time (25%) with
communication apps, such as mobile messaging, followed by mobile web (18%), game apps
(12%), music and radio apps (9%), social network apps (6%), and utility apps (6%). In Results
section, we show that our findings of baseline utilities and satiation levels for app categories
based on the proposed model are quite different from those reported in Table 2 and Figure 1.
Table 1. Mobile Content Categories: Mobile Apps and Mobile Websites
Category
Count
Communication
Mobile Messengers, Mobile
Internet Phone, Email
Entertainment
Book, Cartoon, Adults,
Sports, Travel, Humor,
Magazine
Action, Adventure, Board,
Puzzle, Racing, Role Playing,
Shooting, Simulation, Sports
Map, Navigation
1,071
Weather, News, Education,
Restaurants, Job, Health,
Religion, Fashion
Banking, Stocks, Finance,
Real Estate, Ecommerce
1,808
Game
Map and Navigation
Lifestyle
Personal Finance
247
2,478
337
407
Music and Radio
Radio, Music
304
Photo
Photo Gallery, Camera
264
Portal Search
Portal Site, Search Engine
83
Schedule
Scheduler, Memo, Alarm
Clock
Social Networking Service,
Board, Blog, Microblog
Productivity, Decoration,
Webhard, Widget, Firewall
Multimedia, Broadcasting,
Movie
Websites
1,399
Social
Utilities
Video
Mobile Web
Sum
163
2,241
261
7,944
Example
Kakao Talk, Mypeople Messenger, Phone, GO
SMS Pro, LINE, NateOn UC, LightSMS, Gmail,
LightSMS, Viber, Skype
Naver Webtoon, Live Scores, TIViewer, T store
Book, jjComics Viewer, Naver Books, Score
Center
Rule The Sky, TinyFarm, Smurfs' Village, Shoot
Bubble Deluxe, Hangame, 2012 Baseball Pro,
Angry Birds Space, Jewels Star
T map, Google Maps, SeoulBus, Naver Maps,
Olle Navi, Subway Navigation,
YTN News, MK News, SBS News, Weather,
Bible, Newspapers, JobsKorea
Smart Trading, M-Stock Smart, KB Star
Banking, Auction Mobile, Coupang, Gmarket
Mobile
Music Player, SKY Music, PlayerPro, Mnet,
MyMusicOn, FM Radio, Soribada
Gallery, Camera, Cymera, Photo Editor,
Instagram
Naver, Daum, NATE, Google Search, Junior
Naver
Address book, Alarm/Clock, Calendar, Memo,
Polaris Office, Polaris Office, Docviewer
Kakao Story, Facebook, Twitter, Naver Café,
Daum Café, Cyworld, Naver Blog, Me2day
Calculator, Voice Recorder, HD Browser,
NDrive, Smart App Protector, Battery Widget
TV, Youtube, MX Video Player, SKY Movie,
Afreeca TV, T-DMB, PandoraTV
Naver.com, Daum.net, Nate.com, Google.co.kr,
ppomppu.co.kr, dcinside.com, facebook.com
19,007
14
Table 2. Choice and Time Use According to App Categories
Categories
Choice
Outside Option (other activities)
Communication
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music/Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
Total
Web
18%
Time Use
100.0%
99.3%
65.0%
69.6%
38.0%
68.5%
61.4%
67.5%
90.3%
71.3%
98.7%
69.6%
97.7%
77.7%
99.7%
Hour
Percent
22.23
0.44
0.21
0.03
0.03
0.03
0.04
0.16
0.04
0.07
0.1
0.11
0.11
0.08
0.31
24
92.6%
1.8%
0.9%
0.1%
0.1%
0.1%
0.2%
0.7%
0.2%
0.3%
0.4%
0.4%
0.4%
0.4%
1.3%
100.00%
Communication
25%
Video
5%
Utility
6%
Social Network
6%
Schedule/Memo
6%
Personal
Portal
Photo Music/RadioFinancing
4%
2%
2%
9%
Game
12%
Map/Navigation
1%
Entertainment
2%
Lifestyle
2%
Figure 1. Time Use According to App Categories Excluding Outside Option
15
Model-Free Evidence of Dependence between Mobile Content Choice and
Time Use Decisions
As mobile users increasingly access mobile web and various kinds of apps, their choices become
interdependent. Figure 2 demonstrates that all users use at least two categories of mobile content
during a given week. Approximately 98.5% of users in our data access more than four categories
of mobile content in a given week. These descriptive findings of the joint use of multiple
categories of mobile content lend support to the validity of our econometric model in which we
incorporate the multiple-discrete choice into the continuous time-use decision.
19.7% 19.0%
20.0%
15.6%
14.9%
15.0%
10.3%
10.0%
6.4%
5.0%
0.0%
6.4%
3.3%
1.8%
0.5%0.4%0.6%1.1%
0.0%
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Figure 2. Distribution of Number of App Categories Jointly Used
It is more challenging to make inferences on use time decision across multiple app
categories using simple methods. We do not observe usage time for app categories that are not
selected. As Figure 2 indicates, all categories of mobile content are used in only 6.4% of our
16
weekly data. Therefore, to use simple metrics like a correlation matrix, we should aggregate the
observed weekly data up to monthly or quarterly level or discard observations with zero-time use
observations. To impute such an incidence with non-app use, a correlation matrix of app choice
incidence can be employed using dummy variables which take the value of one if the app is used
and zero otherwise. Similarly, a correlation matrix of app time-use quantity can be established by
imputing no incident with zero value. These approaches are inferior to the proposed approach
because they do not fully leverage the available information.
Tables 3 and 4 show the correlation matrix of app choice incidence variable and the
correlation matrix of app time-use quantity variable, respectively. The observed relationships
among app use incidents are mostly positive and marginal. We observe substantial positive
relationships among communication, utility, and schedule/memo categories. However, we cannot
find any substantial negative correlation. In Results section, we demonstrate that our findings
derived from the proposed model are drastically different from those reported in Tables 3 and 4,
indicating that the simple methods can lead to misleading results in the evaluation of correlation
across mobile content categories. We discuss our modeling approach in detail below.
17
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music & Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
1.00
0.08
0.12
0.06
0.11
0.09
-0.01
0.23
0.11
0.54
0.11
1.00
0.04
0.15
0.09
0.04
0.07
0.10
0.10
0.10
0.08
1.00
0.11
0.20
0.17
0.14
0.16
0.14
0.14
0.11
1.00
0.15
0.06
0.16
0.09
0.12
0.07
0.12
1.00
0.15
0.13
0.18
0.17
0.13
0.14
1.00
0.05
0.11
0.14
0.13
0.11
1.00
0.13
0.11
0.05
0.13
1.00
0.18
0.26
0.22
1.00
0.16
0.13
1.00
0.11
1.00
0.41
0.14
0.00
0.13
0.14
-0.03
0.17
0.14
-0.03
0.10
0.13
-0.05
0.16
0.16
-0.03
0.15
0.09
-0.03
0.08
0.14
-0.03
0.26
0.22
-0.01
0.16
0.16
-0.03
0.41
0.16
0.00
0.11
0.12
-0.03
1.00
0.18
-0.01
1.00
-0.02
1.00
Web
Utility
Video
Web
Game
Communication
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music & Radio
Photo
Portal
Schedule/Memo
Social Network
Communication
Table 3. Correlation Matrix of App Choice Incidence
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music & Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Communication
1.00
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music & Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
0.00
0.03
0.03
0.05
0.01
0.10
0.35
0.09
0.15
0.24
0.08
0.04
1.00
0.02
0.03
0.04
-0.01
0.03
-0.01
0.04
-0.01
0.00
0.04
0.02
1.00
0.03
0.04
0.02
0.00
0.00
0.02
0.08
0.00
0.04
0.03
1.00
-0.01
-0.02
0.03
-0.02
0.06
0.02
0.04
0.01
0.04
1.00
0.08
0.07
0.01
0.08
0.08
0.02
0.08
0.07
1.00
-0.01
-0.02
0.07
0.06
-0.02
0.04
0.02
1.00
0.05
0.06
0.03
0.08
0.01
0.08
1.00
0.06
0.04
0.25
0.03
0.02
1.00
0.09
0.11
0.06
0.09
1.00
0.03
0.09
0.06
1.00
0.02
0.05
1.00
0.02
1.00
Web
0.03
-0.04
0.00
0.00
0.06
0.04
0.01
0.08
0.08
0.01
0.05
0.03
0.09
Communication
Game
Table 4. Correlation Matrix of App Use Time
18
1.00
ECONOMETRIC MODEL
In this section, we present our proposed model to estimate the baseline utility and satiation levels
of different categories of mobile web and apps while allowing for user heterogeneity and crossapp use interdependence even when the number of app categories is large. We then discuss how
we identify our demand system.
Proposed Model
Consumer Utility Function
We observe which app categories are chosen and how much time is spent on each selected app
category in our data. Accordingly, we describe a mobile user’s behavior by virtue of a multiple
discrete/continuous process. Compared to discrete choice models (i.e. Logit or Probit models)
and continuous dependent variable models, multiple discrete/continuous models can handle more
effectively both discrete and continuous natures of observed data within a single utility-based
framework. Because of this advantage, multiple discrete/continuous models have successfully
been applied in several academic fields, including marketing, transportation, and economics
(Kim et al. 2002; Hendel 1999; Bhat 2008). In this paper, we extend Bhat’s (2008) multiple
discrete/continuous extreme value (MDCEV) framework. We specify the latent utility of mobile
content usage as follows:
Uht  10  exp( h 0t )  qh00t   j 1 1hj  hjt {( qhjt  1)
J
 hj
 1},
(1)
where h  1,...,H denotes individual mobile users, j  0 and j  1,..., J denote an outside option
(activities other than mobile content use) and mobile content use categories, respectively, and
t  1,...,T denotes time period (weeks). q hjt ( j  0,..., J ) is time allocated to alternative j by user h
19
in time period t. This specification is referred to as “alpha-profile” in the literature (see Bhat
(2008) for further details on model specification and other possible specifications).  h 0 t is an
user-, alternative-, and time-specific random term associated with the outside option.8 hjt
represents the “baseline marginal utility” (the marginal utility when none is consumed) of
alternative j in time t by user h. When a user decides which category to use first, categories with
large value of hjt have higher probabilities of being selected compared to those with small hjt .
Also, it can be interpreted as a measure of “perceived quality” because higher values of hjt mean
that the alternative confers higher levels of utility from any level of consumption, all else equal.
In addition,  0 and  hj are referred to as satiation parameters in that they determine how the
marginal utility of alternative changes as their consumption quantity increases. To illustrate the
role of the satiation parameter, assume that there are only two choice options – the outside option
(j= 0) and alternative 1 (j= 1) – and that the satiation parameters are fixed to one (i.e.,  0  1 and
h 1  1 ). Then, the utility function (1) becomes Uht  exp( h 0t )  qh 0t  h 1t  qh 1t . The marginal
utilities of the outside option and the alternative 1 are equal to exp( h 0t ) and h1t , respectively.
It should be noted that the marginal utility values do not change with consumption quantities,
resulting in a linear indifference curve (see case (a) in Figure 3). In this case, user h allocates all
her time to the category with the highest perceived quality. If
exp( h 0t )   h1t
( exp( h 0t )   h1t ), as shown in Figure 3(a), the user can attain the highest level of utility by
spending all her time for the outside option (alternative 1) and thus select the outside option
8
The utility function can be classified as a generalized variant of the translated constant elasticity of substitution
(CES) utility function. This specification assumes additively separable utility. We discuss a limitation of this model
specification in detail later in Conclusion. Note that the utility specification of the outside option (the first term in
the right hand side of (1)) is different from those of the other alternatives and this is for the utility function to capture
that the outside option is always consumed.
20
(alternative 1) only. That is,
qh 0 t  Q
and
qh1t  0
( qh 0 t
0
and
qh1t  Q ).
Note that this linear
indifference curve implies that the two alternatives are perfect substitutes.
q h 1t
Uht  exp( h 0t )  qh 0t  h 1t  qh 1t
q h 1t
U ht  10  exp( h 0t )  qh00t
indifference curve
qh 0t  qh 1t  Q
Q
budget line
Q
 1h 1  h 1t {( qh 1t  1) h 1  1}
indifference curve
Q
qh 0t  qh 1t  Q
qˆh 1t
budget line
qh 0t
(a) Two alternatives / No satiation
qˆh 0 t
qh 0t
Q
(b) Two alternatives / Satiation
Figure 3. An Example of How Diminishing Marginal Utility (Satiation) Affects Multiple
Discrete Choices
Now, let us consider a more realistic case where  0  1 and  hj
 1.
Then the utility
function (1) shows diminishing marginal utility. As  hj decreases, the utility function in
alternative j shows more concave patterns, and higher satiation occurs at a lower value of
q hjt .
Due to this diminishing marginal utility, multiple alternatives can become comparable to each
other and they will be chosen together rather than only one option is selected (see case (b) in
Figure 3). In this case, user h consumes the outside option and alternative 1 by
and
ˆ h1t  0 ,
qh1t  q
qh 0t  qˆ h 0t  0
respectively.
According to Bhat (2008), U ht becomes a proper utility function when hjt
0
and  hj
1
for j=1,…, J. To ensure that the baseline utility is non-negative and the satiation parameter is less
21
than 1 regardless of the values of  hj , hj , and  hjt , we specify the baseline utility parameter hjt
and the satiation parameter  hj as the following:
hjt  exp( hj   hjt ), for j  1,..., J ,
(2)
 hj  1  exp( hj ), for j  1,..., J .
 hjt represents idiosyncratic elements in utility. Both  h 0 t and  hjt are known to decision makers,
but unknown to researchers. We assume that these follow Type-I Extreme Value distribution.
User Heterogeneity and Correlation among Mobile Web and App Categories
For user- and alternative-specific  hj , we specify the following factor analytic structure:
 h      Dh    h    h ,
where a ( J  1) vector  h
 [  h 1 ,  h 2 ,...,  hJ ] and 
(3)
is a ( J  1) constant vector. Dh is the ( K  1)
vector of observed demographic variables of user h and

is a ( J  K ) coefficient matrix.

is
a ( J  F ) factor loading matrix and  h is a ( F  1) vector of orthogonal Gaussian factors
( h ~ N ( 0, I F ) ).
  is
a ( J  J ) diagonal matrix and  h is a ( J  1) vector of independent unit-
variance Gaussian random variables (  h
~ N ( 0, I J ) ).
The proposed specification decomposes
individual-level heterogeneity in mobile content choice utility into three parts. The first is what
the observed demographic variables explain. The second part is explained by parsimonious
common factors. In factor analysis literature,  h is referred to as a “common factor”. F elements
in  h influence all J elements in  h . We can understand the main characteristics of the factors by
interpreting the factor loading matrix
 .
Also, it should be noted that, along with
  Dh
, these
common factors generate correlations in  h . The remaining variation in  h is explained by the
22
last term in equation (3),
  h ,
which is referred to as a “specific factor.” Unlike  h , j-th
element in  h influences  hj only.
The factor analytic structure is of interest in our empirical setting for a number of reasons.
First, a factor model allows us to estimate category similarity in unobserved attributes (Elrod and
Keane 1995). We can potentially interpret the factors as inherent mobile users’ traits. Moreover,
the user-specific factor estimates can be used for targeting purposes. Second, the factor model
introduces correlations in the latent baseline utilities across app categories with relatively few
parameters. This characteristic is particularly useful in our context to alleviate the concern for
dimensionality issues inherent in estimating unobserved heterogeneity and their interdependence
when the number of app categories (i.e., alternatives) is exceedingly large. We can reduce the
number of parameters required to estimate a full covariance matrix while remaining highly
flexible and minimizing loss of information. Because of these major advantages, researchers
have applied factor analytic structure in random coefficient Logit or Probit models (Elrod and
Keane 1995; Singh et al. 2005; Hansen et al. 2006). A key methodological contribution of our
proposed model is that it extends factor analytic structure to multiple discrete/continuous models.
From equation (3), we can derive the following covariance matrix of  h :
Cov(  h )     D         ,
(4)
where D is a covariance matrix of Dh . The variance decomposition of equation (4) allows us to
quantify the relative contribution of each part. For example, the proportion of variation in  hj
explained by observed demographic variables is (j-th diagonal element of
   D 
)/(j-th
diagonal element of Cov(  h ) ).
Further, the value of user- and alternative-specific satiation parameter  hj is determined by
23
hj . Similar to equation (3), we specify the following factor analytic model structure to hj :
h      Dh    h     h ,
where ( J  1) vector is h
 [ h 1 , h 2 ,...,hJ ],
(5)
and  is a ( J  1) constant vector.   is a ( J  K )
coefficient matrix.  is a ( J  F ) factor loading matrix and  h is a ( F  1) vector of orthogonal
Gaussian factors (  h ~ N ( 0, I F ) ).   is a ( J  J ) diagonal matrix and  h is a ( J  1) vector of
independent Gaussian random variables (  h
~ N ( 0, I J ) ).
The covariance matrix of h can be
decomposed as follows:
Cov( h )    D         .
(6)
Model Estimation and Identification
Constrained Utility Maximization and Estimation
We derive our demand system by applying the Kuhn-Tucker method to the latent utility of
mobile content use specified in equation (1). By solving the Kuhn-Tucker conditions for
constrained utility maximization, we obtain demand functions wherein a mixture of corner
solutions and interior solutions are a product of the underlying utility structure. The Lagrangian
for the constrained utility maximization problem is given by:
Lht  10  exp( h 0t )  qh00t   j 1 1hj  hjt {( qhjt  1)
 hj
J
 1}   ht ( Q   j 0 qhjt ) ,
J
(7)
where  ht is the Lagrange multiplier, and Q denotes total amount of time given to each mobile
user (i.e. 24 hours per day). The Kuhn-Tucker first order conditions can be derived as follows:
 hj 1
  ht  0, if qhjt  0,
 hj 1
  ht  0, if qhjt  0,
hjt  ( qhjt  1)
hjt  ( qhjt  1)
(8)
24
and
exp( h 0t )  qh 0t  0 1   ht  0
since the outside option is always consumed (i.e. time used for
activities other than mobile app and web uses is non-zero). We use the expression for  ht from
the first-order condition for the outside option to eliminate the Lagrange multiplier form (7) and
then take log in both sides, so the Kuhn-Tucker conditions for the interior and corner solutions
can be written, respectively, as:
 hj  ( hj  1)  ln( q hjt  1)   hjt  ( 0  1)  ln( q h 0t )   h 0t , if q hjt  0,
(9)
 hj  ( hj  1)  ln( q hjt  1)   hjt  ( 0  1)  ln( q h 0t )   h 0t , if q hjt  0.
From this Kuhn-Tucker first order condition, we derive the following probability that any M
of the J alternatives is chosen (see Chapter 4 of Bhat (2008) for detail):
Pht ( q h 0t , q h 1t ,...,q hMt ,0,0,...,0 )

h
 M 1   hj
M !    j 0

q hjt   j

where Vh 0t  (0  1)  ln( qh 0t ) , Vhjt
  M q hjt   j

  j  0 1   hj
 




 e
 e
 hj  (hj  1)  ln( qhjt  1)
M
Vhit
i 0
J
j 0
Vhjt

M 1
(10)
dF (  h ),
for j=1,...,J,  0  0 , and  j
1
for
j=1,...,J. h is a vector of common factors and specific factors in  hj and  hj ( h ,  h ,  h , and
 h ) and F ( h ) is a joint distribution function of h . We use Monte Carlo simulation methods
to calculate the probability (10) and estimate the model parameters by maximizing a likelihood
function derived from equation (10) (see Keane 1993 for detail). To compute the integrals in the
likelihood function, we generate random normal draws for  h ,  h ,  h , and  h from their
distributions, and then take averages of computed integrands. In our application, we use 1,000
random draws. The resulting estimator becomes a simulated maximum likelihood (SML)
estimator. This procedure is the same as maximum likelihood except that simulated probabilities
are used instead of the exact probabilities. Researchers (e.g., Keane 1993) have articulated
thoroughly the properties of SML (i.e. its consistency, efficiency, and asymptotic normality).
25
Identification
Bhat (2008) discusses the identification issues of general MDCEV models and shows the
empirical identifiability of "alpha-profile" specification, which is adopted in the proposed model.
A distinctive feature of our model is that it incorporates the Gaussian factor analytic structure
into a MDCEV model. It should be noted that our Gaussian factor specification is the most
general and widely used specification (Elrod and Keane 1995; Singh et al. 2005; Hansen et al.
2006), but other distributional assumptions can be also used. For the identification of model
parameters, appropriate restrictions on factor loading matrices are required. Following Singh et
al. (2005) and Hansen et al. (2006), we impose a standard triangular restriction on a loading
matrix. It is widely known that this approach imposes the minimum restriction for the parameter
identification. More specifically, with two factors, one of the elements in the second column of
factor loading matrix is restricted to be equal to zero. With three factors, a 3 3 submatrix of
factor loading matrix is lower triangular. We estimate one-, two-, and three-factor versions of the
model specified in equations (3) and (6). The log-likelihood values for the one-, two-, and threefactor models are 97,128, 97,926, and 98,040, respectively. To determine the number of factors,
we use the Bayesian Information Criterion (BIC). The BIC values for the one-, two-, and threefactor models are -191,631, -192,984, and -192,979, respectively. In our case, the two-factor
model is chosen. The estimation results of one- and three-factor models are available upon
request. Accordingly, we report and discuss the results for the two-factor model below.
26
RESULTS
Baseline Utility and Satiation Levels for Mobile Web and App Categories
Tables 5 and 6 show the estimates for the baseline utility parameters (  and
 )
and the
satiation parameters (  and   ), respectively. The results in Table 5 indicate that mobile users’
baseline utility for communication apps is the highest and their baseline utility for personal
finance apps is the lowest among the various mobile web and app categories. The highest
baseline utility for communication apps can be attributed to the high penetration and the wide
use of mobile messaging apps. For example, 93% of smartphone users in Korea use a mobile
messaging app, Kakao Talk (Frier 2013). The lowest baseline utility for personal finance apps
suggests that they remain a niche market. It should be noted that as  increases the baseline
utility increases. The sign for  estimates for all types of mobile content is negative due to the
relatively higher utility of outside options (e.g., mobile users spend more than 22 hours daily
engaging in activities other than mobile content use), which is normalized to zero for model
identification (  h 0
 0 ).
We can also interpret the degree of baseline utility in terms of 
because the sign of the estimates is reversed. However, the relative magnitude of estimates
remains unchanged after the exponential transformation specified in equation (2). Furthermore,
among several demographic variables, age, gender, and education account for substantial
heterogeneity in baseline utilities across mobile users. For example, senior users show a higher
intrinsic preference for portal, schedule/memo, and video apps, but a lower intrinsic preference
for the remaining apps and the mobile web. In addition, women exhibit a higher intrinsic
preference for communication and photo apps and a lower intrinsic preference for entertainment
27
and personal finance apps in comparison with men. Lastly, users with high education exhibit
significantly lower baseline utilities in most categories, except for personal finance apps.
Table 5. Estimates for Baseline Utility Parameters
Constant
(β̅)
Demographic Variables (Π𝛽 )
Age
30's
Age
Female Income Income Education Education
40's
MidUpperHighUniversity
&
class
class
School
Graduates
over
Graduates
-1.30
-0.18
0.34
0.21
-0.13
Communication
-0.03
0.02
-0.05
(0.05)
(0.05) (0.06)
(0.05)
(0.04)
(0.05)
(0.07)
(0.06)
-2.94
-0.13
-0.22
Game
0.04
-0.04
-0.04
0.09
-0.01
(0.05)
(0.05) (0.06)
(0.04)
(0.04)
(0.04)
(0.08)
(0.07)
-2.76
-0.37
-0.26
0.12
Map/Navigation
0.00
-0.02
-0.07
0.06
(0.04)
(0.05) (0.05)
(0.05)
(0.05)
(0.05)
(0.07)
(0.06)
-3.15
-0.35
-0.45
-0.23
-0.21
-0.39
Entertainment
-0.04
-0.09
(0.05)
(0.05) (0.06)
(0.05)
(0.05)
(0.05)
(0.07)
(0.06)
-2.85
0.17
-0.36
-0.21
Lifestyle
0.03
-0.06
0.04
0.02
(0.04)
(0.04) (0.05)
(0.04)
(0.04)
(0.05)
(0.07)
(0.05)
-3.34
-0.12
-0.19
-0.10
0.15
0.33
0.39
Personal Finance
0.05
(0.05)
(0.05) (0.05)
(0.04)
(0.04)
(0.05)
(0.07)
(0.06)
-2.47
-0.41
-0.43
-0.22
-0.38
Music/Radio
0.02
0.01
0.02
(0.04)
(0.04) (0.05)
(0.05)
(0.04)
(0.04)
(0.07)
(0.05)
-2.04
-0.16
-0.27
0.33
-0.17
-0.12
Photo
-0.05
-0.05
(0.04)
(0.04) (0.05)
(0.04)
(0.04)
(0.04)
(0.07)
(0.05)
-2.94
0.15
0.18
0.12
-0.11
Portal
-0.04
0.03
-0.06
(0.04)
(0.05) (0.05)
(0.04)
(0.05)
(0.04)
(0.07)
(0.05)
-1.81
0.16
0.15
Schedule/Memo
-0.05
-0.02
0.04
0.06
0.02
(0.05)
(0.05) (0.05)
(0.04)
(0.05)
(0.05)
(0.07)
(0.06)
-0.40
-0.53
0.10
Social Network
-2.54
-0.07
0.04
-0.11
-0.10
(0.04)
(0.04) (0.05)
(0.04)
(0.04)
(0.04)
(0.07)
(0.05)
-1.60
-0.14
Utility
-0.09
0.00
-0.06
-0.09
-0.10
-0.10
(0.05)
(0.05) (0.05)
(0.05)
(0.04)
(0.05)
(0.07)
(0.06)
-2.76
0.12
0.10
Video
-0.04
-0.01
0.02
0.00
-0.07
(0.04)
(0.04) (0.05)
(0.04)
(0.05)
(0.04)
(0.06)
(0.05)
-1.77
-0.17
-0.11
Web
-0.02
-0.02
0.01
-0.01
0.03
(0.04)
(0.04) (0.05)
(0.05)
(0.04)
(0.05)
(0.07)
(0.05)
Note: Standard errors in parentheses. Bold: significant at the .05 level. Following Bhat (2008), 𝛼0 is
bounded between 0 and 1. The estimated value of 𝛼0 is 0.00 with a standard error of 0.03. To code
discrete demographic variables, we use the following as a reference level: we use age 20’s or less for
age and male for gender, respectively. For monthly income, we create two dummy variables – mid-class
($3,000–$5,000) and upper-class ($5,000 or over), and we use lower-class ($3,000 or less) as a
reference level. Similarly, for education, we create two dummy variables – high school graduates and
university graduates, and we use students (elementary, middle-and-high school students, university
students) as a reference level.
28
Table 6. Estimates for Satiation Parameters
Constant
(λ̅)
Demographic Variables (Πλ )
Age
30's
Age
Female Income Income Education Education
40's
MidUpperHighUniversity
&
class
class
School
Graduates
over
Graduates
1.74
0.31
0.31
-0.22
0.05
0.31
0.23
Communication
0.03
(0.02)
(0.02) (0.03)
(0.02)
(0.02)
(0.02)
(0.03)
(0.03)
1.73
0.09
0.19
0.11
Game
-0.03
0.04
0.05
-0.04
(0.03)
(0.04) (0.04)
(0.03)
(0.03)
(0.03)
(0.05)
(0.04)
3.80
-0.27
0.27
Map/Navigation
-0.03
-0.02
0.04
-0.05
-0.07
(0.04)
(0.03) (0.04)
(0.03)
(0.03)
(0.03)
(0.05)
(0.04)
2.75
0.53
0.68
0.08
0.14
0.23
0.25
Entertainment
-0.05
(0.04)
(0.05) (0.05)
(0.04)
(0.04)
(0.05)
(0.06)
(0.05)
3.82
0.19
-0.22
0.30
-0.13
Lifestyle
-0.02
0.04
-0.02
(0.04)
(0.04) (0.04)
(0.03)
(0.03)
(0.03)
(0.05)
(0.04)
3.84
-0.28
-0.30
0.19
-0.07
-0.26
-0.21
-0.14
Personal Finance
(0.04)
(0.04) (0.05)
(0.03)
(0.03)
(0.04)
(0.05)
(0.04)
2.28
0.44
0.51
-0.10
-0.23
-0.12
Music/Radio
-0.05
0.07
(0.03)
(0.04) (0.04)
(0.03)
(0.03)
(0.03)
(0.05)
(0.04)
3.67
-0.51
0.11
0.17
0.31
0.33
Photo
0.05
-0.03
(0.03)
(0.03) (0.03)
(0.02)
(0.02)
(0.03)
(0.04)
(0.03)
3.59
0.33
-0.08
0.28
-0.12
Portal
0.06
-0.03
-0.01
(0.04)
(0.04) (0.04)
(0.03)
(0.04)
(0.04)
(0.05)
(0.04)
3.32
-0.13
-0.32
0.15
-0.16
Schedule/Memo
0.00
0.04
-0.01
(0.02)
(0.03) (0.03)
(0.02)
(0.02)
(0.02)
(0.04)
(0.03)
2.47
0.30
0.19
-0.48
0.14
0.24
0.25
Social Network
-0.04
(0.03)
(0.04) (0.04)
(0.02)
(0.03)
(0.03)
(0.05)
(0.04)
3.13
0.25
0.19
0.18
0.18
-0.15
-0.08
Utility
0.04
(0.03)
(0.03) (0.04)
(0.02)
(0.03)
(0.03)
(0.05)
(0.04)
2.96
0.08
0.14
0.06
-0.08
-0.12
-0.12
Video
(0.03)
(0.04)
(0.03) (0.04)
(0.03)
(0.03)
(0.03)
(0.05)
(0.04)
2.23
-0.10
0.26
0.25
0.07
Web
-0.04
0.00
-0.03
(0.03)
(0.03) (0.04)
(0.02)
(0.03)
(0.03)
(0.05)
(0.04)
Note: Standard errors in parentheses. Bold: significant at the .05 level. To code discrete demographic
variables, we use the following as a reference level: we use age 20’s or less for age and male for gender,
respectively. For monthly income, we create two dummy variables – mid-class ($3,000–$5,000) and
upper-class ($5,000 or over), and we use lower-class ($3,000 or less) as a reference level. Similarly, for
education, we create two dummy variables – high school graduates and university graduates, and we use
students (elementary, middle-and-high school students, university students) as a reference level.
Table 6 suggests that the satiation level is the highest in the personal finance app category
and the lowest in the game app category. As  increases, the satiation effect also increases. The
29
highest satiation for personal finance apps implies that users access these apps quickly (e.g.,
checking balance information or making a deposit by simply taking a photo of a check). In
contrast, the lowest satiation for game apps suggests that users tend to continue playing games
without growing tired of them. Further, there exists substantial user heterogeneity in terms of
satiation levels. For example, as age increases, satiation with entertainment and music/radio apps
increases, while satiation with personal finance and schedule/memo apps simultaneously
decreases. In addition, women show significantly lower satiation levels for photo and social
networking apps than men do. Moreover, users with high education show significantly higher
satiation levels regarding communication, entertainment, and social networking apps. We can
also interpret the degree of satiation in terms of  . However, unlike  ,  decreases as 
increases after the exponential transformation in equation (2). Consequently, a lower 
represents a higher satiation level.
In Figure 4, we map mobile web and app categories according to the baseline utility and
satiation levels –  and  . The four quadrants of the scatterplot provide intriguing insights
into the multiplicity of mobile content in terms of its choice and time use. For example, the topright, quadrant I, represents mobile content that is used widely as well as extensively, including
communication apps, utility apps, and mobile web. Furthermore, we can distinguish among the
apps with the similar level of baseline utility: low satiation apps from high satiation counterparts.
For example, baseline utility levels are similar to each other for apps in quadrant III (photo, map
and navigation, life style, personal finance, and portal search apps) and quadrant IV (social,
music, video, entertainment, and game apps). However, satiation levels are greater in the former
than in the latter. One notable observation is that hedonic app categories are typically grouped in
30
quadrant IV and tend to be used more extensively than utilitarian counterparts grouped in
quadrant III when their baseline utility levels are relatively low.
Figure 4. Mapping App Categories According to Baseline Utility and Satiation
Variance Decomposition for Mobile Web and App Categories
Tables 7 and 8 present the estimation results for common factors (  and  ) and specific factors
(   and   ). In addition, in the last three columns of Tables 7 and 8, we decompose the overall
variations in baseline utilities and satiations for mobile web and app categories into variations
attributed to the demographic variables, common factors, and specific factors. Interpretation of
the factors is based on estimated factor loading matrices and variance decompositions. For the
baseline utility, the first factor loads strongly with positive signs for major utilitarian app
31
categories (communication, map/navigation, lifestyle, personal finance, schedule/memo), but
with negative signs for major hedonic app categories (game, music/radio, video). A large
positive factor score implies high utility for utilitarian apps, but low utility for hedonic apps. In
contrast, the second factor loads strongly with positive signs for hedonic and social networking
app categories (game, entertainment, portal, social network, video), but with negative signs for
utilitarian apps (communication, schedule/memo, web).
Table 7. Baseline Utility Factor Estimates and Variance Decomposition
Estimates
Common Factor
Variance Decomposition
Specific
Factor
Demographic
0.13
-0.05
0.02
0.69
(0.02)
(0.02)
(0.02)
-0.08
0.10
0.10
Game
0.37
(0.02)
(0.02)
(0.02)
0.08
0.06
Map/Navigation
-0.01
0.75
(0.02)
(0.02)
(0.02)
0.12
0.05
Entertainment
-0.01
0.86
(0.02)
(0.02)
(0.02)
0.07
Lifestyle
0.03
0.03
0.76
(0.02)
(0.02)
(0.02)
0.11
Personal Finance
0.00
0.00
0.64
(0.02)
(0.02)
-0.12
Music/Radio
0.02
0.02
0.86
(0.02)
(0.02)
(0.02)
Photo
0.02
-0.01
0.02
0.99
(0.02)
(0.02)
(0.01)
0.09
0.18
Portal
0.00
0.15
(0.02)
(0.02)
(0.02)
0.10
-0.05
Schedule/Memo
0.02
0.49
(0.02)
(0.02)
(0.01)
0.06
0.04
Social Network
0.03
0.92
(0.02)
(0.02)
(0.02)
Utility
0.02
-0.02
0.02
0.90
(0.02)
(0.02)
(0.01)
-0.06
0.06
0.03
Video
0.29
(0.02)
(0.02)
(0.01)
-0.05
-0.18
0.05
Web
0.08
(0.02)
(0.02)
(0.01)
Note: Standard errors in parentheses. Bold: significant at the .05 level.
Communication
Common
Factor
Specific
Factor
0.29
0.01
0.38
0.26
0.15
0.10
0.12
0.02
0.21
0.03
0.34
0.02
0.13
0.01
0.01
0.01
0.83
0.02
0.49
0.02
0.06
0.02
0.05
0.05
0.63
0.07
0.86
0.06
32
The variance decomposition reveals that, on average, 63% of variation in the baseline utility
can be explained by demographic variables and 33% by common factors. Specific factors
account for the remaining 4%. In marketing literature, researchers assert that demographic
variables often serve as poor predictors of consumer brand preferences that are estimated on
scanner panel data (Singh et al. 2005). However, our results indicate that the demographic
variables are powerful predictors of overall variations in app preference captured by the baseline
utility. This implies that companies can capitalize on strategic advantages that result from their
knowledge on user demographics.
For the satiation parameters, we find that the first factor loads with positive signs for
communication, photo, portal, and social network apps, but with negative signs for game, video,
and the mobile web. A large positive factor score indicates high satiations for communication,
photo, portal, and social network apps and low satiations for game, video, and the mobile web.
The second factor loads strongly with positive signs for all categories, capturing uniformly lower
(higher) levels of usage with a large positive (negative) factor score. The variance decomposition
denotes that only 11% of variation in satiation parameters is explained by demographic variables.
In contrast, 57% of the variation is accounted for by specific factors and 33% by common
factors. Unlike in the baseline utility, demographic variables are poor predictors of heterogeneity
in app satiation. A large variation is explained by specific factors, which indicates that satiation
is a user-specific trait. It is difficult to predict user's app satiations based on her demographic
information. Instead, individual level historical data are required for reliable prediction.
33
Table 8. Satiation Factor Estimates and Variance Decomposition
Estimates
Common Factor
Variance Decomposition
Specific
Factor
Demographic
0.36
0.11
0.02
0.33
(0.01)
(0.01)
(0.01)
-0.05
0.18
0.53
Game
0.03
(0.01)
(0.01)
(0.01)
0.35
0.44
Map/Navigation
-0.01
0.11
(0.01)
(0.01)
(0.01)
0.08
0.90
Entertainment
0.04
0.14
(0.02)
(0.02)
(0.02)
0.05
0.76
0.46
Lifestyle
0.05
(0.02)
(0.02)
(0.02)
0.18
0.76
Personal Finance
0.00
0.09
(0.02)
(0.01)
0.13
0.31
0.84
Music/Radio
0.08
(0.01)
(0.01)
(0.01)
0.46
0.10
0.09
Photo
0.28
(0.01)
(0.01)
(0.01)
0.35
0.83
1.43
Portal
0.01
(0.01)
(0.01)
(0.01)
0.23
0.28
0.45
Schedule/Memo
0.09
(0.01)
(0.01)
(0.01)
0.52
0.12
0.43
Social Network
0.20
(0.01)
(0.01)
(0.02)
0.13
0.45
0.70
Utility
0.03
(0.01)
(0.01)
(0.01)
-0.22
0.47
0.63
Video
0.01
(0.01)
(0.01)
(0.02)
-0.16
0.10
0.85
Web
0.04
(0.01)
(0.01)
(0.01)
Note: Standard errors in parentheses. Bold: significant at the .05 level.
Communication
Common
Factor
Specific
Factor
0.66
0.01
0.11
0.86
0.34
0.55
0.01
0.86
0.70
0.25
0.05
0.86
0.13
0.80
0.69
0.03
0.28
0.71
0.35
0.55
0.48
0.32
0.30
0.68
0.40
0.59
0.05
0.91
Correlation across Mobile Web and App Categories
Tables 9 and 10 present the correlation matrices for the baseline utility and satiation parameters
across mobile web and app categories. Table 9 hints that people who use social apps (e.g.,
Facebook) also frequently use photo apps (e.g., Instagram). Likewise, people who use
entertainment apps (e.g., Naver Cartoon) also frequently consume music/radio apps (e.g., Music
34
Player). These results indicate that social apps and photo apps might be economic complements
and entertainment apps and music/radio apps are frequently used together as complements.9 In
contrast, people who use mobile portal apps (e.g., Google app) infrequently use the mobile web
(e.g., Google on browser), which indicates that portal apps and the mobile web might be
economic substitutes. This result suggests that mobile portal apps offer functionality and
usability similar to those of the mobile web. For example, users can obtain information quickly
and easily regardless of whether they use a mobile search app or visit a mobile search website.
Because people have a limited amount of time to spend on the mobile web and app categories,
substitution could occur in our context.
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music /Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
Communication
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music/Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
Communication
Table 9. Correlation Matrix for Baseline Utility Parameters
1.00
-0.47
0.60
0.00
0.59
0.09
0.21
0.66
0.20
0.78
0.48
0.41
-0.63
0.15
1.00
-0.34
0.40
-0.01
-0.37
0.28
0.09
0.20
-0.63
0.17
0.03
0.48
-0.39
1.00
0.41
0.44
0.24
0.42
0.45
0.00
0.34
0.68
0.60
-0.58
0.10
1.00
0.33
-0.38
0.81
0.35
0.09
-0.40
0.74
0.70
0.06
-0.13
1.00
-0.14
0.33
0.77
0.36
0.19
0.64
0.45
-0.40
-0.06
1.00
-0.62
-0.32
0.14
0.31
-0.17
-0.40
-0.31
-0.24
1.00
0.61
-0.20
-0.30
0.78
0.75
-0.05
0.23
1.00
0.06
0.16
0.80
0.61
-0.45
0.16
1.00
0.16
0.12
-0.19
0.32
-0.84
1.00
-0.08
0.11
-0.59
0.11
1.00
0.71
-0.31
-0.02
1.00
-0.48
0.31
1.00
-0.44
1.00
9
The positive correlation in baseline utility can be explained by not only inherent complementary relationship
among apps but also some external reasons for simultaneous use. For example, in the case of product purchase
decision, bundle promotion (e.g. buy one get one free) can result in correlated demand. When mobile users decide
which app to use, the decision is less likely to be influenced by external factors compared to other purchase or
choice decisions. Thus, we believe that the inherent complementarity plays the major role in our context.
35
Table 10 shows correlations in satiation levels. People who spend a great deal of time on
social apps (e.g., Facebook) also spend a significant amount of time on communication apps
(e.g., WhatsApp) and photo apps (e.g., Instagram), which suggests that social, communication,
and photo apps might be economic complements to each other. For example, many users take a
picture or video, use Instagram to choose a filter to transform its look and feel, and post it to
Facebook or share it on WhatsApp.
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music/Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
1.00
0.07
-0.02
0.23
0.23
0.02
0.31
0.92
0.30
0.27
0.80
0.26
-0.06
-0.07
1.00
0.15
0.06
0.25
-0.05
0.11
0.02
0.14
0.08
0.02
0.16
0.21
0.08
1.00
-0.01
0.52
0.06
0.13
-0.05
0.28
0.34
-0.03
0.32
0.34
0.02
1.00
0.07
-0.09
0.12
0.14
0.06
-0.03
0.15
0.08
0.05
0.03
1.00
0.00
0.28
0.20
0.43
0.41
0.19
0.46
0.46
0.05
1.00
-0.04
0.05
0.04
0.17
0.03
0.01
-0.06
-0.06
1.00
0.25
0.20
0.14
0.24
0.21
0.17
0.03
1.00
0.26
0.27
0.80
0.20
-0.12
-0.11
1.00
0.30
0.23
0.29
0.22
0.01
1.00
0.22
0.28
0.16
-0.05
1.00
0.19
-0.09
-0.09
1.00
0.27
0.02
1.00
0.11
Web
Game
Communication
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music/Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
Communication
Table 10. Correlation Matrix for Satiation Parameters
App-Level Analysis
Our proposed model provides a general empirical framework to analyze the usage data on
various emerging IT artifacts as well as the app time usage data. To illustrate this vital strength,
36
1.00
we conduct an app-level analysis using the proposed model to investigate mobile users' popular
app choices and time usage decisions. We focus on the top-ten apps in terms of total usage time
in our data. We chose the top-ten apps for the sake of the tractability of the model, however, the
number of apps can be increased at the expense of the computational time required to identify
and to estimate the parameters of the model. Table 11 lists the selected top ten apps with their
key descriptive statistics. Note that "Others" represents all of the other mobile content use
excluding the top-ten apps. Column 2 of Table 11 shows that users accessed Kakao Talk (a
communication app) most frequently (94.6%), while they accessed Rule the Sky (a game app)
least frequently (4.1%). Column 5 of Table 11 shows that users spent the most time (20.46%)
using Kakao Talk, followed by Naver (3.67%), Rule the Sky (2.16%), Facebook (1.89%), Tiny
Farm (1.59%), and Mellon (1.38%). In addition, the top ten apps account for 34% of the total
mobile content use and the "Others" category accounts for the remaining 66%.
Table 11. Top-Ten Apps and Summary Statistics
Apps
Outside option (other activities)
Kakao Talk
Naver
Rule The Sky
Facebook
Tiny Farm
Mellon
Daum
YouTube
Mypeople
T map
Others
Total
Category
Communication
Portal Search
Game
Social
Game
Music and Radio
Portal Search
Video
Communication
Map and Navigation
-
Choice
100.0%
94.6%
43.8%
4.1%
26.1%
4.6%
11.2%
11.9%
27.8%
15.2%
12.7%
99.9%
Time Use
Hour
Percent
Percent excld.
outside option
22.23
0.36
0.06
0.04
0.03
0.03
0.02
0.02
0.01
0.01
0.01
1.17
24
92.62%
1.51%
0.27%
0.16%
0.14%
0.12%
0.10%
0.07%
0.05%
0.05%
0.04%
4.87%
100.00%
20.46%
3.67%
2.16%
1.89%
1.59%
1.38%
0.97%
0.66%
0.63%
0.57%
66.00%
100.00%
37
We use the same specification for the model estimation as in the app category-level analysis.
Tables 12 and 13 show the estimates for the baseline utility parameters (  and
 )
and satiation
parameters (  and   ), respectively. The results in Table 12 show that mobile users’ baseline
utility for Kakao Talk (a communication app) is the highest and that their baseline utility for
Rule the Sky (a game app) is the lowest among the top ten apps. The highest baseline utility for
Kakao Talk is consistent with our empirical findings based on the app category-level data. The
demographic variables explain the substantial heterogeneity in the baseline utilities for all mobile
users. For example, older users exhibit a lower intrinsic preference for Kakao Talk (a
communication app), Rule the Sky (a game app), Facebook (a social network app), Tiny Farm (a
game app), Mellon (a music/radio app), and Mypeople (a communication app) but a higher
intrinsic preference for Daum (a portal search app). In addition, women exhibit a higher intrinsic
preference for Kakao Talk, Naver, Mellon, Youtube, and Mypeople and a lower intrinsic
preference for T map (a map/navigation app) and Facebook in comparison with men.
The results in Table 13 show that the satiation level is the highest for YouTube and the
lowest for Rule the Sky among the top-ten apps. A game app (Rule the Sky) shows the lowest
satiation, and this finding is consistent with our results based on the app category-level data. We
find that there is a substantial user heterogeneity in terms of satiation levels. For example,
women show significantly lower satiation levels for most of the top-ten apps than men, but
exhibit a significantly higher satiation level for T map.
38
Table 12. Estimates for Baseline Utility Parameters: App-Level Analysis
Constant
(β̅)
Demographic Variables (Π𝛽 )
Age
30's
Age 40's
& over
Female
Income
Midclass
Income
Upperclass
Education
HighSchool
Graduates
-14.19
-0.27
-0.57
0.41
-0.08
-0.30
Kakao Talk
-0.01
(Communication)
(0.48)
(0.05)
(0.05)
(0.03)
(0.04)
(0.04)
(0.07)
-16.58
0.19
-0.25
0.41
-0.43
Naver
-0.08
-0.08
(Portal Search)
(0.48)
(0.06)
(0.07)
(0.05)
(0.05)
(0.06)
(0.08)
-21.13
-1.70
-2.42
-0.49
-0.58
2.72
Rule The Sky
0.06
(Game)
(0.50)
(0.12)
(0.16)
(0.10)
(0.11)
(0.13)
(0.19)
-16.19
-1.55
-1.37
-0.20
0.17
0.34
-1.09
Facebook
(Social)
(0.48)
(0.70)
(0.08)
(0.05)
(0.06)
(0.06)
(0.09)
-18.94
-1.62
-2.32
Tiny Farm
0.06
0.14
-0.08
0.19
(Game)
(0.48)
(0.11)
(0.15)
(0.08)
(0.09)
(0.11)
(0.14)
-17.87
-0.62
-1.02
0.27
0.25
-0.23
Mellon
-0.06
(Music and Radio)
(0.48)
(0.07)
(0.08)
(0.05)
(0.06)
(0.07)
(0.10)
-19.09
0.74
0.50
-0.84
-0.68
-0.38
Daum
-0.05
(Portal Search)
(0.48)
(0.09)
(0.11)
(0.07)
(0.09)
(0.10)
(0.13)
-17.23
0.12
0.15
-0.13
-0.23
-0.21
YouTube
0.03
(Video)
(0.47)
(0.05)
(0.06)
(0.04)
(0.04)
(0.05)
(0.07)
-18.17
-0.15
-0.59
0.10
-0.15
0.38
Mypeople
-0.06
(Communication)
(0.48)
(0.06)
(0.07)
(0.05)
(0.05)
(0.06)
(0.09)
-18.78
0.26
-0.40
0.35
T map
0.11
0.07
-0.06
(Map & Navigation)
(0.48)
(0.07)
(0.08)
(0.05)
(0.06)
(0.07)
(0.11)
-13.65
0.13
0.09
Others
0.11
-0.06
-0.01
-0.13
(0.48)
(0.06)
(0.06)
(0.04)
(0.05)
(0.05)
(0.08)
Note: Standard errors in parentheses. Bold: significant at the .05 level. The estimated value of  0
Education
University
Graduates
-0.32
(0.06)
-0.37
(0.06)
1.88
(0.17)
-0.91
(0.07)
-0.12
(0.10)
-0.44
(0.07)
-0.12
(0.10)
-0.23
(0.06)
0.35
(0.07)
0.51
(0.10)
-0.27
(0.06)
is -4.26
(standard error: 0.06).
39
Table 13. Estimates for Satiation Parameters: App-Level Analysis
Constant
(λ̅ )
Demographic Variables (Π𝜆 )
Age
30's
Age
40's &
over
Female
Income
Midclass
2.07
0.33
0.53
-0.38
0.06
(0.02)
(0.02)
(0.02)
(0.02)
(0.02)
3.04
0.16
-0.21
-0.17
-0.02
(0.03)
(0.03)
(0.03)
(0.02)
(0.03)
1.19
-0.97
-1.05
-0.29
0.06
(0.13)
(0.12)
(0.16)
(0.11)
(0.10)
2.77
0.35
0.34
-0.28
0.17
(0.04)
(0.04)
(0.05)
(0.03)
(0.03)
1.42
-0.73
-0.86
0.07
-0.16
(0.09)
(0.12)
(0.17)
(0.08)
(0.10)
2.40
0.84
1.18
-0.22
-0.01
(0.07)
(0.08)
(0.09)
(0.06)
(0.07)
2.96
-0.20
0.17
-0.29
-0.10
(0.07)
(0.07)
(0.08)
(0.05)
(0.06)
4.14
-0.41
-0.14
-0.22
-0.08
(0.05)
(0.05)
(0.06)
(0.03)
(0.04)
3.72
0.42
0.49
-0.18
0.27
(0.08)
(0.06)
(0.07)
(0.05)
(0.06)
3.43
-0.82
-1.24
0.37
0.33
(0.09)
(0.07)
(0.08)
(0.05)
(0.06)
1.48
0.11
0.21
0.04
-0.03
(0.02)
(0.02)
(0.02)
(0.02)
(0.02)
Note: Standard errors in parentheses. Bold: significant at the .05 level.
Kakao Talk
(Communication)
Naver
(Portal Search)
Rule The Sky
(Game)
Facebook
(Social)
Tiny Farm
(Game)
Mellon
(Music and Radio)
Daum
(Portal Search)
YouTube
(Video)
Mypeople
(Communication)
T map
(Map & Navigation)
Others
Income
Upperclass
0.00
(0.02)
0.03
(0.03)
0.15
(0.11)
0.08
(0.04)
-0.26
(0.12)
-0.15
(0.07)
-0.04
(0.06)
0.02
(0.04)
0.72
(0.06)
0.36
(0.07)
0.11
(0.02)
Education
HighSchool
Graduates
0.46
(0.03)
0.16
(0.04)
0.67
(0.16)
0.41
(0.06)
-0.14
(0.14)
0.61
(0.11)
0.28
(0.10)
0.37
(0.07)
-0.51
(0.09)
0.67
(0.12)
0.00
(0.03)
Education
University
Graduates
0.47
(0.02)
0.13
(0.03)
0.27
(0.15)
0.62
(0.04)
0.11
(0.10)
0.41
(0.07)
0.24
(0.08)
0.19
(0.06)
-0.63
(0.07)
0.25
(0.09)
0.02
(0.02)
Using the estimation results for common factors (  and  ) and specific factors (   and
  ), we decompose the overall variations in the baseline utilities and the satiations for the top-
ten apps into variations attributed to the demographic variables, common factors, and specific
factors. On average, 39% of the variations in the baseline utility can be explained by
demographic variables and 35% by common factors. Specific factors account for the remaining
40
26%.10 Compared to the variance decomposition results from the category level analysis, the
explanatory power of the demographic variable is weaker. This indicates that the demographic
variables are a more powerful predictor of the overall variation in the category-level preference
than in the individual app-level preference. For the satiation parameters, 40% of the variation is
explained by common factors. Specific factors and demographic variables explain 32% and 28%
of the variation in satiation parameters, respectively.
Kakao Talk
(Communication)
Naver
(Portal Search)
Rule The Sky
(Game)
Facebook
(Social)
Tiny Farm
(Game)
Mellon
(Music and Radio)
Daum
(Portal Search)
YouTube
(Video)
Mypeople
(Communication)
T map
(Map and Navigation)
Others
Others
T map
(Map and Navigation)
Mypeople
(Communication)
YouTube
(Video)
Daum
(Portal Search)
Mellon
(Music and Radio)
Tiny Farm
(Game)
Facebook
(Social)
Rule The Sky
(Game)
Naver
(Portal Search)
Kakao Talk
(Communication)
Table 14. Correlation Matrix for Baseline Utility Parameters: App-Level Analysis
1.00
0.04
1.00
0.01
0.06
1.00
-0.07
0.16
0.12
1.00
-0.16
0.47
0.22
0.17
1.00
0.05
0.34
0.07
0.16
0.26
1.00
0.12
0.57
0.17
0.19
0.46
0.50
1.00
-0.05
-0.02
0.60
0.10
0.21
-0.04
0.03
1.00
0.11
0.47
0.50
0.17
0.48
0.20
0.41
0.47
1.00
-0.12
0.48
0.01
0.32
0.25
0.36
0.44
-0.04
0.36
1.00
-0.12
-0.66
0.07
-0.11
-0.59
-0.34
-0.64
0.15
-0.48
-0.30
1.00
10
Variance decomposition results for the app-level analysis are omitted due to brevity. The results are available
upon request.
41
Table 14 shows the correlation matrix for the baseline utility for the top ten apps. We find
that people who use the Naver app also frequently use its competing portal search Daum's app. In
contrast, people who use T map less frequently use the other popular apps including Kakao Talk,
Mellon, Facebook, Youtube and Mypeople. Table 15 shows the correlation in the satiation
levels. People who spend a great deal of time on Facebook also spend a lot of time on Kakao
Talk, which is similar to our empirical findings based on the category-level analysis. Moreover,
we find that mobile users who spend a great deal of time on Naver also spend a significant
amount of time on Daum.
Kakao Talk
(Communication)
Naver
(Portal Search)
Rule The Sky
(Game)
Facebook
(Social)
Tiny Farm
(Game)
Mellon
(Music and Radio)
Daum
(Portal Search)
YouTube
(Video)
Mypeople
(Communication)
T map
(Map and Navigation)
Others
Others
T map
(Map and Navigation)
Mypeople
(Communication)
YouTube
(Video)
Daum
(Portal Search)
Mellon
(Music and Radio)
Tiny Farm
(Game)
Facebook
(Social)
Rule The Sky
(Game)
Naver
(Portal Search)
Kakao Talk
(Communication)
Table 15. Correlation Matrix for Satiation Parameters: App-Level Analysis
1.00
0.65
1.00
-0.32
-0.26
1.00
0.66
0.56
-0.16
1.00
-0.20
0.13
0.51
-0.14
1.00
0.40
0.23
-0.12
0.51
-0.31
1.00
0.61
0.86
-0.34
0.41
0.19
0.08
1.00
0.22
0.32
0.05
0.18
0.13
0.06
0.34
1.00
-0.09
0.06
-0.37
-0.38
-0.13
-0.35
0.29
0.01
1.00
-0.31
-0.21
0.45
-0.22
0.41
-0.21
-0.20
-0.03
-0.10
1.00
0.48
0.53
-0.18
0.53
-0.03
0.31
0.44
0.18
-0.18
-0.14
1.00
42
APPLICATIONS OF PROPOSED APPROACHES
In this section, we illustrate the usefulness as well as the applications of our proposed model and
findings in several areas, including mobile user targeting and mobile content portfolio/media
planning strategies.
Mobile User Targeting: App Time-Use Prediction
Making inferences about app use time decisions across multiple app categories using a simple
method is a challenge because the omitted variable problem may result in biased and inconsistent
estimates. The proposed approach circumvents this issue by jointly modeling app use time
decision and app choices within a single utility-theory based framework.
Previous research has frequently used demographics or preference estimates based on
individual-level panel data for new customer prediction (e.g. Rossi et al. 1996). In addition,
researchers and practitioners have developed regression-based scoring models to predict
customers’ future behaviors (e.g. Bolton 1998; Malthouse and Blattberg 2005). The purpose of
these approaches is to use the measurements of customers’ prior behavior as predictors of their
future behavior. Similarly, in direct marketing literature, it is a common practice to summarize
customers’ prior behavior in terms of their recency (time of most recent purchase), frequency
(number of prior purchases), and monetary value (average purchase amount per transaction)
(RFM method; Fader et al. 2005). However, when it comes to the case of new customer
prediction, all these established methods can be difficult to apply due to the lack of prior
transactions or behaviors. In this case, an average consumer’s category usage pattern can be used
43
for multiple categories. We can obtain this information from the estimated correlation matrix of
the proposed model.
To add a specific context to our application, we aimed to identify a segment with extensive
social network app use. We divided our 8 week sample data into an estimation period (first four
weeks) and a holdout prediction period (last four weeks). We found that 144 mobile users (out of
1,425 total samples) did not use social network apps during the estimation period but did use
them during the prediction period. We focused on these 144 mobile users (focal group) in our
app time-use prediction.
We first estimated the proposed model using the estimation period data and obtained a
correlation matrix of satiation parameters. Table 16 reports the correlation matrix for satiation
parameters. The satiation pattern presented in Table 16 is similar to the pattern based on the full
8 week data (see Table 10).11 The satiation level of social network apps is positively correlated
with those of photo and communication apps in particular. This indicates that heavy photo and
communication app users are also characterized as heavy social network app users. Accordingly,
we used the average photo and communication app use time during the estimation period to infer
the potential targets of extensive social network app use during the prediction period. It should
be noted that we cannot apply past usage behavior-based targeting approaches (i.e. RFM
method) to the focal group because they never used social network apps during the estimation
period. Our strategy was to use simultaneously the photo and the communication app usage
histories instead.
11
The main findings from the estimation results using the first 4 weeks are the same as those from the full data. The
results are available on request from the authors.
44
Table 16. Correlation Matrix for Satiation Parameters in the App Time Use Prediction
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music /Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
Communication
Game
Map/Navigation
Entertainment
Lifestyle
Personal Finance
Music/Radio
Photo
Portal
Schedule/Memo
Social Network
Utility
Video
Web
Communication
Application
1.00
0.13
0.21
0.51
0.33
0.18
0.22
0.93
0.23
0.41
0.73
0.37
0.05
0.16
1.00
0.12
0.09
0.12
0.04
0.09
0.10
0.09
0.14
0.08
0.13
0.12
0.09
1.00
0.20
0.42
0.28
0.16
0.18
0.24
0.58
0.15
0.39
0.32
0.17
1.00
0.23
0.11
0.16
0.44
0.15
0.28
0.36
0.27
0.10
0.12
1.00
0.20
0.18
0.32
0.23
0.47
0.26
0.36
0.27
0.16
1.00
0.03
0.20
0.12
0.37
0.14
0.19
0.08
0.07
1.00
0.15
0.13
0.17
0.15
0.19
0.21
0.13
1.00
0.19
0.39
0.75
0.31
-0.02
0.11
1.00
0.29
0.16
0.22
0.18
0.12
1.00
0.30
0.44
0.31
0.20
1.00
0.26
0.01
0.09
1.00
0.26
0.17
1.00
0.19
1.00
Table 17 shows the performance of the proposed targeting strategy. We divided the focal
group by the average communication app use time in the estimation period. The average social
network app use time of the top 50% is 0.253 and the bottom 50% is 0.098 during the prediction
period. The average use time for all focal groups is 0.176 and this is what we can expect when
we use a random targeting strategy. When we select 72 users (50%) from the focal group based
on their communication app use time during the estimation period, the average social network
app use time during the prediction period is longer by 44% than that of a randomly selected
group. Similarly, when we select 72 users from the focal group based on their photo app use time
during the estimation period, the gain is 37% in comparison with a random targeting strategy.
The last three columns of Table 17 show the performance of similar targeting applications
based on other strong positive correlations (greater than 0.5 in Table 10). All of these
45
applications show substantial gains in comparison with the random targeting strategy, which
confirms the validity of the proposed targeting strategy. In particular, a strong positive
correlation between entertainment and communication apps (0.51) does not show up on the
correlation matrices based on app use time or app choice incidence variables (reported in Tables
3 and 4). This implies that there is a distinctive advantage of the proposed method over the
simple model-free approaches.
Table 17. Result of Heavy User Targeting Application
Target Category
Social Network
Social Network
Entertainment
Photo
Schedule/Memo
Sorting Variable
Communication
Photo
Communication
Communication
Map/Navigation
0.253
0.098
0.176
44%
144
0.240
0.111
0.176
37%
144
0.049
0.028
0.038
27%
87
0.037
0.025
0.032
17%
37
0.051
0.028
0.040
28%
126
Use Time
Top 50%
Bottom 50%
Average
Gain
Sample Size
Mobile Content Portfolio and Mobile Media Planning Strategies
We further illustrate the strategic value of our proposed method for two application areas –
developing optimal mobile content portfolios and offering optimal mobile media planning.
Companies can build and maintain an optimal mobile content portfolio by making and/or
buying mobile websites and apps. The results of the correlations for the app choices exhibited in
Table 9 provide useful and actionable guidelines and insights for this fundamental issue. App
categories with strong negative correlations in the baseline utility are likely to substitute for one
another. That is, when one category of apps is used, the other category of apps is used less (if not
unused). Hence, a company should acquire (build) apps that substitute for its focal app in terms
of choice to increase the customer base. Strong negative correlations in choice are required for
46
such acquisition (integration) to be valuable. For example, due to the strong negative correlation
in choice between portal apps and the mobile web (i.e., -0.84), a company offering a mobile
portal app can increase its customer base by acquiring (building internally) a mobile-optimized
website, and vice versa. This strategy is effective because the two categories are likely to serve
distinct market segments. Furthermore, the results of the correlation in app time use reported in
Table 10 demonstrate that some categories of apps are used more extensively when they are used
together rather than separately. Therefore, a company should acquire (build) apps that
complement its focal app in terms of time use to maximize the consumption of the combined
apps. For example, due to strong positive correlations in consumption among photo,
communication, and social network apps, a company can maximize the customers’ time spent on
its portfolio of apps by acquiring or combining these app categories together. Facebook’s recent
acquisition of WhatsApp and Instagram is a good manifestation of this case.
The results in this study also provide strategic insights for mobile media advertisers.
Advertisers optimally select media vehicles and determine the size of the target audience and the
number of advertisement impressions. To develop a method for optimal media selection, a
necessary starting point is a model of mobile advertising exposure, which is determined by the
mobile app use time. The proposed model provides an empirical tool to model the usage time of
multiple apps by allowing flexible correlations. Depending on the objectives, advertisers aim to
maximize the proportion of the target audience exposed at least once to the advertising message
(i.e., reach), maximize the number of repeat exposures to the same user among those reached
(i.e., frequency), or both (Rust 1986). According to the results of the correlation in app choice in
Table 9, to maximize the reach of in-app and mobile web advertisements, an advertiser should
use the combination of mobile portal apps and mobile websites as its mobile media vehicles due
47
to the strong negative correlation. Furthermore, according to the results of the correlation for app
time use in Table 10, to maximize the number of incidences of repeat impressions for a user, an
advertiser should include photo, communication, and social network apps together in their
optimal mobile media vehicles.
CONCLUSIONS
The drastic growth of mobile platforms and devices has resulted in an app-centric marketplace,
ushering in the app-based mobile economy that provides new opportunities and challenges for
entrepreneurs, managers, and marketers across all online business segments. Although online
consumers are increasingly migrating from web browsers to mobile apps for their commercial
transactions and social interactions, little was known about how these users behave and make
decisions with respect to their choices, consumptions, and management of mobile apps in diverse
categories. The absence of sound mobile app analytics based on a thorough methodological
approach has impeded our understanding of consumers’ behavioral responses to the choice and
usage of mobile apps.
This study developed a novel framework for mobile app analytics using the theoretical
notions of utility and satiation along with the structured method of factor analytic approach. Our
work contributes to an emerging stream of literature on the economics of the mobile internet and
mobile marketing by being the first study to quantify the baseline utility and the satiation levels
of diverse mobile app categories. Using large-scale, individual-level panel data on mobile app
and web time-use histories, we constructed a unique multiple discrete/continuous model of app
selection and time-use decisions to gain a better understanding on the usage interdependence and
dynamics among diverse categories of the mobile web and apps. The frameworks, mechanisms,
48
and processes discussed in this study offer high flexibility and efficiency in estimating
parameters, particularly when the number of alternatives is exceedingly large, a phenomenon that
is increasingly prevalent in big data analytics. Consequently, our analytical paradigm and
comprehensive computational procedures can supply enlightening ideas for the advancement of
big data analytics in general and mobile app analytics in particular.
Data availability issues suggest that some caution is warranted in the estimation of the
proposed model. For example, some consumers may make their app selection and consumption
decisions and utility maximization as frequently as they wish, i.e., on a daily basis or even an
hourly basis. We had only weekly usage data available for the analysis and therefore could not
fully address the issue. Future research may consider applying our proposed model to the highfrequency data on individual user-level app usage patterns. Moreover, we did not observe
consumers' app purchase behaviors. Future works should be directed towards an integrated
model of the demand for app purchase and consumption, provided that they have individual userlevel data on both app purchase and usage. Furthermore, we postulated that the utility function is
additively separable for the sake of tractability and thus the model cannot explicitly partial out
complementarity or substitution from other factors (Bhat 2008). Recently, Lee et al. (2013)
develop a direct utility model which circumvents this problem, but still the model is restricted to
accommodate a relatively small number of alternatives. Future research may consider tackling
this limitation of MDCEV models. Notwithstanding these limitations, to the extent that mobile
users increasingly access mobile web and various kinds of apps, our utility theory-based
structural model of app selection and consumption may have profound implications for the future
direction of mobile analytics.
49
Online businesses that depend heavily on apps for their growth are confronted with a new
strategic and operational challenge that increases with the volume and complexity of users’ app
consumption. The findings reported in this study may help managers alleviate such challenges
through an enhanced understanding of app selections and usage behaviors. In addition, our study
offers managers direct managerial implications regarding how they should allocate their
advertising resources across diverse app categories, which often complement or compete with
each other. Our empirical results suggest that both positive and negative correlations exist in the
baseline utility and the satiation levels of mobile web and app categories. We also found that
users’ demographic variables and unobservable factors play important roles in explaining app
selection and time-use decisions. These results offer insights related to mobile user segmentation,
targeting, and optimal media planning in the mobile app marketplace, empowering firms to
identify valuable consumers based on their diverse needs, varying demographic characteristics,
and idiosyncratic behavioral patterns.
As we move more deeply into the mobile economy, the adoption and use of effective mobile
app analytics will become a key basis for competition in the mobile-based markets. The number
and complexity of mobile apps will rise above and beyond our prediction and the radical growth
of apps will continue to exercise the mind and intellect of practitioners. The ability to nurture
strong analytical prowess and the capability to drive significant business value from the insights
of mobile app analytics need to be at the core of a firm’s essential mobile strategy. It is hoped
that this study will lay the foundation for future scholarly and practical works on mobile app
analytics.
50
References
App Annie and IDC. 2014. Mobile App Advertising and Monetization Trends 2012-2017: The
Economics of Free. http://www.idc.com/getdoc.jsp?containerId=prUS24771914.
Bart, Y., Stephen, A., and Sarvary, M. 2014. “Which Products Are Best Suited To Mobile
Advertising? A Field Study of Mobile Display Advertising Effects on Consumer Attitudes
and Intentions,” Journal of Marketing Research. Forthcoming.
Bhat, C. 2005. “A Multiple Discrete-Continuous Extreme Value Model: Formulation and
Application to Discretionary Time-Use Decision,” Transportation Research Part B (39), pp.
679-707.
Bhat, C. 2008. “The Multiple Discrete-Continuous Extreme Value (MDCEV) Model: Role of
Utility Function Parameters, Identification Considerations, and Model Extensions,”
Transportation Research Part B (42), pp. 274–303.
Bolton, R.N. 1998. “A Dynamic Model of the Duration of the Customer’s Relationship with a
Continuous Service Provider: The Role of Satisfaction,” Marketing Science (17:1), pp. 45–
65.
Carare, O. 2012. “The Impact of Bestseller Rank on Demand: Evidence from the App Market,”
International Economic Review (53:3), pp. 717–742
Chiang, J. 1991. “A Simultaneous Approach to Whether to Buy, What to Buy, and How Much to
Buy,” Marketing Science (10:4), pp. 297–314.
Chintagunta, P.K. 1993. “Investigating Purchase Incidence, Brand Choice and Purchase Quantity
Decisions of Households,” Marketing Science (12:2), pp. 194–208.
Chintagunta, P.K. 1994. “Heterogeneous Logit Model Implications for Brand Positioning,”
Journal of Marketing Research (31:2), pp. 304–311.
Danaher, P., Lee, J., Kerbache, L. 2010. “Optimal Internet Media Selection,” Marketing Science
(29: 2), pp. 336–347.
Dubin, J.A., D.L. McFadden. 1984. “An Econometric Analysis of Residential Electric Appliance
Holdings and Consumption,” Econometrica (52: 2), pp. 345–362
Einav, L., Levin, J., Popov, I., and Sundaresen, N. 2014. “Growth, Adoption, and Use of Mobile
Commerce,” American Economic Review: Papers & Proceedings (104:5), pp. 489–494.
Elrod, T. and Keane, M.P. 1995. “A Factor-Analytic Probit Model for Representing the Market
Structure in Panel Data,” Journal of Marketing Research (32:1), pp. 1–16.
Emarketer. 2014. http://www.emarketer.com/Article/Mobile-Continues-Steal-Share-of-USAdults-Daily-Time-Spent-with-Media/1010782
Erdem, T., R. Winer. 1999. “Econometric Modeling of Competition: A Multi-Category ChoiceBased Mapping Approach,” Journal of Econometrics (89:1-2), pp. 159–75.
Fader, P.S., Hardie, B., and Lee, K.L. 2005. “"Counting Your Customers" the Easy Way: An
Alternative to the Pareto/NBD Model,” Marketing Science (24:2), pp. 275–284.
51
Flurry. 2013. There’s An App Audience for That, But It’s Fragmented.
http://www.flurry.com/bid/ 96368/There-s-An-App-Audience-for-That-But-It-sFragmented#.U2Q5k_mSxjM.
Forbes. 2014. Facebook Buying WhatsApp - $19B In Cash And Stock,
http://www.forbes.com/sites/benkepes/2014/02/19/breaking-facebook-buying-whatsapp19b-in-cash-and-stock/
Frier, S. 2013. With $200 Million in Revenue, South Korea's Top Messaging App Is All Smiley
Faces, Bloomberg, http://www.bloomberg.com/news/2013-12-22/with-200-million-inrevenue-south-korea-s-top-messaging-app-is-all-smiley-faces.html
Garg, R. and Telang, R. 2013. “Inferring App Demand from Publicly Available Data," MIS
Quarterly (37: 4), pp.1253–1264.
Ghose, A. and Han, S. 2011. “An Empirical Analysis of User Content Generation and Usage
Behavior on the Mobile Internet,” Management Science (57:9), pp. 1671–1691.
Ghose, A., Goldfarb, A., and Han, S. 2013. “How is the Mobile Internet Different? Search Costs
and Local Activities,” Information Systems Research (24:3), pp. 613–631.
Ghose, A. and Han, S. 2014. “Estimating Demand for Mobile Applications in the New
Economy,” Management Science (60:6), pp. 1470–1488
Gordon, M. E. The History of App Pricing, And Why Most Apps Are Free.
http://www.flurry.com/bid/99013/The-History-of-App-Pricing-And-Why-Most-Apps-AreFree#.VAVYBfmSz1B.
Hanemann, W.M. 1984. “The Discrete/Continuous Model of Consumer Demand,” Econometrica
(52:3), pp. 541–561.
Hansen, K., Singh, V., Chintagunta, P. 2006. “Understanding Store-Brand Purchase Behavior
across Categories,” Marketing Science (25:1), pp. 75–90.
Hendel, I. 1999. “Estimating Multiple-Discrete Choice Models: An Application to
Computerization Returns,” Review of Economic Studies (66), pp. 423–446.
Keane, M.P. 1993. “Simulated Estimation for Panel Data Models with Limited Dependent
Variables,” in Rao C.R., Maddala G.S., Vinod D. (eds) The Handbook of Statistics and
Econometrics. Elsevier Science Publishers, Amsterdam, pp. 545–572.
Khalaf, S. 2014. Apps Solidify Leadership Six Years into the Mobile Revolution.
http://www.flurry.com/bid/109749/Apps-Solidify-Leadership-Six-Years-into-the-MobileRevolution#.VAVXI_mSz1C.
Kim, J., Allenby, G.M., Rossi, P.E. 2002. “Modeling Consumer Demand for Variety,” Marketing
Science (21:3), pp. 229–250.
Krishnamurthi, L. and Raj, S.P. 1988. “A Model of Brand Choice and Purchase Quantity Price
Sensitivities,” Marketing Science (7:1), pp. 1–20.
Lee, S., J. Kim, G.M. Allenby. 2013. "A Direct Utility Model for Asymmetric Complements,"
Marketing Science (32:3), pp. 454–470.
Luo, L., B.T. Ratchford, B. Yang. 2013. “Why We Do What We do: A Model of Activity
Consumption,” Journal of Marketing Research (50:1), pp.24–43
52
Luo, X., Andrews, M., Fang, Z., and Phang, Z. 2014. “Mobile Targeting,” Management Science
(Forthcoming).
Malthouse E.C. and Blattberg, R.C. 2005. ‘‘Can We Predict Customer Lifetime Value?’’ Journal
of Interactive Marketing (19:1), pp. 2–16.
Rossi, P.E., McCulloch, R.E., and Allenby, G.M. 1996. “The Value of Purchase History Data in
Target Marketing,” Marketing Science (15:4), pp. 321–340.
Rust, R.T. 1985. “Selecting Network Television Advertising Schedules,” Journal of Business
Research (13:6), pp. 483–494.
Rust, R.T. 1986. Advertising Media Models: A Practical Guide. Lexington Books, Lexington,
Massachusetts.
Singh, V., Hansen, K., and Gupta, S. 2005. “Modeling Preferences for Common Attributes in
Multicategory Brand Choice,” Journal of Marketing Research (42:2), pp. 195–209.
Spissu, E., A.R. Pinjari, C.R. Bhat, R.M. Pendyala, K.W. Axhausen. 2009. “An Analysis of
Weekly Out-of-Home Discretionary Activity Participation and Time Use Behavior,”
Transportation (36:5), pp.483–510.
Taylor, S. and Todd, P. 1995. “Assessing IT Usage: The Role of Prior Experience,” Management
Information Systems Quarterly (19:4), pp. 561–570.
Tellis, G.J. 1988. “The Price Sensitivity of Competitive Demand: A Meta-Analysis of Sales
Response Models," Journal of Marketing Research (25:4), pp. 331–341.
Xu, J., Forman, C., Kim, J., and van Ittersum, K. 2014. “News Media Channels: Complements or
Substitutes? Evidence from Mobile Phone Usage,” Journal of Marketing (78:4), pp. 97–112.
Yonhap News. 2014. Android rules S. Korean market in 2013. http://english.yonhapnews.co.kr/
business/2014/01/21/30/0501000000AEN20140121001000320F.html.
Zhang, J. and Krishnamurthi, L. 2004. “Customizing Promotions in Online Stores,” Marketing
Science (23:4), pp. 561–578
53