Professional Documents
Culture Documents
App 1
App 1
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
MANUFACTURING & SERVICE OPERATIONS MANAGEMENT
Vol. 25, No. 2, March–April 2023, pp. 756–774
https://pubsonline.informs.org/journal/msom ISSN 1523-4614 (print), ISSN 1526-5498 (online)
a
Department of Operation and Supply Chain Management, Monte Ahuja College of Business, Cleveland State University, Cleveland, Ohio
44114; b Department of Management Science, Darla Moore School of Business, University of South Carolina, Columbia, South Carolina 29208;
c
Department of Operations Management, Weatherhead School of Management, Case Western Reserve University, Cleveland, Ohio 44106
*Corresponding author
Contact: m.chung74@csuohio.edu, https://orcid.org/0000-0002-9579-6026 (MC); luv.sharma@moore.sc.edu,
https://orcid.org/0000-0002-1710-4178 (LS); malhotra@case.edu (MKM)
Received: October 20, 2019 Abstract. Problem definition: Initial product design decisions are critically important for
Revised: April 20, 2021; February 9, 2022; mobile apps, which show a relatively short time from launch to peak usage, thus providing
August 12, 2022; November 29, 2022 a narrow window for success and no time for course correction based on market reaction.
Accepted: November 30, 2022 Mobile apps are designed using a highly modular architecture based on software develop
Published Online in Articles in Advance: ment kits (SDKs), with SDK choices being sequentially determined along three dimen
January 13, 2023
sions—multiplicity (total number of SDKs), compatibility (SDK co-occurrence frequency),
and novelty (SDK degree of newness to the developer). We evaluate the consequence of
https://doi.org/10.1287/msom.2022.1181 these decisions on initial market success in the context of mobile gaming app design. Aca
demic/practical relevance: The resulting conceptual framework aids developers in deter
Copyright: © 2023 INFORMS
mining the modularity of digital product development. Methodology: We formulate an
instrumental variables least absolute shrinkage and selection operator regression model to
estimate relationships of interest using a proprietary data set extracted from the application
programming interface server of a leading mobile apps intelligence firm. Results: We find
a negative impact of SDK multiplicity on initial success. High SDK compatibility can miti
gate this negative effect, whereas high SDK novelty can exacerbate the negative effect of
multiplicity. Post hoc analysis shows that business-to-consumer (B2C) communication
features can also mitigate this negative impact. Managerial implications: Prior product
modularity research has predominantly focused on physical products or relied on single-
dimensional modularity measures. Our study conceptualizes modularity as multidimen
sional and investigates how these multidimensional SDK-based modularity choices impact
the performance of a key category of digital products—mobile apps. We demonstrate that
increasing multiplicity, essential in certain markets, negatively affects initial success. How
ever, firms can enhance SDK compatibility, reduce SDK novelty, and use B2C communica
tion channels to mitigate this negative impact.
756
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 757
relies on quickly achieving user critical mass, which has mostly focused on the impact of success factors
relates to in-app advertisements and in-app sales, the such as user ratings (Liu et al. 2014), portfolio strategies
two primary sources of revenue for freemium apps. (Lee and Raghu 2014), and pricing (Garg and Telang
One of the most notable examples of a successful free 2013, Ghose and Han 2014) on app performance. In con
mium product is Angry Birds 2, which accumulated 10 trast, by explicitly focusing on product design as a criti
million downloads in just three days (Grundberg 2012). cal success factor, we contribute to the literature by
The product life cycle for freemium mobile apps dif diving deeper and showing how the actual modularity
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
fers significantly from traditional product life cycles decisions made by the designers impact the initial suc
(Downes and Nunes 2014). This segment has a com cess of the freemium apps.
pressed two-stage product life cycle with exponential To execute our study, we procured access to a market-
growth in the initial stages, followed by gradual decay. leading mobile app analytics firm’s proprietary applica
In most apps, this exponential growth and peak num tion programming interface (API) server and manually
ber of users are achieved within a few days to weeks extracted the empirical data set, which contains daily
from the app launch. This peak number of users repre panel observations of 2,483 top-ranked, free mobile
sents the app’s maximum market penetration, which gaming U.S. iOS apps spanning 3.5 years. These apps
puts an upper limit on the app’s revenue generation are released by 804 developer firms that employ an
potential. Further, the relatively short time from launch average of 11 employees while simultaneously servic
to the peak provides a narrow window of opportunity ing an average of 2.5 apps with an average of 10 SDKs
and no time for course correction based on market reac installed. Based on measures obtained from this data
tion, making prelaunch development decisions critical set, we use two outcome measures to capture initial
to success. Following prior literature, which evaluates success—Bayesian-weighted five-star review ratings and
product performance during early stages in the product the number of daily active users. These metrics capture
lifetime (Song et al. 2011) or takeoff periods (Agarwal different aspects of performance, with review ratings cap
and Bayus 2002), we refer to this exponential growth turing users’ attitudes toward a product (i.e., satisfaction/
period as the initial stage of the app’s lifetime and conse dissatisfaction, perceived quality) and daily active users
quently, retain our focus on ensuring the initial success capturing users’ behavior (Fishbein 1979, Fazio et al.
of an app (i.e., maximizing market penetration in the 1989). On average, the initial peak is reached in 29.8 days
initial stage). since launch. These observations show the resource con
Product design is often considered a critical opera straints for the app development firms (Clutch 2018) and
tional decision that significantly affects the user recep the short time to the initial peak for freemium mobile
tion and commercial success of technology products gaming apps. Our context is grounded in mobile gaming
(Hise et al. 1989, Ulrich 1995, Krishnan and Ulrich apps because they provide a rich setting for examining
2001, Ramachandran and Krishnan 2008, Özkan et al. the impact of product design and market-related factors
2015). Mobile apps exhibit a highly modular product because of the highest degree of competition intensity
design, implemented through software development and shortest average life span (Gordon 2018) among all
kits (SDKs), which are nonproprietary, modularized, app categories. There is also a high degree of perfor
pretested, and reusable codes that enable firms to add mance variance because of this industry’s winner-takes-
specific app features at a low cost (Dalmasso et al. all nature, where popular games generate $2.4 million
2013, Atreyi et al. 2015). Modularity decision refers to per day compared with an average app at $4,000 (Strauss
the choice of these SDK modules in a typical mobile 2013). Within this segment, we focus on top-ranked
app development project and is conceptualized to have mobile apps only because unranked apps are primarily
three dimensions. The first dimension, multiplicity, is composed of nonprofit-focused apps, crude prototypes,
measured as the number of modules in the product and social experiments, which may not have market suc
architecture (Novak and Eppinger 2001, Jacobs and cess as their primary objective.
Swink 2011, Vickery et al. 2016). The second dimension, We estimate the impact of SDK modularity deci
compatibility, is measured as the degree of integration sions using an instrumental variables (IV) regression
knowledge and process overlap among modules (Succi model with the strongest instruments selected using
et al. 1998, Gershenson et al. 2003). The third dimension, the least absolute shrinkage and selection operator
novelty, is measured as the degree of preexisting inte (LASSO). Our results show that a 1% increase in SDK
gration experience or the newness of the module to the multiplicity leads to a 0.55% decrease in daily active
developer team (Griffin 1997, Novak and Eppinger users and a 0.10% decrease in weighted user ratings.
2001, Tatikonda and Stock 2003). However, we find that this negative effect of multi
Our study aims to answer the following research plicity can be mitigated through higher SDK compati
question. How does the product design of mobile apps bility. Specifically, we find that a 1% increase in SDK
along the three dimensions of modularity impact the compatibility significantly mitigates the negative effect
app’s initial stage success? Prior work on mobile apps of SDK multiplicity on weighted user ratings by 0.09%
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
758 Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS
and daily active users by 0.52%. We also find that a 1% research on modularity focuses on product design and
increase in SDK novelty significantly exacerbates the associated production challenges (Vickery et al. 2016,
negative effect of multiplicity on weighted user ratings Dooley et al. 2019), reorganizing/grouping tasks into
by 0.08% but not the effect on daily active users. Choos production cells (Baldwin and Clark 2000, Browning
ing SDKs with high compatibility and low novelty may 2001, Danilovic and Browning 2007), and modularity in
thus partially mitigate the negative impact of multiplic use (Christensen and Rosenbloom 1995, Baldwin and
ity, thus enhancing the app’s performance. We conduct Clark 2000). Modularity research for digital products
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
a post hoc analysis to examine the impact of another focuses on its implication for the firm and software
design decision—business-to-consumer (B2C) commu performance-enhancing flexibility (Nambisan 2002),
nication SDKs. This analysis reveals that including B2C enhancing innovativeness (Sullivan et al. 2001), reduc
communication SDKs can mitigate the negative impact ing cost (MacCormack et al. 2006), and affecting imitability
of multiplicity by 0.01% on weighted user ratings and (Ethiraj et al. 2008). All these prior studies conceptualize
0.19% on daily active users. modularity decisions along the single dimension of the
Our results have multiple implications for theory and degree of modularity and treat modules as homogenous.
practice. First, we propose a multidimensional frame However, this is an oversimplification of reality because
work for modularity decisions, thus extending the lit the modules’ integration difficulty, functional fit with
erature on product modularity that relies on a single other modules, and knowledge requirements are hetero
measure (Nambisan 2002, Lau Antonio et al. 2007, geneous. Accounting for this heterogeneity of modules is
Danese and Filippini 2012, Vickery et al. 2016). Second, important as the integration challenges developers face
we extend the modularity literature into the digital may affect the entire system’s performance.
product design domain, where multiplicity’s negative Although prior research on modularity predomi
effect is more pronounced. Our findings are aligned nantly focused on highlighting the positive effects of
with industry reports (Shoavi 2017) that express con modular design, some studies mention that increased
cerns about excessive modularization in mobile app modularization can lead to reduced innovation, lower
development because of the low cost of adoption for responsiveness to market needs, and suboptimal pro
SDKs and the abundance of SDK options. Third, we duct configurations (Brusoni and Prencipe 2001, Ches
delineate the significance of using multiple performance brough and Kusunoki 2001, Ethiraj and Levinthal
2004, Ernst 2005). This may be exacerbated in mobile
metrics. Although extant literature has primarily relied
apps because developers face challenges that differ
on user ratings only to evaluate app performance, we
from other settings. Most app developers are small
show the implications of modularity choices on both the
start-ups with fewer than 50 employees (Clutch 2018).
attitude of users (captured through weighted user ratings)
Correspondingly, app development budgets are tight
and their usage behaviors (captured through daily active
(generally below $250,000), with development times
users). Finally, our results yield a framework that aids
often less than three months to ensure quick time to
developers in making their modularity decisions. We first
market (Panko 2017). Given these constraints, app
note that the decision to expand the app feature set and
developers rely on SDKs—modularized solutions or
the resulting multiplicity should be carefully made
collections of software code libraries that quickly add
because of the multiplicity’s negative impacts. For low-
desired features to apps (see Online Appendix Table
multiplicity apps, the negative impacts are insignifi WA2) to optimize their effort and return on invest
cant; hence, there is less need for a mitigation strategy. ment. Although modular design enabled by SDKs
However, high multiplicity can be inevitable, espe promotes knowledge reuse, which has been shown to
cially when developing feature-rich apps that target reduce development cost and time (Barnes et al. 1988,
more advanced users. In such situations, we recom Banker and Kauffman 1991) and improve final prod
mend developers prioritize choosing highly compati uct quality (Knight and Dunn 1998), it can also lead to
ble SDKs and then, further supplement them with reduced innovation, lower responsiveness to market
low-novelty SDKs to alleviate the negative impacts of needs, and suboptimal product configurations (Brusoni
multiplicity. Adding B2C communication SDKs will and Prencipe 2001, Chesbrough and Kusunoki 2001,
also help improve app performance in such situations. Ethiraj and Levinthal 2004, Ernst 2005) because of
increased knowledge requirements of the functionality
2. Literature Review of individual SDKs and on calibrating performance at
2.1. Physical and Digital Product the SDK and app level. There may also be a lack of fit
Modularity Research of SDKs within the app architecture or between SDK
Prior literature has extensively studied product mod interfaces, which may severely affect app performance.
ularity for physical and digital products. A review of Therefore, although using SDKs has increased the ease of
research on product modularity is presented in Online integrating features, the resulting feature proliferation
Appendix Table WA1. Regarding physical products, has also increased crashes, viruses, malware, privacy
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 759
breaches, battery drain, and lags (Shoavi 2017, SafeDK 1, steps 1 and 2). This decision determines the total
2018). To aid developers’ app modularity decisions, number of features to be implemented in the app and
there is a need to disentangle the complex effects of defines the multiplicity of the modular system structure.
modularity on a product’s performance and provide Our first modularity dimension, multiplicity, is therefore
prescriptions for maximizing its benefits. measured by the total number of SDKs installed in the
app. Each SDK contributes to an app’s feature. Hence,
2.2. Modularity Dimensions and Sequential increased multiplicity contributes to building a feature-
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
Decision Making in App Development rich app. Feature-rich apps generally attract more en
We propose a multidimensional modularity decision gaged users with a higher lifetime value (Venkatesan
framework that disentangles these opposing effects and Kumar 2004, Dahana et al. 2019). At the same time,
on product performance. Specifically and as already higher multiplicity increases the complications of ensur
mentioned before, we conceptualize modularity to ing fit between SDKs and integration of a more exten
have three dimensions—multiplicity, compatibility, and sive feature set, thus requiring additional resources and
novelty, which are determined sequentially at different development time (Strong and Volkoff 2010, Berente
stages of the app development cycle. To better illustrate et al. 2016). This is especially true in the mobile app
this sequential set of decisions, we present in Table 1 a domain, where third parties, including open-source
typical mobile app development process, along with communities, often develop these SDKs—resulting in
the decisions regarding dimensions and stakeholders poor functional fit, which often necessitates the devel
involved in each critical step of planning, design, devel opment of wrappers and interfaces to ensure seamless
opment, testing, and maintenance/support. This devel integration. High multiplicity also implies more possi
opment process is compiled from case studies (Flora ble interdependencies to manage (Vickery et al. 2016).
et al. 2014, Ghandi et al. 2017), meta-analyses (Jabangwe These interdependencies may cause design changes to
et al. 2018), and anecdotal evidence (Invonto 2022). impact multiple modules simultaneously and impose
Decisions in the planning phase (strategy/analysis further challenges for developers, eventually affecting
and planning) involve determining the app platform the app’s quality. Too many features may also nega
and core feature requirements of the app, which are tively impact the users’ experience because of a steeper
made by the marketing/management of the firm (Table learning curve associated with app usage (Thompson
et al. 2005). Finally, heavy reliance on third-party mod recommend integrating other supporting SDKs, such as
ules may also reduce the innovativeness of the app and OFX, which add capabilities to sell virtual goods, virtual
make them look generic in its appearance and function currency, and downloadable content. As a result, these
ality (Xue et al. 2019). SDK pairs are frequently installed together in an app.
Given the number of features for the app under Another decision developers can consider when
development, developers can influence the ease of SDK coping with SDK integration challenges during the
integration and app performance through a deliberate design and development phase (Table 1, steps 3 and
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
choice of SDKs. During the design and development 4) is to rely on familiar SDKs. Our third dimension of
phase, the design and coding engineers translate these modularity, novelty, captures the developer’s degree
feature requirements into technological code structures of newness of the SDKs installed in an app. Because
and choose/implement the corresponding SDKs (Table 1, novelty reflects the developer’s lack of experience with
steps 3 and 4) (Masi et al. 2012). Integration or coordina using an SDK, it measures the integration challenges
tion of an SDK refers to installing the SDK package resulting in increased development lead time and
library to the integrated development environment and reduced product quality (Clark and Fujimoto 1991,
configuring the various SDK component parameters to Griffin 1997, Novak and Eppinger 2001, Tatikonda
add a particular feature to the app under development. and Stock 2003) when integrating modules into the
The available SDK component parameters and the re product. Reducing novelty means developers tap into
sulting features that can be built vary from SDK to SDK. their prior development experience and choose SDK
This variation in configuration parameters requires devel pairs that they are already familiar with. Even if the
opers to study and learn each SDK to integrate it properly developers assigned to the integration task do not
(Murkin 2021). For example, integration effort may be possess the knowledge, they can interact with their
lower for modules with relatively more standardized colleagues in the firm to indirectly access the required
interfaces—which refers to common, agreed-upon me know-how (Narayanan et al. 2009). However, SDKs
chanisms for interaction (Baldwin and Clark 2000) and with a high degree of novelty will not have the
integration processes (Succi et al. 1998, Gershenson required know-how even among the pooled group of
et al. 1999). An overlap may exist in the configuration developers in a team.
parameters or the configuration user interface between
SDKs, so that knowledge for integration of one SDK
may be similarly applied to another SDK. We refer to
3. Research Hypotheses
this degree of the integration process and interface 3.1. Main Effect of SDK Multiplicity
overlap between a given SDK pair as SDK compatibil SDK multiplicity influences the feature richness of an
ity. Highly compatible SDKs may contribute to the app. Although offering higher multiplicity is viewed to
same app feature cohesively (i.e., perform similar SDK positively impact the users’ perceived quality (Carpenter
logic). These SDKs are often installed together as a et al. 1994, Brown and Carpenter 2000) and achieve dif
standard best practice in building certain app features. ferentiation from competitors (Nowlis and Simonson
Relying on this observation, we capture SDK compati 1996), the negative impacts listed earlier may outweigh
bility as the average degree of SDK pair co-occurrence these benefits for mobile apps. Still, high levels of multi
frequency in the app market for all SDK pairs installed plicity may occur for feature-rich apps that target more
in an app. When a developer chooses highly compati advanced users. From the user perspective, higher mul
ble SDKs, the existence of best practices and tools (e.g., tiplicity requires them to learn about these new fea
interface bridges) to integrate them seamlessly can poten tures, which may cause feature fatigue (Thompson
tially reduce software glitches and required development et al. 2005). Because higher multiplicity increases the
resources. required knowledge and skill levels of the users, it may
As an example of compatibility, chat server support also limit the product appeal to a limited set of users who
SDKs are commonly installed with multiplayer sup are more experienced and have the time and commitment
port SDKs because they jointly facilitate the multi to understanding the complicated interactions between
player environment with user interactions. Because the features (Qiu et al. 2017).
player communication features are essential in a mul From the developer perspective, although using
tiplayer environment, most multiplayer apps also install third-party SDKs allow developers to add features
chat support SDKs, with integration best practices being conveniently, they also result in a low functional fit
well documented. SDK developers even recommend with the developers’ unique needs (Strong and Volk
some SDK combinations with high compatibility. For off 2010, Berente et al. 2016). Additionally, higher mul
example, OpenFeint is a popular mobile app develop tiplicity can potentially compromise product performance
ment SDK that adds multiplayer support such as lead and represent other challenges at the development end.
erboards, achievements, forums, community space, and Even though some end users may prefer apps with many
real-time player communication. OpenFeint developers features and higher SDK multiplicity, they may not like
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 761
the product if it is clunky or slow functioning. Further, high-compatibility SDKs will make the integration
integrated SDKs could have differential user preferences, challenges less pronounced than the former, as develo
data and device resource requirements, and interactions pers will be able to leverage existing best practices and
with other SDKs. Increasing SDK installations without an tools to integrate compatible SDKs seamlessly. Hence,
in-depth understanding of their integration challenges compatibility should have a beneficial moderating impact
may lead to increased crashes, viruses, malware, privacy on the relationship between multiplicity and initial market
breaches, battery drain, and lags (Shoavi 2017, SafeDK success.
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
iOS market over 3.5 years from 2015 to 2018. To obtain 4.1. Variable Description
the sample for our research purpose, we restricted the For each variable description, we include subscripts i
app category to gaming apps in the U.S. app market to denote apps, j for app developers, c for the subcate
(252,568 apps). Then, we selected freemium apps (i.e., gory, t for the initial success date, and l for the app
apps with no download fees), which still account for launch date.
92% of the apps in the category (232,362 apps). We fur
ther narrowed it down to the top 1,000 gross down 4.1.1. Dependent Variable. We argue that initial suc
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
loaded apps. To the top 1,000 ranked apps in 2015, we cess is critical to the app’s overall success. The mobile
added 2,192 apps that entered the ranking between 2016 app market has a compressed two-stage product life
and 2017. Finally, to calculate some of the instruments cycle with rapid exponential growth followed by grad
included in our model, we only included apps devel ual decay. The peak number of daily users achieved
oped by developers that had launched at least two or during this initial stage most likely represents the
more apps in the market during our time frame. This app’s maximum market penetration, putting an upper
dropped 709 apps from our sample. The resulting sam limit on the revenue generation potential for the app. We
ple size consists of 2,483 apps in total. For all apps, we propose a way to identify an app’s initial success and
extracted general information such as file size, in-app define two performance measures—review rating, which
purchase availability, and release dates at the time of captures the users’ satisfaction with the app’s quality, and
launch. As for app performance metrics, we collected logged daily active user, which captures the users’ level
daily/monthly active users and five-star user ratings. of engagement based on actual app execution behavior.
Information on SDKs is compiled by the proprietary mar Quality perceptions or satisfaction toward a product
ket intelligence firm, which downloads each new app form the basis of the users’ attitude. Prior research has
and then decompiles, decrypts, and analyzes the source shown that attitude is a strong precursor of behavior but
code to see which SDKs and development tools are not a sufficient condition (Ajzen and Fishbein 1977, Fazio
installed/uninstalled daily. After deleting apps with et al. 1989). Similarly, the quality of a service or product
missing data, our final sample had 2,282 apps from may be a factor that causes user churn, but there can be
804 developers. other factors (e.g., switching cost) that can prevent the
In this study, we restrict our focus to iOS gaming user from churning (Bolton 1998, Ahn et al. 2006). Using
apps in the United States that attained a top 1,000 the two dependent variables, we can assess whether a
rank at least once during their lifetime. This is done specific modularity decision impacts only attitude or
because the iOS market provides a user experience on both attitude and behavior.
relatively homogeneous iPhone devices (Cuadrado We use the time series of daily active users to iden
and Dueñas 2012). This eliminates concerns about tify the app’s initial success. However, the noisy daily
unobserved user device characteristics from our model fluctuations in time series make identification diffi
estimations. The iOS platform also has extensive guide cult. Therefore, to find the first peak in performance
lines that SDKs must conform to. This results in a highly from the underlying trend in daily active users, we
integrated SDK environment that provides a more ho rely on a commonly used technique in finance where
mogenous user experience. Focusing on a single country the shift in stock price trends is detected by moving
allows us to reduce country-level confounds and over average crossover points (Murphy 1999, Pätäri and
come language barriers in the data. Additionally, by con Vilska 2014). The crossover point is defined as the time
centrating on top-ranked apps, we eliminate those apps point when the short-term moving average crosses a
that tend to be social experiments or crude prototypes long-term average. This crossover signals that the
that are not aiming for commercial success. We focus on momentum is shifting in one direction—signaling a
mobile gaming apps because developers in this category growth or decline for the underlying trend. We define
continuously experiment with novel development pro the maximum before the first crossover that signals a
cesses and adopt cutting-edge technologies to expand the declining trend as the initial success. The time of initial
user base and prolong app life span. This is evident in success t is identified by calculating a short-term 3-day
the average app development cost (Dogtiev 2018) and the moving average and a relatively long-term 10-day
average number of SDKs embedded in gaming apps, moving average of daily active users and then finding
which are the highest across all app categories (Shoavi the first crossover points. The peak is the maximum
2017). Additionally, gaming apps are the revenue-driving daily active user of the raw variable that occurs right
category for the entire market, and there is a high degree before the first moving average crossover point. To
of performance heterogeneity because of this industry’s choose the appropriate moving average lengths, we
winner-takes-all nature. Finally, we identify our time search for the earliest global maximum in the apps’
point of initial success for each app using the method daily active user time series. The app that reached the
described in Section 4.1.1 and convert it into cross- global maximum the earliest was 15 days after launch.
sectional data to conduct the analysis. If we set the long-term moving average to longer than 15
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 763
days, we would lose these observations from our sample. where S(n1 , : : : , n5 ) is the adjusted star score based on
We then tested multiple moving average lengths within the number of k star ratings (nk ), sk is the point value
15 days that accurately capture this earliest global maxi for k stars (i.e., one point for one star, five points for
mum. The maximum identification was robust to 62 days five stars), N is the total number of ratings, K is the
in moving average windows. Contrary to our concerns maximum number of stars (i.e., K � 5 for a five-star
about time series fluctuations, there were minimal fluctua rating system), and zα=2 is the 1 � α=2 quantile of a
tions in the early stage. An alternative method of using first normal distribution (i.e., we use 1.65 for 95% confi
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
differencing (i.e., the first-order derivative of the raw daily dence). We also provide estimation results using the
active user variable) and choosing the peak point where raw review score as a robustness check.
the slope changes from positive to negative yielded consis To capture the user engagement levels, we use
tent results. An example of identifying the initial peak lndaui,t, an app i’s logged daily active users at the time
using two moving average lines is shown in Figure 1. of initial success t. Variable daily active users are users
From the initial success, we define two dependent that open the app at least once during a particular
variables of interest to estimate the impact of the day. In prior literature, the number of downloads
developers’ SDK choice on the user-perceived quality (Garg and Telang 2013, Ghose and Han 2014) and app
of the app and user engagement using wratingsi,t and rankings (Lee and Raghu 2014) were primarily used
lndaui,t. wratingsi,t is the app i’s Bayesian-weighted to measure app demand. In contrast, daily active users
average five-star user rating score at the initial success measure user engagement (i.e., stickiness) of the app
date, t. A five-star user review rating, a commonly used because it captures revisiting users and drops churned
measure for the user-perceived quality of a product, is users. Therefore, daily active users is a performance
used to capture the users’ perception of the app’s quality metric that mobile app practitioners carefully moni
(Lee and Raghu 2014, Wang et al. 2018). We acknowledge tor. Market-leading mobile app analytics tools such
the possibility of measurement error in the raw five-star as Flurry Analytics and Google Play Developer Con
review ratings. Ratings with a small number of reviews sole collect and provide this metric to developers
cannot represent the true quality of the app, and such (Del Gallego et al. 2016).
review scores can be easily swayed by outliers (i.e.,
extremely satisfied dissatisfied users) (Talton et al. 2019). 4.1.2. Independent Variables of Interest. The variable
Hence, we estimate the main effects using the Bayesian multiplicityi,l is the total number of app i’s installed
adjusted review score measure (Miller 2014, Li and Hecht SDKs at launch l. We measure compatibilityi,l as the
2021), adopted by Amazon and several other e-commerce average degree of co-occurrence frequency between
sites (Zhang et al. 2011). This method calculates the Bayes all SDK pairs installed in app i at the time of launch l.
ian adjusted rating score as follows: The concept of co-occurrence is often used in research
S(n1 , n2 , n3 , n4 , n5 )
to infer the degree of skill overlap or relatedness
among individuals, organizations, and industries
XK
nk + 1 based on the outcome state (Teece et al. 1994, Neffke
� sk
k�1
N +K and Henning 2013). We assume that SDKs endorsed
vffi�
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�
ffiffiffi as compatible pairs by the SDK developers or known
u �
u XK 2 nk +1� �PK nk +1�2 to be compatible combinations to implement a cohe
u s � k�1 sk N+K
t k�1 k N+K
sive app feature will frequently co-occur in apps. First,
� zα=2 , (1)
N+K+1 we calculate the co-occurrence frequency matrix
between all SDK pairs for the entire sample. Second,
all pairwise frequencies are normalized. Third, we cre
Figure 1. (Color online) Initial Peak Identification of ate a pairwise matrix for each app that indicates
“Cinderella Fall” by Disney whether an SDK pair is installed. Finally, the normal
ized frequencies are added and averaged per app for
all existing SDK pairs to calculate compatibility.
The variable noveltyi,l is defined as the developer’s
average degree of newness of the SDKs installed in an
app i at the time of launch l. We use a binary variable
to indicate whether a specific SDK is new to the devel
oper during the app development. Next, we calculate
the focal app’s average newness of all SDKs. One par
ticular concern regarding noveltyi,l is the potential con
founding with the novelty of the SDK to the users.
This would happen mainly when the developer intro
duces a novel SDK in their app, which is also a novel
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
764 Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS
SDK to the users. If we do not control for the con visual appeal is an essential factor influencing user
founder, market SDK novelty, omitted variable bias adoption. An app’s file size with heavy graphical fea
may be an issue. We address this endogeneity by (1) tures and three-dimensional modeling can go up to six
adding a control variable for the market-level novelty gigabytes. These visual resources account for most
of the SDK; (2) accounting for developer unobserved of the app’s file size. Therefore, controlling the app’s
heterogeneity that may influence the innovativeness file size allows us to control the aesthetical aspect of
of the developer, such as developer fixed effects, the app (Liu et al. 2014, Wang et al. 2018). We control
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
developer app portfolio size, and number of employ the market hype for the app that may build up before
ees that proxies the resources of the developer; and the app launch by including firsthypei,l, which is the
(3) instrumenting novelty with instruments that are logged number of downloads of an app i during the
orthogonal to user SDK novelty. first launch date l (Lee and Raghu 2014). To control
developer effects, we include empnoj,l, which is the
4.1.3. Control Variables. For control variables, com number of employees at developer j at the time of app
petitionc,t is measured as the Herfindahl–Hirschman launch l. This variable is a proxy for the developer’s
Index in a subcategory c at the initial success t. This resources, potentially affecting app innovativeness
controls the increase in market competition, which and SDK decisions. We also include appnumj,l, which
can affect the demand and quality perceptions (Roma is the number of apps that developer j is operating at
and Vasi 2019). Variable devexpj,l is the cumulative the time of focal app launch l. This variable accounts
number of apps developed and launched by the de for the resource constraints and level of managerial
veloper j at the time of launch l. This variable controls attention that the developer can spend on the newly
for developer experience and brand effects that may launched app (Lee and Raghu 2014, Allon et al.
influence the initial performance of the app (Ghose 2022). To account for market-level novelty, we create
and Han 2014, Lee et al. 2020). Variable ageresi,l is the mktnoveltyi,l by first using a binary variable for each
four classification age-restriction levels coded as one SDK that indicates the first app that adopted the SDK
for “ages 4+,” two for “ages 9+,” three for “ages 12+,” in our sample and then averaging the binary variable
and four for “ages 17+” for an app i at time of launch for each app i to calculate the average market-level nov
l, which is treated as a continuous variable in our anal elty at launch l. We include subcategory fixed effects to
yses. Apps with different age restriction levels target control for potential user behavior heterogeneity. We
different user segments in the market, which may add year and month fixed effects to control for unob
vary in size and preferences (Ghose and Han 2014). served time-correlated shocks. We include developer
Variable multicategoryi,l is the number of subcategories fixed effects to account for developer unobserved
that app i is enlisted at the time of launch l. As the heterogeneity.
number of enlisted subcategories increases, the app Tables 2 and 3 show the summary statistics and corre
can get exposure to a broader user pool through cross lations of the key variables used in our model. All vari
genre listings (Ghose and Han 2014). Variable multi able correlations are below 0.7, and the mean variance
platformi,l is an indicator variable coded as one for an inflation factor (VIF) scores of all estimated models are
app i launched in multiple operating platforms at the below 10. For focal independent variables (i.e., multiplic
time of launch l to control potential spillover effects ity, compatibility, novelty), mean VIF scores are below
across app stores (Ghose and Han 2014). Similarly, five (Kutner et al. 2004). Therefore, we do not find strong
the variable adsi,t controls the developer’s marketing evidence of multicollinearity in the data.
efforts by measuring the number of impressions (expo
sures) made for the app i in the advertisement network 5. Econometric Model
at initial success t. We control the app i’s file size mea We present instrument variables to address endogene
sured in bytes at launch l using appsizei,l. The app’s ity in SDK multiplicity, compatibility, and novelty.
Variable Mean Standard deviation Min Max Variable Mean Standard deviation Min Max
Variables (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15)
Dependent variable
(1) ln(DAU)
(2) Wratings 0.34*
Independent variable
(3) Multiplicity 0.06* 0.02
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
For interactions, we include the interactions of in multiplicity, compatibility, and novelty levels in each
struments to the instrumental variable set. Next, we mean calculation. The logic is derived from previous
address the resulting high dimensionality of the works that identify instruments from average charac
instruments by using the LASSO (Tibshirani 1996) teristics of other products supplied by the same firm
with postdouble-selection method (Belloni et al. 2012, (i.e., capability and resources within the firm), pro
2013), which penalizes weak instruments and instru ducts provided by other firms within the same market
ments correlated with the error term to refine the (i.e., industry standards or conventions) (Berry et al.
instruments. 1995), or studies that use level of market power or
competitive pressure within the market (Berry and Jia
5.1. Endogenous SDK Multiplicity, Compatibility, 2010, Sharma et al. 2020). Mobile app developers’
and Novelty decision on SDK installation can be affected by the
Developers’ SDK choice may depend on unobserved degree of multiplicity, compatibility, and novelty the
app-specific characteristics such as the quality of the developer is comfortable handling, competitive pres
idea, monetization structure, addictiveness, etc. These sures in the market, and cultural/local development
additional characteristics are shaped with the support conventions. However, these capabilities and external
of features enabled by SDKs. However, these charac environments are less likely to be correlated with the
teristics can also directly impact the users’ perceived app-specific unobserved characteristics, which are the
quality and engagement levels. Because we do not sources of endogeneity.
have controls for these characteristics, the estimation
for our variables of interest, multiplicityi,l, compatibilityi,t, 5.2. IV Regression with LASSO Selection
and noveltyi,t, may be biased because of omitted factors We evaluate the impact of SDK modularity decisions
(Wooldridge 2010). To address this endogeneity con on an app’s market performance using an IV regres
cern, we rely on instruments that explain the develo sion model. For app i, subcategory c, developer j at
per’s SDK decisions at the time of launch that are less the time of launch l and time of initial success t, we
likely to be correlated with the omitted factors. We formulate the following system of equations:
report both instrumental variables regression and stan yi,l � β0 + B1 X + B2 Γ + δc + ξj + τl + ɛi,l (2)
dard ordinary least squares (OLS) estimation results for
xi,l � β10 + Θ1 Z + Θ2 Γ + δc + ξj + τl + ui,l , (3)
comparison.
For endogenous SDK choice, we use instruments where yi,l denotes the dependent variables wratings
devaj,l , subcatac,l , and countryak,l , which are defined as the and lndau and B1 is a vector of estimated coefficients
mean values of the endogenous variable a of installed for our endogenous variables of multiplicity, compatibility,
SDKs in apps within the same developer j, subcate and novelty. The vector B2 contains the estimated coeffi
gory c and apps developed in the same country k at cients for control covariates in the vector G. We include
the time of launch l. We exclude the focal app’s subcategory fixed effects, dc , developer fixed effects, jj ,
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
766 Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS
and time fixed effects (i.e., year, month), tl . Errors are As our IV regression model is overidentified, we can
represented by the term, ei,l . For each endogenous vari test whether the excluded instruments are appropriately
able xi,l in vector X, we estimate the first-stage model independent of the error process. We report the test re
where Z is a vector of instruments including devaj,l , sult of overidentifying restrictions using the Hansen J,
subcatac,l , and countryak,l . Coefficients of the instruments which regresses the residuals from the 2-stage least
and control covariates are represented by Q1 and Q2 , squares equation on all instruments (Hansen 1982). We
respectively. For interaction models, all interactions are fail to reject the null hypothesis that all instruments are
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
treated as endogenous. We include interactions of the uncorrelated with the error term and that the overidenti
instrumental variables in the vector of Z (Bun and Harri fying restrictions are valid for all models (p > 0.10).
son 2019). All models are estimated with robust standard Overall, the test results reported in Table 4 jointly sup
errors clustered at the developer level. port the validity of our instruments.
For each endogenous variable, we have three in
struments. For the interaction models, we have nine 6. Empirical Results
instruments for each interaction term. With the high This section presents our results from estimating the
dimensionality of instruments, the validity assumptions instrumental variable regression model and demon
can be easily violated. Therefore, we apply the LASSO strates robustness with alternative specifications.
on the first-stage regression (Tibshirani 1996) using the
postdouble-selection method (Belloni et al. 2012, 2013), 6.1. Estimation Results
which selects strong instruments among the available The results of the estimates are shown in Table 4. The
set. When instruments are added to the first-stage first two columns show effects on weighted user rat
model, LASSO penalizes each addition of an instrument ings, whereas the following two show effects on
by minimizing the following objective function, Q(B, Θ): logged daily active users. In Models (1A) and (2A),
X
n we present estimates of the direct effect of SDK multi
Q(B, Θ) � ‖multiplicityi,l � ΓB � ZΘ‖2 + λ θj , (4) plicity. Compatibility and novelty are included but
j�1
are considered control covariates because their effects
where G is the vector of exogenous regressors with depend on multiplicity, which needs to be examined
estimated coefficient vector B and Z is the vector of in an interaction model. We calculate the elasticities
instruments with estimated coefficient vector Q. The for the estimated effects. We find that a 1% increase in
last term in Equation (4) is the penalization term on SDK multiplicity leads to a decrease in weighted user
each instrument coefficient, θ, with a weight of λ. The ratings by 0.10% (β � �0.024, p < 0.01) and daily active
selection process chooses instruments that are strongly users by 0.55% (β � �0.058, p < 0.01). These results
correlated with the endogenous variable. We report provide support for Hypothesis 1. Despite the benefits
the first-stage estimation results in Online Appendix of increased app features, SDK multiplicity leads to
Table WA3, showing the LASSO-selected instruments. reduced app performance during its initial period.
For the IV regression to produce an unbiased esti Next, we introduce interactions into the model
mate, the instruments must satisfy two validity as to estimate the interaction effects between the focal
sumptions. First, the instrument must be relevant independent variables. As shown in Models (1B) and
such that it is strongly correlated with SDK multiplic (2B), we find a significant positive moderation effect
ity, compatibility, and novelty. We can verify this by of compatibility on the relationship between SDK
analyzing the first-stage regression F statistics of the multiplicity and app performance. Specifically, a 1%
instruments on the endogenous variables (Staiger and increase in compatibility mitigates the negative rela
Stock, 1994). The robust rk Wald F statistics for the tionship between multiplicity and weighted user rat
first-stage regression are 76.53 for the main effect ings by 0.09% (β � 0.130, p < 0.05) and the negative
model and 17.77 for interaction models. These values relationship between multiplicity and daily active
are larger than the critical values of 12.20 for 5% maxi users by 0.52% (β � 0.314, p < 0.05). Thus, Hypothe
mal IV relative bias and 12.33 for 15% maximal IV sis 2 is supported. On the other hand, the modera
size, which suggests that the instruments are not tion effect of novelty is significant and negative on
weakly associated with the endogenous regressor the relationship between multiplicity and weighted
(Stock and Yogo 2005). The significant underidentifi user ratings but insignificant on the relationship
cation of test results for all models (p < 0.01) further between multiplicity and logged daily active users.
support the relevance of the instruments (Kleibergen Specifically, a 1% increase in novelty exacerbates the
and Paap 2006). Second, the instrument must plausi negative effect of multiplicity on weighted user ratings
bly satisfy the exclusion restriction, which means that by 0.08% (β � �0.002, p < 0.05). Therefore, Hypothesis 3
the instruments do not directly influence the weighted is only partially supported. We present the interaction
review ratings and daily active users at the initial peak. plots in Figure 2 with low and high levels of the
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 767
variable calculated as 61 standard deviation from We find that this negative impact is mitigated if the
the mean. Across all plots, we see a negative perfor developers choose highly compatible SDKs or fewer
mance impact of multiplicity. As the number of SDKs novel SDKs (i.e., familiar SDKs). Interestingly, the nov
increases, developmental challenges imposed by nov elty effect is relatively minimal compared with the com
elty can impede the integration of the modules and patibility effects. Also, the moderation effect is only
deteriorate the final app quality and user engagement. significant for weighted user ratings. One explanation is
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
768 Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS
(a) Multiplicity × Compatibility Interaction Effect on Wratings (b) Multiplicity × Novelty Interaction Effect on Wratings
2.6
Low Novelty
2.6 Low Compatibility High Novelty
High Compatibility
2.4
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
2.4
2.2 2.2
2
2
Low high Low high
Multiplicity Multiplicity
6.5
5.5
Low high
Multiplicity
that user ratings are more sensitive to developer design boards, and support chat features (see Online Appendix
decisions than daily active user numbers. In other Table WA1). Prior research shows that B2C comm
words, although user ratings closely reflect the quality unications can stimulate customer sales by increasing
implications of the app feature design, a negative im the available product attribute information (Jang and
pact on ratings does not always lead to user churn. Chung 2015), thus facilitating user learning of the pro
Users may tolerate somewhat lower app quality and duct’s functionality and improving engagement levels
will only churn if the quality degradation is severe (Bol (De Giovanni 2019). Forming a responsive B2C relation
ton 1998, Ahn et al. 2006). For this reason, we may ship through active communication is known to build
observe a significant impact on user ratings rather than attraction, which may lead to technology adoption
daily active users. (Campbell et al. 2012).
To test this user-side multiplicity mitigation effect,
6.2. Post Hoc Analysis we estimate the moderating effect of SDK-based B2C
Our main analysis finds a consistently negative im communication features on the relationship between
pact of SDK multiplicity. We further find that the nov multiplicity and app performance. We operationalize
elty of the installed SDKs can amplify the adverse CommB2C as a binary variable, coded as one if the app
effects. However, these factors only focus on miti has B2C communication-related SDKs installed and
gating the developer-side challenges. From a user zero otherwise. The instrumental variables regression
perspective, handling a high-multiplicity app can be estimation results are presented in Table 5. We find
challenging because of feature fatigue caused by steep that B2C communication significantly moderates the
learning curves and experience requirements. There SDK multiplicity effect on weighted user ratings and
fore, we explore tools at the developer’s discretion to daily active users (Model (3A): β � 0.007, p < 0.05; Model
ease customer adoption barriers. Among the features (3B): β � 0.022, p < 0.05). The significant positive moder
enabled by SDKs, we notice the B2C communication ation suggests that communication features can miti
feature that includes push messaging, announcement gate the negative multiplicity effect on weighted user
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 769
Table 5. Post Hoc IV Regression Results OLS model allows us to examine the direction of the
bias caused by the omitted variable bias and the mag
Model (3A) (3B)
Dependent variable Wratings ln(DAU)
nitude of the correction. We present the non-IV OLS
regression results in Online Appendix Table WA3.
Key independent variables Most of the estimates are qualitatively consistent with
Multiplicity �0.030*** �0.063***
our main findings regarding their magnitudes and
(0.009) (0.016)
directions. We find the main effect and the interaction
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
focus on multiplicity only. Although the main effect of challenges of integration, feature fatigue, and limited
multiplicity is not significant, its magnitude and direc market appeal dominate the user-side benefits of having
tion are consistent with our main findings. This differ more features from SDK multiplicity. However, reducing
ence with the main analysis may be because of the multiplicity may be hard for feature-rich apps that target
meaningful correction from the Bayesian weighting more advanced users. For this situation, we identify that
and the fact that the model does not fully satisfy the the negative impact of multiplicity is mitigated by an in
instrument validity concerning the overidentification crease in compatibility and exacerbated by an increase
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
test. However, the results are much more consistent in in novelty. Furthermore, enabling B2C communication
the interaction model (Online Appendix Table WA6, can further reduce the negative impacts of multiplicity
Model (4B)). The first-order effect of multiplicity and through improved user engagement by keeping them
the interaction effects of compatibility and novelty are abreast of the latest developments in the game, up
consistent with the main model findings from the grades, events, and updates from fellow gamers.
weighted user rating model. Finally, we comment on the significance of the de
pendent variables deployed in this study—daily active
7. Discussion and Conclusion users and user ratings. First, we notice slight differences
in the results between the two dependent variables. In
7.1. Theoretical Implications
particular, we find the moderation effect of novelty to
Extant studies have primarily captured product mod
be significant only in the relationship between multi
ularity along a single composite measure (Nambisan
plicity and user rating. We think this difference can be
2002, Lau Antonio et al. 2007, Danese and Filippini
attributed to the fact that user rating measures the users’
2012, Vickery et al. 2016). In contrast, we capture the
perceptions and attitudes toward a certain app. In con
impact of modular components along the three dimen
trast, daily active users capture users’ actual behavior,
sions of multiplicity, compatibility, and novelty. Instead
such as sustained use. Although attitudes and percep
of treating each module as homogenous, we believe a
tions are significant antecedents of behavior, they may
useful framework for understanding modular design’s
not be a sufficient condition (Fishbein 1979, Fazio et al.
impact on performance lies in conceptualizing each
1989). Unless the decrease in perceived quality is suffi
modular installation as increasing multiplicity while at
ciently low or critical to the app’s performance, users
the same time, influencing joint compatibility and nov may still tolerate the inconvenience and show sus
elty of modules. Our results demonstrate the differen tained use of an app. Second, instead of using the raw
tial impacts of these dimensions, highlighting the need user rating measure, we recommend that researchers
to consider modularity as a multidimensional construct. should make an adjustment because users generally
Second, our proposed framework extends the discus provide ratings when they are either highly satisfied or
sion of modularity, which has primarily focused in the extremely dissatisfied with the app—making raw user rat
past on physical products (Novak and Eppinger 2001, ings a biased estimate of app performance. Also, the infer
Danilovic and Browning 2007, Danese and Filippini ence of app quality may be inaccurate when the app does
2012, Vickery et al. 2016), to the unique context of design not have many reviews. Daily active users, on the other
ing digital products—specifically, mobile apps using hand, captures user engagement with the app, making it
SDKs. Increased penetration of mobile devices is result a more accurate representation of the relationship bet
ing in rapid growth for the mobile apps sector, which ween the choice of an app’s portfolio of SDKs and initial
has unique characteristics that differentiate it from other success. The weighting of the ratings variable and using
software and physical product development contexts. multiple outcomes allow us to triangulate the effects of
The use of SDKs has increased the ease of integrating modularity decisions. Therefore, we urge future research
features to such an extent that apps are witnessing ers and practitioners to assess app performance using
feature proliferation, resulting in increased crashes, multiple metrics, including daily active users.
viruses, malware, privacy breaches, battery drain, and
lags (Shoavi 2017). With limited guidance from the lit 7.2. Managerial Implications
erature, SDK integration in this important sector is To highlight the performance implications of the mo
primarily based on anecdotal evidence. Developing dularity decisions, we present a 2 × 2 framework in
and testing an SDK modularity framework for this
novel context provide much-needed empirical evidence Figure 3. Framework for SDK Compatibility and Novelty
on the link between SDK choice and performance. Combinations
Third, our results shed light on the interrelation
Low Novelty High Novelty
ships and impact of the three modularity dimensions (-1 ) (+1 )
on mobile app performance. Specifically, we find a ne
High Compatibility (+1 ) Quadrant 4 Quadrant 3
gative impact of multiplicity on app performance. As
hypothesized, in the context of SDKs, the developmental Low Compatibility (-1 ) Quadrant 1 Quadrant 2
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 771
Figure 3 that depicts four possible scenarios of combining and novelty. Choosing SDKs without these consid
compatibility and novelty for a given multiplicity level. erations can lead to detrimental outcomes. Moreover,
Each quadrant represents a modularity scenario based on based on the defining role of compatibility, we recom
mean 61 standard deviation split for high and low mend that developers first prioritize choosing high-
levels of compatibility and novelty. Quadrant 1 re compatibility SDKs during the development of high
presents developers that rely on low-novelty SDKs multiplicity apps and then consider reducing the nov
from prior development experience to handle low- elty of chosen SDKs.
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
(a) (b)
8
Quadrant 4 Quadrant 4
2.5 Quadrant 3 Quadrant 3
Quadrant 1 Quadrant 1
Linear Prediction of ln(DAU)
Linear Prediction of Wratings
7
Quadrant 2 Quadrant 2
2
6
1.5
5
1 4
3
.5
0 10 20 30 40 0 10 20 30 40
Multiplicity Multiplicity
Notes. (a) Multiplicity impact on wratings by quadrants. (b) Multiplicity impact on daily active users by quadrants. The difference between quad
rants 3 and 4 is not significant (p > 0.10) at all levels of multiplicity in panel (b). The difference between quadrants 1 and 2 is not significant (p >
0.10) at all levels of multiplicity in panel (b).
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
772 Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS
Korean mobile telecommunications service industry. Telecom Clutch (2018) Top Mobile App Development Companies. Retrieved
munications Policy 30(10–11):552–568. November 14, 2018, https://clutch.co/directory/mobile-application-
Ajzen I, Fishbein M (1977) Attitude-behavior relations: A theoretical developers.
analysis and review of empirical research. Psych. Bull. 84(5): Cuadrado F, Dueñas J (2012) Mobile application stores: Success fac
888–918. tors, existing approaches, and future developments. IEEE
Allon G, Askalidis G, Berry R, Immorlica N, Moon K, Singh A Comm. Magazine 50(11):160–167.
(2022) When to be agile: Ratings and version updates in mobile Curry D (2021) App Revenue Data (2020). Business of Apps. Re
apps. Management Sci. 68(6):4261–4278. trieved January 31, https://www.businessofapps.com/data/app-
Atreyi K, Ye JH, Teo HH (2015) Comparing potential and actual revenues/#1.
innovators: An empirical study of mobile data service innova Dahana WD, Miwa Y, Morisada M (2019) Linking lifestyle to cus
tion. MIS Quart. 39(3):667–682. tomer lifetime value: An exploratory study in an online fashion
Baldwin CY, Clark KB (2000) Design Rules: The Power of Modularity retail market. J. Bus. Res. 99:319–331.
(MIT Press, Cambridge, MA). Dalmasso I, Datta SK, Bonnet C, Nikaein N (2013) Survey, compari
Banker RD, Kauffman RJ (1991) Reuse and productivity in inte son and evaluation of cross platform mobile application devel
grated computer-aided software engineering: An empirical opment tools. Saracco R, BenLetaief K, Gerla M, Palazzo S,
study. MIS Quart. 15(3):375–401. Atzori L, eds. Wireless Communications and Mobile Computing
Barnes B, Durek T, Gaffney J, Pyster A (1988) A framework and eco Conference (IWCMC), 2013 9th International (IEEE), 323–328.
nomic foundation for software reuse. Tracz W, ed. Software Danese P, Filippini R (2012) Direct and mediated effects of product
Reuse: Emerging Technology (IEEE Computer Society Press, modularity on development time and product performance.
Washington, DC), 77–88. IEEE Trans. Engrg. Management 60(2):260–271.
Belloni A, Chernozhukov V, Hansen C (2013) Inference on treat Danilovic M, Browning TR (2007) Managing complex product devel
ment effects after selection among high-dimensional controls. opment projects with design structure matrices and domain
Rev. Econom. Stud. 81(2):608–650. mapping matrices. Internat. J. Project Management 25(3):300–314.
Belloni A, Chen D, Chernozhukov V, Hansen C (2012) Sparse mod De Giovanni P (2019) A feature fatigue supply chain game with
els and methods for optimal instruments with an application to cooperative programs and ad-hoc facilitators. Internat. J. Produc
eminent domain. Econometrica 80(6):2369–2429. tion Res. 57(13):4166–4186.
Berente N, Lyytinen K, Yoo Y, King JL (2016) Routines as shock del Gallego NP, Urag OB, de Dios Bulos R (2016) Predicting daily
absorbers during organizational transformation: Integration, active users for match-3 mobile games. Aviso K, Ilao J, eds.
control, and NASA’s enterprise information system. Organ. Sci. DLSU Research Congress (De La Salle University), 1–6.
27(3):551–572. Dogtiev A (2018) How Much Does App Development Cost?
Berry S, Jia P (2010) Tracing the woes: An empirical analysis of the Business of Apps. Retrieved July 20, 2018, http://www.business
airline industry. Amer. Econom. J. Microeconomics 2(3):1–43. ofapps.com/guide/app-development-cost/.
Berry S, Levinsohn J, Pakes A (1995) Automobile prices in market Dooley KJ, Pathak SD, Kull TJ, Wu Z, Johnson J, Rabinovich E
equilibrium. Econometrica 63(4):841–890. (2019) Process network modularity, commonality, and green
Bolton RN (1998) A dynamic model of the duration of the custo house gas emissions. J. Oper. Management 65(2):93–113.
mer’s relationship with a continuous service provider: The role Downes L, Nunes P (2014) Big Bang Disruption: Strategy in the Age of
of satisfaction. Marketing Sci. 17(1):45–65. Devastating Innovation (Penguin, London).
Brown CL, Carpenter GS (2000) Why is the trivial important? A Ernst D (2005) Limits to modularity: Reflections on recent develop
reasons-based account for the effects of trivial attributes on ments in chip design. Indust. Innovation 12(3):303–335.
choice. J. Consumer Res. 26(4):372–385. Ethiraj SK, Levinthal D (2004) Modularity and innovation in com
Browning TR (2001) Applying the design structure matrix to system plex systems. Management Sci. 50(2):159–173.
decomposition and integration problems: A review and new Ethiraj SK, Levinthal D, Roy RR (2008) The dual role of modularity:
directions. IEEE Trans. Engrg. Management 48(3):292–306. Innovation and imitation. Management Sci. 54(5):939–955.
Brusoni S, Prencipe A (2001) Unpacking the black box of modular Fazio RH, Powell MC, Williams CJ (1989) The role of attitude acces
ity: Technologies, products and organizations. Indust. Corporate sibility in the attitude-to-behavior process. J. Consumer Res.
Change 10(1):179–205. 16(3):280–288.
Bun MJG, Harrison TD (2019) OLS and IV estimation of regression Fishbein M (1979) A theory of reasoned action: Some applications
models including endogenous interaction terms. Econometric and implications. Nebraska Sympos. Motivation 27:65–116.
Rev. 38(7):814–827. Flora HK, Wang X, Chande SV (2014) An investigation into mobile
Campbell DE, Wells JD, Valacich JS (2012) Breaking the ice in B2C application development processes: Challenges and best prac
relationships: Understanding pre-adoption e-commerce attrac tices. Internat. J. Modern Ed. Comput. Sci. 6(6):1–9.
tion. Inform. Systems Res. 24(2):219–238. Garg R, Telang R (2013) Inferring app demand from publicly avail
Carpenter GS, Glazer R, Nakamoto K (1994) Meaningful brands able data. MIS Quart. 37(4):1253–1264.
from meaningless differentiation: The dependence on irrelevant Gershenson JK, Prasad GJ, Allamneni S (1999) Modular product
attributes. J. Marketing Res. 31(3):339–350. design: A life-cycle view. J. Integrated Design Process Sci. 3(4):
Chesbrough H, Kusunoki K (2001) The modularity trap: Innovation, 13–26.
technology phase shifts and the resulting limits of virtual organi Gershenson JK, Prasad GJ, Zhang Y (2003) Product modularity:
zations. Nonaka I, Teece D, eds. Managing Industrial Knowledge: Definitions and benefits. J. Engrg. Design 14(3):295–313.
Chung, Sharma, and Malhotra: Impact of Modularity Design on Mobile App Launch
Manufacturing & Service Operations Management, 2023, vol. 25, no. 2, pp. 756–774, © 2023 INFORMS 773
Ghandi L, Silva C, Martinez D, Gualotuna T (2017) Mobile applica Computing, Applications, and Services, vol. 110 (Springer, Berlin,
tion development process: A practical experience. Alturas B, Heidelberg), 64–79.
ed. Iberian Conf. Inform. Systems Tech., CISTI (IEEE Computer Miller E (2014) Ranking Items with Star Ratings. Evanmiller.org.
Society, Washington, DC), 1–6. Retrieved February 19, https://www.evanmiller.org/ranking-
Ghose A, Han S (2014) Estimating demand for mobile applications items-with-star-ratings.html.
in the new economy. Management Sci. 60(6):1470–1488. Murkin P (2021) What is SDK in a SaaS solution and how to inte
Gordon ME (2018) Benchmarking the Half-Life and Decay of Mobile grate SDK. InAppStory. Retrieved January 26, 2022, https://
Apps. Flurry. Retrieved July 14, 2018, https://www.flurry.com/ inappstory.com/blog/how-to-integrate-sdk.
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
Stock JH, Yogo M (2005) Testing for weak instruments in linear IV regres Thompson DV, Hamilton RW, Rust RT (2005) Feature fatigue:
sion. Andrews DWK, Stock JH, eds. Identification and Inference for When product capabilities become too much of a good thing. J.
Econometric Models: Essays in Honor of Thomas Rothenberg Marketing Res. 42(4):431–442.
(Cambridge University Press, Cambridge, UK), 80–108. Tibshirani R (1996) Regression shrinkage and selection via the lasso.
Strauss K (2013) The $2.4 Million-per-Day Company: Supercell. For J. Roy. Statist. Soc. B 58(1):267–288.
bes. Retrieved July 19, 2018, https://www.forbes.com/sites/ Ulrich K (1995) The role of product architecture in the manufactur
karstenstrauss/2013/04/18/the-2-4-million-per-day-company- ing firm. Res. Policy 24:419–441.
supercell/#23b37cdf6fc1. Venkatesan R, Kumar V (2004) A customer lifetime value frame
Downloaded from informs.org by [124.17.124.215] on 14 March 2024, at 18:33 . For personal use only, all rights reserved.
Strong DM, Volkoff O (2010) Understanding organization-enterprise work for customer selection and resource allocation strategy. J.
system fit: A path to theorizing the information technology arti Marketing 68(4):106–125.
fact. MIS Quart. 34(4):731–756. Vickery SK, Koufteros X, Dröge C, Calantone R (2016) Product
Succi G, Valerio A, Vernazza T, Succi G (1998) Compatibility, stan modularity, process modularity, and new product introduction
dards, and software production. StandardView 6(4):140–146. performance: Does complexity matter? Production Oper. Man
Sullivan KJ, Griswold WG, Cai Y, Hallen B (2001) The structure and agement 25(4):751–770.
value of modularity in software design. Software Engrg. Notes Wang Q, Li B, Singh PV (2018) Copycats vs. original mobile apps: A
26(5):99–108. machine learning copycat-detection method and empirical anal
Talton JO III, Dusad K, Koiliaris K, Kumar RS (2019) How do peo ysis. Inform. Systems Res. 29(2):273–291.
ple sort by ratings? Brewster S, Fitzpatrick G, Cox A, Kostakos Wooldridge JM (2010) Econometric Analysis of Cross Section and Panel
V, eds. Proc. 2019 CHI Conf. Human Factors Comput. Systems Data (MIT Press, Cambridge, MA).
(Association for Computing Machinery, New York), 1–10. Xue L, Song P, Rai A, Zhang C, Zhao X (2019) Implications of applica
Tatikonda MV, Stock GN (2003) Product technology transfer in the tion programming interfaces for third-party new app development
upstream supply chain. J. Product Innovation Management 20(6): and copycatting. Production Oper. Management 28(8):1887–1902.
444–467. Zhang D, Mao R, Li H, Mao J (2011) How to count thumb-ups and
Teece DJ, Rumelt R, Dosi G, Winter S (1994) Understanding corpo thumb-downs: User-rating based ranking of items from an axi
rate coherence: Theory and evidence. J. Econom. Behav. Organ. omatic perspective. Amati G, Crestani F, eds. Conf. Theory
23(1):1–30. Inform. Retrieval (Springer), 238–249.