Economics of Innovation and New Technology

ISSN: 1043-8599 (Print) 1476-8364 (Online) Journal homepage: www.tandfonline.com/journals/gein20

Estimating demand function parameters of mobile

applications

Michael Scholz

To cite this article: Michael Scholz (2017) Estimating demand function parameters of
mobile applications, Economics of Innovation and New Technology, 26:7, 621-633, DOI:
10.1080/10438599.2017.1263444

To link to this article: https://doi.org/10.1080/10438599.2017.1263444

Published online: 04 Dec 2016.

Submit your article to this journal

Article views: 584

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gein20
ECONOMICS OF INNOVATION AND NEW TECHNOLOGY, 2017
VOL. 26, NO. 7, 621–633
http://dx.doi.org/10.1080/10438599.2017.1263444

Estimating demand function parameters of mobile applications

Michael Scholz
Faculty of Business Administration and Economics, University of Passau, Passau, Germany

ABSTRACT ARTICLE HISTORY

The vibrant market for mobile applications has raised awareness of several Received 2 February 2016
professional and also voluntary software developers. The key question Accepted 28 October 2016
especially for professional developers is how to improve the profit
KEYWORDS
gained with a developed app. Recent research provided evidence on the Mobile apps; demand
factors that determine the demand of a mobile app. This paper presents function; network value;
a procedure to estimate demand function parameters that are required willingness to pay
for developing pricing, advertising and also product update strategies.
More specifically, the procedure estimates an app’s maximal willingness
to pay, demand elasticity on price and network value. The procedure is
based on the Fulfilled Expectations Cournot Model and requires
knowledge about the apps being considered as substitutes to each
other. It is applied to a data set consisting of download rank data of
Apple iPhone apps.

1. Introduction
Smart phones showed a high growth rate in the last years due to the availability of a plenty of (useful)
smart phone apps. Around 191.1 million U.S. citizens owned a smart phone in August 2015 (comScore
2015). This corresponds to a penetration rate of 77.1%. The high demand for smart phones has
attracted many app developers. According to Statista (2015), around 1.5 million apps were available
for iOS smart phones and approximately 1.6 million apps were available for Android smart phones in
July 2015. Recent research, however, provides evidence that only a few of these apps generate rev-
enues that allow to operate a professional software development business (Garg and Telang 2013).
Researchers and app developers thus have a keen interest in understanding the drivers of app
demand and estimating demand functions to improve pricing, advertising and updating strategies.
Recent research has investigated several factors that might influence an app’s demand and, for
example, found evidence that pricing strategy, customer reviews, the required operating system
and an app’s file size influence demand (Ghose and Han 2014; Lee and Raghu 2014; Taylor and
Levin 2014). Ghose and Han (2014) investigated the effect of several mobile app characteristics
(e.g. app price, file size, description length, number of screenshots, age restriction) on app
demand and demonstrated that the demand increases with in-app purchase option but decreases
with in-app advertisement. The authors also found that a price discount is more effective in
Google Play compared to Apple App Store. Lee and Raghu (2014) analyzed the effect of different
factors on the survival of an mobile app in the top-grossing-300 charts. They provide evidence
that offering apps across multiple categories is one of the most important success factors. Apps
without an upfront price have a higher probability to survive in the top-300 charts than apps with
an upfront price (Lee and Raghu 2014). Review volume as well as review valence have been also
found to positively influence survival in the top-300 charts (Lee and Raghu 2014) and app

CONTACT Michael Scholz [email protected] Faculty of Business Administration and Economics, University of
Passau, Innstr. 43, 94032 Passau, Germany
© 2016 Informa UK Limited, trading as Taylor & Francis Group
622 M. SCHOLZ

demand (Ghose and Han 2014). Liu, Au, and Choi (2014) showed that the free version of an app
improves demand of the paid version if the free version gets a lot of positive customer ratings.
Carare (2012) provided evidence that consumers’ willingness to pay is approximately $US 4.50
higher for apps ranked in the top-100 charts than for the same unranked apps. In short, these
studies help to explain the demand of mobile apps and show which factors app developers
should focus on in order to improve their apps’ demand. These studies, however, do not estimate
demand function parameters and thus do not allow investigating an app’s network value, maximal
willingness to pay and demand elasticity on price. I close this gap by providing a novel estimation
procedure that enables app developers to estimate demand function parameters.
Apps often are offered at the same price throughout their whole life-cycle. There typically exist
several alternatives to a particular app A that might act as substitutes and that affect A’s demand.
The demand of mobile apps furthermore depends on both quality and network effects that might
change over time (Gretz and Basuroy 2013). Estimating the quality- and network-depended par-
ameters that determine app demand ultimately helps app developers to define successful pricing,
advertising and product update strategies. In other words, demand parameters are of high relevance
for several strategic decisions of app developers.
I contribute to recent research in information systems by providing a procedure for estimating
demand function parameters. More specifically, the procedure computes an app’s maximal valuation,
demand elasticity on price and network value. The procedure is based on the Fulfilled Expectations
Cournot Model (Novshek and Sonnenschein 1982; Katz and Shapiro 1985) and requires knowledge
about the apps being considered as substitutes to a particular app by consumers. With this pro-
cedure, app developers are able to base several of their strategic decisions (e.g. pricing decisions,
decision of when to update an app) on real consumer behavior.
The remainder of the paper is organized as follows. I present the model for estimating demand
function parameters of mobile apps in Section 2. An estimation procedure for this model is presented
in Section 3. Section 4 presents an empirical application of the proposed procedure. I compare the
estimation results of the Fulfilled Expectations Cournot Model to those of a Cournot Model for a
market without network effects and to those of a Fulfilled Expectations Cournot Model with non-
linear network effects in Section 5 and furthermore demonstrate that the estimation is robust also
in the case that only 80% of the data are used for estimation. The findings of the empirical investi-
gation are discussed in Section 7 with respect to their implication for practice and research. I conclude
the paper with a summary of the main findings in Section 8.

2. Model
In this section, I present the Fulfilled Expectations Cournot Competition Model as introduced in Katz
and Shapiro (1985). In the next section, I apply this model to estimate demand function parameters of
mobile applications. I assume a market with mobile apps that can be clustered according to the apps’
features. The apps constituting a particular cluster are assumed to be incompatible substitutes to
each other. I decided to model the app market as a Cournot rather than a Bertrand competition
due to the following three reasons. (1) App developers decide in a first step if they want to offer a
free or paid app. Paid apps in both Apple’s App Store and Google’s Play Store must have a
minimal price of $US 0.99 . The standard deviation of apps’ prices has been furthermore found to
be lower than $US 3 (Ghose and Han 2014). This indicates that app developers that offer paid
apps do have a limited control of their apps’ prices. (2) Recent research cites evidence that an
app’s demand is determined by multiple factors, such as the number of customer reviews, the
average customer rating, the average version age or the length of an app’s description (Ghose
and Han 2014). Thus, development and marketing restrictions (e.g. limited capacity to quickly
develop new versions, limited development experience) seem to mainly drive the demand of
app’s. These restrictions indicate that app developers’ strategic variable is rather quantity than
 ECONOMICS OF INNOVATION AND NEW TECHNOLOGY 623

price. (3) Following (Singh and Vives 1984), the dominant strategy for a firm is to strategically decide
about a goods quantity if goods of the competitors are substitutes.
The demand of an app i belonging to cluster c is denoted as xi . Each cluster of apps hence forms a
Cournot Competition (Negassi and Hung 2014) in which an app’s profit-maximal demand xi depends


on the demand of i’s competitors j = i, j [ c. A cluster’s total demand is defined by zc = i[c xi . The
price pi an app developer charges for app i to obtain a demand of xi is given by the following inverse
demand function for a Cournot Competition.

pi = ai − bzc . (1)

Parameter ai is the maximal willingness to pay for app i among all prospective consumers and b is
the cluster demand elasticity on price. Consumers will base their purchase decision not only on an
app’s basic value but also on its network value ni (xi(exp) ) resulting from the number of expected pur-
chasers xi(exp) of app i. The inverse demand function can hence be formulated as follows (see Katz and
Shapiro (1985) for a more detailed discussion of this model).

pi = ai − bzc + ni (xi(exp) ). (2)

If consumers expect the demand of app i to be xi(exp) , exactly xi(exp) will purchase i in the long-run
equilibrium (Katz and Shapiro 1985; Chen and Shimomura 1998). Thus, I can model a Fulfilled Expec-
tation Cournot Competition by assuming that xi = xi(exp) .
The network effect ni (xi ) is a function monotonically increasing in xi . In line with existing literatures,
I assume a linear network effect (Easley and Kleinberg 2010; Prasad, Venkatesh, and Mahajan 2010;
Wu, Ma, and Lui 2014)

ni (xi ) = ui xi , (3)

where ui is the network value parameter. Figure 1 shows the inverse demand curve for an exemplary
mobile app. The gray curve represents the inverse demand function without network effects. The dis-
tance between the total demand (black curve) and the demand based on only the basic values of the
app (gray curve) is the network effect of the app.
Given that all apps i [ c have an upfront price pi . 0, I can assume that consumers evaluate an
app’s quality, price and network value to make a purchase decision. I further can assume that app i
would get the complete demand for cluster c if i would be available at an price of pi = 0. The maximal
demand for a particular app i with i [ c is thus max(xi ) = zc .
In the next section, I demonstrate how to estimate parameters ai , b and ui based on panel data of
the top-N ranked apps.

Figure 1. Exemplary demand curve for a mobile app.

624 M. SCHOLZ

3. Estimation procedure
I propose a four-step procedure in order to estimate parameters ai , b and ui (Figure 2). The first step
consists of collecting demand data for several days in which the demand is rather stable. Fitting a
regression model based on the collected data with Equations (2) and (3) will likely result in an uni-
dentified model in which parameters ai are set equal to pi , because (i) the daily demand xi,t of app
i is rather a stable proportion of the total daily demand zc,t across all apps being in the same
cluster as i and (ii) prices of a particular app i might be stable over time. I thus propose to estimate
each app’s maximal valuation ai in a third step and finally fit a regression model to estimate the
cluster-specific demand elasticity zc on price and the app-specific network effects ui by using ai as
pre-specified information.


The total daily demand per cluster is given as zc,t = i[c xi,t . App i would obtain a demand of zc,t if
i is offered for free on day t. The inverse demand function for app i at day t is hence given as
pi,t = ai,t − mit xi,t with mi,t = pi /(xi,t − zc,t ) and ai,t = −mi,t zc,t . Parameter ai,t is the maximal willing-
ness to pay across all prospective consumers in cluster c whereas mi,t captures the cluster-specific
demand elasticity on price and the network effect of app i. I repeat this procedure for each day


t [ T. App-specific maximal valuations are finally computed as ai = 1/T Tt=1 ai,t where T is the
number of days.
In the fourth step, I fit the following regression model in order to estimate parameters b and ui .

pi,t − âi = (ûi − b̂)xi,t − b̂ x j,t + 1i,t . (4)
j=i,j[c


The total daily demand per cluster zc,t is here substituted by xi,t + j=i,j[c x j,t in order to get
uncorrelated covariates. I use a restricted maximum likelihood estimator to fit the mixed effects
model specified in Equation (4). This procedure is applied to data about apps for Apple’s iOS in
the next section.

4. Empirical application
4.1. Data
I collected panel data from appshopper.com in order to identify the network value of mobile apps.
Appshopper.com provides a daily updated list of the top-200 Apple iPhone apps. I gathered data
on a daily basis for several apps (listed in Table 1) between 28 October 2014 and 25 April 2015
(180 days). All apps were offered with a fixed upfront price and did not offer additional in-app pur-
chases. I identified three clusters of apps being substitutes to each other based on appcrawl.com
which is a platform that finds alternatives to iOS apps. Each cluster consists of three or four apps. I
assume that consumers’ demand elasticity on price is constant across all apps in one cluster.
I used only those days on which all apps in one cluster have been listed in the top-200 charts in
order to correctly compute the total demand zc for a cluster. Days on which the price of any of the

Figure 2. Proposed estimation procedure.

 ECONOMICS OF INNOVATION AND NEW TECHNOLOGY 625

Table 1. Summary of apps.

Daily demand
Cluster App Price (in US $) Mean SD
PDF Scanners (n = 54) Scanner Pro by Readdle 2.99 1207.89 438.41
TinyScan Pro 4.99 505.66 109.57
TurboScan 2.99 2642.79 824.65
Photo editors (n = 63) Afterlight 0.99 9448.44 2802.50
A Beautiful Mess 0.99 640.55 149.66
Camera+ 2.99 2286.17 722.97
Zombie killing games (n = 56) Call of Duty: Black Ops Zombies 6.99 613.25 138.20
Earn to Die 0.99 692.68 394.71
Plants vs. Zombies 0.99 2154.88 1138.85
Zombieville USA 2 0.99 655.02 161.44

apps in one cluster was changed have been also removed, since I assume that consumers do react to
price changes with some delay in time.1 All three PDF scanners were, for example, listed in the top-
200 charts on 54 days out of the 180 days of the data set (see Table 1). The observations at one day
are considered to be independent from the observations on another day. Figure 3 shows the esti-
mated demand of all PDF scanners at the 54 days used to estimate demand function parameters.
The data set includes each app’s sales rank and price. Sales ranks are based on download data that
are not publicly available. I used the estimation method proposed by Garg and Telang (2013) to esti-
mate app demand based on app sales ranks r. More specifically, the demand for an app i is calculated
as follows:

xi = 52, 958 · ri−0.944 . (5)

The demand xi for app i is assumed to be Pareto distributed (Brynjolfsson, Hu, and Smith 2003;
Chevalier and Mayzlin 2006) with scale parameter b = 52, 958 and shape parameter a = −0.944.
A negative shape parameter indicates that the demand increases with lower (i.e. better) sales rank
positions. The scale parameter defines the demand for the app at position 1. Thus, Equation (5) indi-
cates that the most famous iPhone app gets 52,958 downloads a day and apps at lower positions get
ri−0.944 of the downloads of the most famous app.
The estimated average daily demand of the investigated apps is also listed in Table 1. Afterlight is
with more than 9400 downloads per day the most famous app in the data set.
Table 1 shows that the daily demand is rather stable for each app. I hence can assume that the
proportion of consumers finally purchasing app i is approximately the same as the proportion of con-
sumers that purchased i on a particular day t. The apps are hence in the same product life-cycle phase
throughout the whole data period (Gretz and Basuroy 2013). This assumption allows us to estimate
parameters ai , b and ui with the proposed estimation procedure.

Figure 3. Estimated demand at all days the three PDF scanners were listed in the Top-200 charts.
626 M. SCHOLZ

4.2. Results
I estimated parameters âi , b̂ and ûi (see Equation (4)) independently for each cluster described in
Table 1. I followed the procedure proposed in Section 3 to derive the estimation parameters and
used a mixed effects regression with restricted maximum likelihood estimator in the fourth step.
The results are presented in Tables 2–4. T-tests on the estimated maximal valuations âi indicate
that the estimates for parameters âi are statistically significant. I also conducted t-tests on parameter
b̂ and found that the estimates are statistically significant. Parameter b̂ has been furthermore found to
be positive for all clusters of apps indicating that a higher demand requires a lower price. I conducted
x2 -test for the random effects ûi of each cluster and found these effects to be highly significant
(p , .001). Tables 2–4 show that the overall fit of the models is, with an R2 of at least 0.775 and a
root mean squared error (RMSE) of maximally 0.798, outstanding. Variance inflation factors below
2.5 furthermore indicate the absence of multicollinearity in the models.
The results show that Zombieville USA 2 has the highest maximal valuation â of approximately
$US 8.37. Call of Duty, another zombie killing game has a maximal valuation of only $US 1.19. The
quality of the apps in one cluster is hence rather heterogeneous.
Parameter b̂ is the estimate for the cluster-specific demand elasticity on price. The price for a PDF
scanner needs to be reduced by $US 0.30 in order to attract 1000 additional purchasers (without
regarding the network effects). A reduction of only $US 0.03 is sufficient to acquire 1000 additional
purchasers of a photo editor app (without regarding the network effects). The demand for photo
editors is more elastic on price than the demand for PDF scanners. Price discounts are thus most effi-
cient for photo editors and least efficient for PDF scanners.
The results also show that the network effect û of apps can be positive as well as negative. A posi-
tive network effect such as û = 0.00010 for Scanner Pro by Readdle shows that a price increase by
$US 1 can be compensated by an increased network effect if the app would have 1/0.00010 = 1000
additional purchasers. An increase of the demand for TurboScan by 1/0.00119 = 840, in contrast,
reduces the network value of this app in a way that is equivalent to a price increase of $US 1.

Table 2. Regression results for PDF scanners.

Variable Estimate Standard error p Value
âScannerProbyReaddle 4.22167 0.63911 .007
âTinyScanPro 7.83527 1.73372 .015
âTurboScan 5.72928 0.60123 .003
b̂ 0.00030 0.00005 <.001
ûScannerProbyReaddle 0.00010 – –
ûTinyScanPro 0.00105 – –
ûTurboScan −0.00119 – –
Observations 54 per app
Pseudo R2 0.823
RMSE 0.798

Table 3. Regression results for photo editors.

Variable Estimate Standard error p Value
âAfterlight 4.30206 1.02712 .017
âABeautifulMess 1.04631 0.01503 <.001
âCamera+ 3.70131 0.23708 .001
b̂ 0.00003 0.00001 .016
ûAfterlight −0.00028 – –
ûABeautifulMess 0.00046 – –
ûCamera+ −0.00012 – –
Observations 63 per app
Pseudo R2 0.859
RMSE 0.542
 ECONOMICS OF INNOVATION AND NEW TECHNOLOGY 627

Table 4. Regression results for zombie killing games.

Variable Estimate Standard error p Value
âCallofDuty 1.19127 0.07758 .001
âEarntoDie 2.05740 0.40829 .012
âPlantsvs.Zombies 1.19699 0.07440 .001
âZombievilleUSA2 8.37493 0.52134 .001
b̂ 0.00009 0.00001 <.001
ûCallofDuty −0.00153 – –
ûEarntoDie 0.00027 – –
ûPlantsvs.Zombies −0.00022 – –
ûZombievilleUSA2 0.00028 – –
Observations 56 per app
Pseudo R2 0.775
RMSE 0.267

The demand for TinyScan Pro benefits most from a network effect. An increase of the demand of
1/0.00105 = 952 consumers is equivalent to a price decrease of $US 1 for this app.
I additionally applied the estimation procedure (Section 3) to a pooled dataset including data from
all three categories and added a dummy variable for the category effect in the mixed effects
regression in step four. All three categories significantly affect the demand (pPDFScanner = 0.009,
pPhotoEditors , 0.001, pZombieGames = 0.035), which supports the assumption of cluster-specific effects.

5. Different specifications of the network effect

I have shown the results for a linear network effect in the previous section. In this section, I compare
different specifications of the network effect to the linear specification used so far. More specifically, I
compare the Fulfilled Expectation Cournot Model with linear network effects to a Cournot Model
without network effects in Section 5.1 and to two Fulfilled Expectation Cournot Models with non-
linear network effects in Section 5.2.

5.1. Comparison to a Cournot model without network effects

I computed demand function parameters using the Fulfilled ExpectationCournot Competition Model
in the previous section. This model assumes demand to be a variable depending on (i) consumers’
maximal willingness to pay, (ii) consumers’ demand elasticity on price and (iii) the network value.
In this section, I compare the Fulfilled Expectation Cournot Competition Model (Model 1 in Table 5)
to the Cournot Competition Model for markets without network effects (Model 2 in Table 5). This
comparison tests the robustness of the assumption that an app’s demand also depends on an
app’s network value. The models that will be compared in this section are summarized in Table 5.
The comparison is implemented as a five-fold cross validation in which four parts of the data are
used to estimate the two models and one part is used to compute the predictive validity of the
models. The Fulfilled Expectation Cournot Competition Model (Model 1 in Table 5) is estimated
with the procedure introduced in Section 3 whereas the Cournot Competion Model for markets
without network effects (Model 2 in Table 5) is estimated with an ordinary least squares estimator.
The estimated models allow to predict the demand for the part of data that has not been used for
model estimation. Prediction validity is finally computed as the RMSE for the demand predicted
with the two Cournot Competition Models. The results are presented in Table 6.

Table 5. Models for comparison.

Model Inverse demand function Network effects
Model 1 pi = ai − bzc + ui xi Existent
Model 2 pi = ai − bzc Not existent
628 M. SCHOLZ

Table 6. Predictive validity (as RMSE) of the two Cournot Competition Models.
Model PDF Scanners Photo editors Zombie killing games
Model 1 0.823 0.524 0.281
Model 2 1.907 1.447 0.586
Obs. for model estimation 44 51 45
Obs. for prediction 10 12 11

Table 6 clearly indicates that the proposed model (Model 1) outperforms the Cournot Competition
Model for markets without network effects. The demand of the investigated apps’ thus seems to be
influenced by a considerable (positive or negative) network effect.
The RMSE values for the predictive validity of Model 1 (see Table 6) are furthermore similar to the
RMSE value for the internal validity (see Tables 2–4). This is interesting because the predictive validity
has been calculated based on a holdout dataset (20% of the data were left holdout for estimating the
model and only used for computing the predictive validity). I also found no significant difference
between the demand parameters estimated based on all data and the demand parameters estimated
based on 80% of the data. I hence can assume that the proposed model is robust and also explains
the demand better than a reduced model (Model 2) that neglects network effects.

5.2. Comparison to a fulfilled expectation Cournot model with non-linear network effect
In Section 4, I assumed a network effect that linearly depends on the number of consumers that
downloaded an app. Network effects might be, however, non-linear in real markets (Easley and
Kleinberg 2010). In this section, I compare the Fulfilled Expectation Cournot Model with linear
network effects to two models with non-linear effects. The first non-linear model (Model 2 in Table 7)
assumes a network effect that is described by a concave increase of the network effect in the
number of users xi that downloaded app i. The second non-linear model (Model 3 in Table 7), in con-
trast, assumes a convex increase of the network effect. According to Katz and Shapiro (1985), I assume
network effects vi (xi ) for which vi (0) = 0. All models are described by their inverse demand function in
Table 7. Characteristic network value curves for the three models are furthermore shown in Figure 4.
As in Section 5.1, I conducted a five-fold cross validation to compare the models. All three models
were estimated by using the procedure described in Section 3. Prediction validity is again estimated
as RMSE. The results in Table 8 consistently indicate that the apps’ network effect is rather a concave
curve. In other words, the marginal network effect is decreasing in the demand xi .

Table 7. Models for comparison.

Model Inverse demand function Network effect
Model 1 pi = ai − bzc + ui xi Linear
Model 2 pi = ai − bzc + ui ln(xi + 1) Concave
Model 3 pi = ai − bzc + ui (exi − 1) Convex

Figure 4. Exemplary network value functions.

 ECONOMICS OF INNOVATION AND NEW TECHNOLOGY 629

Table 8. Predictive Validity (as RMSE) of the two Cournot Competition Models.
Model PDF Scanners Photo editors Zombie killing games
Model 1 0.823 0.524 0.281
Model 2 0.365 0.268 0.127
Model 3 2.448 1.859 0.645
Obs. for model estimation 44 51 45
Obs. for prediction 10 12 11

An increase in the demand of one app i is especially important to improve the network effect of i
and to attract more consumers of i if the actual demand of i is rather low. App developers should
hence make use of advertising and pricing strategies to especially attract consumers when their
apps are new on a market.
Parameters ai and ui are exactly the same for Model 1, Model 2 and Model 3. We only found
slightly different demand elasticity parameters b. The correlation between the demand elasticity
on price is 0.724 between Model 1 and Model 2 and 0.570 between Model 1 and Model 3 which indi-
cates that result the proposed estimation procedure are rather robust to changes of the specified
network effect function.

6. Robustness check
The demand parameters have been estimated using three or four apps being identified as substi-
tutes. Identifying substitutes is not a trivial task because apps typically are not perfect substitutes
to each other and each app is indeed different to another one to a certain degree. In this section,
I test the robustness of the results presented in Section 4 when each category consists of only two
or three apps. The following tables present the average estimates when each category consists of
one app less than in Section 4. Again the estimation procedure described in Section 3 was applied
with a restricted maximum likelihood regression in step 4.
The result in Tables 9–11 shows that the direction of the network effects is stable in almost all
cases (except that of ‘Scanner Pro by Readdle’). The estimates for all parameters are different
between the estimation with 3(4) apps and the estimation with 2(3) apps, but they are strongly cor-
related (r = 0.99 for PDF Scanners, r = 0.99 for Photo Editors and r = 0.90 for Zombie Killing Games).

Table 9. Robustness check for PDF scanners.

Variable Estimate (three apps) Estimate (two apps)
âScannerProbyReaddle 4.22167 7.42101
âTinyScanPro 7.83527 13.23101
âTurboScan 5.72928 8.15510
b̂ 0.00030 0.00111
ûScannerProbyReaddle 0.00010 −0.00009
ûTinyScanPro 0.00105 0.00201
ûTurboScan −0.00119 −0.00308
Observations 54 per app

Table 10. Robustness check for photo editors.

Variable Estimate (three apps) Estimate (two apps)
âAfterlight 4.30206 10.69550
âABeautifulMess 1.04631 1.17164
âCamera+ 3.70131 8.82686
b̂ 0.00003 0.00030
ûAfterlight −0.00028 −0.00087
ûABeautifulMess 0.00046 0.00171
ûCamera+ −0.00012 −0.00161
Observations 63 per app
630 M. SCHOLZ

Table 11. Regression results for zombie killing games.

Variable Estimate (four apps) Estimate (three apps)
âCallofDuty 1.19127 1.83028
âEarntoDie 2.05740 1.78543
âPlantsvs.Zombies 1.19699 4.20385
âZombievilleUSA2 8.37493 6.81267
b̂ 0.00009 0.00024
ûCallofDuty −0.00153 −0.00245
ûEarntoDie 0.00027 0.00040
ûPlantsvs.Zombies −0.00022 −0.00036
ûZombievilleUSA2 0.00028 0.00050
Observations 56 per app

Although the estimates are different when reducing the number of apps per category, several
findings are still valid. The ranking of the apps’ maximal willingness to pay is, for example, rather
stable across the two estimations. As a consequence, we are not sure about the concrete maximal
willingness to pay, but we are rather certain about which apps have the highest or lowest
maximal willingness to pay. The same holds for the network effects of the apps. The proposed pro-
cedure thus allows app developers to compare their apps to substitutes in terms of the maximal will-
ingness to pay of consumers and the network effect.

7. Discussion
I presented a procedure to estimate demand parameters of mobile applications. More specifically, the
proposed procedure estimates an app’s maximal valuation, demand elasticity on price and network
value. Based on panel data for iPhone apps, I analyzed the demand function parameters for apps in
three different clusters. The estimated models show an outstanding accuracy and seem to be robust
also if only a part of the data is used for estimation. I also demonstrated that the used Fulfilled Expec-
tation Cournot Competition Model leads to more accurate results than a Cournot Competition Model
for markets without network effects. A comparison of the model with a linear network value to two
models with non-linear network value showed that the app’s network value can be best modeled
using a concave function.
Some apps (TurboScan, Afterlight, Camera+ and Plants vs. Zombies) have been interestingly found
to have negative network effects. These apps are those having the highest demand and being in the
top-200 charts for the longest time. I thus can assume that these apps are in another life-cycle stage
than their substitutes. Gretz and Basuroy (2013) showed that network and quality effects are different
in different life-cycle stages. Network effects can shield to thwart new market entries only in the
growth and maturity phase but not in the following phases (Gretz and Basuroy 2013). Consumers
might furthermore be interested in using new apps which they can recommend to others via
word of mouth (Mowen, Park, and Zablah 2007). User looking for new and unknown apps might
be frustrated when several other users join the network of a particular app they are currently
using. These users will feel a negative network effect if too many other users discover their preferred
apps. A negative network value can also be the result of negative product ratings by consumers that
have a high influence (Domingos 2005). Congestion might be another reason for negative network
effects of apps that synchronize data (e.g. game score or states of a player) with a dedicated server.
The more users such an app has, the more likely the server might be overstrained which ultimately
reduces the value of the app when the number of its users is increasing.

7.1. Implications for research

The study contributes in two aspects to recent research. First, I provide a procedure to estimate
demand function parameters for mobile applications that is based on the demand estimation
 ECONOMICS OF INNOVATION AND NEW TECHNOLOGY 631

procedure by Garg and Telang (2013) and the network effect model by Katz and Shapiro (1985). The
proposed procedure estimates an app’s (i) maximal valuation, (ii) demand elasticity on demand and
(iii) network value. The procedure is dedicated toward the estimation of demand function parameters
of mobile applications. It can be, however, also applied to other goods for which price and demand
are available or estimable and for which an identification of substitutes is possible.
Second, I provide evidence that also mobile apps that do not have network functionalities such as
photo editors or PDF scanners do generate a network value. The network value of these apps might
exist due to product reviews and product recommendations. Review volume has been found to posi-
tively affect demand (Forman, Ghose, and Wiesenfeld 2008). A higher number of app reviews might
improve an app’s trustworthiness and ultimately its (network) value for consumers. Product reviews
form the basis for generating product recommendations. Oestreicher-Singer et al. (2013) demon-
strated that recommendations based on reviews also generate a network value of products that
do not have network functionalities.

7.2. Implications for practice

The results also offer interesting insights for app developers. The proposed procedure is suitable to
predict consumers’ valuation of a developer’s apps and also of the apps being in competition with a
specific app. App developers can use the demand function parameters to reassess their pricing strat-
egy. The results, for example, show that consumers rather have a low valuation for the photo editor
‘A Beautiful Mess’. Lowering this app’s price would lead to a strong demand improvement due to a
rather high network value of this app (Cambell 2013).
App developers can also use the results of the proposed procedure to justify their advertising
strategies. An app offering a high network value can be best promoted by users that like the app
and have many friends to whom they are willing to recommend the app. Apps having a high
maximal valuation offer its users a high quality and/or a high number of useful functionalities. Devel-
opers of these apps can, for example, make use of special offers in order to attract additional users.
Apps with special offers are presented in an extra category on Apple’s App Store and also Google’s
Play Store and might hence create more awareness than apps without special offers.
Finally, app developers can estimate demand parameters over time in order to better predict the
life-cycle phase of their app. Especially the point of reaching the critical mass is of interest for app
developers. The critical mass describes an unstable equilibrium in terms of the demand (Grajek
and Kretschmer 2012). If the demand is less than the critical mass, I have a downward pressure on
the consumption, whereas I have an upward pressure if the demand is above the critical mass
(Easley and Kleinberg 2010). The proposed approach helps to (i) determine if an app indeed has a
network value, (ii) if so, to what extent this network effect exists (this helps to identify the point of
the critical mass) and (iii) if there is a switch from a positive to a negative network effect.

7.3. Limitations
The proposed procedure has some limitations that provide avenues for future research. First, it relies
on the assumption that substitutes for an app are correctly identifiable in order to calculate a cluster’s
total demand zc . Mobile apps are heterogeneous and hence characterized by a limited substitutabil-
ity. I manually classified apps as substitutes to each other and assumed that an app j is either a sub-
stitute to i or not. Using fuzzy regression Wei, Sheng, and Hu (2002) might improve the accuracy of
the proposed procedure. The membership function for the fuzzy regression might be estimable
based on the apps’ descriptions or if available based on recommendation conversion rates (Oestrei-
cher-Singer et al. 2013). The total cluster demand is then computable as weighted sum of the
demand across all apps where the weights are assessed based on the cluster’s membership function.
Second, I collected data within a period of six month. Since demand parameters such as consu-
mers’ maximal valuation or the network value do change over time, I’m not able to identify such
632 M. SCHOLZ

changes in the dataset. I suggest implementing a sliding data approach in which the period of the
data used for estimation will be moved on a daily basis.
Third, I assume that an app’s demand depends on its price, its quality and its network value. In
other words, I assume demand to be a result of completely rational decisions. This might be,
however, not the case especially for apps that offer a hedonic value. Extending the approach to
better estimate an app’s hedonic value might improve the estimation of demand curves (Hew
et al. 2015).
Fourth, I collected the apps download ranks and used the procedure proposed by Garg and Telang
(2013) to infer app demand. The true app demand is not publicly available and the estimated
demand might differ from the true demand in a way that ultimately affects the demand parameters.
The demand rank furthermore depends on the number of consumers who downloaded an app. Poss-
ible in-app purchases are thus not covered in the data set. (Garg and Telang 2013) estimated that the
in-app revenue per download is $US 0.16 which is equivalent to an in-app purchase rate of 16% for an
in-app purchase option of $US 1.
And finally, I assume the observations on day to be independent on the observations on another
day. This helps to identify the regression model specified in Equation (4) but at the costs of maybe
biased estimates for the demand elasticity on price and the network effect. Durbin–Watson tests,
however, did not indicate a significant autocorrelation for the specified regression model and any
of the three app categories. The probability that the estimates are biased due to not captured depen-
dences between the observations is thus rather low.

8. Conclusion
This paper provides a procedure for estimating demand function parameters of mobile phone appli-
cations. The procedure helps to identify consumers’ maximal valuation for an app, an app’s demand
elasticity on price and an app’s network value. These parameters help app developers to implement
better pricing and advertising strategies. App developers can furthermore use the proposed pro-
cedure to identify the life-cycle phase their app currently belongs to. This information finally
allows making better decisions on when to update an app.

Note
1. I, for example, found no significant difference between the demand of the apps Camera+ and Scanner Pro by
Readdle when the developers reduced the price by US $1.

Disclosure statement
No potential conflict of interest was reported by the author.

ORCID
Michael Scholz http://orcid.org/0000-0003-0173-2900

References
Brynjolfsson, Erik, Yu Hu, and Michael D. Smith. 2003. “Consumer Surplus in the Digital Economy: Estimating the Value of
Increased Product Variety at Online Booksellers.” Management Science 49 (11): 1580–1596.
Cambell, Arthur. 2013. “Word-of-mouth Communication and Percolation in Social Networks.” American Economic Review
103 (6): 2466–2498.
Carare, Octavian. 2012. “The Impact of Bestseller Rank on Demand: Evidence from the App Market.” International
Economic Review 53 (3): 717–742.
Chen, Been-Lon, and Koji Shimomura. 1998. “Self-fulfilling Expectations and Economic Growth: A Model of Technology
Adoption and Industrialization.” International Economic Review 39 (1): 151–170.
 ECONOMICS OF INNOVATION AND NEW TECHNOLOGY 633

Chevalier, Judith, and Dina Mayzlin. 2006. “The Effect of Word of Mouth on Sales: Online Book Reviews.” Journal of
Marketing Research 43 (3): 345–354.
comScore. 2015. “comScore Reports August 2015 U.S. Smartphone Subscriber Market Share”.
Domingos, Pedro. 2005. “Mining Social Networks for Viral Marketing.” IEEE Intelligent Systems 20 (1): 80–82.
Easley, David, and Jon Kleinberg. 2010. Networks, Crowds, and Markets: Reasoning About a Highly Connected World.
Cambridge: Cambridge University Press.
Forman, Chris, Anindya Ghose, and Batia Wiesenfeld. 2008. “Examining the Relationship Between Reviews and Sales: The
Role of Reviewer Identity Disclosure in Electronic Markets.” Information Systems Research 19 (3): 291–313.
Garg, Rajiv, and Rahul Telang. 2013. “Inferring App Demand from Publicly Available Data.” MIS Quarterly 37 (4): 1253–
1264.
Ghose, Anindya, and Sang Pil Han. 2014. “Estimating Demand for Mobile Applications in the New Economy.”
Management Science 60 (6): 1470–1488.
Grajek, Michał, and Tobias Kretschmer. 2012. “Identifying Critical Mass in the Global Cellular Telephony Market.”
International Journal of Industrial Organization 30 (6): 496–507.
Gretz, Richard T., and Suman Basuroy. 2013. “Why Quality may Not Always Win: The Impact of Product Generation Life
Cycles on Quality and Network Effects in High-Tech Markets.” Journal of Retailing 89 (3): 281–300.
Hew, Jun-Jie, Voon-Hsien Lee, Keng-Boon Ooi, and June Wei. 2015. “What Catalyses Mobile Apps Usage Intention: An
Empirical Analysis.” Industrial Management and Data Systems 115 (7): 1269–1291.
Katz, Michael L., and Carl Shapiro. 1985. “Network Externalities, Competition, and Compatibility.” The American Economic
Review 75 (3): 424–440.
Lee, Gunwoong, and T. S. Raghu. 2014. “Determinants of Mobile Apps’ Success: Evidence From the App Store Market.”
Journal of Management Information Systems 31 (2): 133–169.
Liu, Charles Zhechao, Yoris A. Au, and Hoon Seok Choi. 2014. “Effects of Freemium Strategy in the Mobile App Market: An
Empirical Study of Google Play.” Journal of Management Information Systems 31 (3): 326–354.
Mowen, John C., Sojin Park, and Alex Zablah. 2007. “Toward A Theory of Motivation and Personality with Application to
Word-of-Mouth Communications.” Journal of Business Research 60 (6): 590–596.
Negassi, Syoum, and Tsu-Yi Hung. 2014. “The Nature of Market Competition and Innovation: Does Competition Improve
Innovation Output?” Economics of Innovation and New Technology 23 (1): 63–91.
Novshek, William, and Hugo Sonnenschein. 1982. “Fulfilled Expectations Cournot Duopoly with Information Acquisition
and Release.” Bell Journal of Economics 13 (1): 214–218.
Oestreicher-Singer, Gal, Barak Libai, Liron Sivan, Eyal Carmi, and Chad Yassin. 2013. “The Network Value of Products.”
Journal of Marketing 77 (3): 1–14.
Prasad, Ashutosh, R. Venkatesh, and Vijay Mahajan. 2010. “Optimal Bundling of Technological Products with Network
Externality.” Management Science 56 (12): 2224–2236.
Singh, Nirvikar, and Xavier Vives. 1984. “Price and Quantity Competition in a Differentiated Duopoly.” The RAND Journal of
Economics 15 (4): 546–554.
Statista. 2015. “Number of Apps Available in Leading App Stores as of July 2015”. Technical report, https://www.statista.
com.
Taylor, David G., and Michael Levin. 2014. “Predicting Mobile app Usage for Purchasing and Information-Sharing.”
International Journal of Retail & Distribution Management 42 (8): 759–774.
Wei, Chih-Ping, Olivia R. Liu Sheng, and Paul Jen-Hwa Hu. 2002. “Fuzzy Statistics Estimation in Supporting Multidatabase
Query Optimization.” Electronic Commerce Research 2 (3): 287–316.
Wu, Weijie, Richard T. B. Ma, and John C. S. Lui. 2014. “Exploring Bundling Sale Strategy in Online Service Markets with
Network Effects.” IEEE INFOCOM 2014 – IEEE Conference on Computer Communications, Toronto, ON, 2014, pp. 442–450.
doi:10.1109/INFOCOM.2014.6847967.