Telecommunications Policy 40 (2016) 3951

Contents lists available at ScienceDirect

Telecommunications Policy
URL: www.elsevier.com/locate/telpol

Can you hear me now? The rise of smartphones and their

welfare effects
Adam D. Rennhoff a, P. Wesley Routon b,n
a
Middle Tennessee State University, P.O. Box 27, Murfreesboro, TN 37132, USA
b
Georgia Gwinnett College, 1000 University Center Lane, Lawrenceville, GA 30043, USA

a r t i c l e i n f o abstract

Available online 17 December 2015 The advent of smartphones has caused a dramatic change in consumers access to infor-
Keywords: mation and media. Using data on consumer telephone service choices, along with geo-
Telecommunications graphic data on cellular network infrastructure and software applications, we estimate the
Smartphones demand for telephone services. We allow for the possibility that consumers purchase both
Cell phones landline and wireless service. Among our results, we nd evidence that smartphones are
Demand complementarity much stronger substitutes to landlines than traditional wireless services, suggesting the
growth of smartphones may speed up the decline of landline services. Using our demand
estimates, we estimate the effect that smartphone availability has had on consumer
welfare. We estimate that average monthly surplus (per-consumer) from smartphone
introduction is approximately $35.50, which implies an aggregate monthly consumer
surplus of approximately $7.03 billion for wireless consumers in the United States. This
welfare increase is due to the expanded consumer choice set, as well as the effect that
smartphone growth has had on improvements in cellular networks and a reduction in the
price of wireless voice services.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction

The U.S. wireless industry has grown dramatically in the past decade. Estimated wireless revenues have more than
doubled from $71.1 billion in 2002 to $178.4 billion in 2012 (CTIA, 2013). In 2011, the average U.S. household spent $1226, or
about 2.5 percent of their total annual expenditures, on telephone services (BLS, 2012).1 This is up from 2.2 percent in 2006
(BLS, 2007). Anecdotally, a signicant portion of this growth is attributable to the development of the smartphone.2
Despite the growing nancial importance of the telephone industry, and the wireless industry in particular, the academic
literature on wireless or mobile telephone service demand is rather small. Ahn and Lee (1999) were among the rst to study
the demand for mobile telephone services, using country-level data to study the relationship among per capita GDP, per
capita xed telephone lines, and mobile phone service demand. In a more recent study, Caves (2011) used U.S. state-level

n
Corresponding author.
E-mail addresses: Adam.Rennhoff@mtsu.edu (A.D. Rennhoff), prouton@ggc.edu (P.W. Routon).
1
This gure includes both landline and wireless services.
2
Smartphones are mobile phones with built-in (mobile) operating systems that allow for more advanced computing capabilities and better con-
nectivity than traditional cellular phones. They combine the functions of a mobile phone, a personal digital assistant (PDA), a portable media player, a
compact digital camera, a portable video camera, a hand-held gaming system, and a GPS navigation system into a single multi-use device. Modern
smartphones also include high-resolution touchscreens and web browsers.

http://dx.doi.org/10.1016/j.telpol.2015.11.004
0308-5961/& 2015 Elsevier Ltd. All rights reserved.
40 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951

data from 2001 through 2007 and found that a one percent decrease in the price for wireless services is estimated to
decrease the demand for landline services by 1.21.3 percent. A key feature of smartphones is their broadband internet
capabilities. Thompson and Garbacz (2011) nd that mobile broadband, such as that provided to smartphones, has an
important direct effect on GDP, especially for low income countries, but that the effect of xed broadband (e.g. home
internet service) is no different than zero. Using a panel of OECD countries, Czernich, Falck, Kretschmer, and Woessmann
(2011) estimate a 10 percentage point increase in broadband penetration raised annual per capita economic growth by 0.9
1.5 percentage points during the period 19962007.
Much of the prior literature on mobile telephone services is comprised of country or region studies. Wallsten (2001)
explores the impacts of privatization, competition, and regulation on mobile operator performance in Africa and Latin
American. The author nds that competition is correlated with increases in per capita services, decreases in price, and that
privatization combined with an independent regulator is positively correlated with telecom performance measures.
Regarding country case studies, Tishler, Ventura, and Watters (2001), Kim and Kwon (2003), Iimi (2005), and Doganoglu and
Grzybowski (2007) examine the Israeli, Korean, Japanese, and German markets for mobile telecom services, respectively.
Alleman and Rappoport (2014) study the entire information and communications technology sector, and discover that
weaker regulation led to higher prices and more investment in Latin America.
The goal of this study is to examine the U.S. market for telephone services, both wired and wireless, in an attempt to
understand consumer behavior and welfare. More specically, the analysis begins with the estimation of demand for both
landline (xed-line, main line, home phone) telephone service and wireless (mobile) telephone services. Smart-
phones and traditional wireless service are differentiated in the consumer's choice problem. A discrete choice model is used,
derived from Gentzkow (2007), that allows for the possibility that consumers purchase both landline and wireless services.3
Demand estimates show evidence that smartphones are much stronger substitutes to landlines than traditional wireless
telephone services. Accounting for unobserved heterogeneity in tastes is found to be important in the context of telephone
demand. Results also indicate that improvements in cellular network infrastructure and especially smartphone features, as
measured by the deployment of cellular towers and the availability of smartphone applications (apps), respectively, increase
consumer utility. This suggests that network quality, which should be directly affected by the increased prevalence of
cellular towers, is an important component in the decision to purchase a wireless phone.
Using the estimates from the demand model, along with telephone pricing information, a counterfactual simulation is
conducted. Specically, this allows for the quantication of the effect that the introduction of the smartphone has had on
consumer welfare. Results demonstrate that monthly consumer surplus, due to the introduction of the smartphone, is
approximately $35.52 per consumer. When aggregated based on the size of the wireless market, this implies a total monthly
consumer surplus of $7.03 billion or $84.36 billion annually. This value is actually quite similar to Hausman et al. (2002)
estimate of the increase in consumer surplus due to the introduction of the cell phone, which ranged from $52.8 billion to
$111 billion annually. Interestingly, it is predicted that approximately 25 percent of early smartphone subscribers were new
wireless customers.4 The growth of the smartphone (at least in the short run) did not come exclusively at the expense of
existing wireless customers.
The remainder of the paper is organized in the following manner. Section 2 describes the market for wireless telephone
services. Section 3 presents the data used for the analysis. In Section 4, the model of consumer demand is described.
Estimation results are presented and discussed in Section 5. Section 6 describes the estimates for the welfare impact of the
introduction of the smartphone. Conclusions are presented in Section 7.

2. The market for cellular telephone services

As of October 2012, 99.5 percent of the U.S. population and 67.8 percent of the U.S. land mass were covered by at least
one wireless service provider (FCC, 2011). The wireless industry is characterized by four nationwide carriers (Verizon, AT&T,
Sprint, and T-Mobile) and a number of smaller regional carriers (e.g. U.S. Cellular).5 Wireless industry revenues have grown
by approximately 33 percent in the last ve years (CTIA, 2013). This growth has come primarily from smartphones and their
required data plans.6 Thus, the increase in industry revenue over the last few years can be directly tied to the popularization
of smartphones. A number of consumers, especially younger consumers, now use their wireless phones as the primary
device by which they send email and even browse the internet.

3
Ida and Kuroda (2009) is the most closely-related wireless telephone study, model-wise. They estimate a discrete choice model of different mobile
telephone services, specically focusing on the choice between second generation (2G) and third generation (3G) mobile phone communication technology
standards. The authors nd that demand substitutability among alternatives is stronger within providers than within like services. For example, the closest
substitute for a provider's 3G service is the same provider's 2G service, not another provider's 3G service.
4
In other words, these were consumers that had purchased only landline services, or no telephone service, prior to the introduction of the
smartphone.
5
The HerndahlHirschman Index (HHI) for the mobile wireless industry was 2873 in 2011 (FCC, 2011). Traditionally, a market with a HHI above 2500
is considered highly concentrated.
6
In fact, total wireless industry voice services revenue fell from $111.9 billion in 2006 to $108.5 billion in 2011 while data services revenue rose from
$15.2 billion to $62.7 billion over the same period (FCC, 2011).
 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951 41

Table 1
Worldwide landline, mobile phone, and mobile broadband subscriptions.
Source: International Telecommunication Union (ITU).

Year Landline Mobile phone Mobile broadband

2006 19.2 41.7 N/A

2007 18.8 50.6 4.0
2008 18.5 59.8 6.3
2009 18.4 68.1 9.0
2010 17.8 77.2 11.3
2011 17.3 85.5 16.6
2012a 16.9 91.2 22.1
2013a 16.5 96.2 29.5

Notes: Values are subscriptions per 100 inhabitants. Mobile broadband refers only to
active subscriptions; smartphone owners without a data plan are not counted.
a
201213 values are estimates.

More than half of wireless consumers are now smartphone users.7 The two leading mobile operating systems are Apple's
iOS and Google's Android. Apple released the rst iPhone in 2007 and Google released the rst Android phone the following
year. Google Android's share of the market grew from 3 percent in May 2009 to 51 percent in March 2012, while Apple's iOS
market share increased from 20 to 32 percent over the same period (FCC, 2011). Thus, these two operating systems have a
collective market share of 83 percent. These operating systems, and others, are available from multiple service providers. An
initial handset exclusivity arrangement between Apple and AT&T ended in 2011 and now iPhone owners may choose their
service provider. One of the more innovative and popular aspects of smartphone technology is the creation of applications
(apps) to be used on the phones. These apps allow smartphone owners to customize and increase the capabilities of their
devices. Apps may be purchased through online (often third-party) sources, such as Apple's App Store and the Google Play
store. It has been estimated that over two million apps were available to smartphone consumers as of the end of 2012
(Deloitte, 2012).
Although this study focuses on the U.S. market, it is worth noting that mobile phone and smartphone adoption has also
grown around the world. Table 1 displays data collected from the International Telecommunication Union (ITU) and shows
the extent of the rise in these services since 2006. The values are number of subscriptions per 100 inhabitants. The data
show that per capita landline subscriptions have been slowly falling over the entire period while per capita mobile phone
subscriptions have more than doubled. Mobile broadband access, as made available by smartphones, was initially popu-
larized in 2007 with the rst iPhone. Per capita subscriptions for this service are on the rise as well and have increased over
700 percent since 2007.
Fig. 1 presents time plots for the percent of the U.S. population that purchases a landline telephone service only, a mobile
service only, both services, or neither service for the period 2003 through 2011. Data from the National Health Interview
Survey was used to produce these estimates.8 Very few Americans, 3 percent or less for the entire period, purchase no
telephone service. Approximately 51 percent of the population were landline only subscribers in 2003 but this number had
dropped to about 10.5 percent by 2011, a stark decline. Only 2.3 percent purchased only a mobile service in 2003 but this
gure had climbed to almost 35 percent by 2011. About 43 percent of Americans chose to purchase both types of service in
2003. This number had climbed to about 57 percent by 2007 but then hit a three year plateau before starting to decline to
52 percent by 2011. This plateau and its following decline correspond to the introduction and popularization of the iPhone in
2007. Anecdotally, it is difcult to overlook the negative correlation between landline service and wireless service in Fig. 1.

3. Data

Data on consumer telephone service choices comes from the Pew Internet & American Life Project, an ongoing project of
the Pew Research Center. The Pew Center is a nonpartisan fact tank that strives to inform the public about issues, attitudes,
and trends that are shaping the world. As part of the Internet & American Life Project, a survey was designed to collect data
on wireless phone ownership, usage, and attitudes. This survey was administered in years 2006, 2011, and 2012, and this
analysis makes use of all three of these survey waves. The timing of these surveys is advantageous for a study on smart-
phones. As previously mentioned, Apple released the rst iPhone in 2007; Google's Android phones hit the market in 2008;
and Microsoft unveiled its next-generation mobile operating system (Windows Phone 7) in 2010. Thus, the combined
dataset contains information on cell phone ownership and use both before and after the introduction of smartphones.

7
By 2012Q2, 55 percent of mobile subscribers used smartphones (FCC, 2011).
8
The National Health Interview Survey does not make a distinction regarding traditional wireless phone ownership versus smartphone ownership.
Further details regarding the NHIS are presented in an appendix.
42 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951

0.6

0.5

0.4
Percent of Americans

No service
0.3
Landline only
Mobile only
Both services
0.2

0.1

0
2003 2004 2005 2006 2007 2008 2009 2010 2011
Year

Fig. 1. U.S. telephone service choices.

Table 2
Means of consumer characteristics by telephone service.

Variable Full sample Landline only Traditional cell only Smartphone only

Male 0.481 0.438 0.536 0.594

Age 46.432 57.288 39.929 32.872
Married 0.507 0.491 0.328 0.319
Education (years) 13.694 12.938 13.003 13.907
Full-time employed 0.548 0.338 0.587 0.738
AfricanAmerican 0.131 0.101 0.171 0.173
Other nonwhite 0.119 0.088 0.141 0.155
Low income 0.387 0.576 0.597 0.332
High income 0.137 0.056 0.053 0.143
N 6203 880 572 383

Notes: Survey weights were used to construct sample means. Low (high) income refers to those earning less than $30,000 (more than $100,000) per year.

The survey asks respondents if they subscribe to a landline telephone service and/or a wireless phone service.
Respondents indicating wireless phone ownership are also asked to indicate which brand of mobile phone they own,
whether their mobile phone is a smartphone, and their phone's operating system (e.g. iOS, Android). In addition to this
product information, the survey collects demographic information on the consumers. The data also contains survey weights
meant to make the data nationally-representative, which adds to the external validity of our ndings.
Table 2 presents sample means for the consumer demographic variables incorporated in the demand analysis. These
variables include a gender (male) dummy, age, an indicator for married individuals, two race dummies (African-American
and Other nonwhite), and an indicator for those that are employed full-time. The dataset also contains categorical variables
pertaining to educational attainment and income. The educational attainment variable is transformed into a continuous
variable that relates to years of schooling. From the categorical income variable, two indicator variables were created, one
for low income earners and one for high income earners. Low income refers to those earning less than $30,000 per year
while high income refers to those earning more than $100,000 per year.9
Means for each telephone service sub-sample are presented in Table 2 alongside the full sample means to show how
consumer characteristics vary across their revealed choice alternatives. For brevity, means for bundled consumers are not
shown. From these simple statistics, it appears that males are quicker to adopt new telecom technology, that is, they make
up an increasing portion of the sample from left (Landline Only) to right (Smartphone Only). The same can be said for the

9
A number of survey respondents declined to provide information on their income level. Analysis determined that there was no statistically signicant
relationship between refusal to provide income information and the respondent's wireless phone status. In other words, it does not appear that those
respondents with only landline service, for example, were more likely to decline providing income information. In order to preserve the size of the survey
sample, the income category (not level) was imputed for those respondents with incomplete records.
 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951 43

young, the unmarried, the educated, the full-time employed, and minorities. These groups may also be more prone to use
smartphones as substitutes to other internet-ready devices and digital cameras, for example. Low income consumers are
more concentrated in the traditional cell phone only group while high income consumers are more concentrated in the
smartphone only group.

4. Model of consumer demand

4.1. Demand specication

The demand model estimated here is based on Gentzkow (2007) model of demand for bundles of goods.10 As shown in
Fig. 1, consumers frequently purchase both landline (home) telephone services as well as either a traditional cell phone or a
smartphone.11 Given this observation, any failure to account for these multiple purchases would inherently mis-characterize
the substitution patterns of consumers.
Consumers are indexed by i f1; ; N g. There are three possible telephone service options, which are indexed by


j Landline; Cellular; Smartphone . The baseline utility that consumer i receives from subscribing to telephone service j in
time t is given by
u ijt j ij  pjt xit j zjt 1

where (ignoring time subscripts for the moment) j is a constant term, ij represents consumer i's unobservable preference
for j, pjt is the price of service j, xit is a vector of observable consumer characteristics, and zjt is a vector of product
characteristics.12 In Eq. (1), , j, and are parameters to be estimated.
The vector xit contains information on consumer i's gender, age, marital status, education, employment status, race, and
income level. Employment status enters the model interacted with urban/rural residence status. Employment (and, thus,
greater amounts of time outside the home) is predicted to increase wireless demand (and perhaps decrease wired demand),
but this effect likely differs across urban and rural employees. j allows the effect of these characteristics to vary by service.
Two service-specic characteristics are also included (zjt): the number of cellular towers in i's local cellular market area and
the number of apps available for the iPhone via Apple's App Store. These measures will be discussed in greater detail below.
It is assumed that consumer (unobserved) preference for each service is distributed as follows:
2 3 0 2 31
land 1
6 7 B 6 7C
4 cell 5  N@0; 4 lc c
2
5A:
smart ls cs 2s
The variance of land is restricted to be equal to 1 for identication purposes. The other elements of the covariance matrix,
including the off-diagonals which allow for possible correlation in unobservable preferences, are unrestricted.
Identifying the price coefcient () requires suitable variation in prices. Ideally, one would observe changes in price
within a market over time, as well as differences in price across markets. Unfortunately, one drawback to the Pew data is
that it does not contain information on either a consumer's telephone service provider or the price they paid for service.13
Therefore, the constant term is redened14 in Eq. (1) as follows:

jt j  pjt : 2

From the baseline utility specication in Eq. (1), the utility that consumer i receives from consuming a bundle r can be
expressed as
X
uirt u ijt r irt 3
jAr

where r is a bundle-specic term which captures the difference between baseline bundle utility and the sum of the stand-
alone utilities of the services in bundle r and irt is consumer i's idiosyncratic preference for bundle r. It is assumed that
consumers face ve inside bundles: three singleton bundles corresponding to the three telephone services, the bundle of
landline telephone service and traditional cellular service, and the bundle of landline telephone service with smartphone

10
Gentzkow (2007) examined the news media choices of consumers in the Washington, DC metro area, focusing on the impact that online news
availability has on subscriptions for traditional print news outlets.
11
For simplicity, the term cell is used to describe traditional wireless service. Smartphones and cell phones refer to distinct products.
12
From an econometric standpoint, can be interpreted as a random coefcient on the product-specic constant () as in McFadden and Train (2000)
mixed logit model.
13
Vogelsang (2010) identies the lack of price variation and information as one of the primary difculties in studying the wireless industry.
14
Eq. (2) implies that even though we assume each product's is constant, we should estimate separate constants for each time period due to price
changes. The relevant question, therefore, is whether there is price variation across time periods. A survey of prices for telecommunications services by
large providers (e.g. Verizon and AT&T) reveals there were no notable price changes between 2011 and 2012. This should, perhaps, not be surprising given
that some telephone services require 12 or 24 month contracts. There were, however, price changes from 2006 to 2011. Two separate values are therefore
estimated for landline phone service and traditional cell phone service, one for 2006 and another for the years 2011 and 2012.
44 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951

service.15 Note that the model does not allow consumers to select a bundle of traditional and smartphone cellular services.
While it may be the case that a number of households subscribe to both traditional cell service and smartphone service, the
Pew survey specically asks about individual choices (not household choices). The Pew data does not identify any indivi-
duals opting for two different cellular telephone services. This option is therefore omitted from the choice set. It is assumed
that irt is distributed i.i.d. type-I extreme value.
Consumers also have the option of a singleton bundle of the outside option, which is denoted r 0. In the model, this
outside option is the no purchase option and utility from this bundle is specied as follows:
ui0t i0t :
The estimated sign of r helps disclose the relationship between the services within bundle r. For example, a positive
(negative) value for r would imply that the value of the two services together is greater than (less than) the sum of
their stand-alone values, indicating that the two services are complements (substitutes). Thus, the sign of r is not
restricted.
As acknowledged by Gentzkow (2007), there are some inherent problems identifying r separately from the elements of
the covariance matrix of unobservables. There are several solutions to this problem, including the use of panel data (which is
not possible in this case). Here, exclusion restrictions are used. More specically, the model allows for several variables
which are assumed to affect only the utility for one good. Recall the two variables included in vector zj. The number of
cellular towers in a market should affect the utility of cell phones and smartphones, but not landline telephone services.16
Similarly, the number of apps available should affect the utility of smartphones but not the utility for either regular cell
phones or landline phones. It should be noted that the extent of the local cellular network affects the utility of both tra-
ditional cell phones and smartphones. Fortunately, the coefcient on cellular towers is able to be identied based solely on
the data for 2006, where smartphones were not part of the consumers choice set. Using only the data for 2011 and 2012
would not allow for identication the effect.
The distributional assumption regarding irt allows us to express the probability that i chooses bundle r during time t as
hP   i
exp j A r jt ij xit j zjt r
irt P hP   i: 4
1 r exp j A r jt ij xit j zjt r

Note that this expression is conditional upon the vector of individual unobservables (i). As discussed in the next section,
numerically calculating this probability, as a function of model parameters, requires integrating over the distribution of .

4.2. Estimation

With the individual choice data, the analysis proceeds with maximum likelihood estimation. Dening y as the N  1
vector of observed consumer choices and as the vector of all model parameters, the likelihood function is given by
Z
     
L x; z; y; yi xi ; z; ; dF ; 5
i
 
where F ; is the multivariate normal distribution of , given . It is well-known that the integral in Eq. (5) does not have
a closed-form solution and must, therefore, be calculated numerically. This probability is simulated by taking S sets of
R    
random draws of i from the multivariate normal distribution. Then, yi xi ; z; ; dF ; in Eq. (5) is replaced with it's
PS   17
simulated value: 1=S s 1 yi;s xi ; z; i;s ; .
Identication of the model presented in Section 4.1 requires information on consumers that do not purchase any tele-
phone services (i.e. those consumers that choose the outside option). By design, however, the Pew survey includes only
those consumers that have telephone service. This problem is not uncommon when researchers aim to estimate demand
based on surveys of past buyers. For example, Langer (2012) and Berry, Levinsohn, and Pakes (2004) both estimate models of
automobile demand based on a survey of recent buyers. Here, Langer (2012) approach is followed by appending our original
sample with consumers who purchased no telephone services.18 This approach requires information on the fraction of non-
buyers in the population, as well as the characteristics of these non-buyers.
In order to obtain information on non-buyers and their characteristics, survey information from the Centers for Disease
Control's (CDC) National Health Interview Survey (NHIS) is gathered. The NHIS asks detailed health and demographic
questions to, currently, approximately 100,000 individuals annually. While the majority of the personal questions relate to
health status and events, the survey also inquires about whether respondents have landline telephone service, wireless

15
This is true of both the 2011 and 2012 samples. In 2006, smartphone service was not part of the consumer choice set. There were, therefore, only
three inside bundles for the 2006 sample.
16
Our hypothesis is that more towers imply a stronger cellular network, which should make owning a cell phone more valuable. It could be the case,
however, that this measure serves as a proxy for cell phone providers investment in a given market. In this manner, it is likely that more cellular towers are
correlated with additional marketing efforts, for example. Nevertheless, these factors are not likely to impact the utility of landline telephone service.
17
We use a Halton sequence of 200 draws in our simulation.
18
Berry et al. (2004) use an alternative approach in which they construct three moment conditions and estimate the choice model via GMM. Given our
preference for the efciency of maximum likelihood estimation, we choose to follow Langer (2012) maximum likelihood approach.
 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951 45

telephone service, or both.19 The NHIS data is used to determine the percentage of non-buyers in each year. Adult NHIS non-
buyer respondents are then randomly sampled and appended them to the Pew data in order to equate the percentage of
non-buyers in our sample with that reected by the NHIS gures. Additional details are presented in an appendix. The nal
augmented sample of telephone subscribers and non-subscribers is used in maximum simulated likelihood estimation of
Eq. (5).

4.3. Endogeneity

Any mention of endogeneity has thus far been avoided. Endogeneity is a frequent problem in consumer choice models as
researchers may be concerned that unobserved factors are correlated with observed product characteristic(s), typically price
and/or various marketing variables. Because price information does not directly enter the consumer choice model (see Eq.
(2)), price endogeneity is not a concern in this case. It may be, however, that two of the other observed product char-
acteristics (the numbers of cell towers and smartphone applications) are endogenous.
To account for this potential endogeneity, the control function approach of Petrin and Train (2010) and Luan and Sudhir
(2010) is adopted. The control function approach essentially turns estimation into a two-stage procedure. The rst stage is
an instrumental variables regression with the endogenous characteristics as the dependent variables. The parameter esti-
mates from the IV regressions are used to derive residuals. These residuals are included in Eq. (3) essentially as additional
explanatory variables. Such a strategy requires valid instruments for both the size of the cellular network in a given location
and the number of smartphone apps available at time t.
A panel of the number of cell towers in each geographic location for the years 2000 through 2012 is assembled.20 Two
instruments are included. The rst instrument is the variance in elevation in the nearby area. The intuition for this
instrument is that areas with numerous peaks and valleys may need to use more cell towers, relative to at areas. The panel
nature of the cell tower data is exploited through the use of the lagged number of cell towers as a second instrument. These
instruments are highly signicant in the IV regression and both, as expected, have positive coefcients. The adjusted R2 in
this rst stage regression is 0.948.
A time series of the number of smartphone apps available at each time t is also constructed. Here, three instruments are
used. The rst two are indicators. The rst indicates the time period post December 2009, while the second the time period
post September 2010. These two dates represent the two biggest pre-2013 changes in app creation and policy. In December
2009, Apple streamlined the app review process (reviewing apps for release went from taking several weeks, on average, to
a couple days) and began giving app creators detailed information as to why their apps were being accepted or rejected
(before, only simple Accepted, Rejected, and Waiting for approval notices were provided). In September 2010, app
programming became more exible (more interpreter languages were allowed, etc.) and the rst App Store Review
Guidelines document was published, making app creation much easier for rst-time programmers. The third instrument is
the lagged number of smartphone apps. All three instruments are highly signicant in the IV regression and have, as
expected, positive coefcients. The adjusted R2 is 0.995.
The appropriate residuals from these two rst-stage IV regressions are included in the consumer choice model. A Wald
test is conducted to determine the need to control for endogeneity. This test fails to reject the null hypothesis that the
coefcients on the rst-stage residuals are jointly equal to zero.21 It is concluded, therefore, that endogeneity is not a
concern in our application. The results presented in the next section, therefore, do not incorporate the control function
approach.

5. Results

5.1. Demand estimates

The maximum simulated likelihood estimates of the coefcients on observable consumer characteristics are presented in
Table 3. All of the included characteristics are found to impact telephone service choices. Males, the full-time employed
(both urban and rural employees), and other nonwhites are less likely to purchase landline services. Rural employees are
additionally less likely to purchase wireless services, while urban employees are more likely to purchase smartphone
services. Older consumers, who often are slower to adopt new technology, are found to be more likely to subscribe to a
landline service and less likely to purchase either wireless service. Marriage and education are both positively associated
with subscribing to each of the three different telephone services. AfricanAmerican consumers are found to be more likely
to purchase wireless services, especially smartphone services. Given our prior descriptive analysis, we might have expected
the same for other minorities, but this is not found to be the case. Unsurprisingly, low income consumers are less likely to
purchase each of the three telephone services while high income consumers are more likely to purchase each of these

19
NHIS survey data on household telephone status has been used in CDC reports such as Blumberg and Luke (2012). We previously cited this data as
the source of Fig. 1.
20
We use the full panel to estimate the rst-stage parameters. We then extract the relevant residuals for 2006, 2011, and 2012.
21
The p value for the Wald test is 0.714.
46 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951

Table 3
Demand estimates: observable consumer characteristics.

Consumer Landline Cellular Smartphone

characteristic

Male  0.402a  0.259 0.052

(0.083) (0.214) (0.242)
Age 0.038a  0.075a  0.128a
(0.003) (0.026) (0.026)
Married 0.671a 0.704c 0.958b
(0.220) (0.368) (0.388)
Education (years) 0.075a 0.351a 0.494a
(0.020) (0.120) (0.119)
Urban-employed  0.445a 0.502 0.934a
(0.102) (0.307) (0.327)
Rural-employed  0.396b  0.739c  1.107b
(0.173) (0.439) (0.485)
AfricanAmerican 0.020 0.989b 1.895a
(0.121) (0.476) (0.491)
Other nonwhite  0.258b  0.380 0.037
(0.125) (0.337) (0.383)
Low income  0.797a  2.100a  3.029a
(0.106) (0.718) (0.705)
High income 0.581a 1.609a 2.558a
(0.147) (0.636) (0.646)
Constant  0.118 4.137b 2.997c
(0.422) (1.612) (1.732)
N 6203 6203 6203

Notes: Standard errors in parentheses.

a
Signicance at the 1% level.
b
Signicance at the 5% level.
c
Signicance at the 10% level.

Table 4
Demand estimates: additional preference parameters.

Cell network variables:

Number of towers 2.0E  04 Number of apps 0.241a
(1.3E  04) (0.051)
Interaction terms ():
Landlinecell: Landlinesmartphone:
Rural-single 0.148 Rural-single  0.788c
(0.322) (0.450)
Urban-single  0.280 Urban-single  1.162a
(0.291) (0.398)
Rural-married  0.451 Rural-married  1.362a
(0.361) (0.490)
Urban-married  0.196 Urban-married  1.177a
(0.336) (0.451)
Nonlinear parameters:
Std. dev. of i;cell 4.351a Correlation of i;land and i;cell 0.237
(1.624) (0.226)
St. dev. of i;smart 5.306a Correlation of i;land and i;smart 0.402b
(1.525) (0.203)
Correlation of i;cell and i;smart 0.911a
(0.148)

Notes: Standard errors in parentheses.

a
Signicance at the 1% level.
b
Signicance at the 5% level.
c
Signicance at the 10% level.

services. The magnitude of the smartphone demand coefcient for low income consumers may be surprising to some, given
these consumers are thought to more often use smartphones as substitutes for other internet-ready devices.
The remainder of the model parameters is displayed in Table 4. The interaction terms (or 's, the bundle-specic terms)
refer to the difference between baseline utility and the sum of the stand-alone utilities of the services in the bundle. In order
to allow for heterogeneity in the value of the bundle, the interaction terms vary based on the consumer's geographic
location (urban versus rural) and their marital status. Estimates provide evidence that consumers receive decreased utility
from bundling landline and smartphone services, implying that the two services are substitutes. This was found to be the
 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951 47

7000

6000

5000
Cell Towers

4000
Tennessee
Utah
Virginia
3000
Maryland
Colorado

2000

1000

0
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

Fig. 2. Cell towers in selected states.

case for all four of our consumer groups: rural-single, rural-married, urban-single, and urban-married. There is no statistical
evidence of additional or decreased utility from the landline-cell bundle. The estimated landline-cell interaction terms are
negative for three of the four consumer groups, suggesting a weak substitute relationship. Vogelsang (2010), in a survey of
the literature surrounding xed and mobile communication services, shows that previous models have been inconclusive
regarding the substitutability or complementarity of these two services (meaning landlines and traditional cell services).
Also reported in Table 4 are the estimated parameters for the cell network variables, number of towers and number of apps.
Both of these choice-specic variables have positive coefcients, though only apps are found to statistically increase con-
sumer utility.
The bottom panel of Table 4 shows the nonlinear parameter estimates, the elements related to the covariance matrix
presented in Section 4.1. There appears to be no correlation in unobserved consumer preferences for landline and traditional
cell services. The estimated correlation between landlines and smartphones, however, is found to be both positive and
statistically signicant meaning that those with unobserved preferences for landline services are more likely to prefer
smartphones and vice versa. Unobserved preferences for the two wireless services appear to be positively correlated which
is perhaps unsurprising since they are similar services.

5.2. Price coefcient

With little variation in prices, consumer price disutility () is unable to be identied as is typically done in discrete choice
models. Hence, an alternative approach to inferring the true value of is applied. The very limited price variation observed
is used to back-out using Eq. (2). Recall that, through the FCC wireless report, it was observed that wireless voice prices
varied between 2006 and 2011. For traditional wireless service, therefore, two separate coefcients are estimated. Using
the 2006 and 2011 estimates for wireless service yields the following two equations:
2006: 4: 136; 552 c  $33:54

2011: 3: 516; 352 c  $41:21

Solving the system of equations for the two unknowns (c and ) yields an inferred price coefcient of 0.0809. In the welfare
calculations that follow, this price coefcient will generate the welfare estimates.

6. Welfare and the smartphone

In this section, a discussion of the welfare impact of the smartphone is presented. In order to quantify the effect, one can
examine how the counterfactual removal of the smartphone from the 2011 sample diminishes consumer surplus. In
addition, the population most likely to benet from the introduction of the smartphone can be discovered by predicting the
counterfactual choices of consumers in the 2006 sample were they to have had the option of selecting a smartphone. An
advantage of using such an approach is that the actual pre-smartphone choices of the 2006 sample are observed.
48 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951

Table 5
Consumer surplus from smartphone introduction.

Monthly surplus per consumer ($):

With no change in towers or prices $18.51
With change in num. towers only $19.81
With changes in both prices and towers $35.52

Aggregate consumer surplus $7.03 billion

Standard error ($804 million)

Notes: NHIS (2011) reports that approximately 87% of adults have some cell phone, which
translates to approximately 198 million adults. We use this gure as the basis for aggre-
gating consumer-level surplus gures to U.S. aggregate values.

Table 6
Changes in service with addition of smartphone.

Consumers with no telephone service:

Continue no phone service 1.90%
Switch to smartphone 22.37%
Switch to smartphone bundle 16.81%
Consumers with landline service:
Continue with landline service 11.47%
Switch to smartphone 11.46%
Switch to smartphone bundle 16.79%
Consumers with traditional cell service:
Continue with cell service 12.53%
Switch to smartphone 26.73%
Switch to smartphone bundle 23.88%
Consumers with cell-landline bundle:
Continue with bundle 43.76%
Switch to smartphone 17.62%
Switch to smartphone bundle 27.14%

Notes: This table reects how choices of 2006 survey respondents would be expected to
change based on the introduction of the smartphone.

The introduction of the smartphone has had three primary effects on consumers: (1) it introduced a new product to
consumers that may be more desirable or attractive than prior products; (2) the smartphone has caused massive investment
in cellular networks which has improved the quality of wireless voice services; and (3) the price of wireless voice service
should be affected by the introduction of smartphones. The third item implies that prices of voice wireless service may be
different and, therefore, even traditional wireless customers not considering a smartphone may be positively impacted by
the introduction of the smartphone. To illustrate this idea, data from FCC (2011) shows that average voice prices have fallen
from $41.21 in 2006 (pre-smartphone) to $33.54 at the time of the report.
With regard to wireless infrastructure investment, FCC cell tower registration data allows for the construction of a time
series of active cell towers in various geographic locations. Fig. 2 shows the number of active cell towers in ve different
states.22 Similar clustering at the beginning of the period and then a more dramatic uptick in the number of towers
beginning around 2007 can be seen. In the welfare calculations that follow, both price changes and the growth of wireless
infrastructure are accounted for.
Given the assumption that the idiosyncratic bundle preference (irt) is distributed extreme value, one can calculate
consumer surplus by converting a consumer i's expected maximum utility into dollars using the price coefcient with the
familiar log-sum:
" #
1 X  
CSi ln exp ub ir 6
rAR

where u b ir is the portion of i's utility from Eq. (3) excluding ir and R is the set of all possible bundles. Notice that while Eq.
(6) has omitted ir they have been integrated-out the remaining utility (u b ) is still a function of the vector of their
unobservable product preferences ().
In order to account for simulation is used to integrate over the distribution of . Rather than giving each set of random
draws equal weight in the simulation, each consumer's observed choice is conditioned on. Each simulation draw is weighted
by a factor that accounts for the likelihood that a given set of draws is consistent with a consumer's observed choice.23 More

22
We have chosen these ve states simply because they are scaled similarly, making the gure easier to read. We could have alternatively focused on
less aggregated geographic areas, such as cellular market areas (CMAs). The gure is meant for illustrative purposes only.
23
This technique is similar in spirit to what Revelt and Train (1999) refer to as conditioning on individual tastes.
 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951 49

Table 7
Overview of predicted smartphone buyers.

Counterfactual choice

Smartphone only (%) Smartphone bundle (%)

Original choice:
No phone service 1.01 0.53
Landline service 23.88 24.53
Trad. cell service 13.68 8.57
Landline-cell bundle 61.43 66.37
Market share 9.42 23.80

Notes: Each column shows the percentage of new smartphone buyers based on their
observed (pre-smartphone) telephone service choice.

specically, the simulation accounts for the expectation of in Eq. (6) as

" #
1 X  
CSi ln exp ub ir
r AR

" #!
1X
S X    
 ln exp u ir s
b  R s jyi : 7
s1 rAR

In Eq. (7), yi is consumer i's observed choice and Rs jyi is the conditional probability of s, given this observed choice. With
demand estimates, one can simulate i's predicted probability of choosing bundle y ( iy ), which allows for the construction
of Rs jyi as follows:

iy s f s
Rs jyi : 8
iy

In order to nd the expected consumer surplus for consumer i in the absence of smartphones, both bundles containing
smartphones from the log-sum in Eqs. (6) and (7) are simply removed. In this manner, one would be summing over the set
of all bundles that do not contain a smartphone. Using differences in these surplus values, one can construct either
aggregate or average welfare changes using survey weights.
In Table 5 welfare values are presented. Using the estimate of , it is found that the average per-consumer surplus due to
the introduction of the smartphone, in dollars, is $35.52 per month. This gure accounts for the fact that prices for tradi-
tional wireless service would be higher and wireless infrastructure development (i.e. cell towers) would be lower.24 The
change in consumer surplus under the assumption that traditional wireless prices and infrastructure remain at their
observed 2011 levels is also calculated. This value is $18.51 per consumer per month. One could think of this surplus as being
a lower limit on the impact of the smartphone, as it allows for an upper bound on the number of cellular towers and a lower
bound on the price level. Additionally, one could interpret this gure as identifying the proportion of the total change in
surplus ($18.51) due solely to choice set expansion. These gures imply that the increase in the choice set accounts for
approximately one-half of the welfare gain from smartphones. Wireless infrastructure effects are relatively modest, while
price effects ($15.71) are nearly as high as that of the increased choice set.
Tables 6 and 7 attempt to answer the question of which consumers benet most from the introduction of the smart-
phone. To answer this question, the pre-smartphone 2006 sample is given the counterfactual option of selecting a
smartphone. The advantage of using the 2006 sample is that pre-smartphone choices are actually observed. Table 7 shows
that it is predicted about 33 percent of consumers purchase a smartphone, with a minority opting to rely exclusively on a
smartphone for their telephone service. The remainder of the percentages in Table 7 tell which product the new smartphone
subscriber had previously chosen.25 The columns depict a similar picture. The majority of new smartphone subscribers
switched from a bundle of a landline with a traditional cellular phone. Interestingly, approximately 24 percent of the
switchers had only previously purchased landline service. This appears to indicate that the growth of the smartphone is not
being driven exclusively through business-stealing. According to the simulations, almost a quarter of new smartphone
owners had not previously subscribed to a wireless service.

24
From an implementation standpoint, we use the pre-smartphone 2006 values for both cell tower deployment and regular wireless prices. In this
regard, our counterfactual uses observed values rather than simulated ones.
25
Note that each column sums to 100 percent. The second column, therefore, shows what the 9.42 percent of smartphone-only consumers purchased
prior to the smartphone introduction.
50 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951

7. Conclusions

The goal of this analysis was to jointly estimate the demand for different telephone services (landlines, traditional cell
phones, and smartphones) so that the impacts of smartphone introduction on this market might be estimated, with a focus
on consumer welfare. Following Gentzkow (2007), a discrete choice demand model is used that allows for consumers to
purchase both types of services, here landline and wireless telephone service. More specically, the model allows con-
sumers to choose a traditional cell phone-landline bundle or a smartphone-landline bundle.
Findings indicate the welfare effect of smartphones to be quite large. The estimate of monthly per-consumer surplus is
$35.52. In this calculation the manner in which smartphone growth has affected the signal strength of cellular networks and
service prices has been accounted for. The expanded consumer choice set accounts for approximately 52 percent of that
welfare gain. Changes in service prices accounts for approximately 44 percent, while increased cellular network strength
only approximately 4 percent of the welfare gain. The aggregate monthly wireless consumer surplus from smartphones is
approximately $7.03 billion (an $84.36 billion annual surplus) in the U.S. Hausman (2002) estimates cell phone introduction
increased consumer surplus by $52.8$111 billion annually, implying smartphones have had a comparable, and perhaps
even a larger effect than its predecessor.
Among the other results, evidence is found that both types of wireless services are substitutes to landline service, though
much more for smartphone services. This nding has implications for the speed at which one should expect wireless
services to eliminate (replace) landline services, given continued growth in smartphone subscriptions. Results show that
low income consumers tend to have either no service or landline service. High income consumers are overwhelmingly more
likely to purchase a smartphone than any other demographic group. The number of smartphone apps appears to be more
important (in terms of indirect utility) to consumers than the cellular network size.
This research suggests many possible avenues for future research on this topic. The most glaring gap in the literature is
the lack of research on competition between wireless providers. Their role is essential in understanding wireless issues.
Unfortunately, data limitations have, thus far, made it relatively difcult to study wireless competition.

Appendix A. The NHIS and the outside option

The CDC's NHIS is currently completed by approximately 100,000 respondents. In an average year, approximately 2000
3000 of the respondents indicate that they receive no telephone services. We wish to use this subsample of respondents to
append to our Pew sample. A number of these respondents have incomplete records, however. While we use imputation to
ll-in a number of the incomplete income records for the Pew data in order to preserve our sample size, such efforts are not
necessary here.26 We eliminate all respondents with incomplete records. This leaves us with approximately 600 (2006) to
800 (2011) complete records per year. We sample enough NHIS non-buyers in order to make the share of non-buyers in our
sample equal to the national share. These non-buyers are randomly selected using sample weights provided by the CDC.
Table A1 compares the average characteristics of non-buyers with those of other purchase groups. The pattern ts what
we might already anticipate: consumers without any telephone service have the lowest average education level and
71 percent of these non-buyers are considered low income.

Table A1
Means of consumer characteristics by telephone service.

Variable No purchase Full sample Landline only Traditional cell only Smartphone only

Male 0.573 0.481 0.438 0.536 0.594

Age 37.302 46.432 57.288 39.929 32.872
Married 0.370 0.507 0.491 0.328 0.319
Education (years) 12.536 13.694 12.938 13.003 13.907
Full-time employed 0.786 0.548 0.338 0.587 0.738
AfricanAmerican 0.144 0.131 0.101 0.171 0.173
Other nonwhite 0.200 0.119 0.088 0.141 0.155
Low income 0.714 0.387 0.576 0.597 0.332
High income 0.032 0.137 0.056 0.053 0.143

Notes: Non-Purchase category observations from NHIS survey. All other values from Pew surveys. Survey weights were used to construct sample means.
Low (high) income refers to those earning less than $30,000 (more than $100,000) per year.

26
Consider a simple back-of-the-envelope calculation. There are approximately 3000 Pew survey respondents each year. If 3 percent of the population
has no phone service then ignoring sample weights we need to append roughly 100 NHIS survey respondents to our Pew data.
 A.D. Rennhoff, P.W. Routon / Telecommunications Policy 40 (2016) 3951 51

References

Ahn, H., & Lee, M. (1999). An econometric analysis of the demand for access to mobile telephone networks. Information Economics and Policy, 11, 297305.
Alleman, J., & Rappoport, P. (2014). Regulation of Latin American's information & communications technology (ICT) sector: An empirical analysis. Pro-
ceedings of ITS 2014, Rio de Janeiro, 30 November3 December 2014.
Berry, S., Levinsohn, J., & Pakes, A. (2004). Differentiated product demand systems from a combination of micro and macro data: the new car market.
Journal of Political Economy, 112, 68105.
Blumberg, S., & Luke, J. (2012). Wireless substitution: Early release of estimates from the National Health Interview Survey, JanuaryJune 2012. National Center
for Health Statistics, December 2012. Retrieved from http://www.cdc.gov/nchs/data/nhis/earlyrelease/wireless201212.PDF.
Bureau of Labor Statistics (2007). Consumer expenditure survey 2006 annual report. Released September 2007.
Bureau of Labor Statistics (2012). Consumer expenditure survey 2011 annual report. Released September 2012.
Caves, K. (2011). Quantifying price-driven wireless substitution in telephony. Telecommunications Policy, 35, 984998.
CTIA (2013). CTIA's wireless industry indices, 19852012. Released 2013.
Czernich, N., Falck, O., Kretschmer, T., & Woessmann, L. (2011). Broadband infrastructure and economic growth. The Economic Journal, 121, 505532.
Deloitte (2012). Technology, media, & telecommunications predictions 2012. Retrieved from https://www.deloitte.com/assets/Dcom-Global/LocalContent/
Articles/TMT/TMTPredictions2012/16264A_TMT_Predict_sg6.pdf.
Doganoglu, T., & Grzybowski, L. (2007). Estimating network effects in mobile telephony in Germany. Information Economics and Policy, 19, 6579.
Federal Communication Commission (2011). 15th Mobile wireless competition report. Released June 27, 2011.
Gentzkow, M. (2007). Valuing new goods in a model with complementarity: Online newspapers. American Economic Review, 97, 713744.
Hausman, J. (2002). Mobile telephone. In M. Cave, S. Majumdar, & I. Vogelsang (Eds.), Handbook of telecommunications economics (pp. 563604). North-
Holland: Emerald Group Publishing Limited.
Ida, T., & Kuroda, T. (2009). Discrete choice model analysis of mobile telephone service demand in Japan. Empirical Economics, 36, 6580.
Iimi, A. (2005). Estimating demand for cellular phone services in Japan. Telecommunications Policy, 29, 323.
Kim, H., & Kwon, N. (2003). The advantage of network size in acquiring new subscribers: A conditional logit analysis of the Korean mobile telephony
market. Information Economics and Policy, 15, 1733.
Langer, A. (2012). Demographic preferences and price discrimination in new vehicle sales. Working paper, University of Michigan.
Luan, Y., & Sudhir, K. (2010). Forecasting marketing-mix responsiveness for new products. Journal of Marketing Research, 47, 444457.
McFadden, D., & Train, K. (2000). Mixed MNL models for discrete response. Journal of Applied Econometrics, 15, 447470.
Petrin, A., & Train, K. (2010). A control function approach to endogeneity in consumer choice models. Journal of Marketing Research, 47, 313.
Revelt, D., & Train, K. (1999). Customer-specic taste parameters and mixed logit. Berkeley working paper, University of California.
Thompson, H., Jr., & Garbacz, C. (2011). Economic impacts of mobile versus xed broadband. Telecommunications Policy, 35, 9991009.
Tishler, A., Ventura, R., & Watters, J. (2001). Cellular telephones in the Israeli market: The demand, the choice of provider, and potential revenues. Applied
Economics, 33, 14791492.
Vogelsang, I. (2010). The relationship between mobile and xed-line communications: A survey. Information Economics and Policy, 22, 417.
Wallsten, S. (2001). An econometric analysis of telecom competition, privatization, and regulation in Africa and Latin America. Journal of Industrial Eco-
nomics, 49, 119.