20030418Staff 2nd Resp - attach I -1 Demand for Access To MTN.pdf

Information Economics and Policy 11 (1999) 297±305
www.elsevier.nl/locate/econbase
An econometric analysis of the demand for access to
mobile telephone networks
a , b*Hyungtaik Ahn , Myeong-Ho Lee
aKorea Information Society Development Institute, 1-1 Juam-Dong,Kwachun,Kyunggi-Do 427-070,
South Korea
bInformation and Communications University, 58-4,Hwaam-Dong,Yusong-Gu,Taejon 305-348,
South Korea
Received 2 February 1999; accepted 1 June 1999
Abstract
In this article, we study the determinants of the demand for mobile telephone networks
using observations on 64 countries. More speci®cally, we estimate the access demand by
exploiting the fact that the subscription rate of a certain country would depend on some
country-speci®c factors where all members in the country face the same values. These
country-speci®c factors are the tariff systems of a country, national wealth, the levels of
technological development and industrialization, ®xed line facilities, and so on. It turns out
that the price effects are not strongly revealed in our study. On the other hand, the
probability of subscribing to telephone networks is positively correlated with per capita
GDP and the number of ®xed lines per person.Ó 1999 Elsevier Science B.V. All rights
reserved.
JEL Classi®cation:C13, C25
1. Introduction
In the last decade, the markets for mobile telephone networks in many countries
have grown rapidly owing to the development of radio communications technolo-
gy, and the growing trend is expected to continue in most countries. Furthermore,
technological development will make it possible to commercialize an advanced
* Corresponding author.
E-mail address:[e-mail] (H. Ahn)
0167-6245/99/$ ± see front matter Ó 1999 Elsevier Science B.V. All rights reserved.
PII: S0167-6245(99)00016-5
298 H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305
form of mobile telecommunications systems which can deliver video as well as
basic data and voice in a few years. A systematic study on the demand for mobile
telephone networks is therefore an urgent problem. In this article, we study the
determinants of the demand for mobile telephone using observations on 64
countries.
In analyzing telecommunications demand, it is important to distinguish between
1access and usage and to take care of the relatedness of the two demand systems.
Our analysis, however, focuses only on the access demand, as the call data are not
available. More speci®cally, we estimate the access demand by exploiting the fact
that the subscription rate of a certain country would depend on some country-
speci®c factors where all members in the country face the same values. These
country-speci®c factors are the tariff systems of a country, national wealth, the
levels of technological development and industrialization, ®xed line facilities, and
so on.
As the subscription rate in a country is a summary measure of individual
decisions on subscribing to mobile telephone networks, the outcome in the
analysis has some implications for individuals' decisions. For instance, price
elasticity can be revealed if the tariff system varies across countries and if all
individuals of a same country face the same tariff system. However, further
inference on individuals' decisions is dif®cult, since our data are aggregated to the
level of a country and aggregation leads to a loss of information.
In the telecommunications literature, some studies use individual data to make
inferences about individuals' access decisions. For example, Train et al. (1987)
study the effect of an individual's calling pattern on the choice among various
local service options. Train et al. (1989) study the residential demand for local
telephone service when individuals face different tariff options. Hartman and
Naqvi (1994) and Tardiff (1995) analyze the determinants of long distance calls
focusing on carrier choices in the United States and in Japan, respectively.
Our work is more closely related to Taylor and Kridel (1990), in which they
study the effect of the local-service rate on residential demand for ®xed telephone
2service using an aggregated U.S. census data set. Both our work and Taylor and
Kridel (1990) proceed with estimating the probabilities of access conditional on
the variables whose values are equal for all members in the same unit of
aggregation but different across the units. However, the estimators of the two
studies are quite different. In Taylor and Kridel (1990), the form of access
probability is speci®ed parametrically, and then the parameters appearing in the
probability are estimated by an iterative method which is similar to maximum
1 More discussions on the theory of telecommunications demand can be found in Chapters 2 and 3 of
Taylor (1994).
2 Their data are aggregated to the level of a census track which is a geographical area.
H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305 299
likelihood. Here, we propose two estimators which are computationally simpler
3than that of Taylor and Kridel (1990).
One of our estimator is based on Berkson's minimum chi-squared method which
is widely used for discrete choice models with grouped data. This estimator is
analytically simple, but it requires specifying the probability of access parametri-
cally. The estimation results from this method can be misleading if the probability
4of access is not correctly speci®ed. Since the form of access probability is not
implied by any economic theory, it is useful to provide an additional estimate
which is robust for the functional form of access probability. Our second estimator
is based on Powell et al.'s (1987) semiparametric weighted average derivatives.
This estimator is more robust for the speci®cation of access probability, but is less
5precise compared to the parametric estimator.
It turns out that the outcomes of both estimates are similar, which may imply
that our speci®cation of the probability function is proper. The effects of tariff
systems are measured by estimating the coef®cients of three prices which are
connection fee, monthly charge, and three-minute local call rate. These price
effects are not strongly revealed in our study. On the other hand, the probability of
subscribing to a telephone network is positively correlated with per capita GDP
and the number of ®xed lines per person.
Section 2 describes our econometric model and two estimators, and Section 3
shows estimation results.
2. Model and estimator
In country i with mobile telephone supply, we observe the sample frequency of
Åsubscription, denoted by d . That is, we observei
Ki1Å]d ;O d , (1)i ikKik51
where K is the size of population in country i, and d denotes an unobservablei ik
binary indicator which equals to 1 if individual k in country i subscribes to a
mobile telephone network and 0 otherwise.
Studies shows that the value of d depends not only on individual k's personalik
or family background such as education, occupation and household income but
3 Note that it is possible to follow their estimation strategy to estimate our model.
4 Taylor and Kridel's (1990) estimator is subject to the same criticism.
5 Often, it is dif®cult to obtain meaningful statistical inference results based on semiparametric
estimators when the sample size is small.
300 H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305
also on country speci®c factors for which all individuals in countryi face the same
values. That is,
d 5 1 f(z ,c)$0 (2)f gik ik i
Here, 1[A] is an indicator function of an argument A, and f(?) is a real-valued
function. The variables z and c denote individual background and countryik i
speci®c factors, respectively. Amongz and c, we observe only a part of c. Letik i iPc;[w ,j ], where w [R is observed and j unobserved.i i i i i
In country i with mobile telephone facilities, the probability of mobile telephone
subscription conditional on c isi
g (;g(w ,j ));E d uw ,j (3)f gi i i ik i i
Loosely speaking,g is the probability that a typical person in country i subscribesiÅto a mobile telephone network. It should be noted from the de®nition of c that di i
converges to g under regularity conditions which imply the law of large numbers.i
We assume that there is a common mechanism which determines the outcome of
the conditional probability g for each country. More speci®cally, we assume thati
g is subject to the following index restriction.i
9g5F(w b 1 «), (4)i i 0 i
where F(?) is a real valued function between 0 and 1,b is a parameter vector,0
and «(;«(j)) is a composite error. By Eq. (4), our goal becomes estimation ofi i
b. Note that b is different from the coef®cients of the usual binary response0 0
model which characterize the individual's utility function. Some parameters of an
individual's utility function may not be identi®ed in our study, since we use
aggregated data. For more discussions, refer to Ameniya (1985).
Below, we outline two assumptions and estimators.
Assumption 1.(a)The function F(?)is known and invertible. (b)The unobserved
conditioning variables j in Eq. (3)are statistically independent from w . (c)i i219E[w w ]exists.i i
By Part (a) of this assumption, we may transform Eq. (4) into
21 9F(g )5 w b 1 «, (5)i i 0 i
6Part (b) implies that «is independent of w . Parts (b) and (c) are suf®cienti i
conditions for identi®cation ofb . That is,0
21219b5E w w ?E w F (g ) (6)f gf g0i i i i
6 The statistical independence implies E[«w ]5 0.i i
H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305 301
The least squares estimator of b can be easily found by replacing the unknown g0iÅby its estimate d in Eq. (6). That is,i
21 21Ã Å9b5Ow w ?O w F (d ) , (7)i i i iF G F Gi i
2In principle, the above estimation strategy is similar to Berkson's minimum x
method for binary response models with grouped data. Unlike Berkson's estimator,
Ãhowever, the sampling variance of b results from two ways; the variance of least
7Åsquares and the variance due to the difference between d and g . The variance duei iÅ Åto using d instead of g is very small, since d is obtained based on the entirei i i8populationKwhich is very large.i ÃAs shown in Eq. (7), Part (a) of Assumption 1 is crucial in calculating b. If
F(?) is not correctly speci®ed, or if it is not invertible, the estimator would not be
consistent in general. Below, we replace this strong condition.
Assumption 2.The function F(?)is not known,but it is continuously differenti-
able with respect to all components in w .i
The continuity of w is a strong requirement in many econometric applications.i
However, it is satis®ed in our case. By taking the derivative of Eq. (4) with respect
to w, we get
­g(w ,j )i i]]]g (w ,j );5 F9(w b 1 «)?b (8)w i i i 0 i 0­w
Hence, for some weight a ;a(w ), the weighted average derivativei i
E a ?g (w ,j )5 E a ?F9(w b 1 «)?b (9)f g f gi w i i i i 0 i 0
is proportional to b . Also, it is clear from the above equations that w should not0i
contain a constant term. Following Powell et al. (1987), it is convenient to set ai
as p(w ), the density of w . Under the same regularity conditions in Powell (1987),i i
it can be shown that the density-weighted average derivative is
d ;E p(w )?g (w ,j)5 2 23E p9(w )?g (10)f g f g0i w i i i i
7 For pedagogical purposes, de®ne
21 21Ä9b5Ow w ?O w F (g )i i i iF G F Gi i
It is straightforward to see that
Ä ÄÃ Ã(b2b)5 (b2 b )1(b2 b)0 0
The ®rst term on the right hand side results from the variation due to the error term«, while the secondiÅterm results from the variation due to the difference between d and g .i i8 ÅIn the following section, we do not adjust the standard error due to d for this reason.i
302 H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305
In the above expression, the unknown density derivative p9(w ) is estimated by thei
nonparametric kernel method. The estimator of d takes the form0
N N21 w 2 wN1i jà ŠÅ]]]]]S D S Dd5OOK9 ?[d 2 d ], (11)P11 i j2hhi51j.i
where K(?) and h are the kernel function and bandwidth parameter respectively.
ÃThe estimator d is consistent to d under proper restrictions on the kernel and0
bandwidth. More speci®cally, we require that the kernel function be symmetric
and integrated into one, and that the bandwidth parameter shrink to 0 as N gets
9large.
3. Estimation results
We extract data from International Telecommunication Union's World Tele-
communication Development Report (International Telecommunication Union,
1998) which contains information on telecommunications for 206 countries. Our
explanatory variables w consist of connection fee (p ), monthly charge (p ), locali1 22 3call rate (p ), per capita GDP (y) and its higher order terms (y ,y ), the number3
of ®xed lines per person (l), and the rate of digitalization (q).
Under rational expectations assumptions, an individual's subscription decision
depends on tariff systems in two ways. First, a combination of connection fee and
a discounted stream of expected future monthly charges would have negative
effect on the subscription decision. Unfortunately, it is dif®cult to construct such a
combination, since interest rates are different across countries and observations on
future monthly charges are not available. In this study, we regard the current
monthly charge as a proxy for the discounted stream of expected future monthly
charges, and include it in w along with the connection fee. Second, thei
subscription decision depends on local call rate indirectly, since demand for access
to the mobile telephone system is connected with demand for use of the system. In
our sample, some countries have both analogue and digital facilities, and have
different values of (p ,p ,p ) between the two facilities. For simplicity, we use1 2 3
the averages of the different values weighted by the rate of digitalization.
The higher order terms of per capita GDP are included, since the relation
between subscription probability and per capita GDP would be nonlinear in
10principle. The number of ®xed lines per person is used, as it may reveal a
country's technological development. It can also re¯ect consumer's decision of
9 Refer to Powell et al. (1987) for detail. Note that our estimator in Eq. (11) is not exactly the same
Åas their estimator, because the unknown g is replaced by d . Again, this difference is negligible in ouri i
study.
10 It is true that the higher order terms of the prices (p ,p ,p ) would be needed for the same1 2 3
reason. Unfortunately, however, our sample size is not big enough to support that many terms.
H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305 303
using between ®xed and mobile telephones. Finally, the rate of digitalization is
used, as it may re¯ect the quality of telephone services.
Among those countries which provide full information on w , we select onlyi
those countries whose per capita GDP's are in upper 80% and which have mobile
11telephone facilities. We end up using 64 observations, and the descriptive
12statistics are shown in Tables 1.
ÅThe table shows that the distributions of d and y are severely skewed to the
right. However, it does not appear that the skewness for these variables is largely
due to discarding lower income countries. Note that the consistency of our
estimators is not affected by the skewness of these variables.
For the parametric transformation model, we specify F(?) as the cumulative
standard normal distribution. The estimates under normal transformation (NT)
would be similar to the familiar probit estimates which could have been calculated
if the individual observations on d were available. The density weighted averageik
derivative estimator (DWAD) requires specifying the kernel and bandwidth. The
kernel function was taken to be the product of the standard normal density
functions. The regressors w were linearly transformed to have the samplei
Table 1
aDescriptive statistics
Variables Mean SD Q Median Q1 3
p 0.1319 0.1575 0.0317 0.0762 0.20301
p 0.0279 0.0190 0.0160 0.0271 0.03612
p 0.0011 0.0013 0.0005 0.0011 0.00133
y 8.7188 11.4315 1.3154 2.9592 10.4733
l 0.2247 0.2002 0.0599 0.1493 0.3126
q 0.4803 0.4143 0.0000 0.5630 0.8710
Åd 0.0461 0.0786 0.0008 0.0110 0.0471
a Notes:SD denotes standard deviation,Q and Q are ®rst and third quantiles, respectively. The1 3
unit of (p ,p ,p ,y) is US$1,000.1 2 3
11 As we select the countries with mobile telephone facilities, our current analysis can only estimate
the probabilty of subscription conditional on the countries with mobile telephone facilities. Hence, the
outcomes of this study may not be proper in explaining the probability without conditioning on the
mobile telephone facilities. This sample selection issue will be one of our future research topics.
Neglecting the lower 20% income countries would not cause an additional sample selection problem,
since most of these countries do not have mobile telephone facilities.
12 The list of selected countries are as follows: Algeria, Argentina, Australia, Belarus, Belgium,
Benin, Brazil, Bulgaria, Cameroon, Colombia, Czech Republic, Denmark, C e d'Ivoire, Ecuador, El
Salvador, Estonia, Finland, France, Gabon, Georgia, Germany, Guatemala, Honduras, Hongkong,
Hungary, Jamaica, Japan, Jordan, Kazakhstan, Latvia, Lebanon, Lesotho, Lithuania, Mauritius, Mexico,
Moldova, Mongolia, Morocco, Namibia, Netherlands, Norway, Oman, Peru, Poland, Portugal,
Romania, Saudi Arabia, Senegal, Singapore, Slovak Republic, South Africa, Spain, Sri Lanka, Sweden,
Switzerland, TFYR Macedonia, Thailand, Tunisia, Turkey, Ukraine, United Arab Emirates, United
Kingdom, Uruguay, Zambia
304 H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305
Table 2
Estimates
Variables NT DWAD
Estimates t-ratios Estimates t-ratios
constant 23.1492 (226.1809)
p 20.2482 (20.7861)20.2482 (20.7321)1
p 26.1001 (22.3671)24.6628 (20.9367)2
p 230.6230 (20.8412)236.4019 (21.0703)3
y 0.2470 (6.2686) 0.0711 (2.2801)
2y 20.0115 (24.9198)20.0039 (21.7205)
3y 0.0002 (4.0853) 0.0001 (1.2559)
l 1.5553 (2.7605) 1.1997 (2.6685)
q 20.0060 (20.0507)20.1443 (20.9829)
covariance equal to an identity matrix, which was done to use a scalar-valued
bandwidth. Finding the right scalar value is still a vexing problem. Powell and
Stoker (1994) have shown the optimal bandwidth for DWAD. However, since this
optimal bandwidth is hard to estimate, we chose the scalar based on the convenient
generalized cross-validation criterion. The estimation results are shown in Table 2.
To make the comparison easier, we rescale the estimates of DWAD so that the
coef®cient of p is same as that of NT.1
All coef®cients of (p ,p ,p ) are negative, but only the monthly charge for NT1 2 3
is signi®cant at the 95% asymptotic con®dence level. Both estimates show that the
probability of subscribing to the mobile telephone network increases with per
2 3capita GDP. Although the coef®cients of (y ,y ) are statistically signi®cant, their
magnitudes are much smaller than the coef®cient of (y), which would indicate that
ignoring these two terms does not signi®cantly change the outcomes. As explained
2 3earlier, the coef®cients of (y,y ,y ) may not exactly reveal the income effect, i.e.,
the increase in probability due to an increase in an individual's income. The
number of ®xed lines per person also has positive in¯uence on the probability of
mobile telephone subscription. On the other hand, the effect of digitalization is not
identi®ed in this study.
Acknowledgements
We are grateful to Y.J. Choi and a referee for their valuable comments.
References
Amemiya, T., 1985. Advanced Econometrics, Harvard University Press, Cambridge, MA.
Hartman, R.S., Zareen, F., Naqvi, Z.F., 1994. Estimation of Household Preference for Long Distance
Telecommunications Carrier. Journal of Regulatory Economics 6, 197±220.
H.Ahn,M.-H.Lee /Information Economics and Policy 11 (1999) 297±305 305
International Telecommunication Union, 1998. World Telecommunication Development Report 1998,
International Telecommunication Union, Geneva, Switzerland.
Powell, J.L., Stoker, T.M., 1994. Optimal bandwidth choice for density-weighted averages. Working
paper, MIT.
Powell, J.L., Stock, J.H., Stoker, T.M., 1987. Semiparametric estimation of index coef®cients.
Econometrica 57, 1403±1430.
Tardiff, T.J., 1995. Effects of presubscription and other attributes on long-distance carrier choice.
Information Economics and Policy 7, 353±366.
Taylor, L.D., 1994. Telecommunications Demand in Theory and Practice, Kluwer Academic Pub-
lishers, Dordrecht, The Netherlands.
Taylor, L.D., Kridel, D., 1990. Residential demand for access to the telephone network. In: de
Fontenay, A., Shugard, M.H., Sibley, D.S. (Eds.), Telecommunications Demand Modelling, North-
Holland Publishing, Amsterdam.
Train, K.E., McFadden, D.L., Ben-Akiva, M., 1987. The demand for local telephone service: a fully
discrete model of residential calling patterns and service choices. Rand Journal of Economics 18 (1),
109±123.
Train, K.E., Ben-Akiva, M., Atherton, T., 1989. Consumption patterns and self-selecting tariffs. The
Review of Economics and Statistics 71 (1), 62±73.