does competition lead to...
TRANSCRIPT
Does Competition Lead to Customization?
Wen-Tai Hsu∗ Yi Lu† Travis Ng‡
December 2011
Abstract
This paper proposes a simple theory of competition and customization. When
firms allocate their production to both custom-made and generic products, the frac-
tion of sales from the former will increase in the face of increased competition. We
test this prediction using a World Bank survey of Chinese firms and find consistent
empirical results.
1 Introduction
Globalization has changed the global playing field substantially in the past few decades.
Significant reductions in tariff and non-tariff barriers and technological advances have
rendered the world’s markets more integrated and more competitive. To cope with the
increased competition they face, firms have experimented with a number of different
strategies, such as the flattening of firm hierarchies (e.g., Thesmar and Thoenig, 2000;
Guadalupe and Wulf, 2010), the geographic fragmentation of production (e.g., Hanson,
Mataloni, and Slaugher, 2005; Feenstra, 2010), innovation (e.g., Aghion, Bloom, Blundell,
∗Department of Economics, National University of Singapore†Department of Economics, National University of Singapore.‡Department of Economics, Chinese University of Hong Kong.
1
Griffith, and Howitt, 2005), and the decentralization of the decision-making function (e.g.,
Bloom, Sadun, and Van Reenen, 2010).
In this paper, we explore another strategy that firms may adopt to cope with compe-
tition, that is, customization. More specifically, we ask whether competition leads firms
to seek the greater customization of their products. We define customization as the costly
alteration of a product to tailor it to customers’ needs or tastes. As the product is custom-
made to some customers, the firm exercises greater market power over these customers,
which constitutes the main source of the gains from customization. Intuitively, without
customization, when there is an increase in competition, the demands faced by individual
firms are necessarily reduced, as is their market power. In this context, customization is one
way to retain or increase market power and profits.
To formalize this intuition and derive formal predictions, we propose a simple theory
of competition and customization by adapting Loginova and Wang’s (2011) model of cus-
tomization, which builds upon Hotelling’s (1929) spatial competition framework. In our
model, the level of competition increases when there is an increase in the number of firms.
We show that the fraction of sales from customization increases when there is increased competi-
tion. Moreover, we show that if this increased competition is induced by a larger market
size, then firms have the incentive to reach out to more customers to accommodate their
needs/tastes.
The main task in this paper is to test our prediction that increased competition leads to
a larger fraction of customized sales. The effect of market competition on customization
has seldom been investigated empirically, primarily due to the difficulty of measuring
customization. Fortunately, a unique World Bank survey of Chinese firms, the Survey of
Chinese Enterprises (SCE), has opened up a window for such investigation. This survey
contains questions asking firms about the proportion of their competitors’ output that is
produced locally and the proportion of their own sales that is custom-made. Considerable
variations in both variables allow us to make inferences.
2
Consistent with our model, we find increased competition to be significantly associ-
ated with a higher degree of customization. To ensure that this finding is not affected by
estimation problems, we estimate the association with the following series of specifica-
tions and robustness checks and find it to remain robust.es:
More Controls: Inclusion of a list of variables, including industry and city dummies, and
firm and CEO characteristics, to address the concern over omitted variables bias.
GMM: Generalized Method of Moments (GMM) estimation with two instruments for
competition to further deal with potential omitted variables bias, the reverse causal-
ity issue, and the measurement error problem.
Tobit Estimation: Alternative estimation to fit the censored data setting.
Outliers: Exclusion of outlying observations to ensure the finding is not driven by par-
ticular observations.
Alternative Measure: Alternative measure of competition to check whether the finding
is sensitive to subjective or objective measurement.
Self-Selection: Exclusion of firms recently relocated to the surveyed city to self-selection
concern.
In the theoretical literature, Loginova (2010) and Loginova and Wang (2011) are the
most closely related to ours. Note that although customization is closely related to hor-
izontal product differentiation, greater product differentiation in a spatial competition
framework is almost equivalent to greater entry, which reduces market power and is not
in line with the purpose of customization. Nevertheless, Shaked and Sutton (1982) have
gone in the vertical direction by showing that quality differentiation helps to relax price
competition.
As previously noted, due to the difficulty of measuring customization, there is rela-
tively little related empirical research. To the best of our knowledge, the studies closest
3
to ours are Mazzeo (2002) and Holmes and Stevens (2010). Mazzeo (2002) shows that the
motels within larger clusters (thus facing greater competition) along interstate highways
in the U.S. tend to differentiate themselves by quality. This paper differs from Mazzeo
(2002) in two ways. First, we examine customization, which is conceptually very different
from vertical differentiation. Second, we go beyond a particular industry to demonstrate
that our prediction generally holds among several manufacturing industries. Holmes
and Stevens (2010) document the greater survival of small plants relative to large plants
in the face of fiercer competition. In particular, when competition becomes fiercer because
of an influx of foreign imports, it is the firms that specialize in custom-made goods that
have a greater chance of survival relative to those producing generic products. Our ob-
ject of examination differs from that of Holmes and Stevens (2010). This paper examines
the allocation of production to customized and generic products within firms rather than
comparing firms that do and do not customize.
The remainder of the paper is organized as follows. Section 2 presents a model that
formalizes our prediction that market competition leads to a larger proportion of custom-
made sales. Section 3 describes the data, specifies our empirical approach, and presents
the results. Section 4 concludes.
2 A theory of competition and customization
We consider a spatial competition model with the possibility of customization by adapt-
ing the model developed by Loginova and Wang (2011). The main idea is that investing in
a customization technology allows firms to offer a subset of customers their ideal product
varieties. Firms can then price-discriminate across the customers who are offered these
customized products. The extra gains in profit due to such price discrimination consti-
tutes the main driving force for customization.1
1The three main differences between our model and Loginova and Wang’s (2011) are as follows. First,to focus on the impact of competition on customization, we do not consider the quality dimension. Second,
4
2.1 Model
Consider a market in which each product i is characterized by xi ∈ (0, 1] on a circumfer-
ence. There are in total of mass D customers who are uniformly distributed in their ideal
variety x ∈ (0, 1]. A customer of type x derives utility v − t |x− xi| − pi from buying one
unit of product i, where v is a positive constant, t is a taste parameter, and pi is the price
of product i. We assume that v is sufficiently large that all customers find a product that
yields a positive payoff in equilibrium. The market includes a large number of ex ante
identical potential entrants. To enter, each entrant needs to pay an entry cost φ. As in
Salop (1979) and Syverson (2004), we assume that all entering firms are evenly spaced;
that is, if there are n firms, then each firm is 1/n distance away from its two neighboring
firms.2 For simplicity, assume that each firm operates with a zero marginal cost of pro-
duction.3 Suppose that the closest firms to a customer of type x are firm A to the left and
firm B to the right, and suppose that the customer’s distance to A is y. Hence the utility
that x enjoys from buying the non-customized, generic products of firm A’s and firm B is
v − ty − pGA and v − t(1/n− y)− pGB, respectively.
Investing in product-customization technology allows a firm to sell customers their
ideal variety beyond the firm’s location. Given an amount of investment by firm (denoted
by i), it can produce a customized product for every customer up to a distance of si away
(on both sides) from its location. Each firm i chooses si ≥ 0, which incurs an investment
cost of c (si). Assume that c (.) is differentiable, strictly increasing, and strictly convex and
that c (0) = 0. Hence, greater customization for more diverse customers is increasingly
costly. For customers beyond a distance of si away from firm i, however, the firm can sell
as data show that firms sell both customized and non-customized, generic products at the same time, weextend the notion of customization in Loginova and Wang (2011) to include customization for only a subsetof customers. Third, we opt for Salop’s (1979) circumference instead of a Hotelling interval, so as to modelthe endogenous increase in competition and the ensuing impact on customization.
2A micro-foundation for Salop’s even spacing is provided by Vogel (2008), who shows that mixed-strategy pricing in an auxiliary game can eliminate the possibility of firms’ undercutting their opponentson price. Under this setup, with the same marginal cost, firms choose to be equi-distant from neighboringfirms.
3It can be verified that assuming a positive marginal cost produces similar results.
5
only a non-customized, generic product. Note that customization allows the firm to set
prices for the customized customers individually instead of applying a uniform price.
The game involves three stages, an entry stage followed by a customization stage and
then a pricing stage. In the entry stage, potential entrants decide whether to enter. In the
customization stage, firms simultaneously decide the amount of customization invest-
ments. These decisions become common knowledge in the pricing stage, in which firms
simultaneously choose prices. Customers decide which products to purchase, and profits
are realized. The subgame perfect equilibrium is solved using backward induction.
2.2 Analysis: market size, competition, and customization
In our SCE data, we observe that firms sell both customized and non-customized prod-
ucts. Thus, to focus our analysis on cases in which there is both customized and non-
customized production, we assume that c (s) increases in s sufficiently fast such that the
optimal si < 1/ (2n) for all i. Because si < 1/ (2n), no two competing neighboring firms’
customization areas overlap.
2.2.1 Pricing stage
Given n evenly spaced firms and customization investments si, i ∈ {1, 2, ..., n}, we can
solve the pricing decisions by looking at a Hotelling duopoly problem in which two firms,
A and B, are located at the two end points of [0, 1/n].4
Customer x ∈ [sA, 1/n−sB] chooses between the two firms in buying a non-customized
product, i.e., max{v − tx− pGA, v − t (1/n− x)− pGB
}. Then, a customer of type x who is
4d’Aspremont, Gabszewicz, and Thisse (1979) show that equilibrium may not exist in the pricing stageif the distance between two neighboring firms is so close that price undercutting neighboring firms is optimal.This issue arises in Salop’s setup if n is too large or t is too small. To avoid this problem, we assume asufficiently large t or a sufficiently large entry cost φ, such that n is small. An alternative is to resort tothe possibility of mixed-strategy pricing in the auxiliary game proposed by Vogel (2008), who proves theexistence of a pure-strategy price equilibrium under such a possibility.
6
indifferent between the two choices is given by
x =1
2n− pGA − pGB
2t. (1)
Note that, for now, x in the above formula may be outside the interval [sA, 1/n− sB]. We
show that, indeed, x ∈ [sA, 1/n− sB] shortly.
For a customer x ∈ [0, sA], in addition to the choice of two non-customized products
from A and B, her additional choice is whether to buy the customized product offered
by firm A. Thus, she solves max{v − tx− pGA, v − t (1/n− x)− pGB, v − pxA
}. In this stage,
c (sA) is sunk, and firmAwould obviously choose to limit price pxA = min{tx+ pGA, t (1/n− x) + pGB
}to capture the sales from [0, sA]. We argue that it must be the case that pxA = tx + pGA, i.e.,
the customizing firm’s pricing of its customized product is constrained by its pricing of its
own non-customized product. To see this, suppose on the contrary that t (1/n− x)+pGB <
tx + pGA, which implies that x < sA, and all customers in (sA, 1/n] will purchase products
from firm B. Then, it is obviously to B’s benefit to increase pGB, as long as x < sA, because
the customers in (sA, 1/n] will continue to buy from it, regardless of the increase in price.
Thus, equilibrium price pGB must satisfy the condition that x ≥ sA, which, in turn, means
that t (1/n− x) + pGB ≥ tx+ pGA for all x ∈ [0, sA]. Hence, pxA = tx+ pGA. A similar argument
for firm B implies that x ≤ 1/n− sB. Thus, we have proved that x ∈ [sA, 1/n− sB].
For now, we assume that the market size (or indeed the density) D = 1 for ease of
exposition. We will make D a general number when we discuss the impact of market
size. Using (1), we write firm A’s profit as a function of pGA and pGB5:
πA(pGA, p
GB) = [x− sA]pGA +
∫ sA
0
pxAdx
=
[1
2n− pGA − pGB
2t− sA
]pGA +
∫ sA
0
(tx+ pGA
)dx (2)
=
[1
2n− pGA − pGB
2t
]pGA +
∫ sA
0
txdx. (3)
5Note that c (sA) is sunk in this stage and hence does not show up.
7
By observing (3), we can see that customization, or the lack thereof, has no effect on how
optimal price pGA is determined, given pGB; the first-order condition for pGA is the same as
that for a standard Hotelling problem, which is similarly the case for pGB. These solutions
are pGA(pGB)= t/ (2n) + pGB/2 and pGB
(pGA)= t/ (2n) + pGA/2. Thus, the equilibrium price
pair is (t/n, t/n), and x = 1/ (2n).
2.2.2 Customization stage
Plugging equilibrium price pair (t/n, t/n) into (2), the equilibrium profit of firm i, i =
A,B, from investing c (si) is thus
πi (si) =
[1
2n− si
]t
n+ts2i2
+tsin− c (si) (4)
=t
2n2+ts2i2− c (si) . (5)
Hence, optimal customization, s∗i , satisfies ts∗i = c′ (s∗i ). Recall that c (.) is strictly increas-
ing and strictly convex, and that c (0) = 0. To ensure that a unique s∗i ∈ (0, 1/ (2n))
maximizes firm i’s profit, it is sufficient that c′ (0) < t and that c increases fast enough in
s to ensure that c(1/ (2n)) > t/ (2n).
As our empirical measure of customization is the share of sales from customization,
we now show that greater competition (a larger n) leads to a larger such share. Divide the sum
of the second and third terms by that of the first three terms in (4), and we can see that
the share of sales from customization is
r =s2in
2 + 2nsis2in
2 + 1, (6)
which is strictly increasing in n. Although an increase in n decreases the sales of both
customized and the non-customized products due to the decrease in prices, the gains
from customization versus no customization, i.e., the second term in (5), is unaffected.
8
These extra gains from customization arise because firms can price-discriminate each of
the customers offered customized products. As the gains from customization are robust
to an increase in competition, the optimal level of customization, s∗i , is unaffected by
increases in n. Meanwhile, as n increases, the number of customers still buying non-
customized products, 1/n − 2s∗i , necessarily decreases, or, put differently, the fraction of
customers offered customized products, 2s∗i /(1/n) = 2ns∗i , necessarily increases.
2.2.3 Market size and customization
What induces an increase in competition? Here, we offer a market size perspective and
examine the impact of such size on customization. Relax D = 1 to a general D > 0. It is
standard for an increase in D to weakly increase the number of firms n (weakly because
n is an integer number). We omit these uninteresting details, and simply denote such a
relation by n∗ (D), knowing that n∗ weakly increases in D.
As the magnitude of D has no impact on pricing decisions, the profit of firm i in the
second stage is
πi(si) = D
{[1
2n∗ (D)− si
]t
n∗ (D)+ts2i2
+tsin
}− c (si) (7)
= D
[t
2 [n∗ (D)]2+ts2i2
]− c (si) .
The first-order condition is thus
Dts∗i = c′ (s∗i ) .
Hence, regardless of the equilibrium entry n∗ (D), the optimal level of customization,
denoted by s∗i (D), must be strictly increasing in D. In terms of the share of sales from
customization, by inspecting (7), we can see immediately that the formula for this share
is the same as (6), except that now n = n∗ (D) and si = s∗i (D). Thus, as D increases, n∗
increases, which leads to a larger r if there is no change in s∗i , for the same reason as that
discussed previously. Nevertheless, because s∗i also increases in D, the increase in r is
9
even larger. We summarize our results in the following proposition.
Proposition 1. An exogenous increase in competition, i.e., an increase in the number of firms n,
leads to a larger share of sales from customization, r, as defined by (6), whereas the level of each
firm i’s customization s∗i remains unchanged, which also implies that the fraction of customers
offered customized products, 2ns∗i , increases in n. An endogenous increase in competition induced
by a larger market size D leads to both a larger share of sales from customization, r, and more
customization activities, s∗i .
3 Empirical analysis
3.1 Data
Our empirical analysis draws on data from the Survey of Chinese Enterprises (SCE), which
was carried out by the World Bank in cooperation with the Enterprise Survey Organiza-
tion of China in early 2003. For balanced representation, the SCE covered 18 prefecture-
level cities in five geographic regions of China: Benxi, Changchun, Dalian, and Harbin in
the Northeastern region; Hangzhou, Jiangmen, Shenzhen, and Wenzhou in the Coastal
region; Changsha, Nanchang, Wuhan, and Zhengzhou in the Central region; Chongqing,
Guiyang, Kunming, and Nanning in the Southwestern region; and Lanzhou and Xi’an in
the Northwestern region.
In each of these cities, the SCE randomly sampled 100 or 150 firms from nine manufac-
turing industries (garments and leather products, electronic equipment, electronic parts
making, household electronics, auto and auto parts, food processing, chemical products
and medicine, biotech products and Chinese medicine, and metallurgical products) and
five service industries (transportation services, information technology, accounting and
non-banking financial services, advertising and marketing, and business services). The
total number of enterprises surveyed is 2,400.
10
The SCE contains two parts. The first is a general questionnaire directed at senior
management that seeks information about the enterprise, such as degree of innovation,
product certification, marketing, relations with suppliers and customers, access to mar-
kets and technology, relations with government, labor force, infrastructure, involvement
in international trade, finance, and taxation, and the information on the CEO and board
of directors. The second questionnaire is directed at accountants and personnel managers
and covers ownership, various financial measures, and labor and training. Most of the
information in the first part of the SCE pertains to the survey year, 2002, whereas that in
the second part pertains to the 2000-2002 period.
As service industries are largely localized and customized, we focus here on the man-
ufacturing firms in the SCE. Our final sample thus contains 1,566 firms.
3.2 Empirical approach
In the spirit of Proposition 1, to test whether increased market competition (x) leads to a
larger fraction of sales from customized products (r), we start with the following linear
equation:
rfic = ζ + β · xfic + εfic, (8)
where f , i, and c index firm, industry, and city, respectively.
The measure of our dependent variable, rfic, comes from the SCE’s question concern-
ing about the percentage of a firm’s sales made to clients’ unique specifications (i.e., its
sales of products that cannot be sold to other clients), and is denoted as Custom-made in
the regression tables. Table 1 shows that of the manufacturing firms in the sample, ap-
proximately 40% of their output is custom-made. The large standard deviation of this
variable suggests that there are substantial variations from which to draw inferences.
The regressor of interest, xfic, concerns market competition. To capture the degree of
market competition, we use the percentage of output produced by a firm’s competitors’
11
Table 1: Summary StatisticsVariable Obs Mean Std. Dev. Min MaxCustom-made 1511 0.406 0.419 0 1Local Competition 1507 0.273 0.337 0 1Firm characteristicsFirm Size 1563 5.040 1.453 0 9.899Firm Age 1566 2.494 0.777 1.099 3.970Private Ownership 1566 0.796 0.389 0 1Labor Productivity 1332 3.503 1.574 -3.761 11.150Skilled Labor 1542 0.026 0.060 0 1CEO characteristicsCEO Education 1553 15.359 2.511 0 19CEO Tenure 1548 6.240 4.580 1 33Deputy CEO Previously 1548 0.280 0.449 0 1Government Cadre Previously 1548 0.036 0.185 0 1Party Member 1524 0.648 0.478 0 1Government-appointed 1544 0.243 0.429 0 1Instrumental variableLocal Clients 1542 0.331 0.362 0 1Local Suppliers 1541 0.373 0.349 0 1Additional controlsClients’ Duration Dummies¡1 yr 1549 0.039 0.193 0 11 to 2 yrs 1549 0.073 0.260 0 12 to 3 yrs 1549 0.125 0.331 0 13 to 4 yrs 1549 0.126 0.332 0 1¿4 yrs 1549 0.637 0.481 0 1Custom-made Component 1444 0.068 0.221 0 1
in the same city. Focusing on the local rather than national market allows us to capture
the effect of market competition, as the product market may not be fully integrated due
to transportation costs and market friction. Recall that we employ the number of firms
to indicate the level of competition in our theoretical model for the sake of simplicity. In
reality, however, with heterogeneous firms, competitors’ output is a better measure. In
the presence of heterogeneous firm sizes in terms of output, excluding firm f itself in
constructing the measure of market competition allows us both to utilize firm-level varia-
tions even within the same city and same industry, and avoid the mechanical correlation
between the regressor and the outcome variable. For ease of exposition, xfic is denoted as
Local Competition in all of the regression tables.
Note that we use a subjective measure of market competition (that is, the firm’s per-
ceived percentage of competitors in the same city) rather than an objective measure (such
as the total number of firms in the same industry and city, as used in Holmes and Stevens
12
(2002) and Henderson (2003)). Our subjective measure has certain advantages over objec-
tive measures based on industry classification. Arguably, a firm’s decision making (with
regard to customization) is based on its perception of the competition. As long as firms
in the same industry and city potentially face different degrees of competition (e.g., large
versus small firms), our subjective measure is able to capture this heterogeneity in re-
sponses. However, it may suffer from the problem of idiosyncratic observational error or
misreporting. We provide details of how we deal with this potential measurement error
problem later in this section. Moreover, as a robustness check, we experiment with an
objective measure used in the literature (that is, the total amount of employment in the
same industry and same city).
To estimate equation (8), we primarily employ ordinary-least-squares (OLS) estima-
tion. However, as the dependent variable is left-side (at 0) and right-side (at 1) truncated,
we also use Tobit estimation as a robustness check. In addition, to deal with the possible
heteroskedasticity, we cluster the standard errors at the industry-city level in all of the
regressions.
Before proceeding to the results, we discuss several potential econometric problems
that may cause biases in estimating equation (8).
Omitted variables. It is plausible that xfic is correlated with the error term εfic in equa-
tion (8), thus biasing the estimation of β. One prominent set of omitted variables includes
industrial differences, such as differences in entry barriers (e.g., φ), customization tech-
nology (e.g., si), and taste (e.g., t). To address this concern, we include industry dummies
in the regression analysis. We also include city dummies to account for any potential city
differences. To further control for variations across industries within a city, we replace the
industry and city dummies with industry-city dummies.
Another prominent set of omitted variables encompasses those related to firm capa-
bility. Picone, Ridley, and Zandbergen (2009) show that firms with a greater ability to
differentiate their products are more likely to cluster strategically. To single out the effect
13
of local competition on customization, we control for a list of firm6 and CEO character-
istics commonly used in the literature.7 To avoid the “bad control” problem, we employ
the lagged values of these variables, as in the regression, wherever possible (Angrist and
Pischke, 2009).
As a result, the estimation equation becomes
rfic = α + β · xfic + Z′
ficγ + υfic, (9)
where Zfic is a vector of control variables.
Admittedly, it is impossible for us to control for all possible omitted variables in the
regression. However, as will be seen, controlling for the aforementioned prominent omit-
ted variables has almost no impact on our estimation of β, in terms of either statistical
significance or magnitude. As a further robustness check, we also resort to instrumental
variable estimation, which we discuss shortly.
Measurement error. Our measure of competition is a subjective measure, which raises
concerns over potential measurement error. Theoretically, it is difficult to say whether
and why firms in more competitive environments are more (or less) likely to misreport
the degree of market competition. However, the long list of controls in the regression
analysis may allow us to control for certain systematic patterns in the measurement error
across firms, although the existence of the white-noise type of measurement error may
drive the estimated coefficient β towards zero against any significant findings. We resort
6The variables related to firm characteristics include Firm Size (measured by the logarithm of 2001 totalemployment), Firm Age (measured by the logarithm of years of establishment), Private Ownership (measuredby the share of equity owned by private parties in 1999), Labor Productivity (measured by the logarithm ofoutput per worker in 2001), and Skilled Labor (measured by the share of workers in 2001 who dealt withadvanced technology).
7The variables concerning CEO characteristics are his or her human capital, including CEO Education(years of schooling), CEO Tenure (years of being CEO), and Deputy CEO Previously (a dummy variable in-dicating whether the CEO was the firm’s deputy CEO before he or she became its CEO), and politicalcapital, including Government Cadre Previously (a dummy variable indicating whether the CEO was a gov-ernment official before he or she became CEO), Party Member (a dummy variable indicating whether theCEO is a member of the Chinese Communist Party), and Government-appointed (a dummy variable indicat-ing whether the CEO was appointed by the government).
14
to the instrumental variable approach to further address this type of measurement error
problem.
More on omitted variables and measurement error. To further address concerns over omit-
ted variables and measurement error, we employ GMM estimation with two instruments
for the key explanatory variable, the measure of market competition.
Krugman (1991) shows that the clustering of manufacturers is positively correlated
with that of consumers due to the demand-supply linkage. As shown in Section 2.2.3, an
increase in the number of consumers leads to an increase in the degree of competition.
However, from equation (6), we can see that this increase in the number of consumers
does not lead directly to a larger fraction of sales from customized products. Hence, the
number of consumers constitutes a good instrumental variable for the degree of market
competition. One question in the SCE asks respondents about the percentage of the firm’s
clients (in terms of sales) that are located in the same city as the firm. We construct our
first instrument, Local Clients, accordingly.
Krugman and Venables (1995) and Venables (1996) further show that the clustering of
manufacturers is also positively correlated with that of their suppliers due to the vertical
linkage. Following these researchers’ insights, we construct our second instrument, Local
Suppliers, which is the percentage of a firm’s suppliers (in terms of sales) that are located
in the same city as the firm.
The validity of GMM estimation relies on the exclusion restriction, which means that
the two instruments can affect the outcome variable (Custom-made) only through the en-
dogenous variable (Local Competition). With regard to the exclusion restriction, note that
with the inclusion of industry and city dummies, the possible correlation between the
instrumental variables and the error term υfic in equation (9) stems largely from firm-
level characteristics. Given the small-sized nature of the sample firms, it is difficult to
see how an individual firm could influence the location decisions of its clients and sup-
pliers, which may be why the Hansen J statistic fails to reject the hypothesis that at least
15
one of our two instruments is valid and why the additional inclusion of firm and CEO
characteristics barely affects our estimated coefficient of β in the GMM regression.
We further conduct two sets of robustness checks for our GMM estimation. First,
we include two variables in the regression to control for the two oft-mentioned chan-
nels through which our instrumental variables may affect the outcome variable rather
than through market competition: Custom-made components (a dummy variable indicating
whether or not the firm’s two major components are uniquely supplied) and Client dura-
tion dummies (dummies of the average duration of the business relationship with clients
in the main business line: less than one year, one to two years, two to three years, three to
four years, and more than four years).
Second, we conduct two falsification tests. The first is based on a unique business
feature of China. Note that, in the SCE data, some firms are engaged in business with
the government (i.e., 301 of the 1,509 firms). In China, a firm’s engagement in business
with the government does not depend on its economic strength but rather on personal
connections (guanxi). Hence, market competition should not change the composition of
production for firms producing for the government. Following this argument, we divide
the entire sample according to whether or not a firm has business with the government,
and we check whether the estimate of β for the former subsample is smaller or even
insignificant.
Our second falsification test follows the spirit of Angrist and Pischke (2009): if some
variables are not supposed to be affected by the endogenous variable, then a reduced-
form regression of those variables on the instrumental variables should result in an in-
significant association. The SCE asks, “According to your tax reporting requirements do
you have to use a cash register or other electronic devices?” It is highly unlikely that
the degree of market competition in a given industry could influence the Chinese tax au-
thority to decide which tax reporting devices to use. Hence, conditional on the controls,
regressing the choice of cash register/electronic devices on our instrumental variables
16
should generate insignificant estimates of the instruments.
Self-selection issue. Even if we obtain a consistent and unbiased estimate of β, it is still
possible that the positive impact of market competition on customization simply reflects
the sorting of firms across locations, i.e., firms with a higher degree of customization
locate in more competitive areas.
Given the cross-sectional nature of our data, it is difficult for us to rule out this “dy-
namic” concern completely. By further exploring the data, however, we can compare
firms that recently moved to the surveyed city with those that have been there for a long
time. If the estimated coefficients of β are similar across these two samples, then self-
selection is unlikely to be a major concern in our analysis. One question in the SCE asks
whether the firm recently relocated from another city. However, as only a few firms an-
swered in the affirmative, it is not sensible to divide the full sample into two based on
firms’ answer to this question and to compare the two estimated coefficients of β. In-
stead, we compare the estimated coefficient of β for the subset of firms answering “no”
to this question with that for the whole sample.
3.3 Empirical results
3.3.1 Main results
Table 2 presents the OLS regression results. As Column 1 shows, Local Competition has a
positive and statistically significant association with the degree of customization, which
is consistent with our theoretical prediction. To gauge the economic significance of this
result, we calculate that a one standard deviation increase in Local Competition is asso-
ciated with an 0.112 × 0.415 = 4.65% increase in the percentage of custom-made prod-
ucts/services, or 11.45% relative to the mean of Custom-made.
To investigate whether the estimation is biased due to any omitted variables, we in-
clude industry and city dummies in Column 2 and industry-city dummies in Column 3.
17
Table 2: Customization and Competition1 2 3 4 5
Dependent variable Custom-madeLocal Competition 0.098*** 0.101*** 0.115*** 0.115*** 0.112***
[0.034] [0.031] [0.031] [0.033] [0.036]Firm characteristicsFirm Size -0.010 -0.004
[0.010] [0.011]Firm Age 0.018 0.007
[0.018] [0.020]Private Ownership 0.039 0.043
[0.030] [0.033]Labor Productivity 0.005 0.004
[0.009] [0.010]Skilled Labor -0.031 -0.056
[0.234] [0.232]CEO characteristicsCEO Education 0.003
[0.006]CEO Tenure 0.004
[0.003]Deputy CEO Previously 0.049*
[0.029]Government Cadre Previously 0.060
[0.073]Party Member -0.055*
[0.033]Government-appointed 0.024
[0.032]Industry dummies - Yes - - -City dummies - Yes - - -Industry-city dummies - - Yes Yes YesObservations 1459 1459 1459 1246 1191R-squared 0.006 0.088 0.153 0.171 0.180p-value for F-test 0.005 0.000 0.000 0.000 0.000Mean[standard deviation]Custom-made 0.404 0.404 0.404 0.403 0.405
[0.418] [0.418] [0.418] [0.415] [0.415]Local Competition 0.275 0.275 0.275 0.267 0.269
[0.337] [0.337] [0.337] [0.329] [0.330]
White-robust standard errors clustered at the industry-city level are reported in brack-ets. ∗, ∗∗, and ∗ ∗ ∗ represent statistical significance at the 10%, 5%, and 1% level. Aconstant term is included in all regressions, but the results are not reported to savespace.
18
Clearly, the estimation coefficients of Local Competition remain positive and highly signif-
icant. Further, the magnitude of the estimated coefficients barely changes across these
specifications.
In Columns 4 and 5, we further include a list of commonly used firm and CEO char-
acteristics to account for any firm heterogeneity that may cause bias in our estimation.
More specifically, we include firm size and age to control for economies of scale and the
seniority effect; the percentage of private ownership to take care of the crucial differences
between state-owned enterprises (SOEs) and private firms in China; and labor produc-
tivity and the skilled labor ratio to deal with technological differences among firms. For
CEO characteristics, we include three proxies of a CEO’s human capital and another three
for his or her political capital. It is clear that the estimation coefficients are always positive
and statistically significant, and their magnitude is almost the same as the corresponding
values in the regressions without these controls.
3.3.2 GMM estimates
Given that the controls used in Columns 2 through 5 largely account for important dif-
ferences across firms, the concern over potential omitted-variable bias is alleviated. To
address the remaining concern that some unobserved heterogeneity may still bias our
results via correlation with Local Competition, we conduct GMM estimation with the two
instruments proposed in Section 3.2, Local Clients and Local Suppliers.
The GMM estimation results are reported in Table 3. As Panel B of Column 1 shows,
both instruments are positively and statistically significantly correlated with the key ex-
planatory variable (Local Competition). The Kleibergen-Paap rk Lm statistic further con-
firms that the instruments are relevant, and the Kleibergen-Paap Wald rk F statistic rules
out the concern over weak instruments.
With regard to the central issue, after being instrumented, Local Competition still has a
positive and statistically significant association with the degree of customization, which
19
Tabl
e3:
Cus
tom
izat
ion
and
Com
peti
tion
:Ins
trum
enta
lVar
iabl
eEs
tim
atio
n1
23
45
6Es
tim
atio
nG
MM
OLS
Sam
ple
Full
Busi
ness
wit
hgo
vern
men
tN
obu
sine
ssw
ith
gove
rnm
ent
Full
Dep
.var
.:C
usto
m-m
ade
Dep
.var
.:C
ash
Reg
iste
rLo
calC
ompe
titi
on0.
200*
0.24
2*0.
253*
-0.1
540.
275*
[0.1
18]
[0.1
37]
[0.1
42]
[0.2
91]
[0.1
57]
Cus
tom
-mad
eC
ompo
nent
0.11
3[0
.081
]C
lient
Dur
atio
nD
umm
ies
Yes
Firm
Cha
ract
eris
tics
-Ye
sYe
sYe
sYe
sC
EOC
hara
cter
isti
cs-
Yes
Yes
Yes
Yes
Indu
stry
-cit
yD
umm
ies
Yes
Yes
Yes
Yes
Yes
Pane
lB,fi
rsts
tage
Dep
.var
.:Lo
calC
ompe
titi
onLo
calC
lient
s0.
375*
**0.
350*
**0.
365*
**0.
280*
*0.
359*
**-0
.039
[0.0
34]
[0.0
35]
[0.0
39]
[0.1
13]
[0.0
44]
[0.0
31]
Loca
lSup
plie
rs0.
122*
**0.
106*
**0.
083*
**0.
113
0.11
5***
0.03
3[0
.029
][0
.032
][0
.031
][0
.086
][0
.033
][0
.032
]C
usto
m-m
ade
Com
pone
nt-0
.023
[0.0
43]
Clie
ntD
urat
ion
Dum
mie
sYe
sFi
rmC
hara
cter
isti
cs-
Yes
Yes
Yes
Yes
Yes
CEO
Cha
ract
eris
tics
-Ye
sYe
sYe
sYe
sYe
sIn
dust
ry-c
ity
Dum
mie
sYe
sYe
sYe
sYe
sYe
sYe
sPa
nelC
Var
ious
econ
omet
ric
test
sfo
rfir
stst
age
Kle
iber
gen-
Paap
rkLM
stat
isti
c64
.96*
**50
.87*
**50
.39*
**15
.24*
**38
.04*
**K
leib
erge
n-Pa
apW
ald
rkF
stat
isti
c10
7.49
65.4
764
.35
9.03
47.2
7H
anse
nJs
tati
stic
2.45
41.
368
1.74
30.
102
0.40
8H
ausm
ante
st0.
505
0.87
20.
902
1.38
21.
345
Obs
erva
tion
s14
3711
8211
1224
891
312
41
Whi
te-r
obus
tsta
ndar
der
rors
clus
tere
dat
the
indu
stry
-cit
yle
vela
rere
port
edin
brac
kets
.∗,∗∗,
and∗∗∗
repr
esen
tsta
tist
ical
sign
ifica
nce
atth
e10
%,5
%,a
nd1%
leve
l.A
cons
tant
term
isin
clud
edin
allr
egre
ssio
ns,b
utth
ere
sult
sar
eno
trep
orte
dto
save
spac
e.
20
is consistent with our OLS estimation results. In terms of magnitude, although the coef-
ficient now rises to 0.200, which is larger than the corresponding OLS estimate (0.115 in
Column 3 of Table 2), the Hausman test (reported in Panel C of Table 3) shows that the
GMM estimated coefficient is not statistically different from the OLS estimated coefficient.
To further verify the validity of our GMM estimation with regard to the exclusion
restriction, we conduct the following four sets of tests.
First, with two instruments for our endogenous variable, we report the over-identification
test by the Hansen J statistic, for which the default hypothesis is that at least one of the
two instruments is valid. As shown in Panel C of Table 3, in all of our GMM estimations,
the Hansen J statistic is always statistically insignificant, which suggests that at least one
of our instruments is valid.
Second, we include firm and CEO characteristics in Column 2. Although we believe
that the small-sized nature of the firms in our data renders it difficult for an individual
firm’s characteristics to influence the location choice of its clients and suppliers, the ad-
ditional control of firm and CEO characteristics can help us to check the validity of our
argument and improve estimation efficiency. As can be seen in Column 2, the estimation
coefficient of Local Competition remains positive and statistically significant. Although the
magnitude increases from 0.200 to 0.242, much of the increase can be explained by the
reduction in sample size.8
Third, we further control for two other oft-mentioned channels in Column 3. One
possible failing of our GMM estimation is that the clustering of suppliers, which is likely
to be correlated with the clustering of firms, might positively affect firms’ incentives to
produce customized goods, because there may be more customized components avail-
able. The SCE contains a question asking each firm whether its two major components
are uniquely supplied to it, which allows us to construct a control variable to address
this concern. Another possibility is that the number of years that a firm has done busi-
8The estimation coefficient for the same sample in Column 2, but without firm and CEO characteristics,is 0.231.
21
ness with its clients may affect both its location and customization strategies. From the
SCE data, we construct five dummies to account for five different business durations with
clients (i.e., less than one year, one to two years, two to three years, three to four years,
and more than four years). As shown in Column 3, controlling these two additional sets
of variables has virtually no impact on our estimation results.
Finally, we conduct two falsification tests. First, as discussed in Section 3.2, we expect
the subsample of firms that have some business dealings with the government to have a
smaller or even insignificant coefficient of β. The GMM estimation results are reported
in Columns 4-5 of Table 3.9 Consistent with our intuition, the estimated coefficient of
Local Competition for the subsample of firms with government business dealings loses its
statistical significance and becomes negative, whereas it remains positive and statistically
significant for those without any such dealings. Second, as market competition is un-
likely to influence the Chinese tax authority’s decision over which tax reporting device to
employ, the regression of the use of a cash register/electronic devices on the instrumental
variables is expected to generate insignificant estimates. Indeed, we find in Column 6 of
Table 3 that our two instrumental variables have no significant association with the use
of a cash register/electronic devices.
Overall, the results in Tables 2-3 confirm our theory that competition increases the
degree of customization. These findings are unlikely to be driven by omitted variables,
measurement error, and/or reverse causality.
3.3.3 Robustness checks
In this subsection, we provide a further series of robustness checks. For ease of com-
parison, we copy the corresponding OLS regression results from Column 5 of Table 2 to
Column 1 of Table 4.
Tobit regression. Our dependent variable varies from 0 to 1, which may raise concern
9The OLS estimation results, which are available upon request, exhibits a similar pattern.
22
Table 4: Customization and Competition, Robustness Checks1 2 3 4 5 6
Estimation OLS Tobit Robust OLS OLS GLS OLSDependent variable Custom-madeSample Full Full Full Local firmsLocal Competition 0.112*** 0.172*** 0.126*** 0.353*** 0.113***
[0.036] [0.002] [0.043] [0.109] [0.037]Log Total Employment 0.058***
[0.015]Firm Characteristics Yes Yes Yes Yes Yes YesCEO Characteristics Yes Yes Yes Yes Yes YesIndustry-city Dummies Yes Yes Yes Yes Yes YesObservations 1191 1191 1190 1227 1191 1167Pseudo R2/R-squared 0.180 0.096 - 0.1067 - 0.181p-value for F-test 0.000 0.000 0.000 0.000 - 0.000
White-robust standard errors clustered at the industry-city level are reported in brackets. ∗, ∗∗,and ∗∗∗ represent statistical significance at the 10%, 5%, and 1% level. A constant term is includedin all regressions, but the results are not reported to save space.
about the validity of OLS regression for this censored data setting. As a robustness check,
we thus employ Tobit regression with a left-side truncation at 0 and a right-side trunca-
tion at 1. The Tobit regression results are reported in Column 2 of Table 4. The estimated
coefficient of Local Competition remains positive and statistically significant, and the esti-
mation becomes more precise.
Outliers. There may also be concern that our estimation results are driven by or biased
due to some outlying observations. To address this concern, we re-estimate our results
using robust OLS regression, which first performs an initial screening based on Cook’s
distance > 1 to eliminate gross outliers and then performs Huber iterations followed by
biweight iterations (Li, 1985). As shown in Column 3 of Table 4, only one observation is
dropped in this robust OLS regression, and our main findings remain robust, thus imply-
ing that outliers are not a major concern in our analysis.
Alternative measure of market competition. In the analysis thus far, we have focused on
a subjective measure of market competition, that is, the perceived percentage of a firm’s
competitors (in terms of output) located in the same city. For a robustness check, we
employ an objective measure used in the literature, that is, the total amount of employ-
ment in the same industry and same city. As can be seen in Column 4 of Table 4, the
alternative, objective measure of market competition still has a positive and statistically
23
significant estimated coefficient.
Self-selection? Finally, it could be argued that the positive coefficient may reflect the
self-selection by firms with more customized goods into more competitive areas rather
than the impact of competition on customization. To alleviate this concern, we re-run our
estimation on a sub-sample that excludes firms that recently relocated from another city
to the survey city. As shown in Column 6 of Table 4, the estimated coefficient of Local
Competition remains positive and statistically significant, and its magnitude is almost the
same as that of the estimated coefficient for the full sample. This result implies that self-
selection is not the main explanation for our findings.
4 Conclusion
As empirical measures of customization are rare, this paper’s main contribution is its an-
swer to the question of whether competition leads to customization based on a unique
data set from a World Bank survey on Chinese firms. Before carrying out a number of
tests, we use a simple Hotelling/Salop model with customization activities to formal-
ize the intuition that competition leads to a larger fraction of sales from custom-made
products, which is precisely our measure of customization. We find this prediction to
withstand the tests, as our results are robust to a variety of empirical specifications.
References
[1] Aghion, Philippe, Nick Bloom, Richard Blundell, Rachel Griffith, and Peter
Howitt. 2005. “Competition and Innovation: An Inverted-U Relationship,” Quarterly
Journal of Economics, 120: 701-728
[2] Angrist, Joshua, and Jorn-Steffen Pischke. 2009. Mostly Harmless Econometrics.
Princeton, NJ: Princeton University Press.
24
[3] Bloom, Nicholas, Raffaella Sadun, and John Van Reenen. 2010. “Does Product
Market Competition Lead Firms to Decentralize?” American Economic Review Papers
and Proceedings, 100: 434-438.
[4] d’Aspremont, Claude, Jean J. Gabszewicz, and Jacques-Francois Thisse. 1979. “On
Hotelling’s ‘Stability in Competition’,” Econometrica, 47(5): 1145-50.
[5] Feenstra, Robert C. 2010. Offshoring in the Global Economy. Cambridge, MA: MIT
Press.
[6] Guadalupe, Maria, and Julie Wulf. 2010. “The Flattening Firm and Product Market
Competition: The Effect of Trade Liberalization on Corporate Hierarchies,” American
Economic Journal: Applied Economics, 2(4) : 105-127.
[7] Hanson, Gordon H., Raymond J. Mataloni, and Matthew Slaughter. 2005. “Verti-
cal Production Networks in Multinational Firms,” Review of Economics and Statistics,
87(4): 664-678.
[8] Henderson, J. Vernon. 2003. “Marshall’s Scale Economies,” Journal of Urban Eco-
nomics, 53: 1-28.
[9] Holmes, Thomas J., and John J. Stevens. 2002. “Geographic Concentration and Es-
tablishment Scale,” Review of Economics and Statistics, 84(4): 682-690.
[10] Holmes, Thomas J., and John J. Stevens. 2010. “An Alternative Theory of the Size
Distribution with an Application to Trade,” Working paper, University of Minnesota.
[11] Hotelling, Harold. 1929. “Stability in Competition,” Economic Journal, 39(153): 41-57.
[12] Krugman, Paul. 1991. “Increasing Returns and Economic Geography,” Journal of Po-
litical Economy, 99(3): 483-499.
[13] Krugman, Paul, and Anthony Venables. 1995. “Globalization and the Inequality of
Nations,” Quarterly Journal of Economics, 110(4): 857-880.
25
[14] Li, Guoying. 1985. “Robust Regression,” in Exploring Data Tables, Trends, and Shapes,
eds. D. C. Hoaglin, F. Mosteller, and J. W. Tukey, 281-340. New York: Wiley.
[15] Loginova, Oksana. 2010. “Brand Familiarity and Product Knowledge in Customiza-
tion.” International Journal of Economic Theory, 6(3): 297-309.
[16] Loginova, Oksana, and X. Henry Wang. 2011. “Customization with Vertically Dif-
ferentiated Products,” Journal of Economics and Management Strategy, 20(2): 475-515.
[17] Mazzeo, Michael J. 2002. “Product Choice and Oligopoly Market Structure,” RAND
Journal of Economics, 33(2): 221-242.
[18] Picone, Gabriel A., David B. Ridley, and Paul A. Zandbergen. 2009. “Distance
Decreases with Differentiation: Strategic Agglomeration by Retailers,” International
Journal of Industrial Organization, 27(3): 463-473.
[19] Salop, Steven C. 1979. “Monopolistic Competition with Outside Goods,” Bell Journal
of Economics, 10(1): 141-56.
[20] Shaked, Avner, and John Sutton. 1982. “Relaxing Price Competition through Prod-
uct Differentiation,” Review of Economic Studies, 49(1): 3-13.
[21] Syverson, Chad. 2004. “Market Structure and Productivity: A Concrete Example,”
Journal of Political Economy, 112(6): 1181-1222.
[22] Thesmar, David, and Mathias Thoenig. 2000. “Creative Destruction and Firm Or-
ganization Choice,” Quarterly Journal of Economics, 115(4): 1201-1237.
[23] Venables, Anthony. 1996. “Equilibrium Locations of Vertically Linked Industries,”
International Economic Review, 37(2): 341-359.
[24] Vogel, Jonathan. 2008. “Spatial Competition with Heterogeneous Firms,” Journal of
Political Economy, 116(3): 423-466.
26