Competition and Innovation:
A Dynamic Analysis of the US Automobile Industry∗†
Aamir Rafique Hashmi‡
and
Johannes Van Biesebroeck§
May 22, 2006
Abstract
We study the relationship between competition and innovation in the US automobile
industry using a dynamic stochastic industry model. We model R&D investment as a strategic decision and estimate the parameters of the model using recently developed techniques
in estimation of dynamic stochastic games.
Key words: Competition and Innovation; Automobile Industry; Dynamic Games
JEL Classification Codes: C73; L13; L62; O31
∗
This draft is preliminary and incomplete, please do not cite. Your comments are very welcome.
†
To be presented at the 40th meetings of the Canadian Economics Association to be held at the Concordia
University, Montreal (Quebec, Canada) from 25th to 28th of May 2006.
‡
Departments of Economics, National University of Singapore and University of Toronto.
§
Department of Economics, University of Toronto.
1
1
Introduction
The US automobile industry has undergone quite significant changes over the last few decades.
Perhaps the most significant of these changes has been an increase in the intensity of competition in the industry. Since 1980s when the Japanese car makers started setting up assembly
plants in the US, the competitive environment in the US automobile industry has changed
quite remarkably with far reaching consequences. Increased competition has reduced overall
profit margins and the market shares of the big-three US car manufacturers (GM, Ford and
Chrysler) have also decreased. In the meantime, the market shares of Toyota and Honda have
steadily increased.
In this paper we investigate how this significant change in competition has affected the
innovative activity in the industry. Apparently, since most of the firms in the industry have been
increasing their R&D expenditures during the 80s and most of the 90s, one might be tempted
to conclude that there is a simple positive relationship between competition and innovation in
the industry. However, this would be a simplistic conclusion. For example, both GM and Ford
were increasing their R&D rapidly until the mid-1990s. Since then, GM has been spending
less and Ford’s spending on R&D has roughly been flat. However, Honda and Toyota have
persistently been increasing their R&D investment. Why is the innovative activity in these two
sets of firms responding differently to increasing competition and if the relationship between
competition and innovation is not monotonic, how exactly are the two related? Overall, the
innovative activity in the industry was increasing rapidly until the mid-1990s and since then
the increase has been slowing down. What are the factors responsible for this change?
These questions are interesting in their own right but they also have welfare implications.
Since competition in the market can be affected by a policy action and welfare of the consumers
is tied to the quality of the products they consume which in turn depends on the innovative
2
activity, a better understanding of the relationship between competition and innovation will
enable us to assess the welfare implications of competition policy.
We try to understand the varied response of firms in terms of their innovative activity to
the changes in competition in the industry with the help of a dynamic model in which firms
produce differentiated products that differ in quality. The firms invest in R&D to improve
the quality of their products. The market share of a firm depends on the relative quality and
price of its product. The outcome of R&D is uncertain and successful R&D leads to increased
technical knowledge of the firm. Increased knowledge translates stochastically into a better
quality product. The investment in R&D is modeled as a strategic decision: a firm takes the
relative position of its rivals and their possible future actions into account before making its
R&D decision. The structural parameters of the model will be estimated using the data from
the US automobile industry. Our objective is to try to match the moments of key competition
and innovation related variables as observed in the data. Once this is accomplished, one can
use the model to do counter factual experiments. For example, one such experiment could be
to see what would be the optimal investment in R&D if the industry were operated by a social
planner. This experiment could shed light on the question of over-investment in R&D. Another
interesting experiment could be to explore how a monopolist would have done in the face of a
potential threat of entry.
We see this paper as a part of the broad literature on the relationship between competition
and innovation. The earlier literature on the subject has been inconclusive as some studies find
a positive relationship between competition and innovation, some find a negative relationship
and some do not find any significant relationship at all.1 There could be many reasons for
1
See Kamien and Schwartz [1982], Cohen and Levin [1989] and Ahn [2002] for comprehensive survey of the
earlier literature.
3
these mixed results but an important one is the huge differences among industries with respect
to this relationship (see section 4, Cohen and Levin [1989]). To avoid this problem, we focus
on one industry and try to uncover the channels through which competition affects innovation
and vice versa. The choice of the automobile industry for our analysis is motivated by at least
three reasons. First, although the interaction between competition and innovation has been
the subject of active research since Schumpeter [1950], to our knowledge this question has not
been explicitly studied for the automobile industry. Second, the competition in the industry
has increased tremendously over the last three decades. This dramatic change in competition
will be helpful in identifying its effect on innovation. Third, the firms in the industry spend
large amounts of money on R&D and innovation plays a key role in determining the relative
success of a firm in this industry. For example, in the year 2000 the Michigan state, which
accounted for 85% of the nationwide vehicle related R&D, spent $13.5 billion dollars on vehicle
related R&D and employed 65,000 people for the purpose (Hill [2002]). The Michigan state
spent around $180,000 per capita on industrial R&D in 1999 and ranked first among all the
states.2 More than three quarters of this amount was spent by the automotive industry.
On the methodological front, we employ recently developed techniques in estimation of
dynamic games. Our study is one of the first to put these recent techniques to a practical
test, demonstrate their usefulness and highlight some practical difficulties associated with their
application.
We start off by showing some trends in competitive environment and innovative activity in
the industry. Each of the following figures (Figures 1-5) is a scatter plot. An observation in
a scatter-plot is marked with a two-letter abbreviation of the company name followed by two
digits representing the year. For example, GM66 is the observation on General Motors for the
2
Source: State Science and Technology Institute (2002), Westerville, OH.
4
year 1966. The data is an unbalanced panel of the following six firms: Daimler-Chrysler (DC),
Ford (FD), Fiat (FT), General Motors (GM), Honda (HM) and Toyoto (TM). The data are
from Standard and Poor’s online database Compustat and cover the time period 1950-2004.
The variables in the graphs are defined in the data appendix.
Figure 1 plots profit margins of six big automobile firms over time. The figure shows a
general downward trend in the profit margins. Since profit margins are a commonly used
measure of competition in the literature (for example see Nickell [1996] and Aghion et al.
[2005]), we interpret this decline in profit margins as an indication of increase in competition.
In addition to overall negative trend in profit margins, at least two other features of the figure
are notable. First, the profit margins are sensitive to business cycles.3 Especially note the big
decline in GM and Ford profits during the recession of 1980-82. Also note low profit margins
during and after 1990-91 recession and again after the 2001 recession. Second, profit margins
also reflect idiosyncratic firm conditions. For example, during the 90s, Ford made above average
profits while Daimler-Chrysler (which was Chrysler until 1998) and Fiat made below average
profits.
In Figure 2 we plot market shares overtime. Until 1980 the market is dominated by GM
and Ford. During 1980s, GM and Ford lose their market shares mainly to Toyota but also to
Honda. During the 1990s the competition gets very intense and by the turn of the century
GM, Ford, Daimler-Chrysler and Toyota have roughly equal market shares followed by Honda
and Fiat. The most striking feature of figure 2 is the convergence of the market shares of GM,
Ford, Daimler-Chrysler and Toyota.
Figure 3 plots the real R&D figures. We see a sharp increase in R&D during the 1980s
3
The grey areas in the figure mark the years in which at least one quarter was part of the NBER’s official
recessions.
5
led mainly by GM and Ford. This trend continues until the mid 90s. Since then, the R&D
investment by GM and Ford has been on decline and by Honda and Toyota on a steep rise.
Figures 2 and 3 relate in an interesting way. To see this, we plot market shares against R&D
in Figure 4. The resulting figure shows a clear inverted-U shape: for the big firms, market shares
and R&D are negatively related and for the small firms they are positively related. The R&D
is the highest for the firms with 20 to 30% of the market share. This inverted-U relationship
is statistically significant and robust to the inclusion of fixed time and firm effects (see Table
1, last column).
Figures 1 and 3 are also related in a similar way. To see this we plot profit margins against
R&D in Figure 5. The inverted-U pattern is there and statistically significant (see Table
1, column 2) but it is not as clear as in Figure 4. Perhaps this is because of sensitivity of
profit margins to business cycle conditions and also to big idiosyncratic variations in the profit
margins.
The main message of these figures is that there is a clear inverted-U relationship between
market share and R&D. There is also an inverted-U relationship between profit margins and
R&D, though it is not as clear as the one between market share and R&D. Our main task in
the following pages will be to explain these patterns theoretically.
This observation of inverted-U relationship between competition and innovation is not new.
Aghion et al. [2005] find a similar relationship for 17 manufacturing industries in the UK.4
They also offer a theoretical model to explain this relationship. However, our work is different
from theirs in at least three respects. First, the focus in Aghion et al. [2005] is more general.
They argue that at the economy level the relationship between competition and innovation is
4
Hashmi [2005] finds a similar inverted-U for 2-digit and 3-digit US manufacturing industries. However, in
Hashmi [2005] the inverted-U is not robust to the inclusion of industry fixed effects.
6
inverted-U. To derive this relationship they combine the traditional Schumpeterian effect (i.e.
more competition hampers innovation) with a new escape-competition effect (i.e. firms innovate
more when faced with more competition because if they do not and their rivals do, they would
be worse off). They then combine these two effects in a general equilibrium setting and derive
an inverted-U relationship between competition and innovation. Our focus instead is on a
single industry and we apply their idea of the interaction between the Schumpeterian effect
and the escape-competition effect to a more detailed model of the industry. Second, unlike
a static duopoly market structure which is essential to keep the model tractable in Aghion
et al. [2005], we allow for a dynamic market structure with entry and exit and an arbitrarily
large number of firms. This feature enables us to incorporate the dynamics of competition into
our model. Third, in Aghion et al. [2005] competition is exogenous and innovative activity
does not have any effect on it. This is not satisfactory because it is possible that a successful
innovator may increase her profits and/or share of the market and hence reduce the level of
competition in the industry. In fact, theorists in this area have long recognized the need for
modeling competition and innovation in such a way that both are simultaneously determined
in equilibrium (see, for example, Dasgupta and Stiglitz [1980]). In our model, competition and
innovation are simultaneously determined in equilibrium. Nonetheless, our work is motivated
by Aghion et al. [2005] and should be considered an application of their idea to a more detailed
model of the industry.
The rest of the paper is organized as follows. Section 2 presents the model, section 3
discusses the estimation techniques and Section 4 presents the estimation results. Section 5
presents the simulation results of the counterfactual experiments and section 6 concludes.
7
2
The Model
The US automobile industry is a mature oligopoly (Klepper [2002a] and Klepper [2002b]) and
one of the most innovation intensive industries. A natural environment to model this industry
is a dynamic oligopoly model along the lines of Ericson and Pakes [1995].
2.1
The Demand Side
Following Berry et al. [1995] and several others studying the automobile industry, we use a
discrete choice model of individual consumer behavior to model the demand side.5 There are n
firms in the industry producing differentiated products. Products differ in quality (g(ω)) that
is positively related to firm’s technological knowledge (ω) i.e. ∂g(ω)/∂ω ≥ 0. There are m
consumers in the market. The utility consumer j gets from consuming good i is
uij = θ1 g(ωi ) + θ2 log(Y − pi ) + νij , i = 1, . . . , n, j = 1, . . . , m.
(1)
g(ωi ) is the quality of the good produced by firm i, Y is consumer’s income, pi is the
price of the good and νij is the idiosyncratic utility that consumer j gets from good i. θ1 and
θ2 are preference parameters. θ1 shows how quality conscious the consumers are and θ2 is a
measure of their price sensitivity. We assume ν are i.i.d. Gumbel distributed with pdf given by
f (x) = exp(−x − exp(−x)) and cdf given by F (x) = exp(− exp(−x)).6 Consumers maximixe
utility and buy only one product each period. There is an outside good that gives zero utility.
Let ũ(g(ωi ), pi ) = exp(θ1 g(ωi ) + θ2 log(Y − pi )). The expected market share of each firm is
σ(ωi , s, p) =
5
1+
ũ(g(ωi ), pi )
P
,
i ũ(g(ωi ), pi )
(2)
Anderson et al. [1992] is a good reference on discrete choice models. They derive a multinomial logit model
very similar to the one described below (see Anderson et al. [1992] pp.39-40).
6
The assumption of i.i.d. Gumbel (or double exponential) distribution ensures that the choice probabilities
are given by the multinomial logit model (see Anderson et al. [1992] Theorem 2.3 p.40).
8
where s is the vector of technological knowledge of all the firms in the industry and p is
the price vector. We assume that marginal cost (mc) is the same for all firms and firms choose
prices to maximize their profits. The first order condition is
Y − pi + θ2 (pi − mc)(σ(ωi , s, p) − 1) = 0.
The unique Nash equilibrium is a price vector p∗ and expected profits are given by
π(ωi , s) = mσ(ωi , s, p∗ )(p∗i − mc).
2.2
(3)
7
(4)
The Supply Side
Time is discrete. At any time there are n ∈ N firms in the industry. Firms differ in their level
of technological knowledge denoted by ω ∈ R+ . The knowledge (ω) is an observable composite
variable that contains all the information about payoff relevant characteristics of a firm. Firms
compete in a spot market and their profits are given by (4). A firm can improve its knowledge
by investing in R&D. We denote R&D investment by x ∈ R+ . The knowledge of the firm
evolves according to the following equation:
ωit+1 = f (ωit , xit ) + it .
(5)
Each period, a firm receives a private offer to sell off its assets for φit . φit is i.i.d. across firms
and over time and follows a well defined probability distribution.8 If the expected discounted
value of profits conditional upon staying is less than φit , the firm will exit the market after
7
Although inventories play an important role in determining automobile prices (Copeland et al. [2005] and
Zettelmeyer et al. [2006]), we abstract from inventory considerations in our model because we are interested in
how average annual profit margins and R&D expenditures are related. Since most inventory-related fluctuations
occur within the year, their inclusion will complicate the model without yielding any positive benefit.
8
The requirement that φ be drawn randomly from a probability distribution is important for existence of
equilibrium. See Doraszelski and Satterthwaite [2005] for details.
9
receiving a one-time payment of φit . The value function of the firm is given by the following
Bellman equation:
V (ω, s) = max
χ∈{0,1}
χ max {π(ω, s) − cx + βEV (ω 0 , s0 )} + (1 − χ)φ ,
x∈R+
(6)
where χ is the exit policy of the firm. If the firm chooses to stay then χ is one, it is zero
otherwise. π(ω, s) is one period profit function that depends on firm’s own knowledge ω and
the distribution of firms by knowledge s. Amount of R&D is x, c is the unit cost of R&D, β is
the discount factor and φ is the sell-off value. The dynamics of ω are determined by (5). There
is an implicit entry function η(s, ωe , ce ) that is taken into account by the firm while forming
expectations about s0 . We describe this entry function in some detail below.
New firms can enter the industry after paying a fixed cost equal to ce . Entrants are identical
in terms of their knowledge. We denote the common knowledge of entrants by ωe ∈ R+ .
Entrants do not produce or invest in R&D in the period of their entry and from the next period
they act like any other incumbent with knowledge ωe . There is no restriction on number of
entrants (η) but as more firms enter simultaneously, the net expected value of entry declines.
To see this, let us first define the net present value of entry as:
Ve (ωe , ce , η) = βEV (ωe , s0 (η)) − ce .
(7)
If there are more entrants, the expected s0 will involve a greater number of firms, reduce expected market share and profits of the entrants and hence result in a lower value of
EV (ωe , s0 (η)). Hence (∂EV (·)/∂η) ≤ 0. We require Ve to be equal to zero in equilibrium (zero
profit condition) and hence given ce , ωe and s, we can determine η endogenously such that one
of the following conditions is satisfied.
η(ωe , ce , s) > 0
η(·)Ve (·)
and
Ve (ωe , ce , η) < 0
=
0.
10
or
(8)
In each period the sequence of events is the following.
1. Firms observe individual and industry states.
2. Production, investment, entry and exit decisions are made.
3. Profits and investment outcomes are realized and firms enter and exit.
4. Individual and industry states are updated.
2.3
The Markov Perfect Equilibrium
The Markov Perfect Equilibrium consists of V (ω, s), π(ω, s), x(ω, s), χ(ω, s), Q(s0 , s) and
η(s, ωe , ce ) such that:
1. V (ω, s) satisfies (6) and x(ω, s) and χ(ω, s) are optimal policies;
2. π(ω, s) maximizes profits in the spot market;
3. η(s, ωe , ce ) satisfies (8);
4. Q(s0 , s) is the transition matrix that gives the probability of state s0 given that the current
state is s;
3
Estimation of the Model
There are at least two ways to proceed from here. The first is to estimate the parameters of
the model from the data and then compute the Markov Perfect Equilibrium (MPE) as defined
above. Benkard [2004] takes this route in his study of the market for wide-bodied commercial
aircraft. The main disadvantage of this approach is that it is computationally very intensive.
For example, the model in Benkard [2004] “requires over 100 CPU-days to solve on a Sun Ultra
11
400 processor”. Despite some recent innovations by Pakes and McGuire [2001], Doraszelski
and Judd [2004] and Weintraub et al. [2005], the computation remains the biggest hurdle in
the use these models.
The second is a two-step approach proposed by Bajari et al. [2006].9 Their main assumption
is that the data we observe represents an MPE. Under this assumption, in the first step, one
can use the available data to estimate the state transition probabilities, equilibrium policy
functions and the value functions. In the second step, the estimates from the first step are
combined with the equilibrium conditions of the model to estimate the structural parameters
of the model. We now explain these steps in some detail.10
We need data on the state variable and the policy actions for each firm in the sample. In our
study, the policy action is directly observable: it is the R&D investment of the firm. However,
the state variable, which is some measure of technological knowledge of the firm, is not directly
observable. We can construct a proxy for the firm’s knowledge from the available data. One
way to do so is to use the patent data and construct a measure of patent stock. Suppose a firm
has a total of Nt patents upto time period t, where t is the first period in our sample. For this
firm we can set ωt = Nt . Next period, supppose the firm gets ∆Nt+1 (= Nt+1 − Nt ) new patents
and its existing stock of patents depreciates by a fraction δ. Then ωt+1 = (1 − δ)ωt + ∆Nt+1 .
Similarly ωt+2 = (1 − δ)ωt+1 + ∆Nt+2 and ωt+T = (1 − δ)ωt+T −1 + ∆Nt+T , where t + T is the
last period in our sample. This way we can construct a time series of patent stocks for each
firm and use this new variable as a proxy for technological knowledge of the firm.
Once we have data on the state variable and the policy actions we can apply the two-step
procedure suggested by Bajari et al. [2006] to estimate the model. In the first stage, we shall
9
Alternative approaches that avoid computing the MPE at each interation include Aguirregabiria and Mira
[2006] and Pakes et al. [2006].
10
The details that follow are sketchy and incomplete. We are still working on them.
12
estimate the transition probabilities and policy functions and use them to obtain simulated
value functions. In the second stage, we shall use the results from the first stage together with
the equilibrium conditions to recover structural parameters of the model. We note that the
demand side in our model is static and hence we do not need the full model to estimate its
parameters. This can be done using standard techniques for such static models as in Berry
et al. [1995].
Since our state variable is continuous, we estimate a state transition function of the form:
ωt+1 = α0 + α1 ωt + α2 xt + t .
(9)
The fixed time- and firm-effects can be added to the above function. To estimate investment
and exit policies we follow Bajari et al. [2006] and use local linear regressions with a normal
kernel.
The next step is to use forward simulations and get a numerical estimate of the value
function given the structural parameter vector θ. To do so, we start from an initial state s0 .
Given s0 , using our demand parameter estimates we find profit and denote it by π0 (s0 ). Given
the estimated policy functions we determine the investment and exit policies and denote them
by x0 (s0 ) and χ0 (s0 ). Next we use our state-transition function to find the state next period
s1 . Using the new state, we can repeat the above steps until either the firm exits or β t becomes
very small. In other words, we evaluate the following equation:
V (s0 |θ) = E
"
e
T
X
χt (st )β t [(πt (st ) − xt (st )] + (1 − χt (st ))β
t=0
Te
#
φT e ,
(10)
where the expectation is over future states and T e is the time period when the firm exits.
If the firm never exits we can stop the forward simulations when t is sufficiently large such that
β t has become very small. We run these forward simulations a large number of times and take
their average as a numerical estimate of V (s0 |θ).
13
In stage two, we use the value function estimates and recover the vector of the structural
parameters of the model θ. The following steps assume that the model is identified and there is
a unique true parameter vector θ0 . Bajari et al. [2006] propose a minimum distance estimator
for this true parameter vector. Let x∗ |s be the equilibrium policy profile. For this to be an
MPE policy profile, it must be true that for all firms, all states and all alternative policy profiles
x|s
Vi (s, x∗ |s, θ) ≥ Vi (s, x|s, θ),
(11)
where x 6= x∗ , provided that we use true value of the parameter vector θ0 . The minimum
distance estimator for θ0 is constructed as follows. For each firm i and initial state s we use
forward simulation method to estimate Vi (s, x∗ |s, θ) and then do the same using alternative
policy profile x|s and compute the difference Vi (s, x∗ |s, θ) − Vi (s, x|s, θ). Let us denote this
difference by d((i, s, x|s)|θ). We can then find d for all i, s and x|s for a given value of θ and
compute the sum of (min{d(i, s, (x|s)|θ), 0})2 . The θ with the smallest sum is our estimate of
θ0 .
14
References
P. Aghion, N. Bloom, R. Blundell, R. Griffith, and P. Howitt. Competition and Innovation:
An Inverted-U Relationship. Quarterly Journal of Economics, 120(2):701–728, 2005.
V. Aguirregabiria and P. Mira. Sequential Estimation of Dynamic Discrete Games. Econometrica, forthcoming, 2006.
S. Ahn. Competition, Innovation and Productivity Growth: A Review of Theory and Evidence.
OECD, Economics Department Working Papers No.317, 2002.
S. P. Anderson, A. de Palma, and J.-F. Thisse. Discrete Choice Theory of Product Differentiation. The MIT Press, Cambridge, MA, 1992.
P. Bajari, C. L. Benkard, and J. Levin. Estimating Dynamic Models of Imperfect Competition.
Econometrica, forthcoming, 2006.
C. L. Benkard. A Dynamic Analysis of the Market for Wide-Bodied Commercial Aircraft.
Review of Economic Studies, 71(3):581–611, 2004.
S. Berry, J. Levinsohn, and A. Pakes. Automobile Prices in Market Equilibrium. Econometrica,
63(4):841–890, 1995.
W. M. Cohen and R. C. Levin. Empirical Studies of Innovation and Market Structure. In
R. Schmalensee and R. D. Willig, editors, Handbook of Industrial Organization, volume 2.
Elsevier Science, Amesterdam: North-Holland, 1989.
A. Copeland, W. Dunn, and G. Hall. Prices, Production and Inventories over the Automotive
Model Year. Working paper, Yale University, 2005.
15
P. Dasgupta and J. Stiglitz. Industrial Structure and the Nature of Innovative Activity. Economic Journal, 90(358):266–293, 1980.
U. Doraszelski and K. L. Judd. Avoiding the Curse of Dimensionality in Dynamic Stochastic
Games. Working paper, Harvard University, 2004.
U. Doraszelski and M. Satterthwaite. Foundations of Markov-Perfect Industry Dynamics: Existence, Purification and Multiplicity. Working paper, Harvard University, 2005.
R. Ericson and A. Pakes. Markov-Perfect Industry Dynamics: A Framework for Empirical
Work. Review of Economic Studies, 62(1):53–82, 1995.
A. R. Hashmi. Competition and Innovation: The Inverted-U Relationship Revisited. Working
paper, University of Toronto, 2005.
K. Hill. Michigan Automotive Research and Development Directory. Michigan Economic Development Corporation, Lansing, MI, 2002.
M. I. Kamien and N. L. Schwartz. Market Structure and Innovation. Cambridge University
Press, Cambridge, UK, 1982.
S. Klepper. The Capabilities of new Firms and the Evolution of the us Automobile Industry.
Industrial and Corporate Change, 11(4):645–666, 2002a.
S. Klepper. Firm Survival and the Evolution of Oligopoly. RAND Journal of Economics, 33
(1):37–61, 2002b.
S. J. Nickell. Competition and Corporate Performance. Journal of Political Economy, 104(4):
724–746, 1996.
16
A. Pakes and P. McGuire. Stochastic Algorithms, Symmetric Markov-perfect Equilibrium and
the Curse of Dimensionality. Econometrica, 69(5):1261–1281, 2001.
A. Pakes, M. Ostrovsky, and S. Berry. Simple Estimators for the Parameters of Discrete
Dynamic Games (with entry/exit examples). Review of Economic Studies, forthcoming,
2006.
J. A. Schumpeter. Capitalism, Socialism and Democracy. Harper, New York, 3 edition, 1950.
G. Y. Weintraub, C. L. Benkard, and B. V. Roy. Markov Perfect Industry Dynamics with
Many Firms. Working paper, Stanford University, 2005.
F. Zettelmeyer, F. S. Morton, and J. Silva-Risso. Scarcity Rents in Car Retailing: Evidence
from Inventory Fluctuations at Dealerships. NBER Working Paper No.12177, 2006.
17
A
The Data Appendix
This appendix defines the variables used in Figures 1-5. The profit margin is defined as:
Profit Margin =
Income after depreciation + R&D Expenditure − 0.085 × Capital
.
Sales
The subscript i refers to the firm and t to the year. The data on Income after depreciation,
R&D expenditure, capital and sales are from Standard & Poor’s Compustat database (from here
on, Compustat) annual series numbers DN0178, DN0046, DN0008 and DN0012, respectively.
The market share (MS) is defined as:
Salesit
MSit = Pn
.
j=1 Salesjt
If the information on a firm’s sales is not available for a given year, sales for that firm-year
are set equal to zero in the sum. The R&D is simply the R&D expenditures as defined above.
18
Table 1: Regression Results (The Inverted-U)
Dependent Variable: R&D
Independent Variable (x):
Profit Margins
Market Share
Coefficient of x
11048.32∗∗∗
(3178.34)
9360.71∗∗∗
(2707.33)
Coefficient of x2
−47227.40∗∗
(19474.78)
−11844.07∗∗∗
(3623.62)
Firm effects
Yes
Yes
Year effects
Yes
Yes
Adj-R2
0.89
0.86
F
27.54∗∗∗
21.08∗∗∗
19
0.15
GM53
GM54
GM60
GM56 GM59
FD59
GM61
GM57
GM66
GM68
GM67
GM69 GM72
GM71
GM73
FD60
FD62
FD61FD63
FD65
GM58
FD64
0.1
FD57
FD56
FD97
FD94
GM77
HM84
GM76
TM85
GM78HM80
HM85
HM79
HM83
FD98
FT99
FD95
FD00
FD96 FD99
FT89
FD90
FD93
GM89
GM88
HM98 HM01
HM02
TM84
FD87
HM72
HM82
HM81
HM73
TM03
TM86
HM97
FD72
HM96
FT85 FT88
HM03
FD77
TM83 FT86
GM79
FT87 TM90FD92
HM99
FD71
HM71
GM83
HM76
HM70
FD69
GM99 TM02
TM88
FD73GM75
FD84
GM95
FD86
HM00
TM82
TM01
FD78
TM89FD91 GM94
FD70
GM00
FD04
HM77
TM87 GM90
HM86
FD76
GM84
GM97
FT90
FD02
HM74
TM91
TM97
FD85
TM00
HM87
GM74
GM04
HM88
HM89
FD03
GM93 TM96TM98
HM69
TM99
DC98
GM03
FD83GM85
GM98
DC99
GM96
HM75 HM78
HM95
HM91
HM90
GM01
FT91 HM94 FT97
GM02
GM70
HM92 TM95
FD74
FD01
FD79
DC88 GM91
GM92
GM86
TM92
TM94
HM93
GM87
FD75
TM93
DC97 FT00
FT92
FT96
DC02DC04
DC89
FT01DC03
GM82
FD67
DC00
GM81
FT95
DC96
FT03
DC91
DC90
FT94
FT02
FD82
FT93
DC94
DC01
GM80
FD66
FD68
0.05
20
Profit Margins
FD88
FD89
FD58
0
FD54
−0.05
DC92
FD81
DC95
FD80
−0.1
1955
1960
1965
1970
1975
1980
DC93
1985
Years
Figure 1: Profit margins over time
1990
1995
2000
1.1
GM51
GM52
GM53
1 GM50
0.9
0.8
21
Market Shares
0.7
0.6
GM60
GM56
GM58
GM55
GM59
GM54
GM67
GM57
GM63
GM62
GM65
GM61 GM64
GM66
GM68
GM71
GM69
GM73 GM76
GM72
GM75 GM78
GM79
GM81
GM74 GM77 GM80
GM70
0.5
GM82
GM83
GM84
GM85
FD70
0.4
0.3
0.2
0.1
1950
GM86
FD74
FD77
FD72
FD75
FD78
FD68
FD73
FD79
FD61 FD64FD66 FD69FD71
FD76
GM87
FD80
FD62
FD81
FD63FD65FD67
DC04
FD57
GM04
FD54
GM88
FD59
GM89
FD55
FD04
FD84
FD82
FD56FD58FD60
GM95
FD83
GM97
GM90
GM93
GM94
GM92
GM96
GM91
FD97
FD96
FD87
FD86
GM01
FD85
GM00
FD88
FD89
FD94
GM99
GM02
GM98
FD00
FD01
FD93FD95 DC98
FD90FD92
FD99
GM03
FD98
FD02
DC00
DC99 DC02
DC03
FD91TM93
FD03
DC01TM03
TM92TM94
TM95
TM96 TM99 TM02
TM91
TM86
TM87
TM98
TM01
TM97
DC91
TM00
TM83
TM88
TM84
TM89
TM90
DC90DC92DC94
DC95DC97
TM82 TM85FT87
DC89
FT90
DC88
FT91DC93 DC96
FT89
HM87
HM03
FT96
FT88
FT97
FT92
FT86
HM02
HM94
FT95
HM93
HM99
HM01
FT98
FT02
HM92
FT00
HM97
FT85
HM80
FT01
HM98
HM85
FT03
HM91
HM82
FT94
HM86
HM96
HM81
HM00
HM90
HM88
HM95
FT99
FT93
HM83
HM89
HM84
HM78
HM79
HM77
HM76
HM75
HM74
HM70
HM73
HM71
HM72
HM69
1955
1960
1965
1970
1975
1980
1985
Years
Figure 2: Market shares over time
1990
1995
2000
6000
GM96
GM95
GM97
5000
GM94
FD96
GM91
FD02
FD95
FD99
GM89
DC90 GM92
FD01
GM93
GM90
GM99
FD03
DC04
GM88
FD00
FD97
FD04
FD98
GM87
GM00
DC03
GM86
4000
R&D
GM98
FD94
FD93
GM85
3000
DC98
DC02
TM03
GM01
DC99
DC00
GM04
GM02
GM03
TM02
DC01
FD92
GM84
22
GM73
GM74
FD91
FD90
GM80
GM79
DC89
GM83
FD89
TM99 TM01
GM78
TM98
FD88
GM81
DC95
GM77 FD79
DC91
HM03
DC88
GM82
DC94
TM00
FD87
GM76 FD78
DC92
FD86
DC96
GM75
HM02
FD80
DC93
DC97
FD77
FD81
HM99
FD85
FD73
FD84
FD82
FD83
HM01
FD74 FD76
GM70
FT91
HM98 HM00
HM94
FT90
FD72
FD75
HM95
HM97
HM96
HM93
FD71
FT92
HM87
HM92
FT03
HM88
FT89
FD70
HM91
FT87
FT88 HM90
FT02
HM89
FT00
FT96 FT98
FT01
FT93 FT95
HM86
FT94
FT97 FT99
HM85
FT86
FT85
HM84
HM83
HM82
HM81
HM80
HM78
HM79
HM77
HM76
HM74
HM75
HM73
GM72
GM71
2000
1000
1970
1975
1980
1985
1990
Years
Figure 3: R&D over time
1995
2000
6000
GM96
GM95
GM97
5000
GM98
DC90
4000
GM94
FD96
FD02
FD95 GM91
FD99
GM92
FD01
GM93GM89
GM90 GM88
FD03 GM99
DC04
FD00 FD97 FD04
FD98
GM87
DC03 GM00
R&D
DC02DC98
FD94
TM03
GM01
FD93
DC99
DC00
GM86
GM04
GM85
GM02
GM03
TM02
FD92
DC01
3000
GM84
23
FD91FD90
DC89
FD89
TM99
TM98
FD88
DC95TM01
DC91
HM03
DC88
DC94
TM00
FD87
DC92
FD86
DC96
HM02
2000
DC93
DC97
HM99
FD85
FD84
FD82
FD83
HM01
HM00
FT91
HM98
HM94 FT90
HM95
HM97
HM96
HM93
FT92
HM87
HM92
FT03
HM88
FT89
HM91
HM90
FT88 FT87
FT02
1000
HM89
FT00
FT96
FT01
FT93
FT98
FT95
HM86
FT99
FT94
FT97
HM85
FT86
FT85
HM84
HM83
HM82
HM81
HM80
HM78
HM79
HM77
HM76
HM74
HM75
HM73
0.1
0.2
GM83
FD79
FD78
FD80
FD77
FD81 FD73
FD74
FD76
FD72
FD75
FD71
FD70
0.3
0.4
Market Share
Figure 4: Market share and R&D
GM82
GM74 GM73
GM80
GM79
GM72GM71
GM78
GM81
GM77
GM76
GM75
GM70
0.5
0.6
6000
GM96
GM95
GM97
5000
GM98
GM94
GM91
FD02
GM92
FD01 DC90GM93
GM90
GM99
FD03
DC04
FD04
GM00
DC03 GM87
GM86
4000
GM89
GM88
FD96
FD99 FD95
FD00
FD98
DC98
FD94
TM03
GM01 DC99
FD93
GM04
GM85
GM02
GM03
TM02
FD92
DC01
GM84
GM73
GM74 FD91
FD90
GM80
GM79
DC89
GM83
TM99
GM78
TM01
TM98
GM71
GM81
FD79
DC91
HM03GM77
DC88
DC94 GM82
TM00 FD78
FD87
GM76
DC92
FD86
DC96
GM75
HM02
DC97
FD77
FD81
HM99
FD85
FD73
FD84
FD82
FD83
HM01
FD74
GM70
HM00
FD76
FT91
HM98
HM94
FT90
FD72
FD75
HM95
HM97
HM96
HM93
FD71
FT92
HM92
FT03
HM88
FT89
FD70HM87
HM91
HM90
FT87FT88
FT02
HM89
FT00
FT96
FT01
FT93
FT95
HM86
FT99
FT94
FT97
HM85
FT86
R&D
DC02
DC00
3000
24
DC95
2000
FD80
DC93
1000
FT85
HM78
HM75
−0.04
−0.02
0
0.02
0.04
HM77 HM76
HM74
0.06
0.08
Profit Margin
Figure 5: Profit margins and R&D
HM82 HM83
HM81
HM80
HM79
HM73
0.1
0.12
HM84
0.14
0.16