View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

Idea Diffusion and Property Rights

WP 20-11

Boyan Jovanovic
New York University
Zhu Wang
Federal Reserve Bank of Richmond

Idea Diffusion and Property Rights∗
Boyan Jovanovic†and Zhu Wang‡
August 29, 2020

Abstract
We study the innovation and diffusion of technology at the industry level.
We derive the full dynamic paths of an industry’s evolution, from birth to
its maturity, and we characterize the impact of diffusion on the incentive to
innovate. The model implies that protection of innovators should be only partial due to the congestion externality in meetings in which idea transfers take
place. We fit the model to the early experiences of the automobile and personal
computer industries both of which show an -shaped growth of the number of
firms.

1

Introduction

Innovation and diffusion are two fundamental drivers for technological progress and
long-run growth. On the one hand, a technological innovation cannot fulfill its potential without being widely adopted, but on the other hand, rapid diffusion may
reduce the incentive to innovate. In this paper, we study the interdependence between innovation and diffusion in an industry setting, and discuss welfare and policy
implications.
The model features an industry in which there is a fixed downward sloping demand
curve for a homogeneous product. Production requires the use of an innovation
referred to as “the idea.” At the outset, a group of homogeneous measure-zero agents
— “firms” — decide whether to pay a sunk cost and innovate immediately or to wait
and imitate later. The idea enables a firm to produce a given quantity at zero cost.
As more imitators arrive and get the idea, price gradually falls and so does the value
of obtaining the idea.
∗

We thank Maikol Cerda and Xi Xiong for helping us with this research.
Department of Economics, New York University. Email: bj2@nyu.edu.
‡
Research Department, Federal Reserve Bank of Richmond. Email: zhu.wang@rich.frb.org. The
views expressed are solely those of the authors and do not necessarily reflect the views of the Federal
Reserve Bank of Richmond or the Federal Reserve System.
†

1

Imitation occurs as ideas are copied in random pairwise meetings between those
that have the idea already and those that do not. Imitation imposes no direct costs,
but the imitator may have to pay a fee to the idea seller. A bargaining parameter determines the share of the idea’s value that its seller receives and in shaping incentives
it plays a role similar to an entry cost.
We study two regimes affecting the payment for ideas, in contrast to the benchmark where ideas are copied for free. Under regime 1, imitators cannot resell ideas
to other imitators. A potential adopter can copy an imitator but the fee goes to an
original innovator and not to the imitator — this is a scenario often seen in the world
of patent licensing or franchising. Under regime 2, by contrast, imitators can resell
ideas to other imitators and keep the proceeds, arguably leading to a faster diffusion
rate. This is a scenario that sometimes arises in the cases of technology transfer or
employee spin-offs. If bargaining allocates all rents to idea sellers, then under both
regimes the initial innovators capture all the discounted rents from their innovation:
Under regime 1 they collect all the remaining discounted rents directly from each and
every imitator, and under regime 2 they capture those rents also indirectly by pricing
in the resale values in their direct sales.
Our analysis yields several key findings. First, our model characterizes the growth
path of a new industry and the impact of diffusion on the incentive to innovate. The
results are shown to depend on the regimes (i.e., on whether imitators can resell the
innovation) and on how much the innovators are compensated for their ideas (through
their bargaining share).
Second, we find that the socially optimal bargain allocation depends on the diffusion process, particularly the meeting rate between idea holders and potential
adopters. In a setting where the meeting rate is fixed and as a result the growth
in the number of producers is constant — which we refer to as a small industry model
— it is socially optimal to allocate all rents to idea sellers. However, in an alternative
setting where the meeting rate is endogenous and leads to a logistic diffusion process
— which we refer to as a large industry model — maximal compensation for ideas
would lead to too many innovators and too few imitators. This is because of the
congestion externality in meetings between innovators and imitators that innovators
ignore, so that the social planner would prefer less innovation and more imitation.
The distinction between the small industry model and the large industry model lies
in the number of potential adopters, and we show the former is the limiting case of
the latter as the number of potential adopters gets large.
Our analysis also sheds light on the debate about policy interventions on diffusion.
Our findings suggest that a policy reducing the speed of diffusion encourages innovation and raises initial capacity, but that it lowers imitation and leads to a slower
growth of capacity. We argue that different takes on such a policy may explain the
geographic pattern of industry development (e.g., Route 128 being overtaken by the
Silicon Valley), and accommodating diffusion can raise welfare.
In the large industry model we find that initial innovators need less than full
2

protection of the revenues that their innovations generate. If the patent system
guarantees that the rents that an innovation generates should all go to the original
innovators, then that protection should be less than perfect. Our analysis is at
the industry level, but a similar result is also found in some recent papers relating
innovation and its diffusion to aggregate growth. For example, Benhabib, Perla and
Tonetti (2019) show that the licensing income received by innovators becomes highly
elastic with regard to the license price when innovators’ bargaining power is too
strong. This may lead to less licensing income which lowers the return to innovation
and the aggregate growth rate. Hopenhayn and Shi (2020) add matching congestion
to the analysis and show that the growth-maximizing bargaining share of innovators
is also sensitive to the parameters in the matching function which they assume has
constant returns.
Our analysis complements this line of macro-diffusion research in several aspects:
First, unlike the existing literature focusing on the steady-state growth rate of the
economy, we solve for full dynamic paths of an industry’s evolution. Second, we consider two policy-relevant regimes where imitators can or cannot resell ideas. We also
specify a meeting process between idea holders and potential adopters that generates
the well-known logistic diffusion curves documented in the technology diffusion literature. Such a meeting process has been commonly used for studying diffusion of
innovation (see Young (2009) for a review), and it corresponds to a quadratic matching function discussed recently by Lauermann, Nöldeke, and Tröger (2020). Finally,
we apply our theory to match the early development of the automobile and personal
computer industries, and discuss policy and welfare implications.
In our model, innovation yields a payoff that depends partly on the use of the
idea in production, and partly on the value it yields when it is sold. The latter
occurs in bilateral meetings and so our model relates to models in which agents
search for a production partner after one has invested, such as Burdett and Coles
(2001), Mailath, Samuelson, and Shaked (2000) and Nöldeke and Samuelson (2015).
In these models, payoffs in a match will depend on their investments and this affects
investment incentives.
In the model, owners of ideas use them to compete in the product market and thus
the flow value of the idea depends on how many others are using it. Manea (2019)
also assumes ideas are sold in bilateral meetings and uses bargaining to allocate rents,
but in his model the flow value of an idea to its user does not depend on how many
others have it.
We fit the model to the expansion path of the number of firms and of output in the
automobile and the personal computer industries, both of which show the familiar shaped growth in the number of producers in the period before the shakeout. We thus
add to the strand of literature on industry life cycles — e.g., Gort and Klepper (1982),
Utterback and Suarez (1993), Jovanovic & MacDonald (1994), Klepper (1996), Filson
(2001), and Wang (2017). Existing studies mainly focus on explaining the shakeout of
firms, while our study explains the expansion of firm numbers prior to the shakeout.
3

Regarding empirical findings, in the benchmark calibration of the automobile
industry we find that the socially optimal compensation share for innovators is 7
percent if imitators cannot resell ideas, or 16 percent if they can. Alternatively, the
social optimum can be achieved by subsidizing 60 percent of the innovation cost. We
then re-calibrate the model to the automobile industry by assuming a larger pool of
potential adopters, and we also apply the model to the personal computer industry.
The results are qualitatively consistent and our analysis explains the quantitative
variation found across different cases.
The basic model is presented next in Section 2. Section 3 applies it to a small
industry to which potential imitators arrive at a constant rate, while Section 4 deals
with a large industry where the meeting rate between incumbents and outsiders depends on their numbers. Section 5 fits the model to data from the automobile and
personal computer industries, and Section 6 concludes. More technical arguments are
in the Appendix.

2

Model

Consider a competitive market environment in continuous time. At date 0, a measure
0 of agents, who we call “innovators,” each invest an amount  in an innovation that
results in a new good. After that, the innovation spreads to other potential producers.
At any date , the measure of agents who have adopted the innovation and whom we
call “incumbents” is  , and other agents are “outsiders.” Diffusion occurs through
meetings between incumbents and outsiders in which an outsider learns and imitates
the innovation. At any date   0, the  measure of incumbents consists of both
innovators and imitators.
Each incumbent produces one unit of output at zero cost, so the total output of
the new good is  at date . The inverse demand function for the new good is
 = − 

(1)

where  is a market size parameter and   0 is the inverse demand elasticity. We
normalize an outsider’s earnings to zero.
An incoming imitator may need to pay a fee to an incumbent for transferring the
innovation. We assume the transfer fee is
 =   

(2)

where   is the value of becoming an imitator at date   0, and  ∈ [0 1] is the
bargaining share of the idea seller. We shall model two regimes. Under the first
one, while an imitator may have learned about the innovation from any incumbent,
she has to pay an original innovator for officially transferring the innovation (e.g.,
a franchising or patent licensing fee). In other words, an imitator is not entitled
to resell the innovation. Under the second regime, an incoming imitator can pay
4

any incumbent, either an innovator or an imitator, for transferring the innovation, a
scenario arising in some cases of technology transfer or employee spin-offs.
If each meeting results in a new incumbent, the number of meetings at date 
is the measure of new adopters, denoted by  . As  rises, output of the new
good rises and its price falls. At  = 0 an agent would take into consideration the
time path of future revenues (i.e., from selling the good and earning transfer fees)
to decide whether investing  to become an innovator or taking the option value of
being a future imitator. We will show, at equilibrium, innovators enter only at date
0; no outsider would want to independently invent the innovation after that.

3

A small industry

We first consider a diffusion process that results in incumbents meeting potential
imitators at a constant rate. This process may fit an emerging industry that is small
compared with the rest of the economy, and for which the pool of potential entrants
is large. Formally, the incumbents meet potential imitators at a constant rate  and
if each meeting results in a new producer, the number of incumbents grows at the
rate  so that

=  
(3)

Conditional on 0 , this implies that at date  industry output and the number of
incumbents is
 = 0  
(4)
One interpretation of this process is that at each date , an incumbent has a
chance  of meeting an outsider to disseminate the innovation. We will show later
that this emerges as a limiting case of the large industry model of Section 4 when the
number of potential adopters goes to infinity. A second interpretation is that each
incumbent has a chance  to generate a spin-off, that is, an employee who leaves the
employer and starts a new firm in the same industry.

3.1

Equilibrium

We now characterize equilibrium under two regimes:
1. Imitators cannot resell ideas to other adopters. A potential adopter can copy
an incumbent imitator but the fee goes to an original innovator.
2. Imitators can resell ideas to other imitators and keep the proceeds.
Each meeting results in the idea being transferred so that Eq. (4) holds in both
scenarios. But agents’ revenues differ.

5

3.1.1

Imitators cannot resell ideas

If an imitator cannot resell the idea, its only revenue comes from selling the good,
and its value   satisfies the Hamilton—Jacobi—Bellman (HJB) equation
 =  +


 
= (0  )− +




(5)

where  is the interest rate. The ordinary differential equation (ODE) has the unique
bounded solution
0− −
 =

(6)

 + 
which decreases at the rate , i.e., it declines faster if the demand curve is less price
elastic or if output grows faster.
Let  be the value of being an innovator at date . An innovator receives revenues
from selling both the good and the idea. The number of ideas sold at  is  and
the total date- revenue from these sales,    is divided among the 0 innovators.
Thus  follows the HJB equation


0− (1−) 
= (0  )− +
+

(7)
  +
0

 + 

Unless  = 0, innovators receive a fraction of revenues from idea sales, we shall
need to restrict the elasticity of demand to be below unity which means that  ≥ 1.
Imposing the boundary condition →∞  ∞ yields the unique solution to Eq. (7):
 =  +

 =

0−
0− −
+
−(−1) 
 + 
( + ) ( + ( − 1))

(8)

Solving for 0 .–Denote by  the option value of becoming a future imitator. At
 = 0 the free entry condition reads
0 − 0 = 

(9)

The pool of outsiders being infinite, an outsider’s chance of meeting an incumbent is
zero so that  = 0 for all , implying that 0 =  Since  decreases over time, no
one would pay  to become an innovator at any   0. Combining 0 =  with Eq.
(8) yields
0−
0−
0 =
+
= 
(10)
 +  ( + ) ( + ( − 1))
Let 0  denote the entry of innovators that solves Eq. (10). Then
¶1
µ
( + ( − 1) + )

=

0
 ( + ) ( + ( − 1))

which is valid as long as  ≥ 1, and for all   0 if  = 0.
6

(11)

3.1.2

Imitators can resell ideas

If imitators can resell the innovation, all the incumbents (be they innovators or imitators) share the same value  . The revenue from an idea sale is  instead of 
and the total date- revenue from these sales,   , is shared equally among all the
incumbents. Then  follows the HJB equation
 =  +  +



= (0  )− +  +




(12)

No restriction on  is needed here because innovators get no revenue from ideas sold
by others and  must decline with . Imposing the boundary condition →∞  ∞
yields the solution
0− −
 =

(13)
 + ( − )
which is unique if   .
Solving for 0 .–Since  decreases over time and, again, innovation occurs only
at  = 0. At  = 0 we again have 0 = , thus
0 =

0−
= 
 + ( − )

(14)

Let 0 denote the entry of innovators that solves Eq. (14). Then
0

3.2

=

µ


( + ( − ))

¶ 1



(15)

Implications of the small industry model

We assume  ≥ 1 throughout this section so that the solutions for  and 0
always exist. Several implications then follow from Eqs. (11) and (15). We begin
with
Proposition 1 0  and 0 both increase with  and  and both decrease with
; moreover,
µ  ¶
½
0
 (1 − ) 2
1 for  ∈ {0 1}
=

(16)
=1+
 1 for  ∈ (0 1)
( + ) ( + ( − 1))
0
The effect of .–Eq. (16) states that the ratio 0  0 is inverted-U shaped
in  and symmetric around  = 12. But under free spillover ( = 0) or full rent
extraction ( = 1), 0 = 0 . In the latter extreme ( = 1), the innovators
capture the entire discounted value of industry output.
The effect of .–The effect of diffusion speed on innovation depends on  and
. Specifically, in Eq. (15), 0 falls with  if    because the effect of entry
7

of competitors on reducing  (summarized by ) exceeds the benefit (summarized
by ) that incumbents derive from selling the idea. The effect of  on 0 turns
positive if   , and it vanishes if  = . Under the regime that imitators cannot
resell (cf. Eq. (11)), 0  falls with  as  ≥ 1  , and the effect of  on 0
vanishes when  =  = 1 These results are summarized in Proposition 2.
Proposition 2 0  decreases in  for  ≥ 1  , while 0 decreases in  for
  .
Proof. Eq. (11) shows 0   0 for  ≥ 1   and 0  = 0 for
 =  = 1 Eq. (15) shows 0  S 0 if  T 
The findings relate to several policy questions:
1. Patent licensing.–Here Eq. (11) is the relevant solution for 0 because the
initial innovators receive the licensing income. As our model illustrates, the investment on innovation highly depends on how innovators’ intellectual property rights get
protected (e.g., through the compensation share  and whether imitators can resell
the innovation).
2. Technology transfer via employee spin-offs.–While our model does not include
labor inputs in production, it is still relevant for explaining employee spin-offs if
we interpret employees as people who have a chance to meet with innovators and
learn about the innovation. For example, they could work in the same company
but do not have to directly produce the new product. In this case, Eq. (15) is the
relevant solution. Conventional wisdom is that spin-offs negatively affect innovation
because founders of spin-off firms may copy innovations from their previous employers
without pay. This is implied by Proposition 2 because in that case 0   0
for    = 0: The higher the spin-off rate, the fewer the entry of innovators.
3. Restricting entry of imitators.–In our model, if an incumbent firm cannot
get sufficient compensation from the imitators, it would prefer a slower diffusion of
the innovation (i.e., a lower ). In reality, researchers and policymakers often debate
whether restricting entry of imitators would be in the public interest. One such debate
concerns whether non-compete contracts should be enforced or banned.1 Recent
research (e.g., Klepper, 2010; Samila and Sorenson, 2011; Cabral, Wang and Xu,
2018) shows that employee spin-offs lead to industry clusters and that banning noncompete contracts is an important contributing factor. According to Saxenian (1994)
and Franco and Mitchell (2008), because California bans non-compete contracts while
Massachusetts enforces them, Silicon Valley overtook Massachusetts’ Route 128 in
developing the high-tech industry.
1

A non-compete contract requires that if an employee leaves the firm, she may not conduct
business to compete against her previous employer for a period of time. Among others, Shi (2018)
analyzes the effects of noncompete contracts which in her model restrict the mobility of managers
and reduce welfare. In practice, the enforcement of non-compete contracts varies substantially across
the 50 U.S. states (See Bishara, 2011).

8

Our model has a mechanism that could have led to such an overtaking pattern.
Suppose that Route 128 and Silicon Valley each specialize in some high-tech subsectors, and the two locations face the same environment (same    and ) except
that, because California bans non-compete contracts, diffusion would be faster so
that  is higher than in Massachusetts. Because Massachusetts enforces non-compete
contracts, the Route 128 area would offer higher incentives to innovators and a resulting higher initial entry rate of firms than Silicon Valley in the early periods (a
higher 0 ). Then Silicon Valley’s higher imitation rate would lead to its overtaking
Route 128.2
The left and right panels of Fig. 1 show the data from Saxenian (1994) and our
model simulation, respectively. In the simulation, we assume  = 0 in two locations.3
´ 1
³

Eqs. (11) and (15) then suggest that 0 = 0 = 0 = (+)
. We then
assume  = 3,  = 1,  = 005, and plot  in two locations: One with a high
diffusion rate ( = 008) and another with a low diffusion rate ( = 004). In each
location,
¶ 1
µ


 = 0  =
 
( + )
As a result, the location with a lower  shows an early dominance in terms of firm
entry, but its industry size later gets surpassed by the other location.

Fig. 1. Industry Overtaking: Data and Small Industry Model Simulation
2

Enforcing non-compete contracts may also increase the bargaining share  of innovators. If that
happens, the entry of initial innovators will be larger in the enforcing location and the timing of
overtaking will be postponed compared with the case where both locations have the same value of
. Note that in the model,  does not affect the growth rate of  and the overtaking is ultimately
driven by a larger diffusion rate.
3
We could think in one location, non-compete contracts are banned so innovators do not receive
any compensation from their employee spin-offs. In another location, non-compete contracts are
strictly enforced and the bilateral negotiation to buy out those contracts is too costly for the parties
involved, so that spin-off entrants are largely blocked. As a result, both locations have  = 0 at
equilibrium but the diffusion speed  differs.

9

3.3

Welfare analysis for a small industry

We now consider the welfare problem of maximizing the discounted consumer plus
producer surplus. For  ∈ (0 1), aggregate utility at output  is
 () =

Z



−  =

0

 1−
 
1−

(17)

Now for  ≥ 1 the above integral is infinite; to ensure consumer surplus
is finite,¢
¡
we put a maximum, − , on the willingness to pay. Let³  () = min − ´−
R
R
R
and define aggregate utility as  () = 0  ()  =  0 −  +  −  . Accordingly, for  = 1 we have
 () = (ln  + 1 − ln )

(18)

and for   1, we have
 () =
3.3.1

 1−
 1−

 
+
−1
1−

(19)

Optimal compensation share

Taking  as given, the social planner chooses 0 to maximize social surplus
Z ∞
0 = −0 +
−  ( ) 
0

Let 0 denote the socially optimal value. With Eqs. (17), (18) and (19), we derive
the first-order condition at any value of  that
0

=

µ


( − (1 − ))

¶ 1



(20)

Proposition 3 In the small industry model, the socially optimal compensation share
is  = 1 for the innovators.
Proof. Comparing (20) to the expressions in Eqs. (11) and (15), 0 = 0 =
0 iff  = 1, The first equality holds for all   0, and the second holds when
 ≥ 1
Thus, allowing the original innovators to extract all the rents from succeeding
imitators would yield the socially optimal incentive for innovation. This result holds
regardless whether imitators can or cannot resell the innovation. We shall return to
this question in Section 4 in which the meeting rate becomes endogenous.

10

3.3.2

Optimal innovation subsidy

Given a compensation share , the planner can offer innovators a subsidy  to achieve
the social optimum. The net entry cost of innovators becomes −. Under the regime
that imitators cannot resell ideas, Eqs. (11) and (20) pin down the optimal subsidy
 , so that
 [ + ( − 1) + ] = ( −  ) ( + ) 
which yields
 =

(1 − )

 + 

(21)

Under the regime that imitators can resell ideas, Eqs. (15) and (20) pin down the
optimal subsidy  , so that
 [ + ( − 1)] = ( −  )[ + ( − )]
which yields
 =

(1 − )

 + ( − )

(22)

Proposition 4 In the small industry model, 0     for any 0    1;

 =  = 0 for  = 1; and  =  = +
for  = 0
Proof. Follows directly from Eqs. (21) and (22).
3.3.3

Optimal diffusion rate

Our model also sheds light on diffusion policies. Suppose that incumbents cannot
force imitators to pay, so  = 0. From a social welfare point of view, should the
planner slow down the diffusion speed  (e.g., by restricting entry of imitators) to
enhance incentives for innovation?
When  = 0 Eqs. (11), (15) and (20) imply
0

=

0 

=

µ


( + )

¶ 1

 0 

(23)

Since 0 and 0 are both decreasing in , one may have conjectured that the
planner would prefer a lower 
That conjecture, however, is not true. Taking the expressions for 0 and 0
in Eq. (23) as given, the planner’s problem is to choose  to maximize social surplus
Z ∞
0 = −0 +
−  ( ) 
0

11

where  ( ) is defined in (17), (18) or (19) depending on the value of .4
µ
¶
⎧
1− 1
1
(+) 
1− 1
− 1
⎪

⎪
 
− ( + )
⎪
(1−)(−(1−))
⎪
⎨
(1−ln )



ln (+)
+ 
0 =
2 − (+) +



µ
¶
⎪
⎪
1− 1
1
1
1
⎪

(+)
1−
−
⎪
⎩     (1−)(−(1−)) − ( + )  +

if  ∈ (0 1) ;
if  = 1;

1−
(−1)

(24)

if   1

It is straightforward to show that for any   0
0
 0

which proves the following:

Proposition 5 Given technology spillover is free ( = 0), the social planner would
not want to slow down diffusion.
Surprisingly, then, even though 0 and 0 are below 0 , the planner does
not want to slow down the diffusion — i.e., does not want a lower  This finding
highlights the important contribution of technology diffusion to welfare. However,
this result is based on the assumption that incumbents meet potential adopters at a
fixed rate. We shall ask this question again in Section 4 in which the meeting rate is
endogenous.

4

A large industry

Our small industry model assumes that the pool of potential entrants is unlimited,
leading to a diffusion process with a constant growth of producers. Evidence shows,
however, that diffusion tends to be -shaped.5 . The following model endogenizes the
meeting rate between incumbents and outsiders and yields -shaped diffusion.
Let  be the population of potential adopters. The model is the same as before,
except that now, instead of being  as in Eq. (3), the number of meetings between
incumbents and outsiders at date  is  ( −  ) where   0 is a parameter. This
matching function is homogeneous of degree 2 so that an agent’s meeting rate is
proportional to the number of potential partners. If every meeting results in an idea
transfer, the number of producers evolves as

=  ( −  ) 

4

(25)

When  = 0,   exists even if   1 So the first line in Eq. (24) is valid.
Griliches (1957) documented -shaped diffusion of hybrid corn on U.S. farms, and among many
later studies, Comin and Hobijn (2004) plot -shaped diffusion curves for major GPTs.
5

12

Accordingly,  follows the classic logistic diffusion curve
 =


  + 

(26)

and  is determined by the initial condition for 0 ∈ (0 ) so that

− 1  0
0

(27)



 + 0 − 1

(28)

=
We then have
 =

The  → ∞ limit.–The small industry model is a limit of logistic diffusion in
the following sense: Define a constant   0 and let
=


→ 0 as  → ∞


Equations (25) and (28) imply that for given  ,
 

=  ( −  ) = (1 −  
)

 + 0 − 1

(29)

Given that the demand curve is downward slopping, 0 has to be finite as  → ∞;
otherwise 0 → 0, and no innovator would enter at date 0. Therefore, Eq. (29)
implies that
¯
  ¯¯
→ 
(30)
 ¯→∞

i.e., the incumbents’ meeting rate converges to a constant, as seen in Eq. (3) for the
small industry model.

4.1

Equilibrium

We, again, characterize the market equilibrium under two regimes: (1) imitators
cannot resell ideas, and (2) imitators can resell. We assume  = 1 in our theoretical
analysis to derive closed-form solutions. We will later relax the assumption in Section
5 when taking our model to data.

13

4.1.1

Imitators cannot resell ideas

Since an imitator cannot resell the innovation, its only revenue comes from selling
the good, and its value   satisfies the HJB equation
 −1  
 
= ( 
) +



+

Solving the ODE and imposing the boundary condition →∞  ∞ yields the solution
 =  +

 =

−

+

 ( + )

(31)

Because an innovator receives revenues from selling both the good and the idea,
the value  of being an innovator at date  follows the HJB equation
 

  +
0

µ
¶

(1 + ) 1



−1
) +
+ 

+
= ( 


2

+
(
+ )
  ( + )


 =  +

Solving the ODE and imposing the boundary condition →∞  ∞ yields
¢
¡
 ( +  + )  + 2 

+

 =

( + )(2 +  )

(32)

Accordingly, we have
0 =


( + )
+


( + )

(33)

With a logistic diffusion, because the outsiders’ meeting rate of incumbents is
endogenous and always positive, the option value of being an outsider,  , is not zero.
Eq. (25) implies that the hazard rate of imitation for outsiders is
 
=  
 − 

(34)

Therefore,  satisfies the HJB equation

 =  [(1 − )  −  ] +
¸
∙
µ 
¶




−
= 
(1 − )
+
−  +


+
 ( + )

The unique bounded solution for  is a constant:
14

(35)

 =

(1 − )

2 + 

(36)

Two forces exactly offset; the meeting hazard  in Eq. (34) rises at the same rate
as  = −1 declines.
At  = 0, the free entry condition requires 0 − 0 = , thus Eqs. (33) and (36)
yield
=
Since  =


0

( + ) −  − 

( + )

− 1, we derive
0 =

( + )

( +  )

The  → ∞ limit.–Define a constant   0, and let  =
Eq. (37) implies that
( + )

0 =
( + )

(37)




→ 0 as  → ∞
(38)

Therefore, everything else fixed, Eq. (29) applies,  → ∞ results in 0 → ∞ and
→ .
the large industry model converges to the small industry model with  

The  → ∞ limit.–As  → ∞,  → , and  →  −1 . Also, the incumbents’
meeting rate of outsiders  ( −  ) → 0 and the outsiders’ meeting rate of incum
bents  → . Eq. (31) shows that  →∞ → 
. This means an outsider gets
(1 − ) ∞ with the exponentially distributed waiting time  with density − 
i.e.,
Z ∞
∞ = (1 − ) ∞
− − 
0


(1 − )

= 2

= (1 − )
 ( + )
 + 

which is implied by Eq. (36).
Comparative statics.–The entry of innovators 0 given by Eq. (37) is very similar
to Eq. (11) when  = 1. Therefore, the comparative statics are similar to what we
derived in Propositions 1 and 2: 0  increases in  and  but decreases in  and
 (except that 0  = 0 when  = 1).

15

4.1.2

Imitators can resell ideas

If imitators can resell the innovation, all the incumbents (i.e., both innovators and
imitators) share the same value  that follows the HJB equation
µ
¶
 



 −1
 +
= (  
) +  
 +

(39)
 =  +



+

+

and the value of an outsider  solves

 =


 


[(1 − ) −  ] +
=  
((1 − ) −  ) +

( −  )


+


(40)

where  = 0 − 1 The free entry condition 0 − 0 =  then pins down the entry of
innovators 0 at date 0 Equations (39) and (40) in general do not have closed-form
solutions, but we can solve particular cases to derive key insights.

4.2

Implications of the large industry model

In the following, we solve special cases for the regime under which imitators can resell
the innovation (i.e., 0 ), and compare them with the regime under which imitators
cannot resell (i.e., 0 ).
First, we solve 0 for the cases where  = 0 and  = 1 (see Appendix I). These
two cases are the most relevant for our welfare analysis. The results show


( + )

= 


0 = 0 =

if

 = 0;

(41)

0 = 0

if

 = 1

(42)

These results confirm our finding from the small industry model in Proposition 1,
0 = 0 for  ∈ {0 1}. The intuition is straightforward: Because the innovators
who enter at date 0 either receive no resale revenues at all (if  = 0) or get all the
revenues (if  = 1), whether imitators can or cannot resell the innovation becomes
irrelevant.
Second, we explicitly solve 0 for the case where  = . Appendix II shows
that
 − 0

µ³
= 
(43)
´2 ³
´ ¶


0


−1
− 
(2−)(− )

0

0

0

which yields comparative statics similar to Propositions 1 and 2: 0 increases in
 and  but decreases in ; moreover, 0  0 for  ∈ (0 1).
16

Fig. 2. Entry of Innovators: Effects of 
The comparative statics are illustrated by Fig. 2, which plots 0 and 0
according to Eqs. (37) and (43) for  =  and  = 1. In the figure, the solid lines
stand for 0 and the dash lines stand for 0 . Cases with different value of  are
plotted in different colors. The figure shows that both 0 and 0 increase in 
and  but decrease in . It also shows that, similar to the small industry model (cf.
Proposition 1), 0 = 0 for  ∈ {0 1}, and 0   0 for  ∈ (0 1) 
Finally, the large industry model yields policy implications similar to the small
industry model:
1. Patent licensing.–Here Eq. (37) is the relevant solution for 0 . Again, the
investment on innovation highly depends on how innovators’ intellectual property
rights get protected (e.g., through the compensation share  and whether imitators
can resell the innovation).
2. Technology transfer via employee spin-offs.–Here Eq. (41) is the solution when
 = 0, which confirms that spin-offs negatively affect innovation (i.e., 0   0).
3. Restricting entry of imitators.–The large industry model also predicts the
overtaking pattern. A location restricting entry of imitators (i.e., a lower ) tends to
enjoy a higher 0 , but its industry size later gets surpassed by other places without
such restrictions (i.e., a higher ).
Figure 3 show simulation results based on the large industry model. Assuming

 = 0,  = 1, Eq. (41) shows that 0 = 0 = 0 = (+)
. Define  = .
We assume  = 3,  = 1000,  = 005, and plot  in two locations: One with
high diffusion rate ( = 008), the other with a low diffusion rate ( = 004). In each
location,
 

 =  

=

+ 0 − 1  + (+)
−
1

We again see an overtaking pattern, similar to the one generated by the small industry
17

model. The location with a lower  enjoys a higher  in early periods, but its industry
size is later surpassed by the other location with a higher .
300

= 0 .0 8
= 0 .0 4

250

k

t

200

150

100

50

0

0

5

10

15

20

25

30

t

Fig. 3. Industry Overtaking: Large Industry Model Simulation

4.3

Welfare analysis for a large industry

We now consider the welfare problem from the social planner’s point of view. The
main difference of the large industry model is that now full rent extraction (i.e.,  = 1)
is not socially optimal because of a congestion effect or rent-grabbing externality that
innovators impose on each other. In the case of lacking intellectual property right
protection (i.e.,  = 0), however, we continue to find that the planner would not want
to reduce diffusion speed. To illustrate these findings, we keep the assumption  = 1.
Accordingly, consumer surplus is still given by Eq. (18), i.e.,  () = (ln +1−ln )
Without loss of generality, we also normalize  = 1 in the welfare analysis so that
Eq. (28) reduces to

 = 

(44)
 + 10 − 1
4.3.1

Optimal compensation share

Assume the technology diffusion parameter  is exogenous. Given  = 1, the planner
chooses 0 to maximize social surplus
Z ∞
− [(ln  + 1 − ln )] 
(45)
0 = −0 +
0

with  given by Eq. (44).
The first order condition requires
Z ∞
£
¤−1
− 02 ( − 1) + 0

=
0

18

(46)

Let 0 denote the socially optimal value that solves Eq. (46). We have the
following proposition.
Proposition 6 In the large industry model,  = 1 is not the socially optimal compensation share for the innovators.

Proof. Eq. (42) shows 0 = 0 = 
when  = 1. We can verify that


0
=  is the solution for Eq. (46) when  = 0. Because the right hand side of

Eq. (46) strictly decreases in , 0 has to be smaller than 
to satisfy Eq. (46)
for any   0 Therefore,  = 1 is not the socially optimal compensation share.

When  = 1, an innovator can appropriate all the value from the meetings that
her presence generates. We have learned from the small industry model that this
yields an equilibrium that is socially optimal. But why this does not hold for the
large industry model?
The key difference of the large industry model is that the meeting probability is
endogenous and there is a congestion effect that an innovator creates — she reduces
the number of meetings that other innovators will have, and this is an effect that
the innovator ignores. To see this, recall in the small industry model the probability
for an incumbent to meet with an imitator is fixed (i.e.,  
= ). In contrast,

in the large industry model, the probability is negatively affected by the number of
incumbents (i.e.,  
=  ( −  )).

One can solve for a   1 that generates the planner’s optimum. Consider an
explicit example where  = . In this case, Eq. (46) can be simplified as
¢2
¡
1 − 0

(47)
= 
1
1

− 1 − ln 

0

0

Under the regime that imitators cannot resell the innovation, Eqs. (37) and (47)
imply that
2(1 − 0 + 0 ln 0 )


=
− 1
(48)
2
(1 − 0 )

Alternatively, under the regime that imitators can resell, Eqs. (43) and (47) imply
that  solves
0
(2−)(1−0 )

µ³

1
1
0

´2

−

³

1
0

´ ¶

=
−1

1 − 0
1
1 
− 1 − ln 

0

(49)

0

Figure 4 illustrates this example where  = . For a given value of () = 03,
Fig. 4(I) plots the relation between 0 and 0 , given by Eq. (45).6 The result
6

Equation (45) shows that  and  are just scaling parameters and they do not affect the maximization of 0 , so without loss of generality we set  = 1 and  = 00001 for plotting Figs. 4(I)-(II).

19

shows that welfare maximizes at 0 = 022. Figure 4(II) shows that this welfare
maximum can be achieved by either setting  = 047 under Regime 1 (i.e., where
imitators cannot resell ideas), or setting  = 057 under Regime 2 (i.e., where
imitators can resell). Figures 4(III)-(IV) extend the results to the full domain of
(). Figure 4(III) plots the relation between 0 and () given by Eq. (47),
and Fig. 4(IV) traces out the relation between 0 and  that satisfies Eqs. (48)
or (49). The negative relation between two endogenous variables, 0 and  , is
induced by changes in () – as () rises, so does 0 but  falls. For
a given value of 0 (or the corresponding ()), the value of  is always
smaller under Regime 1 than under Regime 2.

Fig. 4. Socially Optimal Compensation Share
4.3.2

Optimal subsidy/taxation

Alternatively, the planner could provide a subsidy or impose a tax to internalize
externalities. In the small industry model where the meeting rate is fixed, the planner
always wants to provide a subsidy to innovators for any   1 (cf. Proposition 3). In
the large industry model, however, the planner would want to impose an innovation
tax if  is sufficiently large or provide a subsidy if  is sufficiently small. One can
calculate this tax or subsidy as a function of model parameters, including .
To see this clearly, let us again consider the explicit example where  = . In Fig.
5(I), we plot Eq. (47) using a black solid line and overlay it on Fig. 2 above. The

figure shows that for a given level of 
, 0 always exceeds the market equilibrium


level (i.e., 0
or 0 ) when  = 0, but falls short when  = 1. Moreover, for any
value of  ∈ (0 1), the socially optimal entry 0 can be achieved by adding an
appropriate tax or subsidy to  In the figure, the vertical difference between a market
20

equilibrium path (associated with a particular  and a regime whether imitators can
resell the innovation or not) and the socially optimal path indicates the amount of

adjustment to 
(i.e., by adjusting (−  ) or (− ) ) needed to achieve each
social optimal level of 0 . Figure 5(II) plots the subsidy (scaled by the entry cost
) needed to achieve the social optimum. The figure shows that the scaled subsidies,

  and  , both decrease in 
and , and can turn negative (i.e., becomes

a tax) if  or  becomes sufficiently large. Moreover,  =   0 (i.e., a
subsidy) for  = 0,  =   0 (i.e., a tax) for  = 1, and    for
0    1.

Fig. 5. Socially Optimal Subsidy/Taxation
4.3.3

Optimal diffusion rate

We now explore the implications of the large industry model on diffusion policies.
Consider again the scenario where incumbents are not compensated by imitators, so
 = 0. Should the planner slow down the diffusion?
Recall that under  = 0, Eq. (41) shows that the equilibrium entry of innovators
is given by

0 =

(50)
( + )
regardless of whether imitators can or cannot resell the innovation.
Taking Eq. (50) as given and  = 1, the planner would choose  to maximize
social surplus:
21


0 = −
+
( + )

Z

0

∞

−



µ
µ
¶¶
 ( + )

 − ln  +
−1
 + constant. (51)


0
0
We can then prove 
 0 for any 1  0  0, and 
= 0 when 0 = 1 (see


Appendix III for the proof).
This finding extends our result in Proposition 5 to the large industry model where
the meeting rate between incumbents and outsiders is endogenous, and it lends further
support to public policies that accommodate diffusion. Note that the finding does not
rule out the possibility that policymakers can exploit the welfare gain of temporarily
restricting . For example, policymakers could restrict  initially to achieve the
socially optimal entry of innovators 0 , and then free up the limitation. While
this seems welfare improving, such a policy would not be time-consistent (Kydland
and Prescott, 1977). Presumably policy must apply more broadly, not just to one
instance, but to future products and future instances of .

5

Empirical study

In the empirical study, we calibrate the large industry model to data and relax the
assumption  = 1. We consider two historically important industries: automobile and
personal computer, where idea diffusion played an important role in the industries’
development.7 Using model calibration and counterfactual exercises, we evaluate and
quantify our theoretical predictions.

5.1

Automobile industry

The U.S. automobile industry started in 1890s and grew from a small infant industry
to a major sector of the economy in a few decades. During the process, the industry
output continued to expand, but the number of firms initially rose and later fell.
Starting with 3 firms in 1895, the number of auto producers exceeded 200 around
1910. A shakeout then followed when a major process innovation, the assembly line,
was introduced in the early 1910s and firm productivity rose tremendously ever since.
Eventually, only 24 firms survived to 1930. Figure 6 plots the number of firms and
output per firm in the U.S. auto industry from 1895-1929.
Our theoretical model describes the auto industry development very well for the
pre-shakeout period, and we can also extend the model to incorporate the shakeout
7
For example, Klepper (2010) documents in detail how the spawning of employee spin-offs and
entry by firms in related industries drove the development of the automobile and semiconductor
industries.

22

100

Output Per Firm (1,000)

50

100

0

0

50

Firms

150

150

200

200

without affecting our analysis.8 As shown in Fig. 6, the growth of the auto industry
mainly relied on the extensive margin before 1910. During that period, the time path
of firm numbers followed an -shaped curve and the average output per firm stayed
pretty much constant. In the following analysis, we will calibrate our model to the
auto industry data for the pre-shakeout era (1895-1910) and conduct counterfactual
exercises.

1 89 0

1900

1910

1 92 0

1930

1890

1 9 00

19 1 0

1920

1 9 30

Fig. 6. Auto Firm Numbers and Output Per Firm
5.1.1

Data

The data of U.S. auto industry comes from several sources. Smith (1970) lists every
make of passenger cars produced commercially in the United States from 1895 through
1969. The book records the firm that manufactured each car make, the firm’s location,
the years that the car make was produced. Smith’s list of car makes is used to derive
the number of auto firms each year. Thomas (1977) provides annual data of average
car price and output from 1900-1929. Williamson (2020) provides annual data of U.S.
population, real GDP, and the GDP deflator.
8

Our model can be extended to incorporate the shakeout. Following Jovanovic and MacDonald
(1994), we may assume that the industry expects a disruptive innovation to arrive at a Poisson
rate . This innovation would allow any incumbent firms to make an investment to transform its
product and tremendously increase its output. When such an innovation arrives, a few incumbents
would invest and the rest would exit. In the competitive equilibrium, the present value of an
investing firm (net of its investment) is zero, and the value of an exiting firm is also zero. In the
scenario where imitators can resell ideas, the HJB equation for an incumbent firm would change to
 


 =  +  
 −  + 
  which implies that ( + )  =  +
  +   Similarly, we

can rewrite the HJB equations for other types of firms and for the scenario where imitators cannot
resell ideas. In any case, the original functional forms of our model hold, with just the discount
parameter changing from  to  + 

23

5.1.2

Diffusion and demand estimation

To calibrate the model, we first use the data of firm numbers to estimate the diffusion
parameters.9 With Eq. (28), we can rewrite the diffusion process of  as follows:

=  + ̃
 − 

(52)


= −413 + 053 
 −  (026)∗∗∗ (003)∗∗∗

(53)

ln

0
 and ̃ = 
where  = ln −
0
We assume that the shakeout started after all the potential firms had entered the
industry (and we will adjust this assumption later). So we assume  = 210 and run
the regression model (52). The result shows that

ln

and the standard errors are reported in the parentheses. The estimates of  and
̃ are both statistically significant at 1% level (noted by three stars), and adjusted
2 = 096 Based on the estimates of diffusion parameters, we calibrate ̃ = 053
0
and 0 = 331 (i.e., ln −
= −413).
0
We then estimate an industry demand function using annual data of auto prices
 and output  from 1900—1929. Eq. (1) implies a simple log-log per capita demand
function:

ln(
) =  −  ln( )

In the regression, we control for log U.S. GDP per capita (as a proxy for income) in
the demand intercept  . Both auto price and GDP per capita are in real terms.
To address potential endogeneity of the price variable, we use the output per
firm (lagged by a year) as an instrumental variable to estimate the demand slope
parameter . Output per firm, while is assumed fixed in our theory, did grow over
time in the data due to technological progress. This would have affected the supply
condition but not the demand.10
The first-stage regression result (adj. 2 = 089) is given by:
ln( ) = 856∗∗∗ + 166∗∗ × ln(
(109)

(064)


) − 029∗∗∗ × ln(output per firm)−1 
(003)


9

For simplicity, our model assumes that firms do not exit the industry after their entry. It is
natural to extend the model to incorporate firm exits in the form of exogenous death; the analysis
does not change much (See Appendix V for a detailed discussion).
10
Cabral, Wang and Xu (2018) also estimated the auto demand function for the same sample
period. They used a different instrument variable, the share of spin-off firms in the auto industry.
The idea is that the founders of spin-off firms are more experienced than de novo entrants, so spin-off
firms tend to perform better (Klepper, 2010). They show that their instrument variable performs
well and the estimated demand slope  = 339, which is very close to ours.

24

The second-stage regression result (2 = 083) is in turn given by:
ln(



) = 3237
+ 028 × ln(
) − 333∗∗∗ × ln( )
∗∗∗
(716)
(210)
(038)



Standard errors are reported in the parentheses, with three stars, two stars and one
star representing statistical significance at 1%, 5% and 10% level, respectively.
The IV estimation gives  = 333 and it is highly statistically significant. This
suggests that we can calibrate the value of inverse demand elasticity in our model to
be  = 1 = 03
Figure 7 plots the estimated firm numbers and per capita auto demand, and
compares them with data.
A u t o D e m a n d P e r C a p it a

0

0

.01

50

.02

100

.03

150

.04

200

F irm N u m b e rs

1895

1900

1905

D a ta

1910

1900

E s ti m a te s

1910
D a ta

1920

1930

E s ti m a te s

Fig. 7. Auto Diffusion and Demand Estimation
5.1.3

Model calibration

Based on the estimation results, we parameterize our model and calibrate it to the
auto industry data.
Recall the industry demand function (cf. Eq. (1))
 = − 
and the time path of firm numbers (cf. Eq. (26))


 = 
, where  =
− 1

+
0
As a benchmark, we assume that the industry data is generated by the model with
free spillover (i.e.,  = 0), so it makes no difference whether imitators can or cannot
resell the technology.11 Accordingly, the value of an incumbent firm in the industry
11

Assuming  = 0 is a natural (though not necessarily best) way to taking our model to data.
One could re-do the calibration exercises by assuming a different benchmark value of , but the
method and intuition would be very similar.

25

at date  satisfies

   − 

= (  
) +



+

while the value of an outsider satisfies
 =  +

 =

 



( −  ) +
=  
( −  ) +

( −  )


+


Based on the diffusion and demand estimates above, we calibrate the model parameter values at  = 210  = 053 0 = 331  = 03. Because  is just a scaling
parameter (i.e., only  matters), we normalize  = 1 and assume  = 005
I

1
0 .9

k /N ( m o d e l)

v

t

6

k /N ( d a t a )
0 .8

II

6.5

u

t

t
t

5.5
0 .7
5

0 .6
0 .5

4.5

0 .4

4

0 .3
3.5
0 .2
3

0 .1
0
18 95

1 900

1 905

2.5
1 895

191 0

190 0

19 05

1 910

Fig. 8. Auto Industry: Model Calibration
While a closed-form solution does not exist when  6= 1, we can numerically solve
the time path (   ) and pin down the cost of innovation  by the free entry condition
0 = 0 + Figure 8 plots the calibration results for the time paths of  ,  and  .
The number of firms  grows along a logistic curve. Meanwhile,  decreases while
 increases over time, and the initial difference 0 − 0 pins down the innovation cost
 = 310.
5.1.4

Counterfactual analysis

Given the calibrated model parameter values, we can conduct counterfactual analysis
and evaluate welfare.
Optimal compensation share We first evaluate the effect of the compensation
share . We consider two regimes and label them Regime 1 (under which imitators cannot resell the innovation) and Regime 2 (under which imitators can resell),
respectively.
For any value of compensation share  ∈ [0 1], we numerically solve the equilibrium time paths for    and  (see Appendix IV for the details). Particularly, the
26

free entry condition 0 = 0 + allows us to pin down the counterfactual entry of innovators 0 at date 0. Figure 9(I) shows that 0 strictly increases in  for 0 ≤   071
under both Regime 1 and Regime 2, and 0 is smaller under Regime 2 than under
Regime 1. Figure 9(II) shows that under the two regimes, the free entry condition
holds in terms of 0 − 0 =  when 0 ≤   071, while for  ≥ 071, because of the
corner solution 0 = , the free entry condition holds in terms of 0 − 0  
I

II

10

Regime 1: k /N
0

1

Regime 2: k /N

Regime 1: v

Regime 1: u

Regime 2: v

Regime 2: u

0

8

0

0

0
0

6
0.5

4
2

0

0

0.2

0.4

0.6

0.8

0

1

0

0.2

0.4

III

0.6

0.8

1

IV

1000

1000

800

800

600

600

400

400
Regime 1: W

200

0

200

W

Regime 2: W0

0

0

0

0.2

0.4

0.6

0.8

0

1

0

0.2

0.4

0.6

0.8

1

k /N
0

Fig. 9. Auto Industry: Effects of 
We then assess the welfare effect of changing the value of . Figures 9(III)-(IV)
plot social welfare 0 according to different values of 0 as well as . The results show
that social welfare 0 is hump-shaped in 0 and peaks at 0 = 28 (i.e., 0  = 0133).
This social optimum can be achieved by choosing  = 007 under Regime 1 or
choosing  = 016 under Regime 2. Note that under both regimes, neither  = 0
nor  = 1 is socially optimal: the former yields 0 = 331 and the latter yields
0 =  = 210.
The socially optimal compensation share  depends on the model parameters,
including the inverse price elasticity , the diffusion speed , the innovation cost ,
and the number of potential adopters . Figure 10 plots the comparative statics of
 under Regimes 1 and 2. The results show the following:
27

•  increases in .–A higher  leads to a faster decline in price and discourages 0 . This makes the congestion externality among innovators less of a
concern, so the planner wants to raise the compensation share for innovators.
•  decreases in .– A higher  leads to a faster decline in price, but the
meeting rate is higher, and with it the idea-sales revenue rises fast enough to
boost 0 . This worsens the congestion externality among innovators, so the
planner wants to reduce the compensation share for innovators.
•  increases in .–A higher  discourages 0 . This makes the congestion
externality among innovators less of a concern, so the planner wants to raise
the compensation share for innovators.
•  increases in  (holding  =  fixed).–A higher  leads to faster
price decline that discourages 0 . This, together with a larger pool of potential
adopters , makes the congestion externality among innovators less of a concern, so the planner wants to increase the compensation share for innovators.
This finding is consistent with our theoretical result that the large industry
model converges to the small industry model as  gets large, suggesting that
in that case  should converge to 1.
• Comparison of Regimes 1 and 2.–The socially optimal compensation share for
innovators is higher under Regime 2 than under Regime 1, and the difference
increases in , ,  and .

Fig. 10. Comparative Statics: 
28

Optimal innovation subsidy Given  = 0, the entry of innovators at date 0 is
lower than the socially optimal level. Providing innovators a subsidy , instead of
setting a socially optimal , can also help achieve the social optimality. Note that
with the subsidy,  −  is the net entry cost for innovators. Figure 11 plots the effect
of  on the entry of innovators 0 and welfare 0 . The results show that 0 increases
in , and the social welfare peaks at  = 185 (i.e., lowering the entry cost from 310
to 125, a 60% reduction).
I
1

II

11 0 0

k /N
0

0 .9

100 0

0 .8
90 0
0 .7
0 .6

80 0

0 .5
70 0
0 .4
0 .3

60 0

0 .2
50 0

W

0 .1
0

0

1

2

40 0

3

0

0
1

2

s

3

s

Fig. 11. Auto Industry: Effects of 
Optimal diffusion rate We can similarly evaluate the effects of varying the diffusion rate . Consider again the scenario where incumbents are not compensated by
imitators, so  = 0. Should the planner slow down the diffusion?
I

II

900

k 0 /N

1

W0
850

0.8
800
0.6
750
0.4

700

0.2

0

650

0

0 .1

0.2

0.3

0.4

600

0 .5

0

0.1

0.2

N

0 .3

0 .4

0.5

N

Fig. 12. Auto Industry: Effects of 
Figure 12 shows the effects of  on 0 and 0 . If the planner were to push
down  from the original value where  = 053, the entry of innovators 0 would
29

increase. If the value of  gets sufficiently low (i.e.,  ≤ 001), 0 would reach the
corner solution 0 = . However, social welfare increases in , which suggests that
restricting diffusion would reduce welfare. The intuition is that while slowing down
diffusion could encourage entry of innovators, it would forego too much free learning
in the industry and the welfare effect of the latter dominates.
5.1.5

Model re-calibration: A larger pool of potential adopters

In the benchmark calibration above, we assume that the shakeout started after all the
potential auto firms had entered the industry. Alternatively, we could assume that
the shakeout started in the middle of the diffusion process. Below, we assume a much
larger number of potential auto firms,  = 1 000. This implies that it would take 30
years to reach 99% adoption rate among potential producers had the shakeout not
happened, doubling what was assumed in the benchmark calibration. We will see
that most of our previous analysis goes through, but the optimal compensation share
 increases.
We assume  = 1 000 and re-estimate the diffusion model (52). The result shows
that

ln
= −500 + 030 
(54)
 −  (026)∗∗∗ (003)∗∗∗

and the standard errors are reported in the parentheses.
Based on the estimates of diffusion parameters, we calibrate  = 1 000  = 03
0
and 0 = 669 (i.e., ln −
= −5). In addition, we set  = 03,  = 005, and
0
normalize  = 1 as before. Figure 13 plots the calibration results. It shows that the
entry of firms would reach  = 37744 (i.e.,   = 038) at the end of the sample
period had the shakeout not occurred, and the initial difference 0 − 0 pins down
the innovation cost  = 394.12
I

0.4

k / N (m od e l)
0.35

II

5.5

t

v

5

k / N (d a ta )

u

t

t
t

4.5

0.3

4

0.25

3.5
0.2
3
0.15

2.5

0.1

2

0.05
0
1 895

1.5

19 00

190 5

1
18 95

1 910

190 0

1 905

19 10

Fig. 13. Auto Industry: Model Re-Calibration
12

Comparing Figs. 8(I) and 13(I) suggests that the benchmark calibration fits the data better.
Nevertheless, we conduct the re-calibration exercise for comparison and robustness checks.

30

Figure 14 plots the results of varying the value of , which reveal some differences
from the benchmark calibration above. First, given a much larger pool of potential
adopters, the estimated diffusion rate  becomes smaller to match the data. As a
result, 0 would not reach the full adoption  = 1 000 for any  ∈ [0 1].
I
1.2

II

10

Regime 1: k /N
0

1

Regime 1: v

Regime 1: u

Regime 2: v

Regime 2: u

0

8

Regime 2: k /N
0

0

0

0.8

0

6

0.6
4

0.4

2

0.2
0

0

0.2

0.4

0.6

0.8

0

1

III

3000

0

0.2

0.4

0.8

1

0.6

0.8

1

IV

3000

2000

0.6

2000

1000
1000

Regime 1: W

0

0

W

Regime 2: W0

0

0

0.2

0.4

0.6

0.8

0

1

0

0.2

0.4

k /N
0

Fig. 14. Auto Industry Model Re-Calibration: Effects of 
Second, welfare is maximized at 0 = 100 (i.e., 0  = 01), which can be achieved
by choosing  = 014 under Regime 1 and  = 032 under Regime 2. Compared to the benchmark calibration above, the socially optimal compensation share
 becomes larger under either regime. Our comparative-statics analysis (shown
by Fig. 10) suggests that the larger  is due to the larger , the lower , and
the higher  from the re-calibrated model.
Finally, because of the larger pool of potential producers, the automobile price
would fall more in the long run than it does in the benchmark calibration, which
leads to a higher level of welfare 0 .
Figure 15 plots the effects of a subsidy  on the entry of innovators 0 and welfare
0 . The results show that 0 increases in , and social welfare peaks at  = 256 (i.e.,
lowering the entry cost from 394 to 138, a 65% reduction).

31

I
1

II
2 500

k 0 /N

2 000

0.8

1 500
0.6
1 000
0.4
500
0.2

W0

0
0

0

0.5

1

1 .5

2

2.5

3

3.5

0

0.5

1

1 .5

s

2

2.5

3

3 .5

s

Fig. 15. Auto Industry Model Re-Calibration: Effects of 
Figure 16 plots the results of varying the value of , which are consistent with
our findings in the benchmark calibration. The entry of innovators 0 decreases in ,
but social welfare increases in .
I

0 .3

II

2 000

k /N

W

0

1 800

0

0 .2 5
1 600
0 .2

1 400
1 200

0 .1 5

1 000

0 .1

800
0 .0 5
600
0

0

0 .1

0.2

400

0.3

N

0

0 .1

0.2

0.3

N

Fig. 16. Auto Industry Model Re-Calibration: Effects of 

5.2

Personal computer industry

The personal computer (PC) industry was developed more recently than the automobile industry, but the industry evolution was not much different. Starting with
two firms in 1975, the number of PC producers exceeded 430 in 1992. A shakeout
then started while the industry output continued to expand, but the number of firms
fell sharply. Figure 17 plots the number of firms and output per firm in the U.S. PC
industry from 1975-1999.

32

250
200

400

150
0

50

100

Output Per Firm

300
Firms

200
100
0
1975

1 98 0

1985

1990

1995

2000

1975

1980

1985

1990

1 99 5

2000

Fig. 17. PC Firm Numbers and Output Per Firm
Again, our model describes the pre-shakeout development of the PC industry well.
As shown in Fig. 17, the growth of the PC industry mainly relied on the extensive
margin before the shakeout. During that time, the time path of firm numbers followed
an -shaped curve and the average output per firm stayed pretty much flat. In
the following analysis, we calibrate our model to the PC industry data for the preshakeout era (1975-1992) and conduct counterfactual exercises.
5.2.1

Data

Following Filson (2001), firm numbers in the personal computer industry are taken
from Stavins (1995) and the Thomas Register of American Manufacturers. The data
include desktop and portable computers. Price and quantity information for personal computers is from the Information Technology Industry Data Book. Williamson
(2020) provides annual data of U.S. population, real GDP, and the GDP deflator.
5.2.2

Diffusion and demand estimation

To calibrate the model, we first use the data of firm numbers to estimate the diffusion
parameters. We assume that the shakeout started after all the potential PC firms
had entered the industry (and we will adjust this assumption later). Accordingly, we
assume  = 435 and run the regression model (52). The result shows that
ln


= −549 + 058 
 −  (029)∗∗∗ (003)∗∗∗

(55)

with the standard errors reported in the parentheses. All the coefficient estimates are
statistically significant at 1% level (noted by three stars), and adjusted 2 = 096
Based on the estimates of diffusion parameters, we calibrate  = 058 and 0 = 178
0
(i.e., ln −
= −549).
0
33

We then estimate an industry demand function using annual data of PC prices 
and output  from 1975-1992. As before, in order to address potential endogeneity
of the price variable, we use average output per firm (lagged by a year) as an instrumental variable to estimate the demand slope parameter . We also control for log
U.S. GDP per capita as a proxy for income in the regression. Both PC price and
GDP per capita are in real terms.
The first-stage regression results (adj. 2 = 095) are
− 095∗∗∗ × ln(
ln( ) = 1244
∗∗∗
(023)

(006)


) − 007∗∗∗ × ln(output per firm)−1 
(001)


The second-stage regression results (2 = 095) in turn, are
ln(



) = 14318
−
290
×
ln(
) − 1557
× ln( )
(2900)∗∗∗
(263)
(240)∗∗∗



Standard errors are reported in the parentheses, with three stars, two stars and one
star indicating statistical significance at 1%, 5% and 10% level, respectively.
The IV estimation yields  = 1557 and it is highly statistically significant. Accordingly, we calibrate the value of inverse demand elasticity in our model to be
 = 1 = 006, which suggests that the PC industry is more price elastic than the
auto industry.
Figure 18 plots the estimated firm numbers and per capita PC demand, and
compares them with data.
P C D e m a n d P e r C a p it a

0

0

100

.02

200

.04

300

400

.06

F irm N u m b e rs

1975

1980

1985

D a ta

1990

1975

E s ti m a te s

1 98 0
D a ta

1985

1990

E s ti m a te s

Fig. 18. PC Diffusion and Demand Estimation
5.2.3

Model calibration

Based on the estimation results, we calibrate the model parameter values  =
435  = 058 0 = 178  = 006,  = 005. We also normalize  = 1 Again,
we assume that the industry data is generated by the model with free spillover (i.e.,
 = 0).
34

I

1
0 .9
0 .8

k t / N ( m o d e l)

ut

16
15

0 .6

14

0 .5

13

0 .4

12

0 .3

11

0 .2

10

0 .1

9

198 0

vt

17

k t / N ( d a ta )

0 .7

0
197 5

II

18

198 5

8
19 75

199 0

19 80

198 5

199 0

Fig. 19. PC Industry: Model Calibration
We numerically solve the time path (   ) and pin down the cost of innovation 
by the free entry condition 0 = 0 +  Figure 19 plots the calibration results, which
show the time paths of  ,  and  . The number of firms  grows along a logistic
curve. Meanwhile,  decreases while  increases over time, and the initial difference
0 − 0 pins down the innovation cost  = 684.
5.2.4

Counterfactual analysis

Given the calibrated model parameter values, we then conduct counterfactual analysis
and evaluate welfare.
Optimal compensation share We first evaluate the effect of the compensation
share . Again, we consider two regimes: Regime 1 (under which imitators cannot
resell the innovation) and Regime 2 (under which imitators can resell).
Figure 20(I) shows 0 strictly increases in  for 0 ≤   043 under both regimes,
and 0 is smaller under Regime 2 than under Regime 1. Figure 20(II) shows that under
both regimes, the free entry condition holds in terms of 0 − 0 =  for 0 ≤   043.
For  ≥ 043, because of the corner solution 0 = , the free entry condition holds
in terms of 0 − 0  
Figures 20(III)-(IV) plot social welfare 0 according to different values of 0 and
. The results show that social welfare 0 is hump-shaped in 0 and peaks at 0 = 71
(i.e., 0  = 016). Accordingly, the social optimum can be achieved by choosing
 = 005 under Regime 1 or choosing  = 012 under Regime 2. Note that
neither  = 0 nor  = 1 would be socially optimal: the former yields 0 = 178 and
the latter yields 0 =  = 425.
The finding that  is smaller for the PC industry than for the auto industry
can be explained by the comparative-statics analysis shown in Fig. 10. Compared
to the benchmark calibration of the auto industry, the PC industry has a smaller ,
a higher , and a larger . All these should have led to a larger value of  for
35

the PC industry than for the auto industry. Therefore, the only reason that the PC
industry has a smaller  is because the demand for PCs is more price elastic than
for autos: According to our estimates,  = 006 for the PC industry while  = 03
for the auto industry.
I

1.5

II

30

Regim e 1: k /N
0

Regime 1: v

25

Regim e 2: k /N

Regime 2: v

0

Regime 1: u

0

Regime 2: u

0

0
0

20

1

15
10

0.5

5
0

0

0.2

0.4

0.6

0.8

0

1

III

6000

0

0.2

0.4

0.6

5000

4000

4000

3000

3000

2000

2000
Regime 1: W

1000
0

W
0

0.2

0.4

0.6

0.8

1000

Regime 2: W

0

0

1

1

IV

6000

5000

0.8

0

0.2

0.4

0.6

0.8

0
0

1

k 0/N

Fig. 20. PC Industry: Effects of 
Optimal innovation subsidy Figure 21 plots the effects of a subsidy  on the
entry of innovators 0 and welfare 0 . The results show that 0 increases in  and
the social welfare peaks at  = 58 (i.e., lowering the entry cost from 988 to 408, a
59% reduction).
I
1

II

6 000

k /N
0

0 .9

5 500

0 .8

5 000

0 .7

4 500

0 .6
4 000

0 .5
0 .4

3 500

0 .3

3 000

0 .2
2 500

0 .1
0

0

1

2

3

4

5

2 000

6

W
0

1

0
2

s

3

4

s

Fig. 21. PC Industry: Effects of 
36

5

6

Optimal diffusion rate We can similarly evaluate the effects of varying the diffusion rate . Figure 22 shows that if the planner were to push down  from the original
value where  = 058, the entry of innovators 0 would increase. Eventually, 0
would reach the corner solution 0 =  if the value of  gets sufficiently low (i.e.,
 ≤ 005). However, social welfare 0 increases in , so the social planner would
not want to slow down diffusion.
I

1.2

II

43 00

k0/N

W 0

1

42 00

0.8

41 00

0.6

40 00

0.4

39 00

0.2

38 00

0

0

0 .1

0.2

0 .3

0.4

37 00

0.5

0

0 .1

0 .2

N

0.3

0.4

0 .5

N

Fig. 22. PC Industry: Effects of 
5.2.5

Model re-calibration: A larger pool of potential adopters

We can repeat the exercise by assuming a larger pool of potential adopters. Assume
that the shakeout started in the middle of the diffusion and the number potential
PC firms  = 2 000. In this scenario, it would have taken 30 years to reach 99%
adoption rate among potential producers had the shakeout not happened, doubling
what was assumed in the benchmark calibration. The exercise delivers very similar
results as what we found in the automobile case (See Appendix VI for the details).

6

Conclusion

We modeled an innovation and its diffusion in one industry and discussed policy and
welfare. Capacity constraints imply that licensing raises the revenues of innovators
and that licensing is also socially beneficial to a degree. We showed that the welfare
outcome depends on whether imitators can resell the innovation, and on how much
the innovators are compensated for transferring the innovation.
We found that the socially optimal bargain allocation hinges on the diffusion
process, particularly the congestion externality in meetings between innovators and
imitators. Our analysis also showed that slowing down diffusion encourages innovation and raises initial capacity, but it lowers imitation so that capacity grows more

37

slowly. We argued that this may help explain the overtaking of Route 128 by the
Silicon Valley.
We calibrated the model to data of the U.S. automobile and personal computer
industries. Our empirical findings match well the expansion of firm numbers prior to
the shakeout in each industry and quantify the theoretical predictions of the model.

38

References
[1] Benhabib, J., Perla, J., & Tonetti, C. (2019). Reconciling models of diffusion and
innovation: A theory of the productivity distribution and technology frontier.
Working Paper.
[2] Bishara, N. D. (2011). Fifty ways to leave your employer: Relative enforcement of
covenants not to compete, trends, and implications for employee mobility policy.
University of Pennsylvania Journal of Business Law, 13, 751—795.
[3] Burdett, K., & Coles, M. G. (2001). Transplants and implants: The economics
of self-improvement. International Economic Review, 42(3), 597-616.
[4] Cabral, Luis, Zhu Wang & Daniel Yi Xu (2018). Competitors, complementors,
parents and places: Explaining regional agglomeration in the U.S. Auto Industry.
Review of Economic Dynamics, 30, 1-29.
[5] Chade, H., Eeckhout, J., & Smith, L. (2017). Sorting through search and matching models in economics. Journal of Economic Literature, 55(2), 493-544.
[6] Comin, Diego & Bart Hobijn (2004). Cross-country technology adoption: making
the theories face the facts. Journal of Monetary Economics, 2004, 51(1), 39-83.
[7] Filson, Darren (2001). The nature and effects of technological change over the
industry life cycle. Review of Economic Dynamics, 4(2), 460—494.
[8] Franco, A. M., & Mitchell, M. F. (2008). Covenants not to compete, labor mobility, and industry dynamics. Journal of Economics & Management Strategy,
17(3), 581-606.
[9] Gort, Michael & Steven Klepper (1982). Time paths in the diffusion of product
innovations. Economic Journal, 92, 630—653.
[10] Griliches, Zvi (1957). Hybrid corn: An exploration in the economics of technical
change. Econometrica 25(4), 501-522.
[11] Hopenhayn, Hugo & Liyan Shi (2020). Knowledge creation and diffusion with
limited appropriation. Working Paper.
[12] Information Technology Industry Data Book. 1960—2008. (1998). Washington,
DC: Information Technology Industry Council.
[13] Jovanovic, Boyan & Glenn MacDonald (1994). The life cycle of a competitive
industry. Journal of Political Economy, 102 (2), 322—347.
[14] Klepper, Steven (1996). Entry, exit, growth, and innovation over the product life
cycle. American Economic Review, 86, 562—583.
39

[15] Klepper, Steven (2010). The origin and growth of industry clusters: The making
of Silicon Valley and Detroit. Journal of Urban Economics, 67(1), 15-32.
[16] Lauermann, S., Nöldeke, G., & Tröger, T. (2020). The balance condition in
search-and-matching models. Econometrica, 88(2), 595-618.
[17] Kydland, Finn E. & Edward C. Prescott (1977). Rules rather than discretion:
The inconsistency of optimal plans. Journal of Political Economy, 85(3), 473-492.
[18] Manea, M. (2018). Bottleneck links, essential intermediaries, and competing
paths of diffusion in networks. Working Paper.
[19] Nöldeke, G., & Samuelson, L. (2015). Investment and competitive matching.
Econometrica, 83(3), 835-896.
[20] Samila, S., & Sorenson, O. (2011). Noncompete covenants: Incentives to innovate
or impediments to growth. Management Science, 57(3), 425-438.
[21] Saxenian, A. (1994). Regional advantage: Culture and competition in Silicon
Valley and Route 128. Cambridge, MA: Harvard University Press.
[22] Shi, L. (2019). Restrictions on executive mobility and reallocation: The aggregate
effect of non-compete contracts. Working Paper.
[23] Smith, Philip H. (1970). Wheels within wheels: A short history of American
motor car manufacturing. Funk and Wagnalls, New York.
[24] Stavins, Joanna (1995). Model entry and exit in a differentiated-product industry: The personal computer market. Review of Economics and Statistics 77(4),
571—584.
[25] Thomas Register of American Manufacturers (1958—1999). Springfield, IL:
Thomas.
[26] Thomas, Robert Paul (1977). An analysis of the pattern of growth of the automobile industry 1895—1929. Arno Press, New York.
[27] Utterback, James and Fernando Suarez (1993). Innovation, competition, and
industry structure. Research Policy, 22, 1—21.
[28] Wang, Zhu (2017). Innovation, deregulation, and the life cycle of a financial
service industry. Review of Economic Dynamics, 26, 180-203.
[29] Williamson, Samuel H. (2020). What was the U.S. GDP then? MeasuringWorth.
[30] Young, H. Peyton (2009). Innovation diffusion in heterogeneous populations:
Contagion, social influence, and social learning. American Economic Review, 99,
1899-1924.
40

Appendix I.
Note that with  = 0, we can rewrite Eq. (39) as
 −1 
) +

 + 

which has an unique bounded solution
 = (

 =



+



  ( + )

We can also rewrite Eq. (40) as
¶
µ




+
+
−


 = 


+    ( + )


(56)

(57)

which has an unique bounded solution that is constant over time
 =



( + )

At  = 0, the free entry condition requires 0 − 0 = , thus Eqs. (56) and (57) yield
(1 + ) =  ( + ) 
Given Eq. (27) that  =


0

− 1, we can solve the entry of innovators 0 at date 0:
0 =



( + )

Similarly, we can solve the case for  = 1. We rewrite Eq. (39) as
¶
µ
 −1


) + 
 +

 = ( 

+

+

which has an unique bounded solution :
 =



+

  

We can also rewrite Eq. (40) as
 

 +

 = − 

+

which has an unique bounded solution:  = 0
41

At  = 0, the free entry condition requires 0 − 0 = , thus Eqs. (56) and (57)
yield

Given Eq. (27) that  =


0

(1 + )
= 

− 1, we can solve the entry of innovators 0 at date 0:
0 =





Appendix II.
With  = , we can rewrite Eq. (39) as
µ
¶



+ 
− 1  + (1 + − ) = 0

 +


(58)

which has a general solution
2

 =

¡
¢
 ( +  )
(1−)
 
+
Ψ


+

(2 − )

(59)

Using the boundary condition that →∞  ∞, we can pin down Ψ =
see this, note that Eq. (59) suggests
 =

−
.
(2−)

To

¡
¢
2 + 2 + 2
+ Ψ 1 + − 

(2 − )

Using Taylor expansion, we have



2

=
+
+
(2 − ) (2 − ) (2 − )
µ
¶
1

−
2 −2
1 +  + ( − 1) 
+Ψ
+ 
2
µ
¶

2

+
+ Ψ 
+
=
(2 − ) (2 − )
(2 − )
Ψ
+Ψ + ( − 1)2 − + 
2

To satisfy the boundary condition →∞  ∞, we must have Ψ =
terms involving  sum to 0 and
→∞ =



2
−
=

(2 − ) (2 − )

42

−
,
(2−)

so the

Hence, we can derive the particular solution:
2
¡
¢ ´
 ( +  ) ³
(2−)
 −2
1
−



+

 =
(2 − )

(60)

Accordingly, we have

0 =

¡
¢

( + 1)2 − ( + 1) 
(2 − )

(61)

We can then solve closed-form solution for 0 without involving  . Note that
total discounted revenue is
Z ∞

0 =
−  =
(62)

0
This is shared by the two groups — the initial incumbents 0 and the outsiders  −0 :
0 = 0 0 + ( − 0 ) 0
= 0 + 0 (0 − 0 )
= 0 + 0 
or equivalently,
0 = 0 −  ( − 0 ) 

Then substituting for 0 from Eq. (62),

0 −  ( − 0 ) =




and substituting for 0 from Eq. (61), we get

Since  =


0

¡
¢


( + 1)2 − ( + 1) −  ( − 0 ) = 
(2 − )


− 1, this yields


 −
µ³ ´ 0 ³ ´ ¶
= 
2


0

−1
− 0
(2−)(−0 )
0
Appendix III.
Eq. (51) suggests that
0

=
+

( + )2

Z

0

∞

−



 − 
43

Z

0

∞

−





(+)
−


 +
 +

1



(63)

We first verify that given   0 and   0
µ
¶

 + 
  (+)
 +  −1
 − 1 −  ( + )
=
0
2

[ + (+)
−
1]





(+))
 0).
for any   0 (note that  − 1 −  ( + ) = 0 for  = 0 and ( −1−



Recall Eq. (50) that 0 = (+)
. For any 1  0  0, this requires   +
, so we
have
Z ∞
Z ∞
1
 + +
 + 
−
−







(64)

 + (+)
−
1
0
0


Therefore, Eqs. (63) and (64) yield that
Z ∞
Z ∞
1
 + +
0

−
−

+
  − 




( + )2
0
0
Z ∞
Z ∞


−
  − 
−  −
= 0
=
2 +
( + )
( + )2
0
0
This also implies that for 0 = 1, we have  =


,
+

so

0


= 0.

Appendix IV.
We rewrite the model into the discrete-time version and then solve it numerically.
Model calibration:
 =


 + 

where  =


− 1  0;
0

 = − ;
 =  +
 =

1
+1 ;
1+

1
( +1 + (1 −  )+1 ) 
1+

Welfare:
0 = −0 +

∞
X
=0

1
(1 + )
44

µ

¶
 1−


1− 

Counterfactual analysis:
Regime 1. Imitators cannot resell ideas.
Denote  the value of an imitator,  the value of an innovator, and  the value
of an outsider at date , respectively.

 = 

+

where  =


− 1  0;
0

 = − ;
  =  +
 =

(

 +
 +

 =

1
 +1
1+

1

1+ +1
−1 ( −−1 )

0

+

for  ≥ 1;
1

1+ +1

for  = 0
for  ≥ 1;

1
( (1 − )+1 + (1 −  )+1 ) 
1+

Regime 2. Imitators can resell ideas.
Denote   the value of a new imitator,  the value of an incumbent, and  the
value of an outsider at date , respectively.
 =


 + 

where  =


− 1  0;
0

 = − ;
  =  +

1
+1
1+

for  ≥ 1;

⎧
1
⎨  + 1+ +1
1
+1
 + ( − −1 )  + 1+
 =
⎩
1
= [1 + ( − −1 )]( + 1+
+1 )
 =

for  = 0
for  ≥ 1;

1
( (1 − )+1 + (1 −  )+1 ) 
1+

Appendix V.
We extend the large industry model to allow firms to exit at a constant rate 
every period, and the analysis is very similar.
45

Assume that an innovation takes a logistic diffusion process among  potential
adopters. Each period, a fraction  of firms which have adopted the innovation
exit the industry (i.e., unadopt the innovation). Once an incumbent firm leaves the
industry, it will be replaced by a new entity entering the pool of potential adopters,
so that the total number of potential adopters remains .
The law of motion for firm numbers is

=  ( −  ) −  

Solving the ODE yields the solution of  :
 =

³
1 + ( −
0

 − 
´ ´
´
³ ³
− 1) exp −  −  

(65)

and there exists a steady state →∞ →  where

 =  − 

Substituting  =  +




into Eq. (65) we obtain

 =
1+

³


0


´

− 1 exp (− )

(66)

which is the same as the diffusion equation in our original model except that  is
replaced by  .
Diffusion estimation.–To estimate the extended model, we can rewrite Eq. (66)
as
µ
¶
µ
¶

0
ln
= − ln
+  
(67)
 − 
 − 0

Assume that the industry has achieved the steady-state number of firms before the
shakeout, we assign  = 210 (i.e., the observed peak number of firms). As a result,
we are back to the exactly same regression model as before, and all the estimation
0
results remain unchanged:  = 053 and 0 = 331 (i.e., ln −
= −413).
0

Model calibration.–Denote  the value of an incumbent, and  the value of an
outsider at date , respectively. As before, we assume that the industry data are
generated by the model with free spillover (i.e.,  = 0), so it makes no difference
whether imitators can or cannot resell the technology.
The model is now represented by the following equations:
 =


  

where

=
− 1  0;


  + 
0
46

 = − ;
1−
+1 ;
 =  +
1+
1
( +1 + (1 −  )+1 ) 
1+
We calibrate the model parameter values as before:  = 210  = 053
0 = 331  = 03 We assume the exit rate  = 015. Again, we normalize  = 1
and assume the discount parameter  = 005
 =

I

1
0.9

k /k
t

k /k
0.8

t

ss
ss

II

2.4

(m o d el)

v

2.2

(d ata )

u

t
t

2

0.7

1.8

0.6
1.6
0.5
1.4
0.4
1.2

0.3

1

0.2

0.8

0.1
0
1 895

19 00

19 05

0.6
1 895

191 0

19 00

19 05

191 0

Fig A1. Auto Industry Calibration: Model Extension
Figure A1 plots the calibration results. The difference of 0 − 0 pins down the
innovation cost  = 159 The estimated value of  is smaller than what we got in the
original model calibration because firms take into account the exit rate (i.e., firms
have zero scrape value when they exit).
Welfare.–The welfare equation remains the same as our original model.

Appendix VI.
We assume a larger pool of potential PC firms. The exercise delivers very similar
results as what we found in the automobile case.
Assuming  = 2 000, we re-run the regression model (52). The result shows that
ln


= −609 + 034 
 −  (027)∗∗∗ (003)∗∗∗

and the standard errors are reported in the parentheses.
47

(68)

I

0 .45

II

18

k t/ N (m o d e l)

0.4

vt
16

k / N (d a t a )
t

u

t

0 .35
14
0.3
0 .25

12

0.2

10

0 .15
8
0.1
6

0 .05
0
19 75

19 80

19 85

4
19 75

1 990

19 80

19 85

1 990

Fig. A2. PC Industry: Model Re-Calibration
Based on the estimates of diffusion parameters, we calibrate  = 2000  = 034
0
= −609). In addition, we set  = 006,  = 005, and
and 0 = 451 (i.e., ln −
0
normalize  = 1 as before. Figure A2 plots the calibrated time paths of  ,  and
 . The initial difference 0 − 0 pins down the innovation cost  = 975.
I

1.5
Reg im e 1: k 0/N
Reg im e 2: k /N

Reg im e 1: v 0
Reg im e 2: v

25

0

1

II

30

Reg im e 1: u 0
Reg im e 2: u

0

20

0

15
10

0.5

5
0

0

0 .2

10

0 .4

4

0 .6

0 .8

0

1

III
2

1.5

1 .5

1

1

0.5

0

0 .2

0.4

4

Reg im e 2: W

0

0 .4

0 .6

0 .8

0

1

0

0.6

0.8

1

0.6

0.8

1

IV

Reg im e 1: W

0 .5
W

0.2

10

2

0

0

0.2

0.4

0
0

k0 /N

Fig. A3. PC Industry Model Re-Calibration: Effects of 
Figure A3 plots the results of varying the value of , which is consistent with
our findings from the auto case. First, given a slower diffusion speed , the entry
of innovators 0 does not reach the full adoption unless  ≥ 07. Second, welfare
maximizes at 0 = 300 (i.e., 0  = 015), which can be achieved by choosing
48

 = 009 under Regime 1 or  = 020 under Regime 2. Compared to the
benchmark calibration, the optimal compensation share  becomes larger in either
regime, which is driven by the larger , the lower , and the higher  from the recalibrated model.
Figure A4 plots the effects of a subsidy  on the entry of innovators 0 and welfare
0 . The results show that 0 increases in , and the social welfare peaks at  = 58
(i.e., lowering the entry cost from 975 to 395, a 59% reduction).
I
1

k 0 /N

0.9

10

2.2

4

W0

2

0.8

1.8

0.7

1.6

II

0.6
1.4
0.5
1.2

0.4
0.3

1

0.2

0.8

0.1
0

0.6
0

2

4

6

8

0

2

4

6

s

8

s

Fig. A4. PC Industry Model Re-Calibration: Effects of 
Finally, Fig. A5 plots the results of varying the value of . As before, we find
that the entry of innovators 0 decreases in , while social welfare increases in .
I

1.2

10

k 0 /N

4

W0

1.2 5

1

II

1 .2
1.1 5

0.8

1 .1
0.6
1.0 5
1

0.4

0.9 5
0.2
0 .9
0

0

0 .05

0.1

0.15

0.2

0.2 5

0.8 5

0 .3

N

0

0.05

0 .1

0.1 5

0 .2

0 .25

N

Fig. A5. PC Industry Model Re-Calibration: Effects of 

49

0.3