Economies of Density and Congestion
Economies of Density and Congestion
A Dissertation
Presented to the Faculty of the Graduate School
of Cornell University
by
Haimeng Zhang
August 2022
© 2022 Haimeng Zhang
ALL RIGHTS RESERVED
UNDERSTANDING FIRMS’ STRATEGIC BEHAVIORS
AND THEIR IMPLICATIONS
Haimeng Zhang, Ph.D.
Firms are key participants in the market. Compared to individual consumers, firms,
particularly those with significant market power, are often seen by economists as more
resourceful to overcome information frictions, more likely to retain past information and
less susceptible to bounded rationality. As a result, firms are capable of acting as sophis-
ticated strategic players. Under the assumption of being driven by a clearly defined profit
motive, firms may behave strategically when interacting with other market participants,
including consumers, other firms and governments. Understanding firms’ strategic behav-
iors is a critical step in assessing market efficiency, analyzing the wellbeing of consumers,
This dissertation consists of three essays that analyze firms’ strategic behaviors in
different settings. Chapter 1 studies firms’ strategic interactions in government organized
spectrum auctions. In these ascending-bid auctions, firms as bidders are able to communi-
cate their private information with one another using jump bidding as signals. The signals
are credible since bidders with lower private information incur a higher ex ante cost for
choosing a jump bid with any given size. This prevents the bidders with lower private
information from mimicking those with higher private information. In equilibrium, the
signaling model predicts lower expected revenue to the seller, in this case the government,
than in the “open exit” model in which jump bidding is not allowed. Using data from a
spectrum auction held by the Federal Communications Commission in the United States,
the mean valuation estimated using the signaling model is higher compared to that of
the open exit model. This implies that if bidders are indeed using jump bids as signals,
ignoring it leads to estimates of the mean values that are biased downwards. This result
is consistent with the prediction of the theoretical model that bidders pay lower prices
with jump bidding than in an open exit auction. I estimate that if jump bidding was
prohibited, the government could have had 8% higher revenues from the auction.
agricultural equipment rental markets in India. We observe that private rental firms favor
farmers located in dense areas and demanding higher machine-hours because equipment
needs to be moved in space. Using our own census of 40,000 farmers, we document
that costly delays and price dispersion in rentals are ubiquitous, and that small-scale
farmers are rationed out by private rental firms. This rationing could be detrimental to
aggregate productivity if small farmers have the highest marginal return to capital. A
government subsidized first-come-first-served dispatch system grants small-scale farmers
timely access to equipment at the expense of travel time. In a calibrated model of frictional
rental services, optimal queueing and service dispatch we show that, while the constrained
efficient allocation prioritizes large-holder farmers, small-scale farmers in dense areas are
valuable because they help maximize capacity utilization. Through counterfactuals, we
show that when the induced increase in subsidized equipment supply is high enough,
service finding rates for small-farmers increase relative to large-holders farmers even when
providers prioritize large-scale.
ones. Motivated by the Phoebus cartel, whose reason for engaging in planned obsolescence
cannot be explained by existing theories, I introduce a new theory that centers on an
important concept from behavioral economics: present-biased preferences. I construct
a theoretical model which demonstrates that when consumers are present biased, that
is, when they exhibit time-inconsistent preference in favor of immediate gratification, a
monopolist chooses a profit-maximizing level of durability below that chosen by a perfectly
Haimeng Zhang grew up in China and Singapore. She holds a Bachelor of Arts degree in
Economics and Management from the University of Oxford and a Master of Arts degree
in Economics from New York University.
Before attending graduate school, Haimeng worked for 5 years in investment banking at
UBS Investment Bank and Moelis & Company based in London, covering the telecoms,
media, and technology sectors. She also interned as a research assistant to the chief
economist at the Antitrust Bureau of the New York State Attorney General’s Office.
iii
I dedicate this dissertation to my loving parents, Aihua Li and Yaxin Zhang, who have
wholeheartedly supported every life decision I made.
iv
ACKNOWLEDGEMENTS
I would like to thank my dissertation committee co-chairs, Panle Jia Barwick and Michael
Waldman, and committee members Jörg Stoye, Giulia Brancaccio and Thomas Jungbauer.
Without your invaluable advice and guidance, I would have been at a loss.
I would also like to thank Julieta Caunedo, Namrata Kala, Penny Sanders, Seth
Sanders, Keith Jenkins, Marcel Preuss, Tommaso Denti, Nahim Bin Zahur, Qinshu Xue,
and all participants at various internal and external seminars, workshops, and conferences.
Your constructive feedback has greatly improved the essays in this dissertation.
Last, I would like to thank my family, friends and colleagues for their support. You
have made this Ph.D. journey the most wonderful and memorable.
v
TABLE OF CONTENTS
vi
2.3.1 Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.2 Queue lengths as strategies . . . . . . . . . . . . . . . . . . . . . . . 58
2.3.3 Market cost of provision. . . . . . . . . . . . . . . . . . . . . . . . . 60
2.3.4 Service Providers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4 Symmetric Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.4.1 Social optimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.5 Government regulation in equipment rental markets . . . . . . . . . . . . . 66
2.5.1 Bringing the model to the data. . . . . . . . . . . . . . . . . . . . . 66
2.5.2 Status quo equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.5.3 Accommodating empirically relevant heterogeneity . . . . . . . . . 75
2.6 The value of the subsidies . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
2.6.1 The market prior to the subsidy . . . . . . . . . . . . . . . . . . . . 81
2.6.2 Market Deregulation . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
vii
B.2.1 Proposition 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
B.3 Numerical Solution and Output . . . . . . . . . . . . . . . . . . . . . . . . 131
B.3.1 Value function computation . . . . . . . . . . . . . . . . . . . . . . 131
B.3.2 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
B.4 Additional Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . 132
Bibliography 137
viii
LIST OF TABLES
B.1 Costs of Delays Relative to Optimal Planting Time, Value Added per Acre 132
B.2 Status Quo Equilibrium Outcomes . . . . . . . . . . . . . . . . . . . . . . 134
ix
LIST OF FIGURES
x
CHAPTER 1
JUMP BIDDING AS A SIGNALING GAME
1.1 Introduction
Jump bidding in an ascending-bid auction refers to the action of placing a bid in excess
1
increases by a small fixed amount set by the auctioneer. Bidders choose at which prices
to drop out of the auction irrevocably. The open exit model leaves no scope to study
the behavior of jump bidding, as bidders do not submit bids. In alternative models of
the ascending-bid auction where bidders actively submit bids, jump bidding is weakly
dominated by the strategy of bidding the minimum required price each round in the
absence of two elements: transaction costs and signaling using bids between bidders. The
two main theoretical strands of literature which explain jump bidding are concerned with
these two elements.
This paper focuses on the informational role of jump bidding in an ascending-bid auc-
tion. There is substantial anecdotal evidence that bidders engage in jump bidding to
intimidate their competitors so that they drop out of the auction sooner. In the earlier
spectrum auctions organized by the Federal Communications Commission (“FCC”), jump
bidding was pervasive. For example, in the very first FCC auction (the nationwide nar-
rowband PCS auction) in 1994, 49% of all new high bids were jump bids that exceeded the
high bid by more than two bid increments (Cramton, 1997). While the transaction cost
theory could be applicable in this case, many economists believed that there were other
factors at play given the large sizes of some of the jump bids and the fact that some were
placed towards the end of the auctions. They suspected that bidders were using jump
bids as signals of high valuation or “toughness” (McAfee and McMillan, 1996; Cramton,
1997).
In 1997, the FCC introduced click box bidding which limited the size of a jump
bid. One of the motivations was arguably to reduce the potential anticompetitive effects
of jump bidding. Before this change in auction rules, 15 spectrum auctions had been
conducted by the FCC, bringing in a total revenue of $23 billion. Given the high stakes,
if bidders were indeed using jump bids as signals and were successful in reducing the
prices paid due to early withdrawal of competitors, the revenue effect could be significant.
2
Any attempt to quantify this effect will require a theoretical framework that describes a
signaling equilibrium in an ascending-bid auction.
Avery’s 1998 paper provides an excellent starting point. The paper solves for equilibria
of ascending-bid auctions with two symmetric bidders and affiliated values when jump
bidding strategies may be employed as signals. In this model, each bidder receives a private
observation, which is a noisy signal of her true valuation of the auctioned object. The
private observations are strictly affiliated, meaning a large realized private observation for
one bidder makes the other bidder more likely to have large realized private observations
as well. The paper shows that if bidders are allowed to have a one-off opportunity to
simultaneously choose from a discrete set of jump bids as their opening bids and continue
according to prespecified asymmetric equilibrium in favor of the higher bidder, then there
exists a unique symmetric signaling equilibrium where the size of the opening bid is weakly
monotone in a bidder’s private observation. In other words, the choices of jump bids are
partially separating. The bidder with the higher private observation chooses a weakly
larger jump bid. Each possibility of jump bidding provides a Pareto improvement for the
bidders from the symmetric equilibrium of a second-price sealed-bid auction.
While Avery’s paper provides a useful framework to empirically study the use of jump
bidding as a signaling device, it has a number of stylized features that are incompatible
with most empirical settings. It is a model limited to an auction with 2 bidders. More
important, it only allows bidders a one-off opportunity to send signals through jump
bidding at the beginning of the auction. Bidders’ abilities to send signals in multiple
rounds will change the predictions of the model.
This paper builds upon Avery’s model and extends it to more than two players. By
focusing on the equilibrium in which bidders are willing to drop out of the auctions as
soon as they find out their private observations are not the highest, and extending the
3
jump bid space from a bounded discrete set to a continuous set without an upper bound,
the 2-player 2-stage game is transformed into a first-price sealed-bid auction.
signaling model works like a hybrid model of a first-price sealed-bid auction and an open
exit ascending-bid auction. To structurally estimate such a model involves estimating
each of these two auction formats.
The data used in the structural estimation are from the FCC broadband PCS auction
(C-block), or “Auction 5” that took place between December 1995 and May 1996. In
this auction, the U.S. was divided into 493 regional markets and one license was offered
for each market. The reason I chose this auction is two fold. First, Auction 5 brought
in over $10 billion of revenue, the highest among all 15 FCC spectrum auctions before
jump bids were restricted by the introduction of click box bidding. Second, only small
businesses (defined as those with annual revenues less than $40 million) were eligible to
participate. Compared to other auctions where the sizes of the participating firms are
much more heterogeneous, Auction 5 is more suited to be described with the theoretical
4
model) . One would expect the value of a spectrum license to consist of both a common
value component to reflect market attributes that are valued by all bidders, such as
market size measured by population, and a private value component to allow values to
differ across bidders. This suggests that a common value model, which is nested into the
affiliated value framework of the theoretical model, is the most appropriate,.
auction (Athey and Haile, 2007). Given the challenges with nonparametric identification,
I adopt a parametric approach by choosing a multiplicative specification and making
distributional assumptions following Hong and Shum (2003).
The theoretical model places bounds on the bids observed. While a partial identi-
fication approach may be the most appealing, the implementation of such an approach
is difficult. Haile and Tamer. (2003) point out that the lack of sufficient structure of
an ascending-bid auction makes a mapping between the bids observed and the underly-
ing demand structure challenging. They instead construct bounds on observed bids and
partially identify the distribution functions using an independent private value model.
(1996), Paarsch (1997) and Hong and Shum (2003) where the last bid placed by each
bidder (except the winner of the auction) is equal to the upper bound predicted by the
theoretical model. This approach allows for point identification.
nonlinear least-squares approach following Laffont, Ossard and Vuong (1995) and Hong
and Shum (2003). The estimation involves the computation of the equilibrium strategies
5
for an open exit auction and a first-price sealed-bid auction respectively. The former
has a closed form solution under the parametric assumptions. However, the equilibrium
strategy of a first-price sealed-bid auction does not have a closed form solution. I compute
using jump bids as signals, ignoring it leads to an underestimation of the mean value.
A counterfactual analysis using the estimation results from the signaling model suggests
that if the FCC forbade the action of jump bidding and designed the auction following the
open exit format, the total revenue would have been 8% higher. The result is consistent
with the prediction of the theoretical model: by sharing information and coordinating
among themselves using jump bids as signals, bidders are able to lower the prices they
quences if the seller is the government, as in the case of a spectrum auction. Theoretical
models predict the direction of this revenue effect. It is up to the empirical research to
quantify it. There is a growing body of empirical literature that studies the relation be-
tween auction revenue and jump bidding using reduced form methodologies, which I will
discuss in more details in the next section. In short, these papers find a revenue effect
in the opposite direction as predicted by the theoretical models potentially due to the
The remainder of the paper is organized as follows: Section 2 reviews the literature
6
relevant to this paper; Section 3 introduces the theoretical model; Section 3 discusses the
empirical strategy and identification; Section 4 describes the estimation methodology and
presents the results; Section 5 discusses the counterfactual analysis; Section 6 concludes.
As briefly discussed earlier, there are mainly two strands of the theoretical literature
that explain jump bidding. The first deals with transaction costs. If bidders incur a
transaction cost every time a bid is placed, placing fewer but larger bids could reduce
this cost. The transaction costs could be either pecuniary costs associated with revising
and submitting a bid (Fishman, 1988; Daniel and Hirshleifer, 1998), or time costs due
to bidders’ impatience (Isaac, Salmon and Zillante, 2007; Kwasnica and Katok, 2009).
This paper treats all costs of participating in an auction as sunk and assumes away any
The second strand of literature is concerned with the informational role played by
jump bids when bidders interact strategically. Jump bids are used as signals of private
information either to deter the entry of new bidders (Easley and Tenorio, 2004) or to in-
duce early withdrawal of existing bidders (Avery, 1998; Hörner and Sahuguet, 2007). This
paper contributes to the theoretical literature by extending Avery’s well-known signaling
model to a form that is more compatible with auction settings in real life. It abstracts
away from the effect of entry deterrence of an opening jump bid and treats entry as ex-
ogenous. Instead of information sharing, jump bids could also be used as tools to conceal
information when some bidders have an information advantage over others (Ettinger and
7
vate values is vast. Hickman, Hubbard and Sağlam (2012) presents a comprehensive
survey of structural econometric methods in auctions. In addition, Perrigne and Vuong
(1999) survey on structural econometrics of first-price auctions and Athey and Haile (2007)
This paper is the first to analyze the revenue effect of jump bidding in ascending-
bid auctions using a structural approach. The empirical research on jump bidding in
spectrum auctions is limited and adopts a descriptive approach. McAfee and McMillan
(1996) and Cramton (1997) analyze the first 3 and 6 FCC spectrum auctions respectively.
Both papers document prevalent jump bidding behavior in these auctions but argue that
it had little effect in deterring competition since most of the jump bids were eventually
overtaken.
Based on the signaling model of this paper, whether an auction ends with a jump bid
depends on the realization of the private observations of the top 2 bidders. If they are
sufficiently dispersed, the auction ends with a jump bid with probability 1. Otherwise
the auction ends with a jump bid with a positive probability. While it is true that if
an auction does not end with a jump bid its revenue is the same as the one in an open
exit model, the revenue reduction for those that do end with a jump bid is significant as
demonstrated by the counterfactual analysis.
Jump bidding aside, there are two other papers that adopt a structural approach
8
in estimating value distributions in an FCC spectrum auction. Hong and Shum (2003)
develop an econometric model of ascending-bid auctions with affiliated values and bidder
asymmetries. Fox and Bajari (2013) estimate the deterministic component of bidder
There is a growing body of empirical literature on jump bidding in the Swedish and
Norwegian housing markets, where houses are sold through broker-assisted ascending-
bid auctions (Hungria-Gunnelin, 2018; Sommervoll, 2020; Khazal et al., 2020; Sønstebø,
Olaussen and Oust., 2021). Using a reduced form approach, these papers investigate the
motivation behind jump bidding and its effect on entry and withdrawal of bidders and
the selling prices. The key findings are: 1) bidders’ primary motive of jump bidding
the signaling model of this paper, the third one appears to contradict its prediction. There
are two potential explanations. First, the OLS regressions may not have controlled for all
important attributes of the houses. It is even harder to control for the heterogeneity in
individual tastes. Second, the participants in the housing markets are individuals, who are
likely more susceptible to bounded rationality than the bidders in the spectrum auctions
which are firms.
1.3.1 Timing
9
1. Before the auction, the auctioneer sets a reservations price ρ1 = 0.
3. At the end of round 1, all bidding information, including the value and identity
of the bidder associated with each bid, is made public to all bidders. Each bidder
(except the one(s) with the highest bid) publicly announces whether to drop out.
The auction ends if only 1 bidder remains, who then becomes the winner and pays
her bid.
4. If more than 1 bidder remains, the auctioneer sets the minimum required price for
round r = 2 using the function ρ2 = Pmin (p1 ), where p1 is the highest bid in round
1.3.2 Assumptions
There are n ≥ 2 risk neutral bidders taking part in an ascending-bid auction of a single
object. Each bidder i values the object at Ui , but does not observe Ui directly. Instead,
each receives a scalar private observation X̃i about the object. X̃i ’s are identically dis-
tributed over the support (0, X̄) and are strictly affiliated. Intuitively, strict affiliation
means that large realized values for some of the variables make the other variables more
likely to be large than small1 . Formally, let z and z ′ be points in Rn . Let z ∨ z ′ denote
the element-wise maximum of z and z ′ , and let z ∧ z ′ denote the element-wise minimum.
Strict affiliation requires that for all z and z ′ ,
f (z ∨ z ′ )f (z ∧ z) ≥ f (z)f (z ′ ). (1.1)
1
For a general definition of affiliation, see the Appendix of Milgrom and Weber (1982).
10
Bidder valuations are affiliated in the sense that bidder i’s expected value of the
object Vi is a function of private observations of all bidders Vi = v(xi , {xj }j̸=i ) = E[Ui |
xi , {xj }j̸=i ], where v is continuous and increasing in each argument. By assumption,
the value function v(·) is the same for all bidders. This value function encompasses
the common value case (which consists of both a private value component and a common
value component) and two special cases: the private value case and the pure common value
case. For the private value case, Vi = v(xi , {xj }j̸=i ) = v(xi ), where bidder i’s expected
value only depends on her own private observation. For the pure common value case,
Vi = v(xi , xk , {xj }j̸=i,j̸=k ) = v(xk , xi , {xj }j̸=i,j̸=k ), ∀k ̸= i, that is, private observations
of all bidders (including bidder i) enter into bidder i’s value function symmetrically.
In contrast, when both a private value component and a common value component are
present (i.e. the common value model), a bidder’s own private observation enters her value
function differently from those of other bidders. Assume one places weakly more weight
on her own private observation than her rival bidders’, i.e. Vi = v(xi , xj , {xk }k̸=i,k̸=j ) ≥
Vj = v(xj , xi , {xk }k̸=i,k̸=j ) iff xi ≥ xj .
A jump bid is one that is “substantially” higher than the minimum required price
set by the auctioneer for that round. Formally, a jump bid in round r is defined as
b ≥ ρr + κ̄(ρr ), where κ̄(ρr ) ≫ 0, and is known to all bidders.
Placing a jump bid is not cheap talk because the bid may win the auction. Jump bids are
costly to bidders since by placing a jump bid, bidders forego the possibility of winning at
a lower price. However, if the cost of placing a jump bid (of the same size) differs across
different “types” of bidders, it has the potential of being used as a signaling tool. In a
symmetric equilibrium where the strategy is strictly increasing in the private observation
11
x, a bidder prefers to lose if she can be convinced that there exists another bidder with
a private observation higher than hers because winning requires her to pay above her
expected value. If a jump bid of a given size is more costly to bidders with lower x′ s than
those with higher x′ s, then it can potentially be used as a signaling tool of the underlying
value.
Before presenting the general model, I will first use a simple two-player, two-stage
game taken from Avery’s 1998 paper to illustrate the intuition behind using jump bids as
signals for underlying values when bidder values are affiliated.2 I will demonstrate that
affiliation in private observations makes it more costly ex ante for the player with a lower
x to place a jump bid than the one with a higher x.
The assumptions of this game follow the previous section. Additionally, assume
Pmin (pr ) = pr + ϵ, where ϵ is a small positive number, i.e. the auctioneer raises the
minimum required price by a small amount from round to round. This is a standard
assumption for an ascending-bid auction.
The game (referred to as the “metagame” hereafter) proceeds in two stages. In the
first stage, the bidders simultaneously choose between two choices of bids: an ordinary bid
0 and a jump bid K > 0. The second stage resembles a standard ascending-bid auction,
with the reservation price equal to the maximum bid from the first stage. In this stage,
jump bids are not seen as signals by all bidders.
Let S ∗ (x) denote a symmetric strategy played by both players at the second stage
of the game, which represents the drop-out price in the ascending-bid auction. S ∗ (x) is
assumed to be increasing in x. Note that this strategy does not predict how a bidder
behaves from round to round, but simply at which price the bidder will drop out. One
example of such a symmetric strategy is S ∗ (x) = v(x, x), derived by Milgrom and Weber
2
See Avery (1998), Sections 3 and 4 for a more general discussion and more details.
12
(1982). There is an obvious subgame perfect equilibrium where both players bid 0 in the
first stage, and follow S ∗ (x) in the second stage. However, there exists another class of
subgame perfect equilibria which deliver higher ex ante payoff to the bidders.
Let Sa (x) denote a bidding strategy such that Sa (x) < S ∗ (x) and S ∗ (x) − Sa (x) is
increasing. It can be shown that there exists a threshold x∗ ∈ (0, X̄) such that the
following symmetric strategy is a subgame perfect equilibrium of the two-stage metagame:
in the first stage, bid K if x > x∗ , bid 0 otherwise; in the second stage, play Sa (x) if outbid
by the opponent in the first stage, play S ∗ (x) otherwise. In this equilibrium, if the actions
in the first stage are symmetric/asymmetric, the strategies played in the second stage are
also symmetric/asymmetric. Table 1.1 summarizes the bidding functions of the second
stage as a function of the first stage bidding. In addition, the threshold x∗ is unique for
a fixed two-tuple {Sa , K}.
To prove this strategy is a subgame perfect equilibrium, I will start from the second
stage of the metagame. Suppose it is common knowledge that x1 > x2 , then the strategy
pair (S ∗ (x1 ), Sa (x2 )) form an equilibrium for the second stage. The reason is given x1 > x2 ,
bidder 1 prefers to win since V1 = v(x1 , x2 ) > v(x2 , x2 ) = S ∗ (x2 ) > Sa (x2 ). Similarly,
bidder 2 prefers to lose since V2 = v(x2 , x1 ) < v(x1 , x1 ) = S ∗ (x1 ). Neither bidder has an
incentive to deviate, the strategy pair is therefore an equilibrium of the second stage.
Given the equilibrium of the second stage, it remains to show that there exists a
(partially) separating signaling equilibrium in the first stage with threshold x∗ . Assume
bidder 2 plays the following strategy: in the first stage, bid K if x2 > z, z is fixed and is
public knowledge; bid 0 otherwise. In the second stage, play Sa (x2 ) if outbid by bidder 1
13
S ∗ (x2 )
x’ z
x2
in the first stage; play S ∗ (x2 ) otherwise. Conditional on x2 < z, bidder 1 is able to induce
bidder 2 to shift from a bidding strategy S ∗ (x2 ) to Sa (x2 ) in the second stage by placing a
jump bid K. As shown in Figure 1.1, the yellow path P K (x2 ) = max(K, Sa (x2 )) represents
the price path faced by bidder 1 triggered by an unmatched jump bid. If x2 < x′ , bidder
1 is better off not placing a jump bid and face S ∗ (x2 ) since S ∗ (x2 ) lies below P K (x2 ). If
x′ < x2 < z, bidder i prefers to place a jump bid and faces the lower curve P K (x2 ). Since
X̃1 and X̃2 are strictly affiliated, then intuitively the larger the private observation x1 ,
the more likely that x2 lies between x′ and z, the more likely that bidder 1 chooses to
ϕ(x, z) measures the expected reduction in price that bidder 1 has to pay conditional
on placing a jump bid unmatched by bidder 2. ϕ(x, z) > 0 implies that jump bidding
by bidder 1 would be profitable conditional on bidder 1 winning. Since S ∗ (x) − Sa (x)
14
is increasing, S ∗ (x) − P K (x) is also increasing and continuous. As a result, ϕ(x, z) is
continuous and strictly increasing in both arguments by the property of strict affiliation
between x1 and x2 . In addition, as depicted in Figure 1.1, ϕ(x, z) is negative for x near
zero and positive for large x. By the intermediate value theorem, there is a unique value
x∗ such that ϕ(x∗ , x∗ ) = 0. Update bidder 2’s strategy by replacing the threshold z with
x∗ . To complete the proof, it remains to show that player 1 does not want to deviate from
a symmetric strategy that also uses x∗ as the jump bid threshold in the first stage. There
are 4 scenarios to consider: 1) x1 > x∗ ≥ x2 ; 2) x1 > x∗ , x2 > x∗ ; 3) x1 ≤ x∗ , x2 ≤ x∗ ;
4) x1 ≤ x∗ < x2 . Scenario 1 is straight forward to show. Since x1 > x2 , bidder 1 wins
the auction regardless of her choice in the first stage. However, if she chooses to jump
bid, the expected reduction in price is positive, i.e. ϕ(x1 , x∗ ) > 0. Bidder 1 therefore does
not deviate from the decision to jump bid. For the remaining 3 scenarios, Avery (1998)
The equilibrium in the first stage is partially separating as all bidders with x′ s below
threshold x∗ (“low-value” types) choose one action and those with x′ s above (“high-value”
types) choose the other. Low-value types are not able to mimic high-value types because
they face a higher ex ante cost attributed to the property of strict affiliation of private
observations. If the outcome in the first stage is asymmetric, the bidder that chooses
not to jump bid learns that her private observation is lower than her opponent’s. As a
result, she prefers to lose the auction than to win it. She is therefore indifferent between
strategies S ∗ and Sa .3 When Sa ≤ K, she chooses to drop out right away at the end of the
first stage. The willingness of lower types to drop out at prices lower than the equilibrium
prices without signaling results in lower expected revenue to the auctioneer and higher
expected profit to the bidders.
3
If a small transaction cost of ϵ is introduced every time a bid is placed, the bidder will strictly prefer
Sa .
15
With independent private observations, there remains a unique signaling equilibrium
with jump bids for each two-tuple {Sa , K}. However, this equilibrium is pooling, instead of
partially separating. Both bidders choose the same action in the first stage, always leading
to a symmetric outcome in the second stage. The expected revenue to the auctioneer is
therefore equivalent to that under an ascending-bid auction without signaling. This result
is in line with the what revenue equivalence theorem predicts.4
The simple 2-player 2-stage game has demonstrated that the existence of a partially sepa-
rating signaling equilibrium is directly attributed to the property of strict affiliation. The
same intuition applies when the auction is extended to allow for endogenous jump bids.
Proposition 1.1 by Avery describes the conditions for the existence and the characteristics
of a unique symmetric signaling equilibrium of a 2-player 2-stage game with endogenous
jump bids in the first stage.
Proposition 1.1. (Avery).5 Suppose that the two bidders choose their opening bids en-
dogenously and simultaneously from a fixed set of n possible bids and that they continue
according to a prespecified asymmetric equilibrium in favor of the higher bidder. Then
there is a unique symmetric signaling equilibrium, with strategies identical to those strate-
gies for an n-stage descending signal game with the same set of possible jump bids.
I will broadly outline the idea behind this proposition. For a detailed proof, see Avery
(1998). First consider a fixed set of jump bids {K1 , K2 , ..., Kn } where K1 > K2 > ... >
Kn > 0. Unlike in the two-stage game, bidders may now signal in more than one round.
4
The revenue equivalence theorem specifies a class of auctions that would generate the same expected
revenue to the seller, in particular the independent private values model must apply. See Milgrom and
Weber (1982), Theorem 0.
5
This corresponds to Theorem 4.7 in Avery (1998).
16
In round 1, bidders face a set of {0, K1 }. If at least one bidder chooses K1 , no more
signaling is allowed and bidders proceed to play S ∗ or Sa as in the simple game. If both
bidders choose 0, then they proceed to round 2 and face the new choice set {0, K2 }.
Bidders can now signal up to n rounds. The threshold x∗r in round r is determined the
same way as in the simple game and x∗r is decreasing in r. The subgame that starts from
the last round of signaling is identical to the simple game. Conditional on the existence
of a unique symmetric equilibrium for this subgame, adding one more signaling round to
the beginning of this subgame retains the structure of the simple game and the proof that
a unique symmetric equilibrium exists for the bigger subgame is identical to the proof of
the simple game. Hence by iteratively adding a signaling round to the beginning of the
expanded subgame, it can be shown that the multi-round signaling game with descending
jump bids has a unique symmetric equilibrium.
Next, suppose that bidders again can only signal in the first round, but are now
allowed to choose their opening bids from the descending set {K1 , K2 , ..., Kn , 0}. Consider
the choice of the lowest jump bid Kn in the simultaneous game of the first round. In
equilibrium, the only bidders who will not make at least that bid are those with the
lowest private observations. Suppose the threshold for such a bid is zn . zn is identified
by finding the level of private observation which makes the receiver indifferent between
bidding 0 and Kn , conditional on the opponent’s private observation being below zn . This
condition is identical to the condition for the threshold in the nth round of signaling in
the multi-round signaling game. By inductive reasoning, the n thresholds in this single-
round signaling game are equal to the thresholds in the multi-round signaling game with
While Proposition 1.1 provides an excellent starting point for a structural model for
estimation, there are two issues that need to be resolved. First, the equilibrium changes
with the specification of Sa , the second-stage strategy played by the bidder with the lower
17
jump bid from the first stage. Sa will need to be fixed to avoid multiple equilibria. The
second is that the model needs to be generalized from a model with two bidders to a
model with more than two.
Proposition 1.2. The following symmetric strategy is a subgame perfect equilibrium of the
multi-player ascending bid auction: in the first round of the auction, place a bid following
a symmetric first-price sealed bid strategy S1st (x); if outbid by any bidder in the first
round, drop out immediately, otherwise move on to the second round; from the second
round onward, follow an ascending-bid strategy with drop-out price equal to S ∗ (x | Ω̂),
where Ω̂ denotes all information available up to that point of the auction.
Proof. This equilibrium can be written as {S1st (x), (0, S ∗ (x | Ω̂)}. Proposition 1.1 holds
for any discrete set. Consider an equally spaced discrete set defined by {K, K, δ}, where
K > 0 is the lower bound, K < ∞ is the upper bound, and δ > 0 is the equal spacing.
Fix Sa (x) = 0. This implies that as long as the two bidders do not choose the same
jump bid in the first round, the one with the lower jump bid drops out at the end of
the first round. As K → 0, K → ∞, and δ → 0, the choice set approaches the set of
non-negative (real) numbers. As the private observations are drawn from a continuous
distribution, the probability that the two bidders draw the same private observations
is zero. Hence the probability that they choose the same jump bid approaches 0 and
the probability that the auction ends after the first round approaches 1. In the limit,
the bidders are free to choose any nonnegative bid in the first (and only) round of the
auction. The one with the lower bid drops out and the other one wins the auction and
pays her own bid. This describes a first-price sealed bid auction. Let S1st (x) denote
the strictly increasing symmetric equilibrium strategy for a first-price sealed bid auction,
where S1st (xi ) maximizes E[(Ui − bi )1(S1st (xj ) ≤ bi ) | xi ]. This equilibrium strategy can
be easily extended to when there are more than 2 bidders by redefining xj = maxk̸=i xk .
18
In the limit of K → 0, K → ∞, and δ → 0, the auction beyond the first round becomes
degenerate. Nevertheless, a strategy is needed for a subgame perfect equilibrium. I will
adopt the strictly increasing symmetric strategy S ∗ (x | Ω̂) for 2 or more bidders by
Milgrom and Weber (1982). Note that the key difference between an ascending auction
with 2 bidders and one with more than 2 bidders is that more information is revealed in
the course of the auction in the latter case. When a bidder drops out of the auction, if
the auction does not end (i.e. when there are two or more bidders left), the remaining
bidders are able to infer the private observation received by the bidder who drops out
from her drop-out price. The remaining bidders can therefore update their expectations
since the value function is assumed to be affiliated. Ω̂ denotes the set of drop-out prices
observed up to a particular point of the ascending-bid auction.
as a first-price sealed bid auction with a price floor or a price ceiling (or both). In each
case, there is a probability mass for the first-round bids at 0 or K. The auction beyond
the first round is no longer degenerate.
This section explores an extension of the multi-player model with one round of signaling
to one with multiple rounds of signaling.
Proposition 1.2 describes an equilibrium in which bidders send signals through jump
bidding in the first round only. Beyond the first round, the implicit assumption is that
any further jump bidding will not be interpreted as signals. If this assumption is relaxed,
that is, jump bids in rounds beyond the first round are also seen as signals, then this
provides further opportunities for bidders to increase the expected payoff.
19
Consider a simple case in which bidders can signal in the first 2 rounds. Again
assume K → 0, K → ∞, and δ → 0. Consider the following strategy
{S1st (x), (0, S1st (x)), (0, S ∗ (x | Ω̂))}. The difference between this strategy and the one
in Proposition 1.2 is that in the second round, a bidder plays S1st (x) again if she is not
outbid by any other bidders in the first round. If all bidders follow this strategy, the auc-
tion again ends after the first round. However, this strategy is no longer an equilibrium.
Assume all bidders but i play this strategy. Consider an alternative strategy for bidder
i where she places a reduced bid S̃ = S1st (xi ) − θ > 0 in the first round. Let S1st (xj )
be the highest bid in the first round among the rest of the bidders. If S1st (xj ) > S1st (xi )
(that is, xj > xi ), then bidder i drops out right away and her payoff remains 0. If
S̃ ≤ S1st (xj ) < S1st (xi ), bidder i can move onto the second round with bidder j and
bid S1st (xi ) to win the auction and receive the same payoff. In the last scenario when
S̃ > S1st (xj ), bidder i wins the auction at a lower price. Since the probability of the last
scenario is non zero, the alternative strategy delivers a strictly higher ex ante payoff.
price. The same argument applies when bidders can signal in more than 2 rounds, and in
an unlimited number of rounds.
There are additional benefits for placing a bid lower than S1st (xi ) in earlier rounds
when bidders can signal in multiple rounds. Bidders are able to infer more information
about the private observations other bidders receive from their decisions to drop out or to
stay in the auction using the inverse of the equilibrium strategy. This information affects
both the expected value of the object as well as the expected probability of winning. In
particular, having more information will allow bidders to form an expected value closer
to the true value of the object, thus reducing the winner’s curse typical to a sealed-
bid auction. This ability to collect more information along the course of the auction
20
and update one’s expectations accordingly is what differentiates an ascending-bid auction
from a second-price sealed bid auction. In this case, it also marks the difference between a
multi-round signaling game and a one-round signaling game. I use S̃1st (xi | Ωr ) to denote
the strategy that maximizes E[(Ui − bi )1(S̃1st (xj | Ωr ) ≤ bi ) | xi , Ωr ], where Ωr is the set
of information revealed up to round r − 1.
ing
When multi-round signaling is allowed and the decision to stop placing a jump bid becomes
endogenous, it is complicated to specify a pure-strategy equilibrium. The very nature of
a multi-round auction means the cost for bidders to make a mistake is low, in particular
in earlier rounds. This suggests that there are likely multiple equilibria. It is complicated
to predict which equilibrium is more likely to be played in reality. Instead of trying to
argue which equilibrium prevails, I will focus on predicting some common characteristics
shared by all the equilibria. Note that there is no limit on how many rounds a bidder
can signal through jump bids. The only restriction is that once a bidder stops placing a
jump bid in a round, she will not be allowed to do so in all future rounds. I will call the
earlier rounds where jump bids are placed the signaling stage and the later rounds where
non-jump bids are placed the ascending-bid stage.
Proposition 1.3. For any symmetric (pure) strategy that constitutes a SPE of the multi-
player ascending bid auction with multi-rounds of signaling, the following holds:
1. if bri , the bid in round r by bidder i, is a jump bid, i.e. bri ≥ ρr + κ̄(ρr ), then
2. if bri is not a jump bid, i.e. ρr ≤ bri ≤ ρr + κ̄(ρr ), then S̃1st (xi | Ωr ) ≤ ρr + κ̄(ρr ) and
21
bri < S ∗ (xi | Ω̂);
3. if a bidder i drops out at the end of round r, then either there is a high enough
signal, i.e. brj ≥ ρr + κ̄(ρr ) and brj ≥ S̃1st (x | Ωr ), where brj is the highest jump bid of
round r; or the expected value is lower than the minimum required price of the next
round, i.e. S ∗ (x | Ωr+1 ) < ρr+1 .
Proof. As discussed earlier, in round r, with additional information collected from the
dropout decisions of other bidders, S̃1st (xi | Ωr ) gives the jump bid that makes bidder i
indifferent between placing a jump bid and placing a bid equal to the minimum required
amount. As a result, S̃1st (xi | Ωr ) serves as the upper bound of bidder i’ jump bid in
round r for all equilibria.
When bidder i places a non-jump bid, this means she has endogenously chosen to end
the signaling stage as the smallest jump bid now exceeds the upper bound. She then
moves on to follow the strategy in an ascending-bid auction, which specifies a drop out
price at S ∗ (xi | Ω̂).
A bidder decides to drop out at the end of round r as soon as she receives a signal
that puts the sender’s private observation above her own. This means there must be at
least one jump bid that is higher than the upper bound of bidder i’s jump bid in round r.
There is another reason for the bidder to drop out. This happens when ρr+1 , the minimum
required price for the next round exceeds the drop-out price given by S ∗ (x | Ωr+1 ).
Proposition 1.3 places bounds on the observed bids based on whether a bidder places
a jump bid and her dropout decision at the end of each round. These model predictions
suggest that the signaling model works like a hybrid model of a first-price sealed-bid
auction and an open exit ascending-bid auction. To structurally estimate such a model
22
involves estimating each of these two auction formats.
In 1997, the FCC introduced “click box bidding” in Auction 16. Participants placed bids
by simply clicking on the license numbers and all bids were exactly one increment above
the standing high bid. It was believed that the change in auction rule was to primarily
address the issues of jump bidding and code bidding6 (Cramton and Schwartz, 2000;
Bajari and Yeo, 2009). In later auctions, this rule was relaxed to allow bidders to place
jump bids up to 9 bid increments. Nonetheless bidders’ ability to signal using jump bids
was restricted, making the later auctions less suitable for studying jump bidding.
Among the first 15 auctions where bidders faced no restriction on bid size, Auction 5
(broadband PCS auction (C-block)) was the largest in terms of total revenue raised (over
$10 billion). It was only open to small businesses with annual revenues less than $40
million. Compared to other auctions, the bidders in Auction 5 are more homogeneous
in size, therefore more appropriate to be described as symmetric, which is a key feature
of the theoretical model. In this auction, the U.S. was divided into 493 markets (“Basic
Trading Areas”). One license was offered for each market. 255 small businesses took part
in the auction.
In the empirical exercise that follows, I treat the auction of each of the 493 licenses as an
independent auction and assume each auction follows the rules detailed in the theoretical
section. However, in real life, these auctions ran in parallel following a simultaneous
6
Code bidding refers to the practice of attaching market numbers in the trailing digits of bids to signal
key interest.
23
multiple round format. Each round, bidders simultaneously placed bids on all the licenses
they were interested in. A key difference between the rules of the theoretical model and
those of Auction 5 is that bidders did not need to announce whether they dropped out
of the auction for a particular market at the end of each round. It is therefore unclear
at which point the dropout took place. Following Donald and Paarsch (1996), Paarsch
(1997), and Hong and Shum (2003), I assume this happened immediately after a bidder
placed her last bid in a market and the other bidders made the same inference. While this
assumption may not always hold, the eligibility rule of Auction 5 encouraged participants
to actively bid in all markets they were interested in by making the maximum number of
bids a bidder could place in future rounds dependent on the number of bids she placed
in the current round. If a bidder adopted a “sit-and-wait” strategy in too many markets,
she might eventually lose the eligibility to contest in some of the markets that she stood
a chance to win. The eligibility rule could therefore be interpreted as a milder version of
the irrevocable dropout assumption made in the theoretical model.
The theoretical model defines a jump bid in round r as b ≥ ρr + κ̄(ρr ), where κ̄(ρr ) >>
0, and is known to all bidders. In auctions where bidders can only place a bid that
either equals the minimum required price, or exceeds it by whole multiples of a fixed
increment set by the auctioneer (for example in FCC Auction 17 and some art auctions),
In the data, out of all the bids that exceed the minimum required price, over 20%
exceed the minimum required price by $100 or less, and over 40% exceed it by $1000
24
or less (see Figure A.1 in Appendix A.1). In a signaling model, jump bids are intended
as tools to disclose some information about one’s private observation. However, bidders
could aim to achieve other goals with small jumps. For example, some may simply want to
avoid a tie with another bidder and become the highest bidder of that round. In Auction
5, the highest bidder in a round does not need to raise her own bid in the following round.
Further, since the FCC always sets the minimum required price in multiples of thousands,
i.e. the last three digits are zeros, all code bids (bids with the market numbers attaching
to the end) exceed the minimum required prices. Code bids send signals in a different
way than jump bids.
It is entirely possible that bidders do not agree on how much a bid must exceed the
minimum required price to be interpreted as a jump bid in the way described by the
theoretical model. However, if the majority of them agree on a common definition, is
there a way to identify it? Suppose a bidder believes a jump bid exceeds the minimum
sizes of jumps, at the true κ̄. Figure A.2 is a histogram of the empirical distribution of
jumps. Each bar correspond to a bin with width of 50. We observe distinct spikes for the
following bins with lower bounds at 0, 100 (this bin contains jump sizes from 100 to 149),
and whole thousands. Considering that numbers that are multiples of 100 and 1000 may
contain additional informational value or associate with lower transaction cost, I remove
bids with jump sizes equal to 100, 500, 1000 and 2000. The spikes with lower bounds at
100 and 1000 persist after the adjustment (see Figure A.4). The spike at 1000 become
more prominent in the later rounds of the auction (see Figure A.5).
7 r
ρ + κ̄ − ϵ is strictly dominated by ρr + κ̄ if ρr + κ̄ ≤ S̃1st (·), and is strictly dominated by ρr if
r
ρ + κ̄ > S̃1st (·).
25
Based on the empirical evidence, I define κ̄ at 1000. Compared to 100, 1000 has the
added benefit of excluding all potential code bids (since there are 493 markets, a code
bid exceeds the minimum required price by at most $493). Further, a relatively stringent
criterion for a jump bid will reduce the probability of over stating the revenue effect of
jump bidding.
As discussed earlier, the affiliated value framework of the theoretical model nests the
private value model, the pure common value model and the common value model, which
includes a common value component and a private value component. In the context of a
spectrum auction, one would expect the presence of both a common value component to
reflect market attributes that are valued by all bidders, such as market size measured by
population, and a private value component to allow values to differ across bidders. This
suggests that a common value model is the most appropriate in this empirical context.
The difficulties with nonparametric identification of a common value model are well
recognized in the econometric literature. Laffont and Vuong (1996) show that a common
value model in a first-price sealed-bid auction is unidentified nonparametrically. In fact,
any affiliated value model is observationally equivalent to some affiliated private value
model. Similarly in ascending-bid auctions, Athey and Haile (2002) prove that a common
value model is generally not identified nonparametrically. Given the challenges with
nonparametric identification of a common value model, I adopt a parametric approach
I assume that Ui , the value of the object to bidder i, takes a multiplicative form
Ui = Ai V , where Ai is a bidder-specific private value for i, and V is a common value
component unknown to all bidders. V and Ai ’s are assumed to be independently log
26
normally distributed. Let v ≡ lnV , ai ≡ lnAi , ui ≡ lnUi :
v = m + ϵv ∼ N (m, r02 ),
ai = ā + ϵa,i ∼ N (ā, t2 ),
error term with a standard normal distribution, and s is an unobserved parameter. Let
xi ≡ lnXi , then conditional on ui ,
xi = ui + si ξi ∼ N (ui , s2 ).
As a result, (ui , xi , i = 1, ..., N ) are jointly normally distributed. The joint distribu-
tion is fully characterized by parameters {m, r0 , ā, t, s}. These parameters are common
knowledge among the bidders and are what I will estimate using the signaling model.
1.4.4 Identification
As pointed out by Haile and Tamer. (2003), the lack of sufficient structure of an ascending-
bid auction makes a mapping between the bids observed and the underlying demand
structure challenging. Point identification typically relies on observing at which price a
bidder drops out. Haile and Tamer. (2003) instead construct bounds on observed bids and
partially identify the distribution functions using an independent private value model.
As the theoretical model of this paper also places bounds on the bids observed, a partial
identification approach could naturally follow. However, it is difficult to implement the
approach by Haile and Tamer. (2003) in a common value model due to the interdependence
27
in distributions of bidder information and valuation. In this case, stronger assumptions
become necessary. I follow the approach by Donald and Paarsch (1996), Paarsch (1997)
and Hong and Shum (2003) and assume that the last bid placed by each bidder (but the
winner of the auction) is equal to the dropout prices predicted by the theoretical model.
This assumption allows for point identification. As discussed in Section 1.4.1, whether this
assumption is valid depends on whether the bidder indeed drops out immediately after
placing the last bid. If instead the bidder stays in the auction for a few additional rounds
without placing any bid, the last observed bid is below the real dropout price. While this
scenario could happen in Auction 5 because a formal announcement of dropout was not
required, an eligibility rule was put in place to discourage such a “sit and wait” strategy
by reducing the maximum number of bids a bidder could place in each future round for
bidders who did not place enough bids in a few consecutive rounds.
I use a simulated non-linear least squares estimator following the methodology of Laffont,
Ossard and Vuong (1995) and Hong and Shum (2003). This estimator minimizes the
nonlinear least-squares objective function:
T Nt
1 XX
QT (θ) = (pt − mti (θ))2 . (1.3)
T t=1 i=2 i
pti is the observed log dropout price for bidder i in auction t. For tractability, I number
the bidders in an auction in the reverse of the dropout order. For example, in auction
28
t, the winner is labelled bidder 1, and the bidder who is the first to drop out is labelled
bidder Nt , where Nt is the total number of bidders in this auction. Notice that the bidder
number in Equation 1.3 starts from 2. This is because the dropout price for the winner
is not observed since she never dropped out. mti (θ) is the model-predicted counterpart
of pti , which is the mean of a multivariate truncated distribution. Due to the difficulties
of computing mti (θ) analytically, it is natural to replace it with a consistent simulation
S
1X t
m̃ti (θ) = (b (⃗xs ; θ)), (1.4)
S s=1 i,s
where bti,s (⃗xs ; θ) is the model predicted dropout price for bidder i in auction t for a par-
T Nt
1 XX
Q̄S,T (θ) = (pt − m̃ti (θ))2 . (1.5)
T t=1 i=2 i
Laffont, Ossard and Vuong (1995) show that while m̃ti (θ) is a consistent estimator of
mti (θ), Q̄S,T (θ) is not a consistent estimator of QT (θ) for any fixed number of simulations
S as T goes to infinity. It can be shown that the size of the bias for each auction t is:
S
1 X
∆S (θ) = (bti,s (⃗xs ; θ) − m̃ti (θ))2 . (1.6)
S(S − 1) s=1
29
Equation 1.6 from Equation 1.5:
T Nt S
1 XX 1 X
Q̃S,T (θ) = {(pti − m̃ti (θ))2 − (bti (⃗xs ; θ) − m̃ti (θ))2 }. (1.7)
T t=1 i=2 S(S − 1) s=1
As discussed in Section 1.4.4, bti,s (⃗xs ; θ) in the signaling model is predicted by the
equilibrium strategy of a first-price sealed-bid auction if the observed bid is a jump bid;
and is predicted by the equilibrium strategy of an open exit auction if the observed bid is
not a jump bid. In the rest of this section, I will discuss how each of the two is computed
in auction t with simulated draw ⃗xs . For simplicity, I will suppress the superscript t.
The equilibrium strategy of an open exit auction with symmetric affiliated value follows
Milgrom and Weber (1982). In equilibrium, the price βi at which bidder i drops out of
the auction is defined as follows:
Since bidder valuations are affiliated, the expected valuation of each bidder depends
on the private observations of all bidders. For each bidder i, the private observations can
be partitioned into 3 groups. The first group includes bidder i’s own private observation
Xi . The second group consists of Xj ’s in Equation 1.8, the private observations of the
bidders that remain in the auction. Since these private observations are unobserved by
bidder i, she needs to form expectations on them. Given the assumption of symmetric
bidders, bidder i simply assumes Xj = Xi . The last group includes Xk ’s, the privates
observations of the bidders that dropped out before bidder i. While bidder i does not
30
observe these private observations at the beginning of the auction, she can infer them by
inverting the bidding strategy after observing at which prices these bidders drop out.
Given the parametric assumptions per Section 1.4.3 , the log dropout price has a closed-
′
form solution that can be easily computed with each simulated draw ⃗xs ≡ (x1s , ..., xN
s ).
E[ui | x̂i ] = (ui − µ∗′ Σ∗−1 σi∗ ) + x̂i ′ Σ∗−1 σi∗ , (1.10)
The equilibrium strategy of a first-price sealed-bid auction with symmetric affiliated value
according to Milgrom and Weber (1982) is:
Xi
fY1 (α | α)
Z
bi (Xi ) = ln v(α, α) L(α | Xi )dα, (1.12)
0 FY1 (α | α)
R Xi fY1 (s|s)
where Y1 = max{Xj }, j ̸= i, L(α | Xi ) = exp(− α FY1 (s|s)
ds).
31
This strategy does not have a closed-form analytical solution. To compute it, I employ
a combination of simulation and numerical approximation. First, I divide the integral into
n bins and rewrite the it into a summation:
n
X fY (αs | αs )
bi (Xi ) ≈ ln v(αs , αs ) 1 L(αs | Xi )ωs , (1.13)
s=1 |
F Y 1 (α s | αs )
{z }
ρs,i
where αs is the mid-point of bin s and ωs is the width of bin s. I use ρs,i to denote the
fY1 (αs |αs )
value of bin s conditional on Xi . Next, I estimate components of ρs,i ( v(αs , αs ), FY1 (αs |αs )
and L(αs | Xi )) separately for each bin s. Last, I sum up the values of all n bins and
take the natural log to obtain an estimate for the equilibrium strategy of a first-price
sealed-bid auction.
The number of bins I select for approximation is 10. I conduct tests for robustness
using various bin numbers up to 50. The value estimated is not very sensitive to the choice
fY1 (αs |αs )
of bin numbers. Further, as the number of bins increases, the estimation for FY1 (αs |αs )
tends
to yield undefined values for bins with lower values since FY1 (αs | αs ) is more likely to be
zero. The reason for this will become clear once I discuss how FY1 (αs | αs ) is estimated.
For a given number of bins, an intuitive way of defining the bins is to draw a large
number of observations from the empirical distribution8 , and divide the distance between
the minimum and maximum observations into equal intervals. However, given the nor-
mal distribution, the majority of the observations drawn will fall into the middle bins,
fY1 (αs |αs )
rendering estimations for FY1 (αs |αs )
and L(αs | Xi ) for the bins at the two ends inaccurate
due to a lack of observations. As a result, I define the bins using percentiles of the ob-
servations drawn. In particular, when the number of bins is 10, the bins are defined by
8
The empirical distribution changes in every iteration of the minimization algorithm due to changes
in parameter values. To make sure that the definitions of bins are consistent across iterations, I make
draws from the standard normal distribution to obtain the bounds for each bin, and rescale them based
on the parameter values of a particular iteration. The number of observations drawn is 100 million.
32
the deciles (i.e. the 10th, 20th, ... and 90th percentiles). This makes sure the probability
of a randomly drawn observation falling into any of the bins is the same. I number these
bins from 1 to 10 starting from the lowest decile.
v(Xi = αs , Y1 = αs ) measures the expected value of bidder i when both bidder i and
the bidder with the highest private observation among all other bidders receive a private
observation equal to αs . The analytical solution to v(·) is a multi-dimensional integral
on the joint distribution of X’s, which is difficult to compute. I therefore approximate
v(⃗xs ) by fitting a function on the largest two elements of 500 simulated draws. v(⃗xs )
is evaluated using Equation 1.10. By assumption, v(⃗xs ) is monotone in each element of
vector ⃗xs and the elements are strictly affiliated, v(⃗xs ) is therefore likely to be monotone
The cumulative distribution function (“cdf”) FY1 (αs | αs ) can be re-written as a con-
ditional probability:
To estimate FY1 using simulation, I again draw 500 observations of ⃗xs . I then proceed to
identify all observations that contain at least an element which falls in bin s. Out of these
observations, I calculate the share of observations which do not contain any element that
falls in a bin higher than s. This share provides an approximation of FY1 (αs | αs ). It can
be easily shown that FY1 is increasing in αs . This pattern is generally observed for the
estimates of FY1 for the higher bins. However, for the lower bins, the increasing pattern
is often broken. This is because the probability of drawing an observation of ⃗xs of which
the largest element falls in bin s decreases as s decreases. The accuracy of the estimates
33
therefore decreases as a result of a smaller number of observations for the lower bins. In
some cases, the estimate falls to zero. Unfortunately this issue does not readily go away
when the number of observations drawn increases. The probability distribution function
R XifY1 (s|s)
L(αs | Xi ) = exp(− αs FY1 (s|s)
ds) is an integral. To approximate it, I again rewrite it
into a summation:
Ni
X fY1 (αm | αm )
L(αs | Xi ) ≈ exp(− ωm ), (1.16)
m=s
F Y 1 (α m | α m )
where Ni is the number of the bin into which Xi falls. The computation of this summation
fY1 (αs |αs ) fY1 (αs |αs )
for each bin s makes use of approximations of FY1 (αs |αs )
. However, unlike FY1 (αs |αs )
which
To reflect the fact that Xi falls into bin Ni , but does not necessarily cover the entire
Xi −lbNi
bin, I adjust ρNi ,i by multiplying it with the ratio ubNi −lbNi
, where lbNi and ubNi are the
lower and upper bounds of bin Ni respectively.
The second panel of Table 1.2 shows the estimation results of the multi-round signaling
model. The mean of the common value component m, is parameterized into 2 parts, a
constant and a coefficient on population. I choose this parsimonious parameterization to
reduce the burden of computation. In addition, simple regressions of the winning bids
34
Table 1.2: Simulated nonlinear least-squares estimates
Note:
a Bootstrapped standard errors in brackets, computed from empirical
distribution of parameter estimates from 100 parametric bootstrap
resamples
b Not separately identified from ā
against a selection market attributes, including population, population density and area
suggest that population is the best predictor. The mean of the private value compo-
nent, ā is not separately identified from the constant associated with the common value
component.
when bidders play according to the signaling model but the econometrician fits the data
using an open exit model, I repeat the SNLS estimation using an open exit model with the
same data. In this case, bti,s (⃗xs ; θ), the model predicted dropout price for a random draw
of private observations ⃗xs as defined in Section 1.5.1 is simply equal to the closed-form
analytical solution in Section 1.5.2. The estimation results are shown in the first panel of
Table 1.2.
Comparing the estimates of the components of the m, the mean of the common value,
the values for both parts are higher under the multi-round signaling model than the open
35
exit model. Consider a market with a population of 200k (the median population of all 493
markets is 187k), the mean of the log value of the license estimated by the signaling model
is 13.4, 14% higher than that estimated by the open exit model at 11.8. This suggests
that if bidders are using jump bids as signals, ignoring them leads to underestimation of
the mean.
Note:
a Average model predicted log prices of 1000 simulated draws
In the counterfactual analysis, I assess the revenue impact of jump bidding using the
estimates from the signaling model. In particular, I estimate the changes in revenues from
an auction with multi-round signaling to one with single-round signaling, and to an open
exit auction where no signaling is allowed.
If jump bids were prohibited, the auctions that would have different prices would be
the ones where the winning bids were jump bids. However, the empirical model is not
36
able to make any prediction on the winning bid because it is not the dropout price of the
winner. To get around the limitation, I will instead focus on auctions with second highest
bids being jump bids, and make counterfactual predictions on these bids to approximate
a change in revenue.
In the data, 81 auctions out of 491 fit this description. I refer to these auctions as the
“jump bid” auctions. As shown in Table 1.3, the mean log winning bid for the jump-bid
auctions is at 15.95 and the mean log second highest bid is 15.90. The small difference
between these two numbers suggest that the latter serves as a fairly close estimate to the
former.
The multi-round signaling model predicts the mean log second highest bid to be at
16.58. If we only allow the bidders a one-off opportunity to signal using jump bid, the
The revenues from the 81 jump bid auctions are estimated to increase by 33% when
the auction moves from one that allows multi-round signaling to one that only allows one
round of signaling. The revenues increase by another 35% when signaling through jump
bids are prohibited. The total increase in revenues amounts to 76% of those from the
auction with multi-round signaling. Overall, this increase account for an 8% increase in
total revenues across all 491 auctions.
1.7 Conclusions
The existing empirical research on jump bidding adopts either a descriptive or a reduced
form approach. While both approaches help confirm the pervasiveness of jump bidding
37
and identify its correlation with a number of economic factors, they are unable to quantify
the revenue effect of jump bidding. This paper generalizes a theoretical signaling model
on jump bidding and uses it as the basis for a structural estimation that uncovers the
underlying bidder value distribution. A counterfactual analysis using these estimates finds
an overall improvement of revenue by 8% for the FCC broadband PCS auction (C-block).
38
CHAPTER 2
ECONOMIES OF DENSITY AND CONGESTION
IN THE SHARING ECONOMY
With the intention to reap the benefits from sharing capital services across
small-scale producers, governments in the developing world are increasingly
subsidizing equipment rental markets. Service provision is cheaper in dense
areas and for higher machine-hours contracted because equipment needs to be
moved in space. Using our own census of 40,000 farmers in India, we doc-
ument that costly delays and price dispersion in rentals are ubiquitous, and
that small-scale producers are rationed out by market providers. This rationing
could be detrimental to aggregate productivity if small farmers have the high-
est marginal return to capital. A government subsidized first-come-first-served
(fcfs) dispatch system grants small-scale producers timely access to equipment
at the expense of travel time. In a calibrated model of frictional rental services,
optimal queueing and service dispatch we show that, while the constrained ef-
ficient allocation prioritizes large-holder farmers, small-scale farmers in dense
areas are valuable because they help maximize capacity utilization. Through
counterfactuals, we show that when the induced increase in subsidized equip-
ment supply is high enough, service finding rates for small-farmers increase
relative to large-holders farmers even when providers prioritize large-scale.
2.1 Introduction
From housing services, to construction machinery to car rides, the sharing economy,
39
effects as well as efficiency effects that are not well understood, and that this paper studies.
markets were a stepping stone to that process. But where contract enforcement is weak
and overall wealth is low, rental markets may not emerge, or if they do, they may ration
productive users.
In this paper, we study subsidies to the creation of rental markets for agricultural
equipment, which have the potential to unravel productivity gains and employment real-
location away from agriculture, typical of the development process. Fairness concerns may
lead government supported platforms to servicing farmers in relatively less dense areas,
incurring high transportation costs per service, or to servicing relatively low productivity
farmers in detriment of high productivity ones. Due to the time-sensitivity of agricul-
tural activities demand is synchronous, inducing congestion and delays. By subsidizing
increases in equipment supply the government can lower congestion, but it could also in-
duce marginal farmers to demand equipment services, slowing down service provision for
productive farmers. The balance of these effects dictates the efficiency and distributional
Our paper combines the first available detailed transaction level dataset of equipment
rental markets in the developing world, a census of 130 villages characterizing the entire
demand for services, and a novel model of frictional rental markets to assess market effi-
ciency and the distributional impact of government subsidies. We model and study two
prevalent types of provision of mechanization rentals. The first one is the provision of
services by single owners of equipment, often times farmers who own agricultural equip-
ment and rent their excess capacity out (i.e. “market” providers). Given service capacity
40
constraints, these providers prioritize large orders or orders in densely populated loca-
tions. The second one is a private-public partnership that sets up custom hiring centers
(CHCs) with the objective of providing machinery for small and marginal farmers. These
providers allocate equipment on a first-come-first-served basis, so that order size and lo-
cation is uncorrelated with the timing of service (“fcfs” providers). Marginal farmers are
driven into the government subsidized platform, but the additional transportation costs
associated to moving equipment in space may well overturn the gains in productivity from
timely access to capital for those farmers. Through counterfactual exercises, we quantify
the strength of these channels and argue that general equilibrium forces in prices and
the equilibrium sorting of farmers to providers are quantitatively important to assess the
allocative efficiency of rental markets.
acreage serviced, implement and location. This data is linked to our newly collected
survey data on demographics, farmer assets, productivity and engagement in the rental
markets.First, we document that requests for capital rental services are spread across
space and that there is higher engagement in rental markets by large-scale farmers (more
than 4 acres of farmland) than by low-scale farmers (cropping less than 4 acres). However,
the distribution of these large and small farmers in space is such that once we control
for equipment ownership levels in neighbouring villages, there is no differential access be-
tween large and small scale farmers. Second, we document queuing for service fulfillment
which varies throughout the season. This congestion effect is of first-order importance for
the organic development of rental markets, and a key margin to assess the efficiency and
distributional effects of the markets induced by the government subsidies. Third, because
agricultural activities are highly time sensitive, the delay in the arrival of inputs associ-
ated to these queues is potentially costly. We quantify the costs of delays for productivity
41
and profitability by estimating the optimal date for land preparation during a season. We
choose land preparation because those are the processes for which the rental markets in
Karnataka are most active. We compare profits for farmers that prepared the land away
from that optimal date. We find that the average productivity cost for delay within a
10-day window of the optimal planting date is 8.5% per day.
To assess the merits of different dispatch systems for efficiency and productivity, we
pose a novel model of frictional rental markets for equipment. There are two types of
providers that differ in their order dispatch technology and face capacity constraints for
service. There are two types of farmers that differ in their demand of equipment-hours and
their locations in space. Providers set rental rates and farmers choose in which provider
to queue at, as in a standard directed search model, Shi (2002).1 The main difference to
these models is the interaction of search frictions with service capacity constraints, and the
empirically relevant ability of providers to serve multiple orders (of different types) within
a period. Consistently with the data, providers can serve up to three orders within a day,
posing an interesting combinatorics problem given the provider’s capacity. The main
predictions of the model are that when small and large farmers are equally distributed in
space, small scale farmers should be more likely to approach the fcfs provider than the
market provider. Large farmers should be more likely to approach market providers than
fcfs ones. This sorting induces disparities in service finding rates across the population of
farmers. In equilibrium, rental rates adjust so that farmers face identical expected values
of service from either provider. Heterogeneity in the allocation of farmers in space induces
additional disparities in the cost of provision, and in the incentives of providers to serve
orders, through the opportunity cost of traveled time, per serviced hour.
To bring the model to the data we allow for additional heterogeneity in farmers
1
As highlighted by Lagos (2000) and Sattinger (2002), queueing models are powerful to micro-found
a matching process between, in this case, farmers’ orders and service provision.
42
equipment-hours requests and locations. We solve and calibrate our structural model
of optimal dispatch, using administrative data from the fcfs platform, the Census data
for market providers and our survey data on farmers’ productivity for detailed charac-
waiting times, lower rental costs and increase in service finding rates. However, this dis-
patch system entails equipment transportation costs that often overturn the gains to those
farmers. Large farmers face lower delays when renting from market providers, and their
location is such that when market providers minimize transportation costs, larger farmers
disproportionally benefit from queuing with them.
ment intervention, assuming that market providers (less than 10% of the total supply)
accommodate the entry of subsidized providers. We find that the subsidy generated a
substantial increase in service capacity, which lead to two-to-three fold increases in ser-
vice findings rates, and declines in the cost of service provision of more than 20% per
hour serviced. This additional supply was particularly beneficial to smallholder farmers.
Second, we investigate the nature of the equilibrium when subsidized providers are al-
lowed to behave as their market counterparts, i.e. a market deregulation. The short run
response of the economy (with no endogenous entry or exit of providers) implies higher
service finding rates for large-scale farmers relative to the fcfs provision, consistently with
the profit maximizing strategy of market providers. However, their service finding rates
are below those of small scale farmers that are drawn to the market in response to the
increased supply of services. While service findings rates for small-holder farmers do not
change relative to the baseline, the rental costs increase for all farmers. The long run
43
response of the economy (with endogenous entry and exit) restores rental costs to their
baseline levels, and generates allocations where service findings rates are higher for small
holder farmers than large holder farmers, despite market providers prioritizing the lat-
ter. The reason is that small-scale request are valuable when there are service capacity
constraints, because they allow for better capacity utilization.
The notion that there might be scale economies associated to concentrating production
in certain locations is explored by Holmes and Lee (2012) in the context of crop choices
of adjacent plots. Bassi et al. (2020) documents the workings of rental markets for small
carpentry producers in urban Uganda and argue that frictions in their setup are relatively
limited. Distinctively, agricultural equipment needs to be moved in space to reap its ben-
efits and the time sensitivity of agricultural production makes demand synchronous. We
show that these two margins are key determinants of the efficiency of usage of capital
services through the rental market, and its distributional and productivity impact.
This paper also relates to the literature that emphasized barriers to adoption of tech-
nology in agriculture (Suri, 2011) and of mechanization in particular. Pingali, Bigot and
Binswanger (1988) highlighted the role of contractual enforcement problems while Foster
and Rosenzweig (2017) emphasizes economies of scale. Small poor farmers do not find it
optimal to adopt capital intensive practices when they entail the purchase of equipment, a
relative large expense whose services are used for a limited period in the season. High cost
of capital relative to labor have been emphasized as a deterrent to technology adoption,
(Pingali, 2007; Yang et al., 2013; Yamauchi, 2016; Manuelli and Seshadri, 2014). In this
44
paper we focus on the role of geography and density in assuring access to capital services
through rental markets.
The role of geography for the efficient allocation of factors of production in agriculture
has been studied by Adamopoulos and Restuccia (2018). In their paper they focus on the
characteristics of the soil and the potential yields of different plots to assess the degree
of misallocation of factors of production. Spatial misallocation has also been studied in
the context of the rural-urban gap in labor income by Gollin, Lagakos and Waugh (2013);
Lagakos, Mobarak and Waugh (2015).In this paper we focus on the distributional effects
across land-sizes and locations of equipment rental markets.
We study agricultural rental markets in rural India. The study focuses on the state
of Karnataka, one of the least mechanized states in India (Satyasai and Balanarayana,
2018), and where the majority of agriculture occurs on small farms (less than 4 acres). Our
own census of farming households in 150 villages across the state shows that equipment
ownership rates are low and that farmers rely on informal equipment rental markets within
each village.
In 2016, the state government partnered with the largest manufacturer of agricultural
equipment in India to design and manage a platform through which farmers could rent
equipment. They generated a subsidy scheme for equipment purchases to create custom
hiring centers (CHCs, also known as “hubs”) in 25 districts throughout the state. In
exchange for these government subsidies, the service provider committed to rental rates
45
of about 10% below their market counterparts. Figure 2.1 plots currently active hubs in
space. These CHCs provide rental services in nearby areas and farmers can access these
services through a call-center, via an app on a smart-phone, or in person at the CHC.
Triangles indicate CHCs. We aggregate demand following the village where each farmer is registered.
Green dots correspond to demand for the smallest plots (1 acre or less), red dots correspond to demand
for the largest plots (4 acres or more).
Our first source of data comes from the universe of transaction-level data maintained
electronically by these hubs. This administrative data consists of all the transactions com-
pleted through that platform since 2016 for about 60 hubs in Karnataka (27 hubs entered
in 2016, 29 in 2018 and 10 in 2019). Over the time period covered by the data (Octo-
ber 2016-May 2019), over 17,000 farmers from 840 villages rented equipment from these
CHCs. The data contains information on number of hours requested, acreage, implement
type, as well as farmer identifiers (such as their name, village, and phone number). Equip-
ment available varies across CHCs (hubs) but the median hub provides equipment that
ranges from sprayers to Rotavators or ploughs. Rental rates in the platform are about
10% below market prices on average, as part of the government’s initiative to increase
mechanization access. The service provision is first-come-first-served. When a service
46
is fulfilled, a professional driver brings and operates the equipment at the farmer’s plot.
Equipment arrives within a lapse of two days in most cases. If delays are longer, farmers
are informed at the moment of booking (and may choose to cancel). Table 2.1 reports
the 10 most commonly rented implements for the years 2017 and 2018, the number of
transactions recorded for each implement, their per-hour rental price, and month where
the implement is most commonly rented (has the highest number of transactions).
Table 2.1: Summary Statistics of Commonly Rented Implements from Rental Database
Commonly Rented Implements
Number of Transactions Median Price Per Hour Peak Month
Rotavator 6 Feet 11,239 770 July
Cultivator Duckfoot 7,287 550 April
Cultivator 9 Tyne 5,245 525 May
Plough 2MB Hydraulic Reversible 3,716 450 February
Trolley 2 WD 2,436 250 January
Harvester Tangential Axial Flow (TAF)-Trac 2,048 1800 May
Rotavator 5 Feet 1,811 700 September
Blade Harrow Cross 1,793 360 March
Knapsack Sprayer 20 Litres 1,688 22.5 October
Blade Harrow 5 Blade 1,600 360 June
We use two additional sources of data in Karnataka. The first is survey data collected
in June-July 2019 around a representative sample of the equipment hubs in the main
sample. The survey includes approximately 7,000 farmers, and asks for detailed informa-
tion on input and output per plot which allow us to generate measures of productivity
at the plot level. We also ask about their engagement in rental markets, rental rates,
and perceptions on barriers to participation in the market. The second is a census of
agricultural farmers including more than 40,000 farmers in the villages covered by the
survey data. This census allows us to validate the characteristics of the farmers in the
survey relative to the population. It also allows us to measure inventories of equipment
Finally, we use data from a broader set of Indian states as recorded in ICRISAT’s
household-level panel data. This data which contains detailed agricultural information,
including, season-level agricultural operations, their timing, costs and total revenues. The
data covers eighteen villages over 2009-2014 in Andhra Pradesh, Gujarat, Karnataka,
47
Maharashtra, and Madhya Pradesh.
We start by describing the characteristics of the service demand and farmers equipment
supply. Then, we focus on a handful of outcomes that are informative to the theory that
we describe in Section 2.3. First, because agricultural activities are highly time sensitive,
the timing of demand is synchronous leading to endogenous waiting times as a function
of service capacity. The service capacity includes farmers’ ownership as well as CHCs
capacity. Second, because equipment needs to travel for transactions to take place, the
joint distribution between travel times and the scale of demand, i.e. equipment-hours per
request, is a key input when optimizing service provision. Third, we document substantial
price dispersion in rental rates after controlling for observable household characteristics
and village/market characteristics, consistent with frictional rental markets. Fourth, de-
lays in service provision are costly to farmers, because they affect field productivity. In
own larger equipment such as tractors, or rotavators and cultivators. At the same time,
tractors and cultivators are among the pieces of equipment with the highest equipment-
hours rented. The average hours rented in a season per farmer is 12 hours for tractors and
10 hours for cultivators. These rental transactions mostly entail relational contracts. We
2
Appendix B.4 reports similar statistics using data from the Census.
48
collect information on the typical customer for a farmer that rents out his/her equipment.
We find that 72% of owners report to rent out to people they know from the village or
who they have worked with in the past.
Delays are the most common issue faced by farmers when renting equipment, with 78%
Harvester
Thresher
.3
Cart
Ownership rate
Hand Tools
Plough
.2
Disc
Rotavator
.1
Sprayer
Cultivator
er
er
or
gh
or
er
s
ar
m
ye
to
is
ol
sh
st
at
at
ay
ou
eq
ac
C
D
To
Tractor
ra
ve
tiv
av
re
r
Pl
Tr
Sp
Sp
d
ar
ul
d
ot
Th
lle
an
C
H
ck
pu
H
sa
al
ap
0 5 10 15
m
Kn
i
An
Owned Rentals
The ownership rate is the share of farmers that report to own a given implement relative to the total
population surveyed. Rental hours correspond to the average hours reported for the whole season.
during a season hi , market rental rates, ri and the number of implements i owned or
rented, Ni . Hence, equipment services in a farm k are
X
k= Ni ri hi
i
49
The main hypothesis behind this measure is that differences in rental rates across im-
plements shall reflect differences in the services they provide, and that therefore, more
expensive equipment provides higher services to production. The main challenge in con-
structing such a measure is the availability of data on market rental rates. We exploit
our transaction level dataset to construct mean rental rates per implement at the village
level. Figure 2.3 displays log owned and rented services. Harvesters (the most expensive
implement in our bundle) is reported to be only rented. For those farmers using tractors,
more than 60% of the services available in the farming sector come from rentals whereas
the remaining 40% stem from ownership. Services associated to smaller and cheaper
equipment, such as sprayers, are equally accounted for by rentals and ownership.
It is worth noting that given land holdings, ownership of equipment is not cost-effective
for most farmers. For instance, the rental price of a rotavator is between |750 and |1,000
per hour (including tractor, a driver and fuel) and the average farmer demands about 6
hours of rotavator services in the season or between |4500 and |6000 in services. The
purchase price of a new rotavator is over |110,000 which means that, absent maintenance
costs (which are certainly non-negligible), the average farmer needs 19 years to amortize
the investment. The rental rate for an inferior technology that serves a similar purpose,
i.e. a harrow, is half of the rental rate of the rotavator (|360) and the cost of purchase is
about |50000. Overall, these price differentials are consistent with the observed extensive
The demand for equipment rental services vary by agricultural process and therefore
throughout the agricultural season. The synchronous nature of many of these processes
across farmers induces queuing in the market. Our transaction level data allows us to
50
Figure 2.3: Capital services from ownership and rentals, by implement.
Tractor (770)
Cultivator (547)
Rotavator (769)
Plough (534)
Sprayer (162)
Harvester(1408)
0 .2 .4 .6 .8 1
Rented Owned
Shares of log capital services by implement and ownership/rental status. Average rental rates for an hour
of service (in |) are reported next to each implement.
measure demand fluctuations by computing hours outstanding for service at a daily fre-
quency. We focus on two commonly rented implements for land preparation, rotavators
and cultivators. Indeed, our survey data indicates that farmers are most likely to engage
Figure 2.4 shows hours of unfulfilled orders for each of these implements over the 2018
kharif season. Queueing peaks by the end of July for rotavators and beginning of August
for cultivators. At the peak of the season, the average provider faces 40 hours of demanded
services in queue, which account for over 12 orders on average at a point in time.
Demand moves distinctively between large and small requests, measured in service
hours (Figure 2.4). A large portion of hours outstanding are accounted for by small
orders (less than 4 hours of service), although at pick time the share of hours accounted
for by large and small holder farmers equalizes. This is not explained by a higher number
of large orders but rather by larger orders altogether.
51
Figure 2.4: Hours outstanding in the queue.
Cultivator Rotavator
50
50
40
40
Number of hours
Number of hours
30
30
20
20
10
10
0
0
6/1/2018 7/1/2018 8/1/2018 9/1/2018 10/1/2018 11/1/2018 6/1/2018 7/1/2018 8/1/2018 9/1/2018 10/1/2018 11/1/2018
Date Date
100
100
80
80
Percentage
Percentage
60
60
40
40
20
20
0
Jun. Jul. Aug. Sept. Oct. Jun. Jul. Aug. Sept. Oct.
Time Time
Small orders (< 4hrs) Large orders (>= 4hrs) Small orders (< 4hrs) Large orders (>= 4hrs)
by size by size
Averages hours outstanding in the queue across hubs in Kharif 2018, overall (top panel) and by order
size (bottom panel).
pected that service supply may adjust. To explore these movements we compute service
rates as the fraction of serviced hours within a day divided by the number of service hours
outstanding. On average, we see up to 3 orders being fulfilled during a day per equip-
ment piece. We find that service rates move during the season, and that they positively
correlate with hours demanded (see Figure 2.5). However, they do not move enough to
avoid queueing and congestion in service provision, likely due to capacity constraints or
52
Figure 2.5: Service rates.
4
3
Standardized Hours Served
0 1
-1
-2 2
-2 -1 0 1 2 3 4
Standardized Service Rate
We first document that service delays are negatively associated with cultivated size sug-
gesting that even if the productivity costs of delays are of same magnitude between small-
and large-holder farmers, the incidence of those delays is disproportionally bear by those
with small plot sizes, columns (1) and (3) in Table 2.2. It is possible that these delays
are explained by the geographical location of plots since equipment needs to travel to
generate services. Columns (2) and (4) Table 2.2 show that delays have an important
That is, in the surroundings to a particular village, small and large farmers face similar
delays, but if this clustering is not accounted for, smaller farmers face longer delays.
53
Table 2.2: Delays as a Function of Land Area and Location Fixed Effects
Delays (Sum of Average Delays Over the Season)
Log(Area) -0.215* -0.144 -0.319** -0.128
(0.115) (0.0926) (0.145) (0.108)
Observations 5,615 5,615 4,345 4,345
R-squared 0.002 0.182 0.003 0.252
Village Fixed Effects No Yes No Yes
Mean Delays 2.158 2.158 2.789 2.789
Estimated coefficients from a regression of reported delays in service provision and the log(area) owned.
The first two columns include those that report zero delays whereas the last two columns only focus on
those that report positive delays.
But why are there delays to begin with? Is this a consequence of low ownership rates and
service capacity, or rather the consequence of frictions in the rental market that prevent
farmers and providers to contract services when desired? There are two features of the
The first one is that the current supply of equipment seem adequate to serve mar-
ket demand. To compute supply we turn to a Census of 150 villages from the same
area, which includes information on over 40,000 farmers. We assume that the equipment
has a catchment area of about 10km, since transporting equipment over large distances
is time-consuming and expensive (particularly for farmers whose main activity are not
rentals). For a median village, the number of available cultivators within a 10 km radius
can serve 2016 orders per season, while average demand is 1190 orders. For rotavators,
available supply can serve 1008 orders in the season while market demand is 450 for the
median village.3 Hence, the observation of pervasive delays in service provision paired
with adequate supply within each geographical market suggest that congestion does not
necessarily arise due to supply shortages.
3
We assume a six-week season, and that each piece of equipment serves three orders a day. The number
of orders served in the day is consistent with services per equipment per driver at peak utilization that we
observe in the transaction dataset. Even when assuming a shorter 4 week season, supply would account
for 1344 per season for cultivators and 672 orders per season for rotavators, well under seasonal demand
for each equipment.
54
The second one is the presence price dispersion in rental prices of equipment within a
10km catchment area of each village. Using information from our survey of 7,000 farmers,
Figure 2.6 shows the distribution of residualized rental rates during land preparation from
village effects, i.e. that is the variation in rental rates per hour serviced across farmers
within a village. The interquartile range is 1.7 while the coefficient of variation is 0.7.
Burdett and Judd (1983) were the first to show that price dispersion can arise in an
environment with identical agents where consumers/farmers find it costly to search for
providers. Price dispersion can also be related to informational asymmetries Varian (1980)
or to consumer preferences for certain providers over others Rosenthal (1980). Overall,
the exchange of identical goods for heterogeneous prices is typically a sign of frictions in
the market, which we entertain through the structural model that we study next.
Hourly rent distribution (residualized by location)
.002
allow for multiple orders to be served within each period. Second, we build in capacity
55
constraints for service providers. The first feature is important to assess implications for
travel times across orders, i.e. optimal service routes and the costs associated to them.
The second feature is important to quantify congestion and the equilibrium service delays
as a function of the composition of the service queue in terms of the size and geographical
location of the plots being served.
2.3.1 Environment
Consider an economy populated by F farmers that use capital services for production.
Farmers differ in their demand for equipment-hours and in their geographic location.
A fraction s of them are “large-scale” farmers and demand ks hours, while the remaining
(1−s) fraction are “small-scale” farmers, and demand ks− hours. We think of the demand
farmers. For simplicity, we focus on ex-post shocks related to delays in service provision,
which are idiosyncratic to the farm.
There are H rental service providers that can serve up to o orders a day. Providers j
differ in their service capacity, up to k̄j machine-hours a day, and in their technology for
in providers’ geographical location, i.e. the expected travel time for service provision
4
Albeit h and H are assumed exogenous, both of them can be easily endogenized with a costly set up
of providers and an associated free-entry condition.
56
conditional on the machine-hours demanded is the same for both providers.
F
Denote the ratio of farmers to service providers, f = H
, and focus on the case where
the market is large, i.e. F, H → ∞ and neither side is infinitely larger than the other,
f ∈ (0, ∞). Providers post prices rij and a selection criteria (with commitment) simulta-
neously at the beginning of each period. Geographical considerations for service provision
are included into the opportunity cost of moving equipment from a provider to the plot.
Then farmers decide whether and which provider to approach, generating queues for each
available provider. Finally, providers decide which orders to serve given the selection crite-
ria and farming production takes place. Given the large number of providers and farmers
we focus on a symmetric mixed-strategy equilibrium where ex ante identical providers and
farmers use the same strategy and farmers randomize over the set of preferable providers.
The key assumption is that agents find it difficult to coordinate their decisions in a large
market.
A type i-farmer’s strategy is a vector of probabilities Pi ≡ pi,fcfs , ....; pimkt , .... where
pij is the probability of applying to each type j-provider. A type-j provider’s strategy
consists of rental rates per hour serviced, rij and a selection rule χj ∈ [0, 1] for the market
provider. The selection rule applies only when a provider receives requests from both type
of farmers, in which case the provider prefers the large scale farm if χj = 1, prefers a small
scale farm if χj = 0, and he is indifferent between them for χj ∈ (0, 1). When the provider
receives requests from a single type of farmer type, he randomly selects one farmer for
service. Those that request a service from the fcfs provider face the same probability of
being first in line irrespective of the machine-hours demanded.
57
2.3.2 Queue lengths as strategies
Each farmer maximizes expected profits from farming trading off the probability of ob-
taining a rental service and the cost of such a service. The characteristics of their demand
(or ex-ante productivity) and geographical locations are as in Assumption 1.
ks > ks− . The expected travel time to servicing small-scale farmers is weakly higher than
that for large-scale farmers, E(ds ) ≤ E(ds− ).
A convenient object for analysis is the queue length, i.e., the expected number of
farmers requesting a service from a given provider.5 Let qij be the queue length of type
i farmers that apply to a type j provider, where i ∈ {s, s− } and j ∈ {fcfs, mkt}. Then,
qsj = sF psj and qs− j = (1 − s)F ps− j . The probabilities of approaching different providers
H hqs,fcfs + (1 − h)qs,mkt = F s (2.1)
H hqs− ,fcfs + (1 − h)qs− ,mkt = F (1 − s) (2.2)
(o − 1)(ks + E(ds )) <= k̄j and (o − 1)(ks + E(ds )) + ks− + E(ds− ) > k̄j
5
From a theory standpoint, when the number of providers and firms grow large, the probability of
requesting a service to a given provider approaches zero and it is inconvenient to work with. From an
empirical standpoint, queues are directly observable from our rental requests data while probabilities are
not.
58
(o − 1)(ks− + E(ds− )) + ks + E(ds ) <= k̄j and (o − 1)(ks− + E(ds− )) + 2(ks + E(ds )) > k̄j
In other words, for tractability we assume that the service capacity k̄j , is enough to serve
at most “o” orders; and that if the provider serves only large-scale orders, it can serve
“o-1” orders. Finally, we assume enough capacity to serve o − 1 small scale orders and 1
large scale order.
A farmer of scale i that requests service from provider j gets served with probability
∆ij . This conditional probability depends on the provider’s selection criteria and the
number of orders it can potentially serve within each period, i.e. its capacity k̄j ; as well
as on the machine-hours demanded ki and the expected travel time for service E(di ). We
assume the empirically relevant service capacity which implies that, on average, o = 3
orders can be fulfilled within the period.6 Hence, ∆ij is the sum across all possible number
of orders of type i being served, ōi , of the probability of servicing ōi type i farmers , ϕij (ōi ),
˜ ij (ōi ),
times the probability that a certain farmer of type i is chosen, ∆
X
∆ij = ˜ ij (ōi ).7
ϕij (ōi )∆ (2.3)
ōi ∈{1,2,3}
Using the definition of the probability of service, equation 2.3, it is possible to show
that the probability of type i being served (weakly) declines in the queue length of type
i′ ̸= i farmers. For the first-come-first-served provider the result is straightforward because
service probabilities decline with the number of machine-hours in the queue, irrespective
of their type. For the market provider with a selection criteria that favours type i′ farmers,
the decline in the probability of service for type i ̸= i′ is strict as the number of type i′
farmers in the queue increases. The service probability for type i′ farmers is independent
of the queue length of type i ̸= i′ due to the selection criteria.
6
This is consistent with the median number of orders served within a day in our administrative data.
7
The full derivation can be found in Appendix B.1.
59
2.3.3 Market cost of provision.
We follow Burdett, Shi and Wright (2001) and describe a farmer’s decision as a function
of the market price it would get for the rental service, rij , which in turn determines its
expected “market” profit, Ui . The agents take the value of the market profit as given
when the number of agents in the economy is large, F, H → ∞. Let qj ≡ {qsj , qs− j } be
the queue at provider j. Each farmer chooses the service provider to minimize costs given
8
Ui and the production technology:
Ci (rij , qj , ki ) = min
′
Ci (ki , rij ′ , qj ′ ) ≡ min
′
rij ′ ki
j j
subject to
∆ij πij (ki , rij , qj ) ≡ ∆ij E(z(∆ij ))kiα − rij ki ≥ Ui ,
deviations from this optimal timing to productivity costs. The realization of the land
preparation date is a random draw, θ, from a known distribution G(θ̄(∆ij )) with mean
∂ θ̄(∆ij )
θ̄(∆ij ) that depends on provider j’s probability of service. We assume that ∂∆
<0
so that a high probability of service induces shorter average wait time to service. If the
realization of the preparation date differs from the optimal, the farmer faces a productivity
cost proportional to the delay relative to the optimal date as follows, z = 1−η(θ−θ⋆ )Iθ⋆ ≤θ ,
8
Notice that the farmer takes the capital demand as given for convenience. We could trivially model
the link between land-holdings and capital demand through a Leontief production function between
capital and land.
60
where η is the productivity cost per delayed service day. Expected productivity is
Given that the productivity function is asymmetric, the expected productivity is in-
dependent of the choice of provider whenever the expected wait times are relatively low,
i.e. the probability of service is high. We assume this is the case in the reminder of the
analysis, i.e. E(z(∆ij )) = 1. Appendix B.2 includes the analysis when there are expected
losses from timing at the beginning of the season. Finally, because the draw of the service
provision is idiosyncratic, there is no aggregate uncertainty in the economy and factor
A type i farmer requests a service from a type j firm with positive probability if the
expected profits are weakly larger than Ui . The strict inequality cannot hold because then
a type i farmer would apply to that provider with probability 1, yielding qij → ∞ as the
number of farmers grows large. Then, ∆ij → 0 contradicting ∆ij πij (rij , ki ) > Ũi . The
farmers’ strategy is
This expression summarizes the tradeoff between lower provision cost and higher prof-
its; and a lower probability of service. Given the shape of the probability function, there
exist a unique queue length q(rij , Ui ) that satisfies the problem of the farmer. The farmer
decides his queueing strategy as a function of his capital demand, ki and market prices
rij .
61
2.3.4 Service Providers
A service provider with capacity j maximizes expected returns. The stock of machine-
hours available to a provider are exogenously given at k̄. For simplicity, we assume the
same service capacity across providers, no depreciation or capital accumulation and no
maintenance costs.9 The cost of servicing a farmer depends on its location relative to
the provider. The location of each plot is indexed by di and dq̂t is a vector collecting
the locations of the orders completed within a period, q̂t . Providers choose the cost of
service rij taking the the machine-hours demanded by each type of farmer as given. For
a vector {Ui , ki , E(di )}i=s,s− , he chooses the queue lengths by picking the cost of service
and service strategy, i.e. whether to prioritize certain farmer types. The cost of travel
time includes the foregone services that could have been provided if the equipment was
not travelling as well as the opportunity cost of the driver, which commands a wage w per
hour. Finally, the queue length is reset each period and therefore the service provision
problem is static.10
queue qfcfs depends on the numbers of orders of each type being served, {ōs , ōs− } and the
revenue per type net of labor and transportation costs, {ri,fcfs ki − wki − wE(di )}.11 The
9
In mapping the model to the data, each provider correspond to a piece of equipment and therefore it
is reasonable to assume that service capacity per machine is the same irrespective of the dispatch system
used.
10
This feature allow us to handle the high dimensionality of the combinatorics problem when providers
are allowed to prioritize certain farmer types.
11
We assume the revenue is separable in the number of orders and relax this assumption in the quanti-
tative exercise when the provider optimizes service provision in space, i.e. minimizes transportation cost
across orders.
62
value for a first-come-first-served provider is
V (qfcfs ) = max Ṽ {ōs , ōs− }qfcfs , {ri,fcfs ki − wki − wE(di )}i=s,s− (2.6)
{ri,fcfs }i=s,s−
X
ks (i) + E(ds (i)) ≤ k̄fcfs . (2.7)
i∈qfcfs
Consider now the problem of a market provider who, in addition to choosing the cost
of provision, rmkt , chooses a selection criteria χ. This choice in turn determines the type
of orders being served and their quantity, given service capacity. The value of a market
provider is
V (qmkt,t ) = max Ṽ {ōs , ōs− }(qmkt ,χ) , {ri,mkt ki − wki − wE(di )}i=s,s− (2.8)
χ,{ri,mkt }i=s,s−
X
ks (i) + E(ds (i)) ≤ k̄mkt . (2.9)
i∈qmkt
F
Consider the ratio of farmers to hubs, H
, as well as the fraction of providers that serve on
a first-come-first-served basis, h, as exogenously given. A symmetric equilibrium consists
of farmers expected profits Us , Us− , provider strategies rij , χ, and farmer strategies, qij for
i = {s, s− } and j = {fcfs, mkt}, that satisfy:
63
1. given Us , Us− and other providers’ strategies, each type provider maximizes value,
equations 2.6 or 2.8;
3. the values Us , Us− , through qij , are consistent with feasibility, equations 2.1 and 2.2.
Proposition 2.1. In all symmetric equilibria where providers serve both types of farmers,
the selection process is χ = 1 and the per period profit from farmers of type i is Vij :
j
Vij = γ1i j
(zksα − wks − wE(ds )) + γ2i (zksα− − wks− − wE(ds− )),
j j
where γ1,i , γ2,i are non-linear functions of the queue lengths and the elasticity of the service
The expected per period value of servicing large-scale farmers is higher than for low-
scale farmers, Vsj > Vsj− . If the surplus from large-scale orders is sufficiently larger than
from small-scale orders, the expected profit from service for large-scale farmers is greater
than for small-scale farmers, Us > Us− .12
A few characteristics are worth highlighting. First, differences in location and the cost
of travel explain disparities in the incentives to serve farmers operating different scales.
In other words, for two plots located at the same distance to the provider, the marginal
cost of service is lower for larger scale farmers. Second, the farmers’ expected profit from
equipment services depends on the return to his own demand for services and on the
equilibrium rental rates. The surplus from service provision is the difference between the
64
In equilibrium, farmers that are served from both provider types shall be indifferent
between them. For the expected profits to be equalized across providers, the product
between the probability of service conditional on machine-hours demanded and the cost
χ, {qij }i=s,s− ,j=fcfs, mkt to maximize total surplus in the economy, i.e.
X X
∆ij zkiα − w(ki + E(di ))
W = max (2.10)
χ,{qij }
j=fcfs,mkt i=s,s−
subject to the service capacity constraints, equation 2.7 and 2.9, and the feasibility con-
straints, equations 2.1 and 2.2. The probabilities of service are defined analogously to
those described in the decentralized allocation, equation 2.3.
This result follows from the directed nature of the search in our environment, mimick-
ing previous results in Shi (2002). Giving priority to large-scale orders among providers
that have a technology to select farmer’s types is optimal because those orders have the
lowest marginal cost of service per unit of time traveled. The queue lengths are op-
timal because the expected profits of the farms participating in the market equal the
farms’ social marginal contribution, which account for their contribution to output and
the displacement of other farms from service provision. One way to see this is to no-
tice that by maximizing surplus, the planner effectively is maximizing the sum of the
65
expected profits to the farm and the expected value of service for each provider, i.e.
W = j=fcfs,mkt i=s,s− ∆ij (πij⋆ + V̄ij⋆ ) which follows from the definition of the equilibrium
P P
expected profits in the decentralized allocation (see Appendix B.2.1). Because farms max-
imize profits and providers maximize value, these values can be taken as constants. It is
trivial from here that the queueing behaviour that solves the decentralized problem by
definition maximizes (2.10) which proves the result.
In this section, we bring the model to the data to characterize how allocations change with
The quantitative assessment of the impact of the government intervention in the rental
markets for equipment consists of two blocks. The first block solves the model in Section
2.3 for the equilibrium market rental rates and queue lengths, given the empirical supply
and demand for equipment services. The second block simulates queues and service
66
provision strategies for farmers with different scale and geographical location.
Solving the first block requires taking a stand on the heterogeneity in machine-hours
demanded. We construct two groups of farmers following their average machine-hours
requests in the transaction data: those with requests of more than 3.5 machine-hours per
order are denominated large-scale while those with requests of less than 3.5 machine-hours
are denominated small-scale. Then, we solve for an equilibrium in which both type of
farmers are served by both types of providers, as in the data. We call this equilibrium
The second block involves finding the expected delay and subsequent productivity
costs as well as provider profitability under alternative dispatch systems using equilibrium
rental rates and queue lengths from the first block. In theory, the queue length itself yields
the expected wait time by farm type. However, we recognize that empirically, farmer
heterogeneity is richer than the one accommodated by the stylized theoretical model both
in terms of machine-hours demanded and in the spatial allocation of demand. We simulate
300 paths of queues of length q ⋆ and composition (qs⋆ , qs⋆− ) as dictated by the equilibrium
of the selection model. The actual sample paths for queues (qs⋆ , qs⋆− ) are drawn from the
joint empirical distribution of machine-hours and geographical location. Then, given the
equilibrium rental rates and the technology for dispatch, we let the provider optimize
service delivery. The optimization of service provision in space is effectively the solution
Data
Consistently with the evidence in Section 2.2 we use data for the Kharif season (June to
October) in year 2018. We exploit four sources of data: (1) detailed transaction data from
the government subsidized service provider, (2) our own survey of farmers, (3) our own
67
census of farming households in the catchment area of the subsidized service providers
and (4) high-frequency data from ICRISAT.
village to the farmer’s service request. We focus again on commonly used implements,
i.e. rotavators and cultivators; and we narrow the set of provider hubs to those with more
than a 100 transactions within the season. We are left with 15 hubs for analysis. We use
information on the closest village to each request as well as service hours requested to
compute the empirical joint distribution of machine-hours demanded and service travel
time in the catchment area of each hub. We geolocate villages and hubs and compute
travel times using information from a commonly used Application Programming Interface
(API), which factors in road connectivity across locations. Rural areas are not always
well connected and while geographical distances may seem short, travel times can increase
The second source of data allows us to characterize the productivity (output per acre)
of the farmers being served. We compute the empirical joint distribution of productivity
and production scale (i.e., cultivated area) within the catchment area of each provider
hub. The survey covers more than 7500 households along 180 villages and includes detailed
information on inputs and outputs in farming.
The third source of data allows us to characterize equipment supply, including not only
the government subsidized equipment supply but also equipment owners’ supply of rental
services. We also use the census to measure the total demand of services by production
scale within the catchment area of each hub, including demand towards market providers.
The census covers more than 40000 households along 150 villages that overlap with the
catchment area of the hubs, and entails a brief set of questions about rental market
68
engagement and farming engagement.
The fourth source of data allows us to measure the productivity costs of delays. The
data entails 6200 plots in 18 villages in India during 2009-2014 with daily detailed mea-
sures of inputs and output in farming.
Parameterization
There are 10 parameters per hub that need to be calibrated, as shown in Table 2.3. 8 of
these parameters are calibrated directly from the data while the remaining 2 are calibrated
internally by solving the model. From those parameters measured in the data, 4 of them
are common across hubs: the providers’ discount factor β, and their opportunity cost of
moving equipment in space w, the productivity cost of delays η, and the curvature of
the farming profit function α. The remaining 6 parameters are hub specific and include
and standard deviation of productivity and the correlation between productivity and
machine-hours), the ratio of farmers demanding service to the providers in the catchment
area of each hub f , as well as the share of large farmers in the population of farmers
demanding equipment in the catchment area of the hub s. The latter two model-calibrated
parameters are chosen to match the queue length of small-scale farmers at first-come-first-
served providers, and to make sure the equilibrium displays positive queues of small and
large-scale farmers with both providers, as we observe in the data. In addition to these 10
parameters, we feed the distribution of plots in space (and their corresponding travel-time)
as measured from the platform data.
We set the discount factor to β = 0.96 with an implied daily discount rate of 4%.
69
Table 2.3: Parameterization
The opportunity cost of travel time equals the hourly wage of a driver which is directly
observed from the platform data, at w =|75. The curvature of the profit function is set
to 0.6, as estimated from our own survey data on farm profitability. We exploit the fact
that farming profits are proportional to this parameter, i.e. πi = (1 − α)yi and estimate
α from the average ratio of profits to value added as reported by farming households.
To discipline the productivity costs of delays, η, we use high frequency data from
ICRISAT. We study profitability and value-added per acre of about We define an optimal
planting time as the date that maximizes the profits per acre in a given village year.
Then, we define the cost of the delay as the difference in average value added per acre or
profit per acre (depending on the variable of interest) as we move away from the optimal
planting date. Formally, we estimate
where X are controls for plot characteristics, farmer, village and time fixed effects. Stan-
70
dard errors are clustered at the village level. Our estimates for the costs in value added
per acre are reported in Table B.1. They indicate that within a 5-day windows, each
additional day away from the profit maximizing date entails a cost of 3.4% in terms of
value added per acre. If we compute the cost at a 10-day window, the cost raises to 8.5%
per day. For our assessment we focus on the former, more conservative estimate of the
cost of delay, η = 3.4%.
Then, we calibrate hub-specific parameters. We use our census, to compute the share
of machinery available from government-subsidized hubs and that available from machine-
owners (i.e. we count inventory of implements per hub and implements owned by farm-
ers within the catchment area of each hub).To characterize the productivity of farmers
requested. Their correlation ranges from -0.28 to 0.35 displaying the wide-heterogeneity
in demand characteristics across hubs, column (5) in Table 2.4. When machine-hours
requested are proportional plot size a negative correlation between output per acre and
machine-hours follows from the negative correlation between productivity and farm size,
as has been documented by others in the literature, e.g. Foster and Rosenzweig (2017).
A positive correlation is consistent with more productive mechanized farms. Our data
is rich enough to display both patterns. We assume that the distribution of productiv-
ity is log-normal, ln(z̄) ∼ N (µ, σ) and fit the empirical distribution of value-added per
acre for survey farmers in the catchment area of each hub via maximum likelihood. The
travel time to services from the platform data for each hub using B-splines, akin to a
71
non-parametric estimation of the distribution. On the travel dimension, the distribution
is typically bimodal, with orders bunching at less than 10-minutes travel time from the
hub and 30-minutes travel time.13
Table 2.4: Hub specific characteristics.
Notes: Hub-specific parameters for each hub-implement combination, “Hub” in Column (1). Hubs labeled
5-7, 11 and 13 correspond to Cultivators while the remaining hubs contain information for Rotavators.
Information for hubs labeled 6-8 are correspond to different implements in a single government subsidized
hub, and therefore demand characteristics are the same. Column (2) reports the share of first-come-first-
serve providers in the total equipment supply within each catchment area. Columns (3)-(5) report demand
characteristics for each hub, including the characteristics of the distribution of productivity across farmers
and its correlation between hours demanded. Columns (6)-(7) report parameters calibrated jointly in the
model.
Parameters calibrated jointly include the ratio of farmers to providers in the catchment
area of each hub, as well as the the share of large-scale requests in that catchment area.
We minimize the distance of the latter to its empirical counterpart, while generating an
equilibrium allocation that displays service request from both types of farmers to both
type of providers (as in the data). The former is calibrated to minimize the distance
between the model predicted queue of small-scale farmers at the fcfs provider and the
13
We could have alternatively calibrated a joint distribution of productivity, machine-hours requested
and travel time. However, the overlap of the survey data and platform accounts for 20% of the survey
data and we therefore benefited from including the fullness of the distribution of machine-hours and travel
time. The latter is a key input into the costs of service and therefore the incentives to service large and
small scale farmers.
72
Table 2.5: Moments
Notes: Calibration moments, data and model counterparts, Columns (2-5). Untargeted queue length for
large-scale farmers relative to small-scale farmers, Columns (6-7).
data. To do so, we take the stand that small-scale farmers are those with requests of up
to 4 machine-hours per order, and large-scale farmers are those with orders of more than
4 machine-hours per order. These two parameters are calibrated jointly for each hub.
The calibrated ratio of farmers to providers ranges from 2.3 to 4.9, where a provider
should be interpreted as a piece of equipment, column (7) in Table 2.4. The calibrated
shares of large-scale farmers are higher than in the data (0.31 vs. 0.11 on average across
14
hubs). When there are few large-scale requests, The model generates queue lengths
for small farmers that are broadly in line in the data, Table 2.5. Queues that are “too
short” (less than 2 orders) fail to generate equilibria where both type of farmers request
service from both providers because a queue length of 1 order for the market provider
implies that small farmers are served with probability one there, given capacity. If the
queue length is instead “too long” (more than 5 orders) the model benefits an equilibrium
14
Alternatively, we could have targeted the queue of large-scale farmers in the first-come-first-serve
providers which we currently use as an untargeted moment. Results are qualitatively similar to those
reported here and available upon request.
73
or small-farmers only request service from fcfs providers, which is inconsistent with the
data.
For completeness, we report the (untargeted) ratio of queue lengths of large-scale and
small-scale farmers. On average, this ratio is lower in the data than in the model, but
this difference is mostly driven for hubs where the differences in queue lengths between
small and large-scale farmers is small. Otherwise, the model performs well in explaining
the relative length of large-scale queues relative to small-scale queues across hubs.
We solve for the rental rates and queue lengths when both types of farmers have access to
both types of providers, i.e. the status quo equilibrium. Each type of farmer is indifferent
between being served by a fcfs provider or by a market provider (see equation B.14 in
The rental rates for both types of farmers are lower at the first-come-first-served
provider than at the market provider to compensate for the longer queues (see Table B.2
in Appendix B.4). Rental rates are particularly lower for small farmers queueing with fcfs
providers, which makes these providers attractive. Consistently, the service finding rate
for small-scale farmers is larger with the fcfs than with the market providers, while the
opposite is true for large-scale farmers, as show in Figure 2.7. The level of the rental rates
are higher for small-scale than for large-scale farmers due do the higher cost of service per
machine-hour rented (i.e. machinery travel costs).
74
Figure 2.7: Equilibrium
2
small large Small mean
mean mean Large mean
.98
1.5
Service finding rate, fcfs to mkt
Rental rate, fcfs to market
.96
1
.94
.5
.92
0
0 5 10 15 0 5 10 15
hub hub
The stylized equilibrium queueing model does not capture for full extent of the observed
heterogeneity in location and machine-hours demand. We accommodate this heterogene-
ity through simulation exercises where service queues to each provider are drawn from
the empirical joint distribution of location and machine-hours observed in the catchment
area of each hub. In other words, the queue lengths of small and large scale orders are
the equilibrium ones, but their composition is allowed to vary following the empirical
distribution of machine-hours and location observed in the data. We take the stand that
a small order corresponds to less than 4 machine-hours. We sample, with replacement,
1000 queues per provider. Each order in the queue is a three dimensional object that
includes the machine-hours demanded, the location of the plot, and the productivity of
the farmer that requested service. This productivity level (output per acre) is then used
to compute the cost of equilibrium delays in service provision.15
15
As robustness, we simulate outcomes when we assume no correlation between farm productivity and
order sizes, and when we flip the sign of the empirical correlation between machine-hours and productivity
within the catchment area of a hub. These results are available upon request.
75
Figure 2.8: Demand characteristics by provider
2.5 2
travel time for service (mean)
3
acres per farmer (mean)
1.5
2
1
1
.5
0
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
On the supply side, we solve for service dispatch system through two possible delivery
routes. One follows a “hub and spoke” pattern, under which the equipment must return to
the CHC between two consecutive orders. The other allows for a solution to a “Traveling
Salesman Problem (TSP)”, where the implement travels optimally from order to order
within the day. Under fcfs, the provider follows the route that minimizes the travel time
within a given day for a given set of requests (and their order) in the queue. Under the
market allocation, the type of requests being served is jointly determined with the best
service route. The value of an order in the market allocation depends on the density of
orders around them, and the size of the order relative to the CHC serving capacity.
We fix the number of orders each provider can serve in a day to 3, in line with the
maximum number of orders that we observe being served by a driver in the administrative
data.16 We Then we estimate the value function for each provider, i.e. a function that
maps any queue of orders to their service value, conditional on the dispatch system and
the delivery route.
16
As we explain in Appendix B.3, this is a high dimensional problem, and the number of possible
combinations of orders to be served within a period grows exponentially with the number of orders in
the queue and its characteristics (including hours serviced and location, i.e., latitude and longitude).
76
Farmer allocation across providers
We start by describing the sorting of farmers into different providers, classifying farmers
by order size, i.e. the acreage of the plot for service; and by location, i.e. the travel time
for service, see Figure 2.8.
The average order size served by a market provider is 3 acres while the average order
size served by a fcfs provider is 2.2 acres. This differential is a consequence of the dispar-
ities in the queue composition discussed before. At the same time, there are systematic
differences in the travel time to locations. During a service day, market providers travel
on average .8 hours (48 minutes), while fcfs providers travel twice as much. This differen-
tial travel time reflects not only differences in the queue composition along the location
dimension, but also disparities in the ability to prioritize orders . While equal-access-
concerns may favor a fcfs service arrangement, it is possible to improve upon the baseline
allocation by allowing government subsidized hubs to optimize service delivery within a
day. Figure 2.9 displays travel times when providers are allowed to solve a TSP among
the orders served within each day. This option is particularly beneficial to fcfs providers
that are not allowed to prioritize order sizes. The average travel time (as % of service
time) is 11% for the market provider and it declines to 9% once optimizing routes. The
average travel time (as % of service time) is 15% for the fcfs provider and it declines to
11.5% once solving the TSP.
Figure 2.10 compares waiting times for different dispatch system and demand character-
istics. On average, the mean wait times faced by farmers queueing with market providers
is longer that with fcfs. This feature is in part due to differences in the composition of
the queue. Market providers have higher service rates for large-scale farmers, but those
77
Figure 2.9: Travel times
FCFS Market
without TSP with TSP without TSP with TSP
travel time (% of service time), FCFS
.25
.2
.15 .2
.1
.1 .05
0
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
farmers are only 30% of the population of farmers in the economy on average. For the
reminder of the farmers, market providers have lower service rates than fcfs providers,
which is reflected in higher waiting times among those queueing with market providers.
Despite the improvements in waiting times for farmers queueing with fcfs providers,
these providers face equipment transportation costs (in terms of the opportunity cost of
time) that can double those of the market providers. In other words, their inability to
prioritize order sizes also affects their ability to service demand in space, traveling “too
much” relative to their market counterparts. It is not surprising then that the value of
optimizing service routes is higher for fcfs providers (as we have seen from the differential
Indeed, the value of participating in the market is higher for market providers than
for fcfs providers. In other words, if given the choice, subsidized providers would choose
to operate the technology that allows them to prioritize large-scale orders. One way to
measure the willingness to pay for this technology is to compute the difference in net
present value of market participation across providers, which we show in Figure 2.11.
78
Figure 2.10: Time management by hub
.3
fcfs market fcfs market
4
.2
2
.1
1
0
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
fcfs market
.3
Provider Value, small scale farmer
.2
.1
.1
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Delays in service provision are costly for farmers. Delays are an equilibrium outcome in our
economy given the nature of search in the market for rental equipments and the disparities
in service provision across providers. Figure 2.12 left panel reports the productivity cost
per acre across providers. These are measured as the decline in revenue per acre relative
to the average revenue per acre in the catchment area of a hub.
79
Figure 2.12: Farmer and Providers’ costs
fcfs market
.4
fcfs market
cost per acre, p.p. of average productivity
8
.2
4
.1
2
0
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Small farmers queueing with fcfs providers face higher service rates, but these providers
travel larger distances (relative to service time) than market providers. These travel costs
may well overturn the productivity gains to small-holder farmers from timely access to
service. Those costs should be valued at the opportunity cost of time, which corresponds
to the wages of a driver. Figure 2.12 right panel reports the travel cost per hour serviced,
80
Figure 2.13: Relative Surplus
1.8
small large
0 5 10 15
hub
The cost per acre of service is driven by the delays in service provision and the productivity
costs to farmers. While the additional transportation costs from a fcfs arrangement are
sizable, the cost of additional time (in terms of the wages of the driver) is relatively low.
However, the opportunity cost of travel time is indeed the productivity costs of the delays
induced on others. A direct comparison across different dispatch system could be done
by computing the surplus generated across providers per day of service, e.g. the terms
summing into the surplus of the economy, equation 2.10.
We now study how the status quo equilibrium compares to the equilibrium prior to the
intervention. While data for the pre-subsidy market is unavailable, we can account for this
effect by running a counterfactual analysis where we shut down the supply of equipment
81
stemming from fcfs, i.e. the government subsidized supply. This counterfactual is valid
under the assumption that market participants accommodated the entry of new supply
in the market. This is plausible scenario given the relatively low ownership of equipment
in the population and the desire of most farmers to have some engagement in the market
for equipment.
The substantial increase in equipment supply due to the subsidy implies an increase
in the service finding rates, moving from 15% prior to the subsidy to between 40% and
55% after the subsidy depending on the provider, see Table 2.6. Gains in service findings
rates are mostly accounted for small farmers, which prior to the subsidy faced a service
finding rate of 6% and after the subsidy face a service finding rate of 56%, with a bit
more than half of it being accounted for the fcfs providers. Market providers’ service
probability also increase in equilibrium because the cost of service declines in response
to stronger competition. Rental rates falls by 18% for small scale farmers and by 5%
for large farmers. The differential effect is a consequence of the implicit priority given to
small-scale farmers by the fcfs subsidized provider.
One of the findings in Section 2.5.3 is that fcfs providers could benefit for operating a
technology that allows them to optimize equipment in space as well as optimize the type of
82
orders being served. We study the effect of a market deregulation through counterfactuals.
We first explore a scenario in which the technology that allows the market providers to
prioritize large farmers is made available to the fcfs providers. Because the prioritization
is costless and the fcfs providers are at least as well off as before (i.e. they can now
prioritize the high marginal return orders), a profit driven fcfs provider would choose to
adopt the technology, i.e. h = 0.17 In other words, there is no longer any differentiation
between these two types of providers. The nature of the equilibrium may however change.
We study the equilibrium effects in two scenarios: first we do not allow providers to exit
or enter the market, i.e. the short-run; and second, we allow entry/exit until the expected
value of market participation for farmers equals zero, i.e. their outside option.
Table 2.7: Effect of Market Deregulation
Benchmark Deregulation
fcfs mkt short-run long-run
Service Finding Rate
all 0.41 0.56 0.49 0.51
small 0.3 0.26 0.31 0.33
large 0.11 0.3 0.18 0.17
Rental Rate
small 119.3 124.5 125.5 119.6
large 92.2 93.5 94 92.3
The short run response of the economy (without endogenous entry or exit of providers)
implies higher service finding rates for large-scale farmers relative to the fcfs provision,
consistently with the profit maximizing strategy of market providers. However, their
service finding rates are below those of small scale farmers that are drawn to the market
in response to the increased supply of services. While service findings rates for small-
holder farmers do not change relative to the baseline, the rental costs increase for all
farmers. The long run response of the economy (with endogenous entry and exit) restores
rental costs to their baseline levels, and generates allocations where service findings rates
are higher for small holder farmers than large holder farmers, despite market providers
17
As we pointed out before, the value of service for market providers is always above the one for first-
come-first-served providers, and therefore each provider has incentives to adopt a dispatch system that
prioritizes large orders.
83
Figure 2.14: Impact of Deregulation
.3
4
.2
2
.1
1
0
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 6
1
2
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
prioritizing the latter. The reason is that small-scale request are valuable when there are
service capacity constraints, because they allow for better capacity utilization.
Figure 2.15 plots the change in average farm sizes served and the travel time for service,
i.e. the change in the distribution of served location, as the market deregulates. Exit of
providers induces an increase in the average size of the farm served by each provider, and
a reduction in the travel time to service, consistently with providers prioritizing services
with low marginal cost of provision. For most hubs, the travel time as a proportion of the
service time more than halves.
84
2.7 Conclusion
Rental markets hold considerable promise in expanding mechanization access and increas-
ing productivity in the farming sector. However, the spacial distribution of demand in
space and its synchronous nature, as well as the fixed supply capacity, pose interesting
trade offs between efficiency and market access. The returns to these rental markets de-
pend crucially on factors such as density, i.e. the proximity of suppliers to farmers, the
overall supply capacity, and the ability to optimize traveling equipment time. In this
paper, we document and quantify how these factors determine the allocative efficiency
and distributional effects of rental markets.
We find that when the government increases service capacity by subsidizing the pur-
chase of equipment from rental service provision, and at the same time imposes a first-
85
CHAPTER 3
MADE TO BREAK? PLANNED OBSOLESCENCE WITH
PRESENT-BIASED CONSUMERS
3.1 Introduction
In April 2022, the Biden administration announced new energy efficiency regulations
that would phase out incandescent light bulbs in the United States. This means after
dominating the market for over a century since its commercialization in the late 1800s,
the incandescent light bulb will finally give way to its more energy-efficient substitutes.
The history of how the structure of the incandescent light bulb market evolved is a
fascinating one. In particular, it witnessed the formation of arguably the first cartel on an
international scale - the Phoebus cartel. It was formed in December 1924 by all the major
manufacturers of incandescent light bulbs in the world. Evidence suggests that during its
16 years of operation, the Phoebus cartel successfully carried out planned obsolescence.
86
production technology.1 After coming into operation, the Phoebus cartel steadily brought
down the average life span of a light bulb from around 1,800 hours to 1,200 hours for
reasons other than cost reduction. The profit motive behind such an action appears to
Planned obsolescence by firms with market power - in the extreme case, by a mo-
nopolist - is well studied by economists. The existing theoretical literature offers two
main reasons why a monopolist may choose a lower level of durability. The first is to
the durable good more like a non-durable one, the severity of the commitment problem is
reduced. The second reason applies when services delivered by the new product units and
the used product units are not perfect substitutes (Waldman, 1998; Hendel and Lizzeri,
1999). Specifically, the quality of services provided by a new unit is superior to that of a
used unit. In this case, the monopolist reduces durability to further lower the quality of
the used units in order to charge higher prices for the new units and increase the overall
profits.
Absent these two reasons, that is, if the monopolist is able to commit not to lower
prices in the future by means such as drawing up future contracts or establishing the
relevant reputation, and if the services delivered by the new and used product units are
perfect substitutes, Swan (1971, 1972) finds that the monopolist chooses the same level
of durability as a perfectly competitive market with the same production technology in
87
is that the choice of durability has no impact on the present value of the revenue stream
received by the monopolist. Consumers pay for the flow of services delivered by the durable
good. The level of durability only determines how often the revenue is collected. With
time-consistent discounting, the present value of the total revenues to perpetuity remains
unchanged regardless of the level of durability. As a result, the monopolist chooses the
cost-minimizing level of durability, which coincides with the one chosen by a perfectly
by enforcing strict quantity quotas on its members. During the operation of the cartel,
there were no substantive technological advances in the design of the light bulb, therefore
services from a new unit could be considered a perfect substitute for services from a
used unit. Since neither of the two reasons that justify planned obsolescence applies
to the Phoebus cartel, the profit-maximizing level of durability for the Phoebus cartel,
according to Swan (1971, 1972), should have been the same as in a perfectly competitive
market. The existing theories fail to explain why the Phoebus cartel reduced the life span
of a light bulb by a third.
behavior of the Phoebus cartel. It rests on an important concept from the field of be-
havioral economics - time-inconsistent preferences. Specifically, instead of assuming that
consumers apply time-consistent discounting when making intertemporal decisions as in
existing theories of planned obsolescence, this paper assumes that consumers are present
biased.
88
be time consistent. A consumer’s relative preference for her well-being at an earlier date
over a later date is the same no matter when she is asked. However, empirical evidence
shows that this relative preference gets stronger as the earlier date gets closer. Thaler
(1981), Loewenstein and Thaler (1989), among many others, provide empirical evidence
that discount rates are lower for longer delays. These findings suggest that consumers
are present biased. They tend to favor immediate gratification by giving bigger weight to
payoffs that are closer to the present time when considering trade-offs between two future
periods. Such a feature could be described by a hyperbolic discounting model.
This paper finds that under perfect competition, consumers’ present-biased preferences
have no effect on durability choice. The reason is that the level of durability is determined
by cost minimization while a change in consumer’s preferences only affects demand and
therefore revenues. However, in the case of monopoly, the present value of the revenue
stream is no longer independent of the durability choice when consumers are present
biased. Consumers assign a bigger weight to the utilities in the present, which coincides
with the period of the purchase action. By reducing durability, the monopolist can induce
consumers to repeat the purchase action more often, thus increasing the overall utilities,
which is directly translated into higher willingness to pay and thus higher revenues.
The rest of the paper is organized as follows. Section 2 briefly reviews the literature
on planned obsolescence and present-biased preferences; Section 3 offers a brief history of
the Phoebus cartel; Section 4 introduces the theoretical model; Section 5 concludes.
This paper is closely related to two strands of literature. The first is theoretical works on
planned obsolescence. The commitment problem, or time-inconsistency problem faced by
89
a monopolist of a durable good was first recognized by Coase (1972) and later formalized
by Bulow (1982). The problem arises as future prices affect the future values of units
sold today. This effect acts as an externality. In the absence of the ability to commit, the
monopolist does not internalize this externality and sets prices which are too low in the
future. In a simple two period model, if the monopolist sells the durable good, which lasts
exactly two periods, at the monopoly price in the first period, she will have an incentive
to lower the price and serve the residual demand in the second period. Unless she can
commit herself not to lower the prices in the second period, consumers’ willingness to pay
in the first period is reduced, and overall monopoly profits fall. One way to resolve this
commitment problem is to reduce the durability of the good. In the two period case, if
the durability of the good is reduced from 2 periods to 1, the durable good effectively
becomes a non-durable one and the externality is eliminated. In a multi-period model, or
when time is continuous, the shorter the product lasts, the more limited the externality
is.
1998; Hendel and Lizzeri, 1999). Here, durability is defined as how fast the quality of
services provided by a product deteriorates. A secondhand market for used units exists.
By making the used unit deteriorate faster, the monopolist makes it less appealing to
the high-value consumers thereby allowing the monopolist to charge a higher price for
new units and improve the overall profit level. This explanation shares similar intuition
with the literature on pricing a product line (Mussa and Rosen, 1978; Maskin and Riley,
1984; Moorthy and Png, 1992). The monopolist’s choice of a sub-optimal durability level
for the used unit is analogical to the choice of sub-optimal quality for products sold to
low-valuation consumers.
90
a reputation in a repeated game setting, or through contractual arrangements such as
price guarantees, and if the services from the old and new units are perfect substitutes,
then Swan (1971, 1972) shows that a monopolist chooses the same level of durability as
a perfectly competitive market. The reason is that the present value of total revenues
to perpetuity does not depend on the choice of durability when consumers have time-
consistent preferences. Consumers are paying for the services derived from the product
and will adjust their willingness to pay when durability changes. From the revenue per-
spective, durability simply affects the frequency of revenue collection. It does not affect
the present value of total revenues to perpetuity. Therefore the monopolist chooses the
level of durability which minimizes the present value of the total costs to perpetuity, which
is exactly how a perfectly competitive market chooses the level of durability. A monop-
olist can still exercise its market power by charging higher prices and selling a smaller
quantity, but the level of durability is the same as under perfect competition.
However, it has long been criticized for being unrealistic (Loewenstein, 1992). Ample
empirical evidence has emerged that indicates that consumers are present biased (Thaler,
1981; Loewenstein and Thaler, 1989). The discount rate between the present and the near
future is higher than that between the near future and the far future. In other words,
consumers are more impatient over short delays than over long delays. For example,
Thaler (1981) finds that participants in his experiment were indifferent between receiving
$15 now and receiving $30 in 3 months, implying a 277% annual discount rate; the same
91
participants were also indifferent between receiving $15 now and receiving $100 in 3 years,
implying a 63% annual discount rate.
parameter model, the (β, δ)-preferences, that modifies the exponential discounting model
to capture the most basic form of present-biased preferences.
Scientific research that led to the invention of the incandescent light bulb could be traced
back to the work by the English scientist and inventor Ebenezer Kinnersley, who demon-
strated in 1761 that a brass wire became red hot when passing electricity through it
(Blake-Coleman, 1992). Further research was actively conducted throughout the 19th
century. One of the earliest patents for an incandescent lamp was received by Frederick
de Moleyns of England in 1841 (Friedel, Israel and Finn, 2010) . By the end of the 19th
century, active effort to commercialize the incandescent light bulb was carried out across
the Atlantic in both Europe (mainly Britain) and the United States.
The markets for incandescent light bulbs in the late 19th century and the early 20th
century were concentrated, largely attributed to patent protections. Cartels, or industrial
alliances at national, even continental levels were common. For example, General Electric
92
fixed prices with Westinghouse Electric Company, its primary competitor, through cross
licensing (United States v. General Electric Co., 272 U.S. 476 (1926)). Formed in 1903, the
Verkaufsstelle Vereinigter Gühlampenfabriken was a European cartel of carbon-filament
lamp manufacturers (Krajewski, 2014). However, such cartel arrangements were often
short lived due to constant innovation. For example, in 1906, two European companies
introduced a new light bulb with filament made from tungsten paste. This new bulb lasted
longer and was brighter than the carbon-filament lamp, thus rendering the Verkaufsstelle
Vereinigter Gühlampenfabriken superfluous (Krajewski, 2014).
Technological advancement in incandescent light bulbs reached its peak when General
Electric introduced a light bulb with the filament made of tungsten metal and filled
with a noble gas (“the basic bulb”). Alliances formed based on the patent for the basic
bulb were interrupted by World War I. As soon as the war ended, a new cartel, the
Internationale Glühlampen Preisvereinigung, was established in 1921 to control prices for
much of continental Europe (Krajewski, 2014). This is the predecessor of the Phoebus
cartel.
In December 1924, the Phoebus cartel, or formally the Phoebus S.A. Compagnie Indus-
trielle pour le Développement de l’Éclairage, was created in Geneva. All major light bulb
manufacturers in the world, including Germany’s Osram, the Netherlands’ Philips and
France’s Compagnie des Lampes, were members. General Electric was represented by its
British subsidiary, International General Electric, as well as the Overseas Group, which
consisted of its subsidiaries in Brazil, China and Mexico (Krajewski, 2014).2
2
Other members included Hungary’s Tungsram, the UK’s Associated Electrical Industries, and Japan’s
Tokyo Electric.
93
The officially stated goals of the Phoebus cartel were “securing the cooperation of all
parties to the agreement, ensuring the advantageous exploitation of their manufacturing
capabilities in the production of lamps, ensuring and maintaining a uniformly high quality,
increasing the effectiveness of electric lighting and increasing light use to the advantage
of the consumer” (Krajewski, 2014).
The Phoebus cartel was intended to be dissolved in 1955. However, its dominant
position was weakened by the expiration of General Electric’s patents for the basic bulb
in 1929, 1930 and 1933. Conflicts among its members and legal battles in the United
States also threatened the cartel’s stability. After the outbreak of World War II, the
cartel’s agreement was nullified in 1940 (Krajewski, 2014).
The Phoebus cartel exercised its market power through strict quantity control. The cartel
divided the world market into national and regional zones, and assigned a sales quota to
each of its member companies. Companies that exceeded their quotas were fined. While
the Phoebus cartel did not directly fix prices, the quantity control allowed it to maintain
stable prices despite falling manufacturing costs (Krajewski, 2014). The strictly enforced
quota system over the years ensured that the Phoebus cartel was not subject to the
commitment issue. Consumers therefore would not anticipate the cartel to reduce prices
in the future in an attempt to sell more.
no major improvement in the design of the bulbs. Therefore the services received by a
consumer from a new bulb was arguably identical to the services from the old one.
94
The absence of the commitment issue and perfect substitutability between the services
of new and used units suggests that the Phoebus cartel would not have any incentive to
shorten the durability of the light bulb, according to Swan (1971, 1972). However, the best
known legacy of the Phoebus cartel today is arguably the first successful implementation
of planned obsolescence in modern manufacturing.
Before the creation of the Phoebus Cartel, the average life of an incandescent light bulb
was around 1,800 hours. From 1925 to 1934, the average life decreased steadily to around
1,200 hours (Chhatre, 2021) (see Fig. 3.1). This reduction in durability was achieved by
the Phoebus cartel through rigorous research, close monitoring and strict enforcement.
In the corporate archives of the cartel member Osram in Berlin, “meticulous correspon-
dence” between the cartel’s factories and the laboratories on research to shorten the life
span of the bulbs was found. The measures included modifying the filament and adjust-
ing the current or voltage at which the bulbs operated. Each factory bound by the cartel
95
Figure 3.2: Archived document showing average operational hours of incandescent light
bulbs produced by member factories of the Phoebus cartel from 1926 to 1934
Source: Krajewski (2014) 96
agreement had to regularly send samples of its bulbs to a central testing laboratory in
Switzerland, where they underwent thorough tests against the cartel standard, including
the life span. Fig. 3.2 is a piece of historical document discovered in the Osram archive.
It documents detailed testing results of the average lives of the light bulb samples sent by
member factories from 1926 to 1934. The document confirms that such tests took place
on an annual or biannual basis on a large scale. The testing results depict a decreasing
trend in the average life span of an incandescent light bulb. Last, the cartel enforced the
shortened life span by fining any factory that submitted samples lasting longer or shorter
than the regulated standard (Krajewski, 2014).3
There is no evidence suggesting that the reduction in durability led to lower production
costs. Even the cartel members did not claim that cost reduction was the goal. Instead,
they argued that the shortened life span of the light bulb was the unfortunate byproduct
of the attempt to offer the consumers a light bulb that was “of a higher quality, more
efficient and brighter” (Krajewski, 2014). However, quotes from top management of a
cartel member suggest that this may not be the genuine reason behind the reduced life
span. After discovering an incidence where some cartel members attempted to secretly
introduce a longer lasting bulb, Anton Philips, head of Philips warned that “after the
very strenuous efforts we made to emerge from a period of long life lamps, it is of the
greatest importance that we do not sink back into the same mire ... supplying lamps that
The history of the Phoebus cartel provides us with a prime example of “made to
break” - the manufacturer makes a product that breaks down faster in an attempt to
increase profitability. The Phoebus cartel succeeded in engineering such a product, but
there is little information to confirm if such a move indeed improved profitability. During
3
Krajewski (2014) does not provide direct evidence that shows at which level this regulated standard
was. However, from other evidences such as Fig. 3.2, it can be inferred that this regulated standard was
below the average life span of light bulbs at the beginning of the formation of the cartel.
97
the years the cartel was in operation, it lost market share to cheaper bulbs imported
mainly from Japan (Krajewski, 2014). While this is likely a result of the quantity quota
and high prices, the effect of a shorter life span is unclear. A simple way to reconcile
the discrepancy between the behavior of the Phoebus cartel and the existing theories on
planned obsolescence is that the cartel made a mistake. Its profitability would have been
higher if it kept the durability unchanged. In this paper, I will argue that the reduced level
of durability was indeed the profit maximizing level and propose an alternative theory to
explain the behavior of the Phoebus cartel.
3.4.1 Assumptions
Let the structure of the market for a certain durable good be a monopoly. The good has
a “sudden death” property. Specifically, each unit of the durable good provides 1 unit of
homogeneous service in all discrete time periods t < N and provides no service in any
t ≥ N , where N is an integer and N ∈ [1, N̄ ]. There are no fixed costs of production.
Marginal cost c(N ) is a function of the durability of the product where c′ (N ) > 0. For a
fixed durability N , marginal cost c(N ) remains constant for any quantity Q > 0 produced.
At the beginning of period t = 0, the monopolist chooses a durability N . The monopolist
then announces a sale price PN (the rental option is not available). The monopolist is
able to commit to the price announced for all subsequent periods. The discount rate faced
by the monopolist is constant at δ.
The same durable good market could alternatively operate under perfect competition.
The large number of perfectly competitive firms have the same cost structure as the
98
monopolist: zero fixed cost and marginal cost c(N ).4 The perfectly competitive firms also
chooses a durability N , where N is an integer and N ∈ [1, N̄ ] at the beginning of period
t = 0. They do not set prices. Instead, they act as price takers and sell at p = c(N ).
of 1 unit of service per period and values it at v. v is sufficiently large such that at the
beginning of each period in which a purchase decision needs to be made (either to buy
the product for the first time, or to replace the used unit which broke down at the end of
the last period), each consumer purchases 1 unit of the product in equilibrium regardless
the market structure.5 The quantity sold at the market level is therefore 1 in period
t = 0, N, 2N, ...; and 0 in all other periods. Consumers are assumed to exhibit time-
a consumer’s intertemporal utility from the perspective of period t. Formally the (β, δ)-
preferences can be represented by:
For all t,
T
X
U t (ut , ut+1 , ..., uT ) ≡ δ t ut + β δ τ uτ ,
τ =t+1
where 0 < β, δ ≤ 1.
periods relative to the current period t. The exponential discounting is a special case of
(β, δ)-preferences with β = 1.
4
To make the comparison between two market structures - monopoly and perfect competition - easy,
the monopoly could be seen as when the firms under perfect competition collude as a cartel with no new
entry.
5
Demand is perfectly inelastic up to a certain price.
99
3.4.2 Durability choice under perfect competition
First, I will examine the durability choice by a perfectly competitive market with the
same production technology.
As discussed in Section 3.4.1, in each period t = 0, N, 2N, ..., 1 unit of the durable good
is produced and sold. The present value of the total costs of production to perpetuity is
c(N )
κ(N ) ≡ c(N )(1 + δ N + δ 2N + ...) = . (3.1)
1 − δN
Under perfect competition, firms choose a level of durability N pc such that it minimizes
the present value of the total costs to perpetuity κ(N ), i.e., N pc = argmin κ(N ). For
any fixed δ, N pc always exists since κ(N ) is continuous in N and N is bounded. The
uniqueness of N pc will depend on the functional form of c(N ). A sufficient condition
that guarantees the uniqueness of N pc is that κ(N ) is convex. The intuition behind a
convex κ(N ) is that while the per period marginal cost c(N ) increases in N , the present
value of the total costs may be reduced by increasing N and postponing production to
a further future. To simplify the discussion, I assume that κ(N ) is convex. For further
As illustrated in Fig. 3.3, the present value of the total costs to perpetuity κ(N ) is
convex (−κ(N ) is concave and is maximized at N pc ). Since the equilibrium market price
is fixed at P pc = c(N pc ) and quantity is equal to 1, the revenue each period and the
present value of the total revenues to perpetuity are independent of N . As a result, π(N ),
the present value of the total profits to perpetuity is an upward parallel shift of −κ(N )
100
Figure 3.3: Choice of durability by a perfectly competitive market
and is maximized at N pc . At its maximum, π(N ) = 0 due to the zero profit condition.
Next, I will determine the level of durability chosen by a monopolist when consumers have
(β, δ)-preferences. I will first examine the special case with β = 1, i.e., when consumers
have time-consistent preferences, and then analyze the general case with 0 < β < 1 when
consumers are present biased.
101
Like in a perfectly competitive market, 1 unit of the durable good is produced and
sold in each period t = 0, N, 2N, ..., as long as the price charged by the monopolist does
not exceed V (N ), the present value of the total value derived from the flow of services
from 1 unit of the good to the consumers with (β, δ)-preferences. To maximize profits,
the monopolist prices the good at V (N ), i.e.,
P m (N ) = V (N )
N
X −1
=v+β δtv
t=1
β(δ − δ N )
= v(1 + ).
1−δ
β(δ−δ N )
The present value of the total revenues to perpetuity γ(N ) = v(1 + 1−δ
)( 1−δ1 N ).
Note that this is a present value to the monopolist, whose preferences are time consistent
with a single discount factor δ. The present value of the total profits to perpetuity is
β(δ − δ N ) 1 1
π(N ) = v(1 + )( ) − c(N )( ). (3.2)
1−δ 1 − δN 1 − δN
factor δ. This is usually guaranteed by a perfect capital market. When the monopolist
and the consumers are equally patient, the present value a consumer places on the services
derived from one unit of the product in any period is exactly equal to the present value of
the revenues received by the monopolist for that period. As a result, the level of durability
N only determines how often the revenues are collected, not the total level measured by
102
the present value. If, for example, the monopolist is more patient than the consumers,
i.e. δm > δc , γ(N ) is no longer independent of N . This gives rise to the opportunity for
the monopolist to manipulate the present value of the total revenues to perpetuity by
When the monopolist and the consumers face the same discount factor, the choice
of N when β = 1 is once again a cost minimization problem and the monopolist would
choose the same level of durability as the perfectly competitive market, i.e.,
N m always exists for the same reasons discussed for N pc . It is also unique if π(N ) is
∂ 2 γ(N ) ∂ 2 κ(N )
< . (3.3)
∂N 2 ∂N 2
Equation 3.3 again relies on the convexity of κ(N ). In addition, κ(N ) must be “more
convex” than γ(N ) in order to guarantee the concavity of π(N ). I assume that π(N ) is
6
For a detailed discussion on relaxing the assumption that the monopolist and the consumers having
the same discount rate, see Schemalensee (1979).
103
concave for simplicity.
As shown in Fig. 3.4, the present value of total profits to perpetuity π(N ) is concave
by assumption. π(N ) is no longer a parallel upward shift of −κ(N ), the negative of the
present value of the total costs to perpetuity, like in the case of perfect competition. Since
the present value of total revenues to perpetuity γ(N ) decreases in N , π(N ) is maximized
at a smaller N than −κ(N ). In other words, with present-biased discounting, the mo-
nopolist chooses a lower level of durability compared to the one chosen by a perfectly
104
competitive market, i.e., N m < N pc .
The above analysis demonstrates that when consumers have time-consistent prefer-
ences, the monopolist is not able to increase the present value of total revenues to perpe-
tuity by manipulating the level of durability. If the monopolist shortens the durability of
the product, consumers will adjust their willingness to pay accordingly since what they
are paying for is the stream of services delivered by the product. This results in no change
in the present value of the total revenues to perpetuity. The monopolist therefore chooses
the level of durability to minimize costs, which is exactly how the level of durability un-
der perfect competition is determined. However, when consumers are present biased, they
assign a bigger weight to the “present” periods, defined as the periods in which a unit
of the product is purchased. By reducing the durability level therefore inducing the con-
sumers to repeat the purchase action more frequently, the monopolist can achieve higher
revenues measured by the present value, thus justifying the decision to carry out planned
Next, I will illustrate the reduction in the level of durability by a monopolist when con-
sumers are present biased using a simple numerical example. Let marginal cost c(N )
take on the form of a power function, i.e. c(N ) = N a , where a ∈ (0, 1). As discussed in
Appendix C.1, this specification of c(N ) on the given parameter value ranges guarantees
the uniqueness of the solution to the minimization problem of κ(N ). Assign the following
To find the level of durability chosen by a perfectly competitive market, i.e., the cost
minimizing level of durability, take the first order condition of Equation 3.1 and obtain
105
the following function:
N δ N log δ
a(N ) = − .
1 − δN
This function is invertible for the given parameter values (see Appendix C.1 for details).
While there is no closed-form solution for N , N can be solved numerically and the integer
solution is N pc = 6.7
Solve for N numerically and the integer solution is N m = 5.8 The durability compared
to N pc is reduced by 16.7%. The reduction increases with consumer’s per period value v
and a corner solution of N m = 1 is reached when v increases to 2.44 or higher.
It can be verified that the second order condition is negative, confirming that the
solution is a local and global maximum. Fig. 3.5 illustrates the monopolist’s profit level
at different levels of durability.
7
The real number solution to the minimization problem is N = 5.63.
8
The real number solution to the maximization problem is N = 4.92.
106
Figure 3.5: Profit of a monopolist at different durability levels
Parameter values: a = 0.5, β = 0.8, δ = 0.8, v = 1.
3.5 Conclusions
The Phoebus cartel is remembered today not only for being the first international cartel,
but for giving us one of the finest examples of a product that was “made to break”. The
cartel members carefully engineered a light bulb with a shorter lifespan in an attempt
to improve profits. While there was not enough information to confirm whether the
Phoebus cartel succeeded in it, should they have done so, the existing theories on planned
obsolescence appear to be unable to explain why. The Phoebus cartel was able to adopt
and enforce a strict quota system, which removed members’ incentives to reduce prices
in the future. With the ability to commit to future prices at high levels, the cartel did
not need to reduce the product’s durability to protect its pricing power. The slowdown in
107
technological innovation before and during the years that the cartel operated suggests that
the services from new and old units of the light bulbs were perfect substitutes. Hence
making the old units breakdown sooner and incentivizing consumers to buy new units
will not improve overall profitability. When consumers’ preferences are time-consistent,
they adjust their willingness to pay accordingly with any change in the durability of
products. The present value of total revenue to perpetuity is therefore independent of the
Behavioral economists have presented ample empirical evidence to show that con-
sumers often exhibit present-biasedness when making intertemporal decision. When con-
sumers are present biased, they assign a bigger weight to the “present” periods in which
they make a purchase. This paper shows that by assuming consumers are present biased,
a monopolist is able to increase revenues and the overall profitability by reducing the
product’s durability from the level chosen by a perfectly competitive market. This alter-
native theory on planned obsolescence provides a plausible explanation for the Phoebus
cartel’s decision to carry out planned obsolescence.
108
APPENDIX A
APPENDIX TO CHAPTER 1
109
Figure A.2: Density distribution of jump bids by size II
110
Figure A.4: Density distribution of jump bids by size IV
111
APPENDIX B
APPENDIX TO CHAPTER 2
Consider a farmer of type i that requests a service from a provider j given that other
farmers request services with probability pi,j . The probability that a given farmer is
being served given that the provider chooses one farmer of type i and he is queuing with
˜ ij (1), i.e.
this provider is ∆
fi −1
X fi − 1 1
˜ ij (1) =
∆ pnij (1 − pij )fi −1−n ,
n=0
n n + 1
where fi is the number of farmers of type i searching for a provider, fs = sF and fs− =
(1 − s)F and fin−1 = n!(ffii−1−n)!
−1!
. Hence,
fi
˜ ij (1) = 1 − (1 − pij ) .
∆
fi pij
As the number of agents in the economy gets large, and using the definition of queue
−qij
˜ ij (1) = 1 − e
∆ .
qij
That is, a given farmer of type i is served if at least one farmer of type i has requested
a service, which occurs with probability 1 − e−qij , divided by the number of requests of a
112
given type, qij .
Next, consider the probability of a given farmer being served when the provider serves
˜ ij (2). Similar computations to those above yield a service
ō = 2 orders of type i, ∆
probability as follows
−qij
˜ ij (2) = 2( 1 − e
∆ ) − e−qij .
qij
Finally, consider the probability of a given farmer being served when the provider
˜ ij (3), which follows
serves ō = 3 orders of type i, ∆
−qij
˜ ij (3) = 3( 1 − e
∆ ) − 2e−qij − e−qij qij
qij
In what follows we characterize the probability that a provider of type “j” services a
farmer of type “i” given that a farmer of type “i” is standing in the queue.
First-come-first-served. The fcfs provider only considers feasibility and the posi-
tion in the queue. Let the probability of serving ō farmer of type i be ϕi,f cf s (ō). Given the
qs +qs− qs +qs− !
queue lengths at this provider, there are Po = (qs +qs− −o)!
possible permutations for
the o-tuple, (the provider identifier has been dropped for notational convenience). Under
Assumption 1, a fcfs provider serves a single large-scale farmer if one of the large-scale
farmers are among the first three positions in the queue and at least one has applied. Let
q −
s P2
this probability be ϕ̂s,f cf s (1) ≡ 3qs qs +qs− P
.
3
113
gle farmer of type s has requested service. To this probability we should add the
probability of service when less than o = 3 farmers apply for service, ψ̂s,f cf s (1) ≡
e−qs,f cf s (e−qs− ,f cf s + qs− ,f cf s e−qs− ,f cf s ) which is the probability of service of large scale
order when there are no other service request or there is exactly one additional order
requested.
A fcfs provider services 2 large-scale farmers if there are two or more large-scale orders
in the first o positions of the queue and at least two large scale farmers have applied . Let
qs −2+q −
s P1
the first probability be ϕ̂s,f cf s (2) ≡ 3qs (qs − 1) qs +q −
s P3
where ψs,f cf s (2) = (1 − e−qs,f cf s − e−qs− ,f cf s qs e−qs,f cf s ) is the probability that there are
at least three orders in the queue conditional of a farmer of type s requesting service,
2
of which at least two are of type s (including the one requesting service). To this
probability we should add the probability that there are only two large-scale farmers in
the queue ψ̂s,f cf s (2) ≡ qs,f cf s e−qs,f cf s e−qs− ,f cf s .
Given feasibility, the fcfs provider never serves 3 large-scale orders, ϕs,f cf s (3) = 0.
A fcfs provider serves a single small-scale farmer if there is one of them in the first o
positions of the queue. This probability is defined analogously to its counterpart for large
scale orders, exchanging indexes,
and adding the probability ψ̂s− ,f cf s (1) when there are less than three orders.
2
This is the probability that at least another large scale and at least one small scale farmer request
service, or at least two other large scale farmers request service.
114
A fcfs provider services 2 small-scale farmers if at least two small-scale orders in the
qs− (qs− −1)qs
first o positions of the queue, ϕ̂s− ,f cf s (2) ≡ 3 qs +q −
s P3
where ψs− ,f cf s (2) = (1 − e−qs,f cf s )(1 − e−qs− ,f cf s ) is the probability that there are at
least three orders in the queue conditional of a farmer of type s− requesting service, of
which at least one is of type s and at least two are of type s− (including the one requesting
the service). To this probability we should add the probability that there are only two
orders in the queue, ψ̂s− ,f cf s (2) defined analogously than for large-scale farmers.
A fcfs provider services 3 small-scale farmers if there are three small-scale orders in
q −
s P3
the first o positions of the queue. This probability is defined as ϕ̂s− ,f cf s (3) = qs +q −
s P3
where ψs− ,f cf s (3) is the probability of having at least two other small scale requests, i.e.
ψs− ,f cf s (3) = (1 − e−qs− ,f cf s − qs− ,f cf s e−qs− ,f cf s ).
3
X
∆i,f cf s = ˜ i,f cf s (ō),
ψ̂i,f cf s (ō) + ϕi,f cf s (ō)∆ (B.1)
ō=1
The main difference in the probability of service for large and small relies on the queue
lengths. If the queue lengths are identical, then a first-come-first-served provider serves
both types of farmers with the same probability, 3ō=2 ϕs− ,f cf s (ō) = 3ō=2 ϕs,f cf s (ō).
P P
115
Market. The market provider has a technology that allows him to prioritize farmers
of either type. The probability of interest is the probability that exactly ō farmers of type
i are served conditional on the farmer under consideration having applied and at least 3
(1 − χ)ϕ̃s,mkt (1) = (1 − χ)ϕs,f cf s (1); or if the provider prioritizes large scale farmers
and no other large-scale farmer requested service, χψs,mkt (1) = χe−qs,mkt (1 − e−qs− ,mkt −
qs− ,mkt e−qs− ,mkt ). These service probabilities add up to,
where we should the events when there are less than three orders ψ̂i,mkt (1) = ψ̂i,f cf s (1)
for any i=s, s− by definition.
Two large-scale farmers are served by a market provider if he does not prioritize
large orders and they stand in the first 3 positions, which happens with probability
(1 − χ)ϕ̃s,mkt (2) = (1 − χ)ϕs,f cf s (2); or if the provider prioritizes those orders and there
is at least one additional large-scale service request, which happens with probability
χψs,mkt (2) = χ(1 − e−qs,mkt − e−qs− ,mkt qs e−qs,mkt ).3 These service probabilities add up
to
ϕs,mkt (2) = χ(ψs,mkt (2)) + (1 − χ)ϕ̃s,mkt (2).
to this probability we add ψ̂s,mkt (2) defined analogously than for the fcfs provider.
Feasibility prevents three large-scale orders to be served within the period and there-
3
If there are more large-scale orders the provider still serves two because of its capacity constraints.
116
fore, ϕs,mkt (3) = 0.
Analogous arguments can be used to describe the probabilities of service of small scale
farmers. A single small-scale farmer is always served by a market provider (conditional
on a request) if it prioritizes high-scale requests and at least two large scale farmers
have requested service, which occurs with probability χψs− ,mkt (1) = χ(1 − e−qs,mkt −
qs,mkt e−qs,mkt ; or if the provider does not prioritize high-scale requests and there is a single
small-scale order among the first three orders in the queue, (1 − χ)ϕ̃s− ,mkt (1), where
ϕ̃s− ,mkt (1) = ϕs− ,f cf s (1). The reason for always serving a small scale order even when
prioritizing large scale is that capacity constraints allow the provider to served at most
o − 1 orders leaving always and idle slot. To these probabilities we add those associated
to the event when there are strictly less than two orders in the queue.
Two small-scale farmers are served by a market provider if it prioritizes high-scale re-
quests and exactly one large-scale farmer requests service and at least another small scale
farmer requests service, which occurs with probability χψs− ,mkt (2) = χqs,mkt e−qs,mkt (1 −
e−qs− ,mkt ). Alternatively, two small-scale farmers are served if the provider does not pri-
oritize large-scale orders and there are two small-scale orders among the first three orders
in the queue.
ϕs− ,mkt (2) = χ(ψs− ,mkt (2)) + (1 − χ)ϕ̃s− ,mkt (2).
To these probabilities we add those associated to the event when there are strictly less
than two orders in the queue, ψ̂s− ,mkt (2).
Three small-scale farmers are served by the market provider if it prioritizes high-
117
scale requests and no large-scale farmer requests service and there are at least three
small requests, which occurs with probability χψs− ,mkt (3) = χe−qs,mkt (1 − e−qs− ,mkt −
qs− ,mkt e−qs− ,mkt ), or if it does not prioritize them and there are three small-scale orders
3
X
∆i,mkt = ˜ i,mkt (ō).
ψ̂i,mkt (ō) + ϕi,mkt (ō)∆ (B.2)
ō=1
Following equations B.1 and B.2, the probability of being served for a given type i
(weakly) declines in the queue length of the other type of farmers. In the first-come-first-
served provider the result is straightforward. For the market provider, the decline in the
probability of service is strict for the small scale farmers and independent of the queue
length of small-scale orders when the provider prioritizes large-scale orders.
service for each provider. We consider alternative scenarios, i.e. when the provider serves
at capacity (o = 3 orders) and when the provider serves less than capacity.
Φ̂ = (1 − e−qsj ) + (1 − e−qs− j )
118
The first-come-first-served provider can serve three orders of small scale (given feasi-
bility) with a service probability of
q−
−qs− ,j 1 2 s ,jP
3
Φs− ,f cf s (3) = 1 − e (1 + qs− ,j + qs− ,j ) qs,j +q − ;
2 s ,jP
3
or to serve two orders of one type and one of another, with probability
The provider can also serve two orders of large size, (either because he received only
two orders, or because all orders in the queue are of large scale)
or it can receive exactly two orders of small size and serve those,
1
Φ̃s− ,f cf s (2) = qs2− ,j e−qs− ,j (e−qs,j ).
2
Finally, the provider can serve two orders, one of each type
119
The probabilities for the market provider are similar to the ones above, except that
we need to account for the market provider’s ability to select large scale orders.
1
Φ̄s− ,mkt (3) = χe−qs,mkt (1−e−qs− ,mkt −qs− ,mkt e−qs− ,mkt − qs2− ,mkt e−qs− ,mkt )+(1−χ)Φs− ,f cf s (3),
2
(B.3)
or the orders of large scale and one small,
Φ̄s,mkt (2) = χ((1 − e−qs,mkt − qs,mkt e−qs,mkt )(1 − e−qs− ,mkt ) + (1 − χ)Φ̄s,f cf s (2), (B.4)
Φ̄s− ,mkt (2) = χqs,mkt e−qs,mkt (1 − e−qs− ,mkt − qs− ,mkt e−qs− ,mkt ) + (1 − χ)Φ̄s− ,f cf s (2). (B.5)
When there are less than three orders in the queue there is no need to prioritize orders,
and therefore the probabilities of service are identical to those characterized for the FCFS
120
us to compute the expected value of service provision:
Ṽ {ōs , ōs− }qfcfs , {(ri,fcfs − w)ki − wE(di )}i=s,s− ≡
X h i
Φ̂−1 {
Φ̄i,f cf s (2) 2 (ri,f cf s − w)ki − wE(di ) + (ri′ ,f cf s − w)ki′ − wE(di′ ) +
i=s,s−
Φs− ,f cf s (3)3((rs− ,f cf s − w)ks− − wE(ds− )) +
Φ̃i,f cf s (11 ) (ri,f cf s − w)ki − wE(di ) +
+Φ̃i,f cf s (12 ) (ri,f cf s − w)ki − wE(di ) + (ri′ ,f cf s − w)ki′ − wE(di′ ) +
Φ̃i,f cf s (2)2 (ri,f cf s − w)ki − wE(di ) } (B.6)
Ṽ {ōs , ōs− }(qmkt ,χ) , {(ri,mkt − w)ki − wE(di )}i=s,s− ≡
X h i
Φ̂−1 {
Φ̄i,mkt (2) 2 (ri,mkt − w)ki − wE(di ) + (ri′ ,mkt − w)ki′ − wE(di′ ) +
i=s,s−
Φs− ,mkt (3)3((rs− ,mkt − w)ks− − wE(ds− )) +
Φ̃i,mkt (11 ) (ri,mkt − w)ki − wE(di ) +
+Φ̃i,mkt (12 ) (ri,mkt − w)ki − wE(di ) + (ri′ ,f cf s − w)ki′ − wE(di′ ) +
Φ̃i,mkt (2)2 (ri,mkt − w)ki − wE(di ) }. (B.7)
where the first two terms in either expression correspond to the expected value of serving
three orders or different types, while the remaining terms correspond to the expected
returns of serving strictly less than three orders. Also note that the scaler Φ̂−1 is in-
consequential to the optimality conditions of the problem (and therefore omitted in the
derivations that follow).
121
B.2 Proofs
B.2.1 Proposition 1
First, we solve for the equilibrium value of service when both farmers queue with both
providers. Then, we show the value when the service provider serves a single type of
farmers. Then we show that the guess that the selection criteria for the market provider
should be to prioritize large scale orders. Finally, we show that the expected value of
service is higher for large scale farmers.
Proof. Using the definition of the expected value of service provision (equations B.6 and
B.7) and rearranging terms, as well as the participation constraint for the farmers, equa-
tion 2.4, the problem of the provider is
X Ui
max Φi,j (2)[zkiα − − w(ki + E(di ))] +
qs,j ,qs− ,j ∆ij
i=s,s−
X Ui′
Φi,j (1)[zkiα′ − − w(ki′ + E(di′ ))] +
∆i′ j
i=s,s−
U s−
Φs− ,j (3)3[zksα− − − w(ks− + E(ds− ))],
∆s− j
where Φi,j (2) = (Φ̄i,j (2)2 + Φ̃i,j (2)2 + Φ̃i,j (11 ) + Φ̃i,j (12 )) and Φi,j (1) = (Φ̄i,j (2) + Φ̃i,j (12 )).
Let V̄ij be the profit per order of type i for provider j, i.e. V̄ij ≡
(ri,j − w)ki − wE(di ) . The optimality condition swith respect to the queue length of
122
large-scale and small-scale farmers are
!
X ∂Φi,j (2) ∂Φi,j (1) ∂Φs− ,j (3)
V̄i + V̄i′ + 3 V̄s− +
−
∂qs,j ∂qs,j ∂qs,j
i=s,s
!
X ∂ V̄i ∂ V̄i′ ∂ V̄s−
Φi,j (2) + Φi,j (1) + Φs− ,j (3)3 = 0, (B.8)
−
∂qs,j ∂qs,j ∂qs,j
i=s,s
!
X ∂Φi,j (2) ∂Φi,j (1) ∂Φs− ,j (3)
V̄i + V̄i′ + 3 V̄s− +
∂qs− ,j ∂qs− ,j ∂qs− ,j
i=s,s−
!
X ∂ V̄i ∂ V̄i′ ∂ V̄s−
Φi,j (2) + Φi,j (1) + Φs− ,j (3)3 = 0, (B.9)
∂qs− ,j ∂qs− ,j ∂qs− ,j
i=s,s−
where " # !
∂ V̄i V̄i + w(ki + E(di )) − zkiα ∂∆i,j
=− .4
∂qi,j ∆ij ∂qij
Let the elasticity of the probability of service with respect to the queue length be ζq∆ (o) ≡
∂∆sj qsj ∂ V̄ qij
− ∂q ∆sj (o)
, let the elasticity of the value of service to the queue length be ζV̄ q ≡ ∂qij V̄
∂ψ q
and that of the probability of arrival of o orders to the queue length be ζqψ (o) ≡ ∂q ψ(o)
.
Then, the envelope condition indicates that the elasticity of the value of service to the
Equations B.8 and B.9 form a system of linear equations that can be solved for the
4
If weh account for the i cost of expected delays, then the envelope condition is
∂∆i,j 1 Ui ∂z ∆ij α
− ∂qij ∆ij ∆ij − ∂∆ij z zki
123
two unknowns V̄s− , V̄s as a function of the queue lengths.5
V̄s zksα
− w(ks + E(ds ))
Γ = a
α
V̄s − zks− − w(ks + E(ds ))
− −
Γ1 Γ2
where Γ = , for
Γ3 Γ4
124
and in the LHS
Φs,j (2) Φs− ,j (1) Φs,j (1) Φs− ,j (2) Φs− ,j (3)
a11 a12 ζ∆s qs ( qs,j + qs,j ) ζ∆s− qs ( qs,j + qs,j + 3 qs,j )
a= = Φ − (1) Φ − (2) Φ − (3)
Φ (2) Φ (1)
a21 a22 ζ∆s qs− ( sq −,j + qs,j− ) ζ∆s− qs− ( qs,j− + sq −,j + 3 sq −,j )
s ,j s ,j s ,j s ,j s ,j
The last vector on the LHS corresponds to the surplus from trade for each farmer type.
Standard matrix algebra implies that expected value to the providers satisfies
j j
V̄sj = γ1s (zksα − wks − wE(ds )) + γ2s (zksα− − wks− − wE(ds− )) (B.10)
V̄sj− = γ1s
j α j α
− (zks− − wks− − E(ds− )) + γ2s− (zks − wks − wE(ds )) (B.11)
j Γ1 a22 −a12 Γ3 j a21 Γ1 −a11 Γ3 j a11 Γ4 −Γ2 a21 j a12 Γ4 −Γ2 a22
for γ1s − = Γ1 Γ4 −Γ2 Γ3
and γ2,s − = Γ1 Γ4 −Γ2 Γ3
while γ1,s = Γ1 Γ4 −Γ2 Γ3
and γ2s = Γ1 Γ4 −Γ2 Γ3
.
Notice that the denominator of each of the γ parameters shifts depending on the provider
as a function of the probability of service. This heterogeneity changes the value for the
derivatives in Γ.
If a provider j attracts only large-scale farmers, i.e. qs− j = 0, then the expected
per period profit of the provider satisfies
where the second term corresponds to the surplus associated to the transaction and γ ∈
(0, 1) is a non-linear function of the queue length, the elasticity of the service probability
with respect to the length of the queue, ζ, and the share of capital in farming production.
125
Proof. The problem of the supplier when it only receives large scale orders is
max ψ V̄s
qsj ,rsj
subject to
˜ sj πs (rsj , ks ) ≥ Us
∆
X
ks (i) + E(ds (i)) ≤ k̄j
i∈q̂j
where ψ = 2 1 − e−qs,mkt (1 + qs,mkt ) +e−qs,mkt qs,mkt because there are no small-scale orders
and either the supplier serves one or two orders of large scale.
Using the definition of profits to the farmers, we can replace the cost of capital into
the objective function. Replacing the rental price of capital as a function of the expected
profits, the provider solves
" #
Us
max ψ zksα − − w(ks + E(ds ))
qsj ∆sj
Note that the properties of the probabilities ψ and ∆ (decreasing and convex in the queue
length) imply that the first order conditions to the problem are necessary and sufficient
for an optimum. The optimality condition for the queue length is
" #
∂ψ ˜ sj (2)
V¯s + w(ks + E(ds )) − zksα ∂ ∆
V̄s − ψ = 0. (B.12)
∂q ∆˜ sj (2) ∂q
Let the elasticity of the probability of service with respect to the queue length be
∂∆sj qsj
ζq∆ (o) ≡ − ∂q ∆sj (o)
and let the elasticity of the probability of arrival of o orders to the
126
∂ψ q ζq∆
queue length be ζqψ (o) ≡ ∂q ψ(o)
. Finally, let γ(qs,j , ζ∆q , ζψq , α) ≡ ζqψ +ζq∆
,
which proves the result. For the value to be positive we require γ > 0 which is true by
construction. The provider takes a fraction of the surplus from the transaction.
If a provider j attracts only small-scale farmers, i.e. qs,j = 0, then the expected
per period profit of the provider satisfies
where γ ∈ (0, 1) is a non-linear function of the queue length, the elasticity of the service
probability with respect to the length of the queue, ζ, and the share of capital in farming
production.
The derivations when the provider serves only small-scale providers follow the same
∂Πmkt
The market provider wants to prioritize large scale orders: Compute ∂χ
,
which are strictly positive, given the definition for the unconditional probabilities of ser-
vice, B.3 to B.5, and the value of the provider, equations 2.8 and B.7. Then, the optimal
Expected profits to the farmers The expected profits to the farmers depend on the
equilibrium being realized, i.e. whether providers serve both type of farmers or providers
specialize in a single type. The reason is that the expected profits to the farmer depend
127
on the cost of service, which can be in turn expressed as a function of the value of service
using the definition of the value per period, zkiα − Ui
∆ij
− w(ki + E(di )) = V̄ij
j j
Us = ∆sj ((1 − γ1s )(zksα − wks + wE(ds )) − γ2s (zksα− − wks− + wE(ds− )))
j α j α
Us− = ∆s− j ((1 − γ1s − )(zks− − wks− + wE(ds− )) − γ2s− (zks − wks + wE(ds )))
When the providers specialize in service provision, they determine the expected
profits to the farmer.
economy should be satisfied.6 We describe the indifference condition for large scale farm-
6
If they choose to reach out to a single provider, then the equilibrium queue length is determined by
128
ers, the ones for small scale farmers are analogous.
fcfs
∆smkt (1 − γ1s )(zksα − wks − wE(ds )) − γ2s
fcfs
(zksα− − wks− − wE(ds− ))
= mkt mkt
∆sfcfs (1 − γ1s )(zksα − wks − wE(ds )) − γ2s (zksα− − wks− − wE(ds− ))
When small scale farmers queue only with fcfs providers, the indifference condition for
the large farmer is
fcfs
∆smkt (1 − γ1s )(zksα − wks − wE(ds )) − γ2s fcfs
(zksα− − wks− − wE(ds− ))
=
∆sfcfs (1 − γsmkt )(zksα − wks − wE(ds ))
These indifference conditions jointly with the feasibility constraints of the economy,
equations 2.1 and 2.2, yield the optimal queue lengths by provider and type.
Rental rates The rental rates can be computed from the definition of U once the
optimal queues have been solved for.
Value of service for farmers Us ≥ Us− whenever the different in the surplus of
service for large-scale providers is large enough. Because in equilibrium the market value
of service is the same irrespective of the provider, it is w.l.o.g. to use the values from the
fcfs providers.
U s − U s− ≥ 0
j j α
(∆sj (1 − γ1s ) + ∆s− j γ2s − )(zks − wks − wE(ds )) −
j j α
(∆sj γ2s + ∆s− j (1 − γ1s − ))(zks− − wks− − wE(ds− ))) ≥ 0
If the surplus is weakly higher for large scale farmers, zksα − wks − wE(ds ) ≥ zksα− −
feasibility only, which in turn determines the expected value for farmers.
129
wks− − wE(ds− ), then it is sufficient that
j j
(∆sj (1 − γ1s ) + ∆s− j γ2s −) (zksα− − wks− − wE(ds− ))
j j ≥ . (B.15)
(∆sj γ2s + ∆s− j (1 − γ1s − ))
(zksα − wks − wE(ds ))
The above is a condition on the elasticities of the probability of service to the queue
lengths (in γ) relative to the values of the surplus.
Value of service for providers The value of serving large farmers is higher than
the value of serving small-farmers for any provider For V̄s > V̄s− it is sufficient that
(γ1s − γ2s− ) > 0 and (γ2s − γ1s− ) > 0, see equations B.10 and . This is the same as
If the queue lengths are the same Γ4 + Γ3 = Γ2 + Γ1 and also a11 > a21 and a12 > a22
because the elasticity of the probability of service to the queue large farmers is higher than
for the queue of small farmers. By continuity, if the queue lengths are not too different
the above result holds. Intuitively, the reason is that if there are no systematic differences
in travel time across farmers, then the provider’s marginal cost of provision is higher for
the smaller farmers and therefore the provider finds them less valuable.
130
B.3 Numerical Solution and Output
The value function maps an ordered queue to the expected present value of this queue.
Each order i in the queue comprises two dimensions: hi , the number of hours demanded
discretized to 6 bins, and di the travel hours to and from the hub that represents a variable
cost of service. For a queue length equal to 3, the value function is a mapping from R6
to R1 .
The relatively high dimensionality of the problem prompts us to implement the sparse-grid
method proposed by Smolyak (1963) (see Judd et al. (2014) for details). The grid points
are selected for an approximation level of 2, which results in 85 grid points. We then
B.3.2 Simulations
We simulate the expected wait time and productivity cost under the fcfs arrangement and
the market arrangement respectively for three cases: when productivity is uncorrelated,
length equal to the number of actual orders, from a log normal distribution where the
131
parameters are obtained by fitting the actual productivity information to a log normal
distribution by hub. We then choose a sequence for each simulation case that produces
a correlation with hours demanded in the data that is the closest to a target correlation
for that case, and assign that sequence to the actual orders. In the uncorrelated case,
the target is zero; in the negatively correlated case, the target is the actual correlation
for that hub; in the positively correlated case, the target is symmetric to the negative
correlation.
We use bootstrap sampling of the actual orders for the simulation and assume each
bootstrap sample represents an actual queue.
We compute the wait time for the first three orders in each bootstrap sample under the
fcfs arrangement and the market arrangement respectively. In any period if one or more
out of these three orders are not served, the queue is filled going through the bootstrap
sample. To avoid having a large order ”jamming” the queue, we assume every order
is feasible. This implies that the maximum wait time under the fcfs arrangement is 3.
We cap the wait time at 5 for the market arrangement. The productivity cost is then
calculated by multiplying the simulated productivity by a percentage loss as described in
Table B.1: Costs of Delays Relative to Optimal Planting Time, Value Added per Acre
Cost per day, value added per acre
Whole Sample 5-day around optimal 10-day around optimal
Before After Before After
β1 -41.97 391.1** -215.7 1,166*** -931.1***
(26.33) (140.2) (146.9) (338.7) (298.5)
132
Ownership rate
0 .2 .4 .6
Th
r es
he
H r
ar
ve
st
e r
D
is
c
Pl
ou
gh
Sp
ay r
R er
ot
av
at
C or
fig1
ul
tiv
at
Kn or
Owned
ap Tr
sa ac
ck to
r
Sp
ra
ye
An r
im
al
pu C
ar
lle t
d
eq
H m
an t
d
To
ol
s
133
Rental rate
0 .2 .4 .6 .8
Th
r es
he
r
Sp
r ay
total population surveyed.
H er
ar
ve
H st
e r
fig2
C
ar
t
Pl
ou
Figure B.1: Ownership and rentals by implement.
gh
Rentals
D
is
An R c
i m ot
av
al at
pu or
lle
d
eq
m
C t
ul
tiv
at
or
Tr
ac
to
r
The ownership (rental) rate is the share of farmers that report to own a given implement relative to the
Table B.2: Status Quo Equilibrium Outcomes
(a) Average Queue Length (b) Rental Rates, |
small large small large
fcfs mkt fcfs mkt fcfs mkt fcfs mkt
1 2.0 2.6 1.0 1.7 1 125 126 92 93
2 3.0 1.8 1.0 2.1 2 108 117 89 93
3 3.5 1.4 2.0 5.4 3 123 131 101 101
4 4.0 1.6 2.1 4.6 4 115 122 92 92
5 3.5 1.5 1.1 3.0 5 127 133 100 101
6 4.3 1.9 1.1 2.6 6 118 123 91 91
7 3.3 1.8 1.2 2.4 7 113 119 90 90
8 4.3 2.0 1.1 2.5 8 116 120 90 91
9 2.0 2.2 1.1 2.2 9 137 138 101 101
10 2.0 4.8 1.2 0.1 10 142 144 104 104
11 2.0 2.6 1.0 1.6 11 131 132 92 93
12 2.5 1.4 1.1 2.8 12 114 123 90 91
13 3.5 1.4 1.1 3.3 13 102 111 89 90
14 4.0 1.6 1.2 3.9 14 109 114 84 84
15 4.0 1.9 1.0 2.3 15 118 124 91 96
Panel (a) reports equilibrium average queue lengths by hub for farmers of different scale and
different providers. Panel (b) reports the equilibrium rental rates per hour for the relevant
implement.
134
APPENDIX C
APPENDIX TO CHAPTER 3
The optimal level of durability is determined by the cost minimization problem under per-
fect competition and under monopoly when consumers have time-consistent preferences.
A sufficient condition for the uniqueness of the solution to the cost minimization problem
c(N )
is for the present value of the total costs to perpetuity κ(N ) = 1−δ N
to be convex on the
relevant parameter value ranges. Whether this condition holds depends on the specific
functional form of the marginal cost c(N ). By assumption, c′ (N ) > 0. One class of func-
tions that satisfies this assumption is the power functions c(N ) = N a , where a > 0. Let
Na
κ(N ) = 1−δ N
, where N ∈ [1, N̄ ], a > 0, and δ ∈ (0, 1). Assume local extrema (maxima or
minima) exist, then at these points,
N δ N log δ
a=− . (C.1)
1 − δN
There is no closed-form solution for N at the local extrema. For the local extrema
to be unique, Equation C.1 needs to be invertible on the specified parameter ranges for
∂a
N and δ. It can be shown that ∂N
is negative for N ∈ [1, N̄ ] and δ ∈ (0, 1). The strict
monotonicity guarantees the invertibility. This implies that there exists a unique local
extremum for κ(N ) when δ ∈ (0, 1) and a falls in its range within which Equation C.1 is
invertible. We have a > 0 by assumption. To find out the upper bound of a, it can be
similarly shown that for any fixed N ∈ [1, N̄ ] and δ ∈ (0, 1), a increases in δ, and a → 1
135
as δ → 1. Thus there exists a unique local extremum for κ(N ) when δ ∈ (0, 1), a ∈ (0, 1)
and N ∈ [1, N̄ ].
The last step is to verify the unique local extremum is a minimum for the specified
parameter ranges. The second order condition is
Substitute Equation C.1 into the second order condition, it can be simplified to
∂ 2 κ(N )
= −N a−1 δ N log δ(1 − δ N )−3 (−1 + δ N − N log δ)
∂N 2 | {z }| {z }
>0 ?
Let f (δ) = −1 + δ N − N log δ. The sign of f (δ) is not immediately clear. Take the
first derivative,
∂f (δ) 1
= N (δ N −1 − ) < 0
∂δ δ
for δ ∈ (0, 1) and N ∈ [1, N̄ ]. In other words, f (δ) decreases in δ for the specified
parameter range. Further, as δ approaches 1, f (δ) approaches 0 from above. This shows
f (δ) > 0 for δ ∈ (0, 1) and N ∈ [1, N̄ ]. The second order condition is therefore positive
and the extremum is a minimum.
136
BIBLIOGRAPHY
motivational states within the person, Cambridge University Press. Pages: xvi, 440.
Ainslie, George, and Nick Haslam. 1992. “Hyperbolic discounting.” In Choice over
time. 57–92. Russell Sage Foundation.
Ainslie, George, and R. J. Herrnstein. 1981. “Preference reversal and delayed rein-
forcement.” Animal Learning & Behavior, 9(4): 476–482. Place: US Publisher: Psycho-
nomic Society.
Athey, Susan, and Philip A. Haile. 2002. “Identification of Standard Auction Mod-
els.” Econometrica, no. 6(November): 2107–40.
Athey, Susan, and Philip A. Haile. 2007. “Chapter 60 Nonparametric Approaches to
Bajari, Patrick, and Jungwon Yeo. 2009. “Auction design and tacit collusion in FCC
spectrum auctions.” Information Economics and Policy, 21(2): 90–100.
Bassi, Vittorio, Raffaela Muoio, Tommaso Porzio, Ritwika Sen, and Esau
Tugume. 2020. “Achieving Scale Collectively.” C.E.P.R. Discussion Papers CEPR Dis-
cussion Papers 15134.
137
technology. Harwood Academic.
Bulow, Jeremy I. 1982. “Durable-Goods Monopolists.” Journal of Political Economy,
90(2): 314–32.
15(1): 143–149.
Cramton, Peter. 1997. “The FCC Spectrum Auctions: An Early Assessment.” Journal
of Economics and Management Strategy, no. 3(September): 431–95.
Cramton, Peter, and Jesse A. Schwartz. 2000. “Collusive Bidding: Lessons from the
FCC Spectrum Auctions.” Journal of Regulatory Economics, 17(3): 229–52.
Daniel, Kent D., and David A. Hirshleifer. 1998. “A Theory of Costly Sequential
Ettinger, David, and Fabio Michelucci. 2015. “Hiding Information in Open Auctions
with Jump Bids.” The Economic Journal, 126(594): 1484–1502.
138
Ettinger, David, and Fabio Michelucci. 2016. “Creating a winner’s curse via jump
bids.” Review of Economic Design, 20(September): 173–186.
Fishman, Michael J. 1988. “A Theory of Preemptive Takeover Bidding.” The RAND
ary): 100–146.
Friedel, Robert D., Paul Israel, and Bernard S. Finn. 2010. Edison’s electric light:
the art of invention. Johns Hopkins introductory studies in the history of technology.
Haile, Philip A., and Elie Tamer. 2003. “Inference with an Incomplete Model of
English Auctions.” Journal of Political Economy, no. 1(February): 1–51.
Hendel, Igal, and Alessandro Lizzeri. 1999. “Adverse Selection in Durable Goods
94(1): 1–19.
Hong, Han, and Matthew Shum. 2003. “Econometric Models of Asymmetric Ascend-
139
ing Auctions.” Journal of Econometrics, no. 2(February): 327–58.
Hungria-Gunnelin, Rosane. 2018. “An Analysis of Auction Strategies in Apartment
Sales.” Journal of European Real Estate Research, no. 2(August): 202–23.
Khazal, Aras, Ole Jakob Sønstebø, Jon Olaf Olaussen, and Are Oust. 2020.
“The Impact of Strategic Jump Bidding in Residential English Auctions.” Journal of
Property Research, no. 3(May): 195–218.
Lagakos, David, Mushfiq Mobarak, and Mike Waugh. 2015. “The Macroeconomics
of Rural-Urban Migration.” manuscript.
140
Lagos, Ricardo. 2000. “An Alternative Approach to Search Frictions.” Journal of Po-
litical Economy, 108(5): 851–873.
Loewenstein, George. 1992. “The fall and rise of psychological explanations in the
economics of intertemporal choice.” In Choice over time. 3–34. Russell Sage Foundation.
Loewenstein, George, and Drazen Prelec. 1992. “Anomalies in Intertemporal
Choice: Evidence and an Interpretation.” The Quarterly Journal of Economics,
107(2): 573–597.
Loewenstein, George, and Richard H. Thaler. 1989. “Anomalies: Intertemporal
Choice.” Journal of Economic Perspectives, 3(4): 181–193.
Milgrom, Paul R., and Robert J. Weber. 1982. “A Theory of Auctions and Com-
petitive Bidding.” Econometrica, no. 5(September): 1089–1122.
Moorthy, K. Sridhar, and Ivan Png. 1992. “Market Segmentation, Cannibalization,
O’Donoghue, Ted, and Matthew Rabin. 1999. “Doing It Now or Later.” American
Economic Review, 89(1): 103–124.
Paarsch, Harry J. 1997. “Deriving an Estimate of the Optimal Reserve Price:
141
Perrigne, Isabelle, and Quang Vuong. 1999. “Structural Econometrics of First-
price Auctions: A Survey of Methods.” Canadian Journal of Agricultural Economics,
47(3): 203–223.
Pingali, Prabhu, Ives Bigot, and Hans P. Binswanger. 1988. “Agricultural mech-
anization and the evolution of farming systems in sub-Saharan Africa.” The World
Bank.
Sattinger, Michael. 2002. “A Queuing Model of the Market for Access to Trading
Partners.” International Economic Review, 43(2): 533–547.
Satyasai, KJS, and M Balanarayana. 2018. “Can mechanization in agriculture help
Sommervoll, Dag Einar. 2020. “Jump Bids in Real Estate Auctions.” Journal of Hous-
ing Economics, , (September): 101713.
142
Suri, Tavneet. 2011. “Selection and Comparative Advantage in Technology Adoption.”
Econometrica, 79(1): 159–209.
Swan, Peter. 1972. “Optimum Durability, Second-Hand Markets, and Planned Obsoles-
Sønstebø, Ole Jakob, Jon Olaf Olaussen, and Are Oust. 2021. “Opening
Bid Strategies in English Auctions.” Journal of Real Estate Research, no. 1(Jan-
uary): 123–43.
659.
Waldman, Michael. 1993. “A New Perspective on Planned Obsolescence.” The Quar-
terly Journal of Economics, 108(1): 273–283.
Waldman, Michael. 1998. “Durable Goods Pricing When Quality Matters.” The Jour-
nal of Business, 69(4): 489–510.
Yamauchi, Futoshi. 2016. “Rising real wages, mechanization and growing advantage of
143