pay-per-bid auctions with an exit option: an experimental ... · pdf fileabstract online...
TRANSCRIPT
Pay-Per-Bid Auctions with an Exit Option: An
Experimental and Empirical Investigation
Author: Tai Lam
Supervisor: Dr. Ben Greiner
A thesis submitted in partial fulfilment of the requirements for the degree of
Bachelor of Commerce/Economics (Honours)
School of Economics
24th October 2011
1
Declaration
I hereby declare that this thesis submission is my own work. Any contributions or
materials by other authors which are used herein have been appropriately acknow-
ledged. Furthermore, this thesis has not been submitted to any other university or
institution as part of the requirements for another degree or award.
Signed,
Tai Lam
24th October 2011
2
Acknowledgments
First and foremost, I would like to thank my supervisor, Dr. Ben Greiner, for his
knowledge and guidance throughout the year. If not for his encouragement and
passion for the topic, I would surely not have been able to see this thesis all the way
through.
I would like to acknowledge the generous support afforded to me by the NSW
Treasury Honours Scholarship and the grant from the ASBLab Small Projects Fund
without which there would not be any experiment at all.
Thanks must also go to Dr. Paul Pezanis-Christou, for the many helpful discus-
sions and insights provided in our meetings, and for acting as my discussant at the
Honours Colloquium. Much appreciated insights and comments were also received
from Dr. Bill Schworm. To all the other participants at the Honours Colloquium
and other honours presentations, thank you for your comments and suggestions.
To Dr. Andreas Ortmann and again Dr. Ben Greiner, thank you for igniting my
interest in behavioural and experimental economics.
Many thanks to my family for their support and understanding during this busy
year. Special thanks to my brother, Loune, for teaching me programming all those
years ago, it has been immensely helpful in many unexpected circumstances not
the least of which was this thesis, and for troubleshooting any and all programming
issues I encountered.
To the 2011 economics honours cohort, thank you so much for the many great
memories and fun moments. It is you people’s company, whether it is during lunch
eating noodles, or in the honours area talking about economics concepts so advanced
that not even a nobel laurete could follow, or just plain ranting on about thesis, that
have defined my economics honours experience.
3
Contents
1 Introduction 9
2 Auction Details 13
3 Literature 15
4 Model 18
4.1 Baseline Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Exit Option . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5 Empirical Investigation 24
5.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.2 Duration Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.2.1 Platform Comparison . . . . . . . . . . . . . . . . . . . . . . . 26
5.2.2 Exit and No Exit auctions . . . . . . . . . . . . . . . . . . . . 28
5.3 OLS Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4 Auctioneer Profits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.5 Bidder Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6 Experimental Investigation 38
6.1 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
6.2 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.3 Duration Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.4 Probit Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.5 Auctioneer Profits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.6 Bidder Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
7 Discussion 53
8 Conclusion 56
A Empirical Regression Outputs 59
4
B Experimental Regression Outputs 60
B.1 Probability of Bidding . . . . . . . . . . . . . . . . . . . . . . . . . . 60
B.2 Probability of Exiting . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
C Experimental Documentation 62
C.1 Experimental Script . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
C.2 Experimental Instructions . . . . . . . . . . . . . . . . . . . . . . . . 64
C.3 Comprehension Test . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
C.4 Post Experimental Questionnaire . . . . . . . . . . . . . . . . . . . . 70
5
List of Figures
1 BidRivals.com website . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2 Payoff Schedule with Exit AV Option . . . . . . . . . . . . . . . . . . 22
3 Payoff Schedule with Exit BV Option . . . . . . . . . . . . . . . . . . 22
4 Kaplan-Meier Survival Estimates - Platform Comparison . . . . . . . 27
5 Hazard Rate Estimates - Platform Comparison . . . . . . . . . . . . . 27
6 Kaplan-Meier Survival and Hazard Rate Estimates - BidRivals.com
(giftcard) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
7 Kaplan-Meier Survival and Hazard Rate Estimates - BidRivals.com
(non-giftcard) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
8 Percentage Profit Histograms - Platform Comparison . . . . . . . . . 34
9 Revenue and RRP Scatter Plot - BidRivals.com . . . . . . . . . . . . 34
10 Auctioneer Gross Bid Revenue . . . . . . . . . . . . . . . . . . . . . . 35
11 Bidder Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
12 Experimental Auction Decision Screen . . . . . . . . . . . . . . . . . 40
13 Kaplan-Meier Survival and Hazard Rate Estimates - Experimental
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
14 Average Predicted Probabilities of Decision to Bid . . . . . . . . . . . 48
15 Auctioneer Profits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
16 Auctioneer Profits and Cost of Goods . . . . . . . . . . . . . . . . . 51
17 Bidder Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
List of Tables
1 Summary Statistics - Empirical Data . . . . . . . . . . . . . . . . . . 30
2 OLS on Log-Transformed Final Bidding Rounds . . . . . . . . . . . . 32
3 Bidder Gains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4 Summary Statistics - Experimental Data . . . . . . . . . . . . . . . . 41
5 Bid Level Data Description . . . . . . . . . . . . . . . . . . . . . . . 45
6
6 Probit Marginal Effects of Decision to Bid . . . . . . . . . . . . . . . 46
7 OLS on Unique Bidders in an Auction: Coefficients . . . . . . . . . . 59
8 Probit Regression on Bidding Decision: Coefficients . . . . . . . . . . 60
9 Probit Regression on Exit Decision: Coefficients . . . . . . . . . . . . 61
7
Abstract
Online pay-per-bid or "penny" auctions are a relatively new type of auction
market which is rapidly gaining popularity online. The auction is character-
ised by small fixed bid increments, pay-per-bid bidding fees and a resetting
countdown timer with similarities to other all-pay auctions and the war-of-
attrition game. Empirically, estimates suggest that the auction allows the
auction operator (the sole seller of goods in the market) to collect average rev-
enues of up to 155% of the retail price, with average ending prices of around
25% of the recommended retail price. This paper seeks to extend the existing
literature by experimentally and empirically analysing the effect of an exit op-
tion, introduced by some operators, that allows participants to offset the cost
of purchasing the item at retail price with the sunk costs of previously paid
bidding fees. Our empirical analysis indicates that the exit option extends the
length of giftcard auctions while having an ambiguous effect on non-giftcard
auctions. Experimentally, we find the exit option provides auctioneers with
the potential for increased revenue through increased quantity sold per auc-
tion, along with lower volatility in earnings. Bidder welfare is also found to
be significantly higher, which suggests that the exit option may be pareto
improving.
8
1 Introduction
This paper investigates a relatively new type of online auction format that, as em-
pirical evidence suggests, allows the seller of a good to collect average revenues of
up to 155% of the value of the goods auctioned. The basic structure of the auction
is as follows:
• The price starts at zero and rises in fixed increments, the bid increment, (e.g.
$0.014) whenever a bidder places a bid.
• A fixed cost, the bidding fee, (e.g. $0.75) must be paid each time the bidder
places a bid.
• A count down timer, which resets (e.g. to 15 seconds) every time a bid is
placed, marks the close of the auction (once it reaches zero), the ending price
and the winning (last) bidder.
This auction was first introduced by Swoopo.com which began operations on Septem-
ber 2005. The mechanics of the pay-per-bid auction (also known as penny auctions
and pay-to-bid auctions) has parallels with the all-pay auctions and the war-of-
attrition game format. As with the classic all-pay auction (Shubik 1971), all bidders
pay a positive sum if they choose to participate in the auction, regardless of whether
they win or lose. The costly action required to become the winning bidder is also
reminiscent of the war-of-attrition game format. Key differences exists however, as
the main cost incurred by bidders come in the form of bidding fees, which are indi-
vidually small, and bidders are not required to place a bid every round in order to
stay in the auction.
However, what really sets this particular auction format apart from other non-
traditional auctions is the apparent success of its real world implementation, which
appears to be highly profitable for the website operator/auctioneer (who is the sole
seller of goods). As with other all-pay auctions, particularly the dollar auction, the
presence of sunk costs incurred from the bidding process seems to be a significant
9
factor in creating a dislocation between game theoretic predictions and actual ob-
served behaviour. Given the small bidding increments, empirical ending prices for
the auctions are typically 25% of the recommend retail price (RRP) of the item,
which when prominently advertised serves as effective enticement for potential bid-
ders. In effect, the losing bidders’ sunk bids subsidise the winning bidder’s payment
for the item with the remaining amount going towards the auction operator. The
auction can also be considered comparable to a lottery, where each bid is a lottery
on the chance that the timer will countdown to zero, or essentially the chance that
no one else places a bid/buys another lottery ticket.
Our paper focuses on one particular operator of this type of auction, Bid-
Rivals.com, which features a unique variation of the pay-per-bid auction. Bid-
Rivals.com offers the standard pay-per-bid auction with an exit option, the effect
of which has not been thoroughly studied within the existing literature.1 The exit
option allows bidders to purchase the item at the recommended retail price (RRP),
using their sunk bids in the auction to offset the cost of purchasing the item. This is
a unilateral option available to all bidders at any time during the auction and after
the auction ends, with the auction continuing unaffected should any bidder choose
to exercise it.
The exit option effectively caps the potential loss of the bidder to the RRP of
the good less their valuation and dramatically changes the characteristics of the
auction. Considering a bidder who would otherwise intend to purchase the item at
a retail store, the website allows the bidder to choose between purchasing the item
at a physical store (the inherent outside option, potentially below the RRP); or
some unknown chance of attaining the item at a price below RRP in the auction or
at a maximum of the RRP. The difference between the store price and RRP would
effectively be the cost of entering into the prospect.1The literature has thus far, with the exception of Byers et al. (2010), has not explicitly
addressed the implications of this exit option. Existing studies have thus far only concerned theoriginator of the auction format - Swoopo.com.
10
Our main questions deals with how bidder behaviour is affected by the presence
of the exit option and how this translates into differences in auction outcomes.
Secondarily, such differences may help explain why BidRivals.com, among other
operators, have decided to implement this feature.2
Given the particulars of this mechanism and its potentially exploitative design,
it is not surprising that there have been concerns about whether this type of market
should be treated as a form of gambling and regulated as such. The potential for
unlimited losses, in the standard auction, and an unknown chance of winning the
item would give weight to this particular argument (although it is worth considering
that these factors are not determined by the auctioneer but rather the actions of
the other bidders). The introduction of the exit option may represent a response to
such criticism.
Alternatively, operator profits may simply be higher for platforms which offer
this exit option. This would hold true to the extent that increases in revenue from
bidding and bidder participation offset bid revenues refunded from bidders exercising
the option.
Since this development in the growing online pay-per-bid auction space has not
yet been thoroughly explored and the feature dramatically changes the functioning
of the auction, it represents an interesting and important avenue of research.
This paper approaches the research question by utilising a mix of methods, firstly
by making theoretical predictions regarding the effect of the exit option on bidding
behaviour, then through the collection and analysis of empirical data on the auction
platform and lastly in utilising complementary experimental data to account for the
limitations of the theoretical and empirical analysis.
We find that the exit option does indeed influence bidding behaviour as predicted
by the theoretical model, resulting in longer lasting auctions. Bidder welfare is
unambiguously improved by the introduction of the exit option but auctioneer profits
from offering the exit option are heavily dependent profit margins of goods auctioned.2It is worth noting that Swoopo.com introduced this exit option in 2009 and a number of other
penny auction clones have also adopted this mechanism in recent years.
11
While Exit auctions provides the autioneer with lower bid revenues, more items are
sold per auction compared to No Exit auctions. Given a low enough cost of goods
sold, the Exit auction has the potential to be welfare pareto improving.
In Section 2, we provide further details regarding the auction. Section 3 presents
the existing literature on pay-per-bid auctions that we draw upon for our analysis.
Section 4 sets up the theoretical framework from which we form our predictions for
the exit option. In Section 5, empirical data of pay-per-bid auctions with an exit
option is compared to existing studies of the standard pay-per-bid auction. Section 6
details our experimental setup and analysis of experimental data. Section 7 contains
a discussion of our results and Section 8 concludes.
12
2 Auction Details
In this section we provide some further details regarding the functioning and prac-
ticalities of BidRivals.com’s pay-per-bid auction website.
Potential bidders are able to view the progress of all currently running auctions
from the main web page where up to 10 auctions may be active at any one time.
Bidders may also view details of recently concluded auctions. The types of items
auctioned at any time range from $15 gift cards to high priced electronics like tele-
visions and smart phones (all brand new). Typically, auctions of high priced items
are interspersed with a greater quantity of low priced items. New items are seeded
on the site with a high initial timer, allowing bidders to see what potential items
may be actively bid upon in the future.
If bidders wish to place a bid, they must first create an account and pre-purchase
bidding packs of 20, 50 or 100 bids (at US$0.75 per bid). Every time a bid is placed,
the auction price increments by US$0.015 and the timer of that auction resets to 15
seconds if it is currently below 15 seconds.3 If the timer reaches 0 without any bids
being placed, the last leading bidder wins the auction and pays the auction price.
At any time, bidders are provided with information regarding the last 10 bidders
(username, price bid was placed and whether a BidAgent was used) with the timer
being updated in real time. An example is shown in Figure 1.
Figure 1: BidRivals.com website
As the pay-per-bid auction platform is accessible to any individual with an inter-3Since the price rises in fixed increments, it is not possible to engage in jump bidding.
13
net connection, the potential pool of bidders is essentially worldwide, with auctions
running 24 hours of the day. This also means that auctions can be in active bidding
(in the 15 second countdown phase) for a prolonged period of time. As such, an
increase in new entrants would likely lead to auctions lasting longer and a lower prob-
ability of winning. Country specific webpages (portals) provide localised prices but
differ only in the interface, allowing access to the same pool of auctions worldwide.
BidRivals.com, as is common practice among other operators, allow the use of
a BidAgent, which will automatically place bids based on parameters set by the
bidder. The BidAgent will place bids at approximately 10 seconds left in the timer,
for as many bids as the bidder dictates or until a maximum auction price set by the
bidder is reached, whichever is reached first. Multiple BidAgents within the same
auction will bid sequentially in different rounds of the auction.
At any time during or after the auction ends, the bidder can elect to exercise
the exit option. The RRP, total auction bids placed and net payment required to
purchase the item is calculated for the individual and shown on a separate section
of the bidding screen.
A shipping and handling fee of approximately 10% of the RRP must be paid if
an item is won or bought using the exit option regardless of whether the item is
physically delivered or not.
Some auctions also features special additional characteristics. No BidAgent auc-
tions disallow the use of a BidAgent, bonus auctions have a reserve price under
which bids are refunded at a rate of 150%, beginner auctions which only allow bid-
ding from users who have not won an item before and no-buy-now auctions which
do not allow the use of the exit option.
14
3 Literature
Since the pay-per-bid auction has only existed since 2005, the literature for this
auction format is still developing. Several papers have emerged in the last few years
to investigate the various theoretical and empirical aspects of the auction design.
To date, all papers have focused on Swoopo.com. In this section we present a brief
summary of the literature.
Augenblick (2009) presents one of the most comprehensive analyses of the pay-
per-bid auction format. The author characterises a theoretical model based on
fully symmetric Markovian bidders with a common valuation. A mixed equilibrium,
where bidders bid such that their expected profits from bidding equals their expected
losses, exists where bidders and auctioneers make expected profits of zero. It is then
extended through the consideration of naive bidders who experience regret and are
susceptible to the sunk cost fallacy. While the base model fails to account for the
amount of bidding that is empirically observed, the extensions to the model do
assist in explaining the over-bidding that occurs. Examination of the theoretical
and empirical hazard rates4, which diverge as the number of bids placed increases,
also leads the author to extrapolate that the sunk cost fallacy might play a role in
driving bidder behaviour in the auction. Further, supply side issues of the auction
design, like the examination of optimal supply given the number of active bidders
and the significant start-up costs of attracting a large user base serve as a barrier
to entry to most clones of the original site are discussed.
Ockenfels, Tillmann & Wambach (2010) augment their own base theoretical
model which is structurally similar to Augenblick’s (2009) and consider reference-
dependent preferences and loss aversion as potential explanations of the observed
bidding behaviour of participants within the auction. The authors find that the
reference-dependent model assists in explaining the abnormal bidding behaviour
seen within the empirical data. Additionally, an experimental form of the auction
game is conducted in order to test the reference-dependence hypothesis and the4The probability that an auction terminates at a given number of bids placed.
15
predicted effects of changing the number of bidders and the bidding increment.
Increase in bidders permitted to participate in an auction appears to increase the
overall bidding, but reduces per-participant bidding. The authors’ empirical and
experimental research also leads them to conclude that no obvious optimal strategy
exists for the auction format.
Byers, Mitzenmacher & Zervas (2010) take an alternative approach by utilising a
number of simulation models, which incorporate asymmetric information, to explain
the observed over-bidding. In particular, they find that asymmetries in bidding fees
and perceived number of bidders are responsible for the significant profits of the
auctioneer. The authors are also one of the few in the literature to consider the
impact of shill bidding and the exit option on the auction platform. They consider
the exit option under a situation where two separate classes of bidders exist to form
the bidder population. When one subset of bidders commits to purchase the item,
and to the extent that their bidding behaviour does not affect the bidding beha-
viour of non-committed players, overall bidding and auction length should increase.
The longer the player commits to stay in the auction, the higher the profit for the
auctioneer, although this may be offset by the cost of supplying an additional item
on the exercising of the exit option. Where two or more players commit to stay in
the auction, the auction devolves into a game of chicken which results in ambiguous
predictions for the profit of the auctioneer. Under the particular conditions imposed
in the system, the exit option has the potential to lead to higher auctioneer profits.
Platt, Price and Tappen (2010) observe that a significant majority of bidding
behaviour can be accounted for even under assumptions of risk neutrality. Bidding
in high price video game consoles is found to be the exception which would require
the presence of risk-loving participants. The authors equate the auction to an in-
cremental king-of-the-hall contest and discuss the possibility of using the auction to
model rent seeking behaviour.
Nanney (2010) provides a comprehensive theoretical exploration of the pay-per-
bid auction which includes the consideration of an asymmetric equilibrium, imperfect
16
information between bidding groups and a basic model of auction entry. The author
also considers the auction in the form of lottery to assist in explaining the over
bidding observed.
Since pay-per-bid auctions are essentially a type of all-pay auction, it is worth
noting the contribution of Shubik (1971) who laid the foundation for non-standard
auction formats with the introduction of the dollar auction. Consistent over bidding
in all-pay auction experiments as in Gneezy and Smorodinsky (2004) hold parallels
with the over bidding in pay-per-bid auctions. Despite this, theoretical support from
the existing all-pay literature is limited due to a distinct difference in the structure
of pay-per-bid auctions. Whereas standard all-pay auctions and war of attrition
games can be modelled as a static game (Baye, Kovencock and Vries (1996) and
Bulow and Klemperer (1999)), characterisation of the pay-per-bid auction is more
complex. This is due to the way in which bidding fees are incurred. In a war of
attrition game, for example, costs are a direct function of the number of rounds in
which participants choose to remain in the game. This allows for the reduction of
the game to a decision concerning the number of rounds to remain contingent on the
bidder’s valuation. In the pay-per-bid auction, the bidding decision can be made in
every round of the game regardless of whether they chose to bid previously or not.
17
4 Model
In this section, we briefly replicate the standard theoretical model developed by Au-
genblick (2009). Others, for example Platt et al. (2010), Ockenfels et al. (2010) and
Hinnosaar (2010), developed similar models. Despite the inability of the standard
model to explain the profitability of this auction format, it serves as an informative
baseline to which we can anchor our empirical and experimental analysis. Although
a complete model of the pay-per-bid auction with an exit option is beyond the scope
of this thesis, we also discuss the theoretical implications of this option on bidding
behaviour.
4.1 Baseline Model
We consider fully symmetric, rational and risk-neutral bidders who have full inform-
ation about the game. Let us define n as the number of bidders in the auction, each
of which share the same valuation v. The bidding fee f is paid each time a bid is
placed and accepted as the current highest bid. For simplicity, we assume that when
more than one bid is placed in any round t, the bid which is actually accepted is
randomly chosen with equal chance. We denote the bidding increment, the incre-
ment by which the ending price rises every time a bid is accepted, as m. As such,
the current ending price in any round t to be given by tm.
Within the theoretical framework, the auction begins at a price of 0 and in each
round t, all non-leading bidders make a decision whether to place a bid or not. If
all non-leading bidders do not place a bid, the game ends and the leading bidder in
the last round wins the auction and pays the ending price. Since the ending price
increments as the auction proceeds, there is a final round of the auction where no
bidders will bid. This occurs when the value of the item gained from winning v less
the bidding fee f is lower than the ending price of the auction that must be paid
tm. This is simply given by v − f = tm, which allows us to denote the final round
of t̄ = v−fm − 1.5
5Note that this only occurs if mod(v − f,m) �= 0. Where mod(v − f,m) = 0, there exist many
18
Knowing the last round of the auction allows us to employ backwards induction
to solve for the Subgame Perfect Nash Equilibrium. Since all bidders know that
no one will bid at the t̄th round as all players gain zero payoff from bidding, n− 1
bidders will bid with certainty in round t̄− 1 as being selected as the leading bidder
allows them to win with certainty. Given this, no one will bid in round t̄ − 2, and
the iterative process repeats until the start of the game where all bidders bid with
probability 1 at round 1 and cease bidding or no one bids, depending on whether k̄
is odd or even. However, this leads to a trivial equilibrium which unsatisfactorily
provides a prediction of no bidding.
Alternatively, we can look to a mixed equilibrium where bidders bid with a
probability such that they are indifferent between bidding and not bidding. This
mixed equilibrium is considered useful within the literature as it predicts positive
probability of bidding such that any round before the final round can be reached.
Since the end of an auction occurs when no bidders submit a bid, we can express the
termination probability in terms of each bidder’s probability of bidding. Each bidder
i’s Markov strategy set consists of a bidding probability for every period where he
is not a leader {pi0, pi1, pi2...pit̄} where pit ∈ [0, 1]. The hazard rate or termination
probability can therefore be given by:
h̃(t, lt) =
�ni=1(1− pit)
1− pltt
Where lt denotes the leader at time t. As no leader exists at t = 0, the hazard rate
at that time is given by h̃(t, lt) =�n
i=1(1− pit).
With the probability of bidding and probability of termination defined, we can
now state the point of indifference where the payoffs available to each bidder from
bidding and not bidding are equal given their bidding probability as h̃(t, lt).(v −
tm) = f . More succinctly, the hazard rate when the indifference condition is satisfied
is given by h̃(t, lt) = fv−tm . Which simply states that the probable payoff from
winning the auction must equal the cost of bidding in the auction. As all non-
equilibria.
19
leading bidders are symmetric, each individual bidder’s probability of bidding can
be backed out by the probability a bid is placed, split between each n−1 non-leading
bidders, pit = n−1
�1− f
v−tm .
The model also simplifies the situation by assuming that bidding will occur with
certainty in the first round. This is driven by the intuition that if no bidding occurs,
the auction operator can simply re-list the item costlessly and thus the auction never
actually terminates at t = 0.
As such, in equilibrium, the following holds:
pit =
1 for t = 0
n−1
�1− f
v−tm for 0 < t ≤ t̄
0 for t > t̄
and
h̃(t, lt) =
0 for t = 0
fv−tm for 0 < t ≤ t̄
1 for t > t̄
The revenue predictions for the model can be understood intuitively. As the ex-
pected profits for bidders in each round are zero, auctioneer profits from each round
are similarly zero. This particular aspect of the framework and its congruency with
the traditional auction literature on revenue equivalence adds to the attractive-
ness of this particular equilibrium. Augenblick’s (2009) theoretical model therefore
provides us with an informative baseline of bidding behaviour at which no profits
can be elicited.
4.2 Exit Option
As discussed previously, the exit option allows any bidder at any time during the
auction to unilaterally purchase the item at the recommended retail price (RRP)
20
less their sunk bidding costs in that particular auction.
Efforts to implement the exit option within the existing theoretical framework
yielded unsatisfactory results, with extensions leading to increased complexity and
associated difficulty in finding an equilibrium. This is due in part to payoffs from ex-
ercising the exit option changing based on past history which affects the probability
of bidding asymmetrically for individual bidders.
That said, there are a few predictions that can be made. In instances where
the RRP (i.e. the exit option price x) is lower than the value (we refer to this
situation as Exit AV (exit price above valuation)), and there is no other outside
option available to a bidder, then exiting the auction before t > t̂i = x−bifm is a
weakly dominated strategy (where bi represents number of bids placed by bidder i).
In other words, bidders should continue to bid in the auction, such that the Hazard
rate must equal 0 until t = max(t̂i). Things become more complicated when making
the more realistic assumption that the RRP is larger than the value, or if there exists
a viable outside option. Then, a bidder is essentially choosing between purchasing
the item at a physical store; or some unknown chance of either a) attaining the
item at some price below RRP in the auction, b) buying at the RRP after a certain
amount of bidding costs have been accumulated, thereby capping the losses, or c)
making a loss in the size of a smaller amount of bidding costs. We refer to this
situation as Exit BV (exit price below valuation).
Figures 2 and 3 on the following page depict the payoff schedule for bidders as
they incur sunk bids under Exit AV and Exit BV conditions, respectively. At each
state of sunk bids, the potential payoffs of winning the auction, losing the auction
(not bidding) and exercising the exit option are shown. The parameterisation as-
sumes a valuation of $20, a bidding fee of $0.50 and exit prices of $16 and $26,
respectively, for Exit BV and Exit AV.6 Payoffs from winning and losing decline
linearly as the number of sunk bids increase whereas payoffs from exercising the exit
option are invariant to sunk bids, as the latter are used towards offsetting the exit6We ignore bidding increments in this instance as the effect on the payoff schedule is minimal.
21
price payment.
Figure 2: Payoff Schedule with Exit AV Option
The exit option essentially caps the potential losses that can be incurred by
bidders. In the Exit AV example, once a bidder has incurred more than 12 bids,
the exit option can be exercised to cap losses at $6, which allows the bidder to bid
costlessly past that point. Bidding should therefore be a weakly dominant strategy
past this kink point (until gains from winning fall below the capped losses), which
we refer to as the Valuation Exit Threshold (b̂i = v−xf ).
Figure 3: Payoff Schedule with Exit BV Option
In the case of Exit BV auctions, bidding is a weakly dominant strategy from the
22
beginning of the auction. However, past 30 sunk bids, payoffs from bidding actually
falls below payoffs from exiting. Although bidding remains a weakly dominant
strategy till that point, given the diminishing gains from staying in the auction, cost
of effort may lead bidders to exit before reaching that point.
23
5 Empirical Investigation
Similar to other online auction operators and operators of the pay-per-bid auction
format, BidRivals.com displays a range of real time data for each auction in pro-
gress. On the auction level, each auction web page contains information about the
product being auctioned (product name, item description, RRP) and auction spe-
cific information (auction ID, auction type). On the bid level, real time information
is pushed through to the web page, with the current count down timer, current
ending price and information regarding the last 10 bidders (user id, time stamp, bid
agent indicator) being displayed.
Our analysis methodology involves the replication of the duration analysis un-
dertaken by Augenblick (2009). Although Swoopo.com and BidRivals.com feature
different bid increments, bidding fees and product composition, comparison of met-
rics across the two auction platforms is undertaken as a reference point to exist-
ing literature. Duration analysis is employed to characterise differences between
Exit and No Exit auctions and ordinary least squares regressions are carried out to
parametrically isolate the effect of the exit option. Additionally, our analysis also
highlights the shortcomings of the empirical data, and motivates the complementary
experimental analysis undertaken in Section 6.
In this section, we first describe the procedure undertaken to collect our data
and follow it up with a comparison between Swoopo.com and BidRivals.com. We
then investigate the differences between exit and no exit option auctions within
BidRivals.com. Lastly we provide some tentative analysis of the auctioneer profits
and bidder gains.
5.1 Data Collection
Since bid level data is displayed and discarded as new bids are placed, and auction
level data also expires relatively quickly, it was necessary to program a spider to
continuously monitor, track and record the data available on the website. The spider
was developed in C# using Microsoft Visual Studio 2010, and incorporates the use
24
of a SQL Server Database for the storage of data. Specifically, the spider monitors
JSON strings transmitted from the BidRivals.com’s server to the web browser and
parses the data stream to track updates to auctions and the bids placed within those
auctions. The information is then collated by the program and stored within the
database. By utilising the information sent from the server in its raw form, just as
the web browser does when it fetches information from the server, our dataset has
the advantage of ensuring that the data captured is consistent with the information
on which the auction operator actually acts upon. Consequently, our dataset should
possess a better signal to noise ratio in respect of timing and ordering data when
compared to other datasets employed in the literature which have been captured
second hand from the web browser.
Collection began 1 July 2011 and ended on 31 August 2011. The complete
dataset consists of 34,402 auctions comprising of 2,743,145 individual bids from
29,176 unique bidder accounts. Collection was undertaken through the Australian
web portal.7 Primary and backup servers were utilised simultaneously to minimise
the chance of missed or corrupted data. Missing data arising from hardware and
software failures were corrected by the merging of the separate databases. Essentially
all auctions recorded held complete records of all bids placed. In instances where
both the main and off site servers failed, incomplete observations (0.01%) were
removed from our analysis dataset.
For our duration and regression analysis, we disregard all “in house” goods auc-
tioned by BidRivals.com given the unique nature of the goods. These “in house”
goods include bid packs and discount cards for other auctioned items. Although
such items, which comprise of 83% of all items auctioned and are priced between
$15 to $100, are likely to influence bidder behaviour, we restrict our analysis to
auctions concerning tangible real-world goods for clearer interpretability of results.7Essentially all items accessible from the Australian web portal are items accessible from every
other country specific portal. The only issue of note is that No Exit auctions are not available forthe US portal. To the extent that the bidder base is internationally diverse, this should not posea significant issue, but this remains a potential source of bias for our analysis.
25
5.2 Duration Analysis
In this subsection we compare the hazard rates predicted by the theoretical frame-
work with empirical hazard rates observed in our dataset. Duration analysis seeks to
characterise the probability that auctions reach a particular ending price and allows
comparisons between theoretical and empirical hazard rates. Survival rate refers
to the percentage of auctions in the sample still undergoing bidding at a specific
normalised round. Hazard rates refer to the probability of an auction ending at
a specific normalised round conditional on it reaching that normalised round. We
approximate the empirical survival and hazard rates within our auctions using the
Kaplan-Meier product-limit estimator.8 Comparison of theoretical and empirical
survival estimates and hazard rates serves as a useful first point of analysis as it
provides an aggregate indication of the extent to which bidding behaviour differs
across auction platforms and from theoretical predictions.
5.2.1 Platform Comparison
Figure 4 on the next page presents Augenblick’s (2009) survival estimates for Swoopo.com9
(left-hand side graph) and survival estimates for auctions within our BidRivals.com
dataset (right-hand side graph) which disallowed the exit option. Despite Swoopo.com
and BidRivals.com featuring different bid increments, bidding fees and goods being
auctioned, a comparison can be made with the assistance of the theoretical baseline
(dashed line).
The analysis undertaken by Augenblick (2009) shows that auction survival rates
far exceed those predicted by the author’s theoretical model. Interestingly, our
BidRivals.com survival estimates do not lie entirely to the right of the theoretical
prediction. For normalised rounds less than 20, the proportional of auction surviving
up to that round is actually less than suggested by theory. For normalised rounds8Auction data is normalised by setting each auction’s ending number of bids relative to their
RRP value normalised to 10, see Augenblick (2010). This normalisation allows comparison ofhazard rates across goods of different values.
9Survival estimates reported here pertains to auctions (n=7861) with a bid increment of $0.01(Augenblick 2009). Underlying data for the Augenblick graphs could not be obtained.
26
Figure 4: Kaplan-Meier Survival Estimates - Platform Comparison
Swoopo.com (Source:Augenblick 2009) BidRivals.com (No Exit auctions only)
Theoretical (Dashed), Empirical (Solid)
greater than 20, the survival rates deviate in a manner more similar to Swoopo.com.
Figure 5: Hazard Rate Estimates - Platform Comparison
Swoopo.com (Source:Augenblick 2009) BidRivals.com (No Exit auctions only)
Theoretical (Dashed), Empirical (Solid)
Figure 5 reports Swoopo.com and BidRivals.com’s respective hazard rates for No
Exit auctions. Smoothed hazard estimates show that the termination probability
at any normalised round is less than predicted.10 The lower than predicted survival
probabilities observed previously are no longer present after smoothing, although
early normalised round hazard rates remain relatively high. Both platforms’ hazard
rates seem to maintain a termination probability around half that of the theoretical
predictions.
For the most part, results for Swoopo.com and BidRivals.com are fairly similar,
with both exhibiting survival and hazard rates that deviate from the theoretical10The sharp increase in hazard rates above the theoretical prediction past 150 normalised rounds
is attributed to the lack of observations at that level.
27
prediction.
5.2.2 Exit and No Exit auctions
We now turn to differences between Exit and No Exit auctions within BidRivals.com.
For this analysis we separated auctioned items between categories and observed sig-
nificantly different effects of the exit option for giftcards being auctioned in com-
parison to the other items on offer. Giftcards auctioned range from $15 to $100 and
are redeemable at a range of prominent online retailers like Amazon.com. These
giftcards are essentially pure common value items which allows for a cleaner ana-
lysis.11 Non-giftcard auctions are in contrast subject to less predictable valuations,
with the analysis of the category complicated by a wide range of consumer goods
with uncertain valuations.
We report survival and hazard rate estimates for giftcard auctions in Figure 6.
Figure 6: Kaplan-Meier Survival and Hazard Rate Estimates - BidRivals.com (gift-card)
The effect of the exit option is clear, survival estimates across all normalised
rounds are strictly higher for Exit auctions compared to No Exit auctions. The
effect on hazard rates show that termination probabilities in the Exit auction are
initially much lower than the No Exit auction. However, as the auction progresses
the Exit auction hazard rates proceed to cross over and back under those of the11Recall that since shipping and handling must still be paid, the giftcard auction is an Exit AV
auction. The value of the giftcard may also be worth less to the individual depending on theirpropensity to shop at that particular retailer. The pure value nature of the items also closelyresembles our experimental implementation in Section 6.
28
No Exit auction. The difference between No Exit and Exit auctions are significant
according to a Mantel-Cox test with a p-value of 0.0000.12 A possible explanation
would be that bidders are initially attracted by the fail-safe provided by the exit
option, which drives termination probability down. This may be a result of increased
participation by bidders or increased in bids per bidder. Bidding is then maintained
by bidders entering the Valuation Exit Threshold until bidders begin exercising the
exit option. From the perspective of the No Exit auctions, the naive sunk cost
fallacy argument proposed by Augenblick (2009) appears to apply with termination
probability decreasing, relatively to the theoretical prediction, as bids are placed.
The cross over of hazard rates would seem to suggest that an exit option encourages
initial bidding, but nullifies the sunk cost effects experienced further into the auction.
It is however important to note the tentative nature of such an analysis as normalised
rounds represent the aggregate number of sunk bids distributed non-evenly among
many potential bidders.
Figure 7 depicts survival and hazard rate estimates for non-giftcard auctions.
Figure 7: Kaplan-Meier Survival and Hazard Rate Estimates - BidRivals.com (non-giftcard)
In contrast to gift cards auction, the exit option for remaining items (consist-
ing mostly of consumer electronics worth $100 - $1500) significantly reduce survival
probability (p-value of 0.0236). Hazard rates for Exit auctions are also consist-
ently higher than those of the No Exit auction. For the No Exit auctions, survival12In the absence of censored data, as in our sample, the Mantel-Cox logrank test is equivalent
to a Mann-Whitney U test.
29
probabilities are strictly higher than predicted. A possible explanation for the dif-
ference may be that the predominately higher priced items in this analysis leads to
a stronger susceptibility to the sunk cost effect, which overshadows any effects of
the exit option in encouraging bidding.
5.3 OLS Analysis
To further isolate the effects of the exit option, we undertake ordinary least squares
parametric analysis on the number of bids (round) reached by each auction. Sum-
mary statistics broken up by giftcard and non-giftcard auctions are displayed in
Table 1
Table 1: Summary Statistics - Empirical DataObs. Mean Std Dev. Min Max
Giftcard auctions 3664Bids 63.09 125.82 2 2127RRP 28.21 19.10 15 100
Beginner Feature 0.05 0.21 0 1Bonus Feature 0.00 0.02 0 1Exit Feature 0.16 0.37 0 1
No BidAgent Feature 0.17 0.38 0 1Unique Bidders 4.62 3.35 1 40
BidAgent Bids % 0.43 0.31 0 0.99BidAgents % 0.44 0.32 0 1
Non-giftcard auctions 2107Bids 484.72 1211.75 2 19909RRP 261.15 291.47 26 1599
Beginner Feature 0.02 0.12 0 1Bonus Feature 0.28 0.45 0 1Exit Feature 0.87 0.17 0 1
No BidAgent Feature 0.00 0.05 0 1Unique Bidders 18.41 27.63 1 395
% of BidAgent Bids 0.63 0.26 0 0.99% of BidAgents 0.34 0.19 0 1
******** ******** ******** ********Note: % of BidAgent Bids refers to the percentage of bids made by BidAgents
Note: % of BidAgents refers to the percentage of bidders who used a BidAgent at least once
Of note in our summary statistics, exit option auctions are far more prevalent in
non-giftcard auctions (87%) compared to giftcard auctions (16%). The value of the
30
goods are also a point of distinction between the two classes of auctions. BidAgent
bids also make up a larger percentage of non-giftcard auctions (63% vs. 43%) despite
less bidders employing the use of a BidAgent (34% vs. 44%).
Our regression analysis on log-transformed (non-normalised) final bidding rounds
(Table 2) corroborates the effects found in our duration analysis. Models (GR1) and
(NR1) isolate the effect of having an exit option for giftcard and non-giftcard auc-
tions, respectively, whilst models (GR2) and (NR2) incorporates separate dummies
for each observed type of auction (where one auction may contain more than one
feature). For giftcard auctions, the presence of an exit option increases the length
of the auction by a significant 30.16% holding all other variables constant. For
non-giftcard auctions, the effect is not significant. Auctions with both an exit and
no BidAgent feature increases final bidding round by a significant 142% in giftcard
auctions.
We also observe that the ending round increases by 4.2% for every 1% increase
in the proportion of BidAgent bids. Conversely ending round decreases by 2.4% for
every 1% increase in the proportion of bidders who employ a BidAgent.
31
Table 2: OLS on Log-Transformed Final Bidding RoundsGiftcard Non-giftcard
(GR1) (GR2) (NR1) (NR2)RRP 0.0197*** 0.0173*** 0.0022*** 0.0020***
(0.0010) (0.0010) (0.0001) (0.0001)No BidAgent 0.8227***
(0.0711)Exit 0.3501*** 0.3016*** -0.4024** -0.4980
(0.0507) (0.0539) (0.1650) (1.1306)Exit and No BidAgent 1.4153*** 3.3351
(0.1273) (1.2202)Bonus 0.1148
(1.1362)Bonus and Exit -0.3207
(1.1283)Beginner -0.3706***
(0.0967)Beginner, Bonus and Exit -0.4620 -0.5751
(0.7451) (1.1488)% of BidAgent Bids 0.0384*** 0.0422*** 0.0438*** 0.0447***
(0.0009) (0.0010) (0.0012) (0.0012)% of BidAgents -0.0261*** -0.0235*** -0.0247*** -0.0242***
(0.0009) (0.0009) (0.0017) (0.0017)Constant 1.9912*** 1.6558*** 2.7287*** 2.7489**
(0.0447) (0.0625) (0.1835) 1.1330
Adjusted R2 0.4350 0.4722 0.5490 0.5630N 3664 3664 2107 2107
Note: * p-value < 0.10, ** p-value < 0.05, *** p-value < 0.01
Additionally, a OLS regression on the number of unique bidders show that the
Exit auctions attracts an extra 2.29 unique bidders for giftcard auctions (p-value
of 0.000) and non significant numbers for non-giftcard auctions (regression output
available in Appendix A on page 59). This result is consistent with the exit option
being an attractive feature for bidders as it caps any potential losses.
5.4 Auctioneer Profits
An obvious metric of interest for auction market mechanisms and particularly pay-
per-bid auctions is the revenue and profits that can be obtained.13 We restrict our13Since BidRivals.com has its headquarters in the US, we restricted our analysis to the US Dollar.
To the extent that a majority of bidders originate from the US, this should not greatly affect our
32
calculation to revenues given the difficulties associated with determining the op-
erating costs and wholesale per unit costs of obtaining the items for the operator
(with the exception of the Swoopo.com and BidRivals comparison). On the surface,
our estimations indicate that as a whole, bids and ending price payments provide
revenues of 155% of the RRP, which appears consistent with the level of profitab-
ility reported in the no exit pay-per-bid literature.14 To arrive at this number, we
multiplied all bids placed in our dataset by the bidding fee of AU$0.70 and added
on revenues owing to the final ending price paid by the winner.
It is worth noting there exists clear limitations to calculations of the profitability
of (most) pay-per-bid auctions. Given the international nature of a website auction
and localised portals for individual countries, revenue can be influenced by currency
fluctuations. Country specific portals charge localised prices for bidding fees but no
information is available on the origin of bidders. Additionally, items of equivalent
value may be substituted in localised auctions despite bids being pooled across all
portals. As such, revenue calculations may be exaggerated or understated.
Even more importantly, with our focus of the exit option, the use of this exit
option is not reported by the auction operators. The effect of the exit option on
profitability would not only depend on the frequency of use and when individuals
exercise the option but also on the secondary effects of enticing (or discouraging,
if bidders feel there will be increased competition) bidders to place additional bids.
Since an exercise of the exit option does not necessary reduce the revenues from the
bidder if the sunk costs are less than the difference between the wholesale value and
the RRP (exit price), it is not clear whether our calculations actually overstate or
understate total revenues and profits.
Figure 8 on the next page shows Swoopo.com and BidRivals.com have markedly,
different profit percentage profiles and this is likely owing the differing item compos-
analysis. For this aggregate analysis only, we also include “in house” items previously excluded toadjust for bids won rather than bought.
14Augenblick (2010) reported auction revenues of over 150% of the lowest market price of thegoods (from amazon.com). Nanney (2010) computed an average 70% profit margin in the Byerset al. (2010) dataset.
33
Figure 8: Percentage Profit Histograms - Platform Comparison
Swoopo.com (Source:Augenblick 2009) BidRivals.com
ition presented on the respective platforms. BidRivals.com features a larger range
of low price (filler items) which more easily end with no bids being placed, which
accounts for the spike seen in the graph. The notably noisier distribution may be a
symptom of the ending behaviour of the distinctly different classes of goods on offer
(high price electronics versus low price filler items).
Figure 9: Revenue and RRP Scatter Plot - BidRivals.com
Figure 9 provides a representation of the composition of items auctioned in Bid-
Rivals.com in terms of RRP and ending prices reached. A large proportion of items
auction are below US$500. It is also clear that the pay-per-bid auction format
provides the potential for extremely high earnings for auctioneers on individual
products. Case in point, the highest data point in Figure 9 comes from the auction
34
of a high end “iPad” which held a RRP of US$1,199. That particular auction ended
at an ending price of US$312.48, with 20,893 bids having been placed, leading to
revenues of US$15,936.48 for the auctioneer. The winner of that particular auction
placed a total of 89 bids, at a cost of US$66.75 which when combined with the end-
ing price paid results in a price saving of approximately 58%. This is of course an
extreme example, but it highlights the potential for profits in this auction for both
the auctioneer and the winner at the cost of the losers.
Figure 10: Auctioneer Gross Bid Revenue
In terms of Exit and No Exit auctions, revenues as a percentage of good value
is as expected given the preceding analysis of bids. Exit auctions for giftcards earn
auctioneers higher bid revenue compared to No Exit auctions.
5.5 Bidder Gains
A variation on the investigation of revenues earned by the auction operator is the
gains and losses incurred by the participating bidders. Figure 11 on the following
page depicts the profits and losses incurred by each unique bidder (ordered by net
gain ascending from left to right) within our dataset, over the course of collection.
We assume all bidders value the item at the RRP in calculating their net gain or loss.
Since we do not observe the use of the exit option, net losses are likely over-stated
in this analysis. As expected given the auction format, a large majority of bidders
(96.12%) incur losses and only a small minority makes gains, although such gains
can be significant for a proportion of such bidders. A significant majority of bidders
35
place very few bids and appear to never bid again.
Figure 11: Bidder Gains
Net gain of zero (Dashed Line)
Interestingly, bidders at the extremes of gains and losses in our sample have
similar characteristics in terms of bids placed and value of items won. Given the
limited time frame during which we captured data, it may be possible that these
“experienced” bidders fluctuate around incurring large losses and large gains. Over
the long term, this may be indicative of a bidding strategy and outcome that reflects
bidding as predicted by the theoretical model. As such the transient bidders who
place a low volume of bids and win few items, may actually form the bulk of the
operator’s revenue base. These characteristics are consistent across both giftcard
and non-giftcard auctions as well as Exit and No Exit auctions.
36
Table 3: Bidder GainsObs. Mean Std Dev. Min Max
Giftcard No Exit auctions 1521Cost of Bids 109.40 359.53 0.75 6748.50
Value of Items Won 99.05 286.51 0 3554.00Net Gain/Loss -10.35 169.14 3194.50 1028.50
Giftcard Exit auctions 721Cost of Bids 156.25 407.18 .75 4452.75
Value of Items Won 113.52 406.50 0 4510.00Net Gain/Loss -42.75 216.11 -2202.75 1785.50
Non-giftcard No Exit auctions 3874Cost of Bids 264.78 1009.51 0.75 18820.50
Value of Items Won 191.74 1120.42 0 28560.00Net Gain/Loss -73.04 558.12 -12797.25 13764.75
Non-giftcard Exit auctions 820Cost of Bids 123.25 323.92 0.75 4997.26
Value of Items Won 71.04 448.15 0 7252.00Net Gain/Loss -52.22 319.57 -2092.00 4497.75
******** ******** ******** ********
Summary statistics for bidder gains separated by giftcard and non-giftcard auc-
tions and Exit and No Exit auctions are reported in Table 3. While mean bidder
losses are larger for Exit auctions compared to No Exit auctions for giftcards, the
opposite holds true for non-giftcard auctions. However, few solid conclusions can be
drawn from this when bidder’s use of the exit option is not observed.
37
6 Experimental Investigation
Given the key limitations of the empirical study of pay-per-bid auctions, notably,
the lack of data on use of the exit option, it is difficult to empirically analyse the true
effects of this change in auction mechanism. As such, an experimental study serves
to complement the existing analysis through the examination of factors of interest
not typically observable in the empirical data, whilst allowing for the control of
factors outside the scope of our analysis.
As standard with experimental studies, our analysis of experimental data will
focus on treatment effects of the control and with exit option treatment as opposed
to metrics within treatments. Our conclusions should be robust to the extent that
inherent differences in conditions within a laboratory environment and the abstrac-
tions made in the implementation of the auction format are not correlated with our
variable of interest, the exit option.
In this section we begin by detailing our experimental design and implementation
and presenting summary statistics of interest. Then we undertake auction level
duration analysis which is designed to describe aggregate behaviour in the auctions,
bid level probit analysis which isolates individual drivers of bidding behaviour and
conclude with a short discussion of the experimental results.
6.1 Experimental Design
A basic treatment-and-control between-subject design is utilised to experimentally
test the effects of the exit option on outcomes of the pay-per-bid auction format.
In our experimental implementation of the pay-per-bid auction, participants were
randomly assigned into groups of 5 to bid for a coupon, which is redeemable for
$20.00. Following the theoretical model, bidding took place in discrete stages, with
participants being given an initial 10 seconds, which was reduced to 5 seconds after 5
rounds, to make a decision whether to place a bid or not. The time limit is reflective
of the decision time constraints faced by bidders in the online auction and also limits
round time and associated participant fatigue. At the end of each discrete round,
38
one out of all bidders who opted to place a bid is selected to be the leading bidder
and is charged a bidding fee of $0.50. Each bid increments the ending auction price
by $0.05.
Participants were given an initial endowment of $15 from which they could spend
on bidding fees. Earnings consisted of participant’s remaining endowment, their
gains from attaining coupons, and a $5 show-up fee. Potential losses for participants
were therefore restricted and participants were paid a minimum of $5 for their time
in the lab. Due to the $15 endowment, bidding is likewise limited to 30 bids per
participants.
In the exit option treatment, participants were allowed an additional option to
“purchase” the coupon at any time at an exit price less the amount of bidding fees
they have already incurred in the auction. Exit prices ranged from $16 to $26
and were fixed for the duration of an individual auction. Having both auctions
with exit prices above and below valuation allows us to better isolate the effects of
the exit option for a range of items. In the Exit AV auctions, once a bidder has
placed bidding fees up to the difference between their valuation, $20, and the exit
price, and the Valuation Exit threshold, individual bidding aggressiveness should
increase. Likewise, for Exit BV auctions, this should hold true immediately for all
participants.
Time permitting, participants took part in up to 8 auction matches. To alleviate
potential wealth effects from distorting behaviour across auctions, a random pay-
ment mechanism selected one auction out of all the auctions played to be paid out
at the end of the experiment. Although participants remained in the same groups
throughout all auctions, Bidder IDs were randomised across auctions.
Consistent with the theoretical model, participants are aware of the symmetric
nature of the auction and are also provided with full information regarded bidding
histories and exited players.
The experiment was programmed and conducted with the software z-Tree (Fisc-
hbacher 2007) with all auction interactions conducted on lab computers. Figure 12
39
depicts the auction decision screen participants made their decisions in. Recruitment
was undertaken via ORSEE (Greiner 2003), with 185 students from the University
of New South Wales participating. The average payment for each participant for
the 2 hour session was $21.73. Our sample consisted of 48% females, 53% business
faculty students and 82% international students.
Figure 12: Experimental Auction Decision Screen
A short comprehension test was conducted to assist participants in understanding
the auction and an exit questionnaire was used to elicit demographic variables and
measures of risk aversion.15 Experimental instructions, experimental script and
post-experimental questionnaires are available in Appendix C.
Rather than designing the experimental auction such that it more closely rep-
licates the real world implementation (e.g. with the presence of a timer countdown
and allowing for continuous bidding rounds), we instead implement a simplified ver-
sion of the auction which closely mirrors the theoretical framework discussed above.
The advantages of this method are two fold: firstly, this allows us to more closely
test the theoretical predictions of the model within a controlled environment and
in turn eliminate theoretical simplifications of the auction as a source of deviation;
and secondly, the discrete nature of the theoretical model lends itself to more eco-
nometrically palatable data collection and analysis.15We use the standard instrument per Holt and Laury (2002).
40
The use of a coupon with a common valuation simplifies the analysis and rep-
resents a realistic replication of a range of goods offered in the online platform with
well known retail values (gift cards, consumer electronics). The restriction of auction
groups to 5 also limits our analysis on the dimension of participation. In the online
auction, bidding behaviour is likely to be influence by the potential entrance of new
participants. Never-the-less, experiment participants can still profit by not parti-
cipating in the auction and simply taking home their endowment and this should
account for some variability in participation.16
Although limiting spending on bids to their initial endowment has the poten-
tial to distort bidding behaviour, ethical concerns and the practical implications of
requiring participants to pay out their losses took precedence.17
6.2 Summary Statistics
Table 4: Summary Statistics - Experimental DataObs. Mean Std Dev. Min Max
No Exit Auction 110Unique groups 19Ending round 33.52 35.75 0 146
Exit BV Auction 36Unique groups 18Ending round 45.56 35.01 0 123
No. exited 3.80 0.84 2 4Bids refunded 28.06 28.86 0 114
Exit AV Auction 45Unique groups 18Ending round 42.60 38.52 0 126
No. exited 1.51 1.19 0 4Bids refunded 18.82 27.34 0 99
****** ****** ****** ******
We present auction level summary statistics in Table 4. Both types of Exit16Ockenfels et al. (2010) found overall bidding increased when group sizes increase from 2 to 4
but less bids placed per participant.17Ockenfels et al. (2010) noted resistance in getting participants who experienced net losses to
pay the experimenter back.
41
auctions result in higher average ending rounds when compared to the No Exit
treatment. As expected, the number of exiting bidders and bids refunded per auction
are higher in the Exit BV auction than in the Exit AV auctions.
6.3 Duration Analysis
In our duration analysis, we examine aggregate auction level metrics to tease out
differences that occur across auction types, our treatment and control. Kaplan-Meier
survival estimates are presented for our experimental data in Figure 13, stratified
by treatment type. We report survival and smoothed hazard estimates for the exit
and no exit treatments, separating the exit treatments according to auctions with
exit prices above valuation and exit prices below valuation. Theoretically computed
survival estimates and hazard rates for the no exit option auction are also displayed.
Figure 13: Kaplan-Meier Survival and Hazard Rate Estimates - Experimental Data
Our aggregate survival estimates illustrate that at essentially every round, sur-
vival rates for our exit option auctions dominate that of the No Exit auction. This is
consistent with expectations that the presence of an exit option makes entry into the
auction relatively more attractive. Interestingly, survival rates for the No Exit auc-
tion in our experiment do not deviate as much from theoretical predictions as those
in the empirical analysis. Particularly, auction survival rates for the early rounds
are much lower than predicted, although this deviation reverses as additional bids
are placed. A Mantel-Cox logrank test for equality of survivor functions finds no
statistically significant differences between the theoretical and experimental survival
42
estimates (p-value of 0.26).18
Focusing on the more readily interpretable hazard rates, which have been smoothed
using a 20 bid running average, both Exit auction variations display a distinctly lower
hazard rate at the early stages of the auction comparative to the no exit control.
Auctions with exit prices above valuation maintain a consistently lower hazard rate
than No Exit auctions throughout the analysis time while auctions with exit prices
below valuation cross both other hazard curves around 50 bids placed.
The Exit AV and No Exit difference in hazard rates is consistent with our theor-
etical predictions about the effect of an exit option on bidding behaviour. The pres-
ence of an exit option increases bidder participation and bidding, resulting in lower
hazard rates, which is maintained once bidders enter the Valuation Exit Threshold
(not directly observable in total bids placed).
For auctions with exit prices below valuation, where bidding is a strictly dom-
inant strategy, termination probability in the beginning are lower comparative to
both No Exit and Exit AV auctions as expected. Relative to Exit BV auctions, Exit
AV auctions should initially display higher termination rates as bidding remains a
potentially costly action, but then exhibit similar termination rates as individual
bidders reach their exit-price-valuation gap. Interestingly however, the termination
probability exceed those of the No Exit and Exit AV auctions past 50 total bids
placed in the auction. This may be attributed to participants recognising that bid-
ding is likely to continue till all participants have exhausted their endowment, at
which point the payoff for winning is essentially equal to exiting immediately. Where
remaining potential gains outweigh cost of effort required to stay in the auction, par-
ticipants become more likely to exit.
The Mantel-Cox logrank test, stratified across No Exit, Exit BV and Exit AV
auctions, do not find statistically significant different survival rates (p-value of 0.19)
between the three auctions.19 Pairwise comparisons between No Exit, Exit BV18In the absence of censored data, as in our sample, the Mantel-Cox logrank test is equivalent
to a Mann-Whitney U test.19In this application the Mantel-Cox logrank test is equivalent to a Krushal-Wallis test.
43
and Exit AV also yielded no significant differences.20 Alternatively, semi-parametric
testing through a Cox proportional hazard regression, clustered at the group level,
is conducted to better isolate the effects of each auction type on hazard rates. With
No Exit auctions as the baseline, Exit AV auctions had a proportional hazard rate
of 0.77 with a p-value of 0.0950. This means that, at any round, Exit AV auctions
are 33% less likely to terminate than No Exit auctions.21
6.4 Probit Analysis
In our probit analysis, we utilise the experimental bid level data to investigate
how participant behaviour with regards to the decision to bid changes across our
treatment and control. A key advantage of experimental data over the empirically
collected data is the degree to which we can actually observe and capture data
regarding decisions made by participants. Since participants make a distinct decision
to place or to not place a bid each round, a probit analysis is feasible. We also
briefly examine the decision to exit and the probability of a bid being a winning
bid. Table 5 on the next page presents the bid level variables collected within our
experiment. Each data point represents a participant’s decision in one specific round
of an auction.
We estimated a variety of probit models to analyse how the probability of bidding
was influenced by auction related factors. We account for potential within-group
interaction effects by clustering standard errors at the group level. Observations
where bidders were leading or had exited were excluded from the analysis as the
bidding choice is not made in those circumstances.
The marginal effects of these models at sample means are presented in Table 6 on
page 46 (regression coefficients are available in Table 8 in Appendix B.1). The base
probit model (B1) incorporates basic auction and bidder state variables including20P-value of 0.1683 between No Exit and Exit BV, 0.1397 between No Exit and Exit AV and
0.8636 between Exit BV and Exit AV (paired test).21A Schoenfeld residuals test of the proportional hazards assumption, however, rejects a pro-
portional hazards relationship between No Exit and Exit BV auctions which precludes Exit BVauctions from the analysis.
44
Table 5: Bid Level Data DescriptionVariable Type DescriptionSubject Discrete Subject Unique IDGroup Discrete Group Unique IDBid Binary 1 if the subject placed a bid, 0 otherwiseWin Binary 1 if the bid placed was a winning bid, 0 otherwiseExit Binary 1 if the subject chose to exit, 0 otherwise
Sunk Bids Discrete Current number of own bids placed by the bidderRound Discrete Current round / total number of bids already placed in the auctionMatch Discrete The current auction match being played
Exit Option Binary 1 if the auction allowed an exit option, 0 otherwiseExit Price Gap Discrete Valuation less exit price
Bid Lag Binary 1 if the subject placed a bid in the previous stage, 0 otherwiseActive Bidders Discrete Number of bidders in the group who placed a bid last roundLeading Lag Binary 1 if the bidder was the leading bidder in the previous round, 0 otherwiseNo. Exited Discrete Number of subjects who have exitedThreshold Binary 1 if the subject has reach the Valuation Exit Threshold
the number of sunk bids submitted by the bidder, total auction bids placed, auction
type, the valuation relative exit price and the round of auction to control for learn-
ing/order effects across auctions played.22 Model (B2) has the addition of lagged
(previous stage) variables, to account for behaviour influenced by past information
and behaviour, and the Value Exit Threshold (the point from which bidding is a
weakly dominant strategy - see Section 4.2) and (B3) accounts for non-linearity in
the main continuous variables. Finally, model (B4), from which we draw our analysis
henceforth (unless stated otherwise), controls for any residual exit option treatment
effects.23 Each iteration of model is tested for goodness of fit and explanatory power
using the Wald test (p-values displayed in Table 6).24
22Note that with the inclusion of the Exit Price Gap, we necessarily force opposing effectsbetween Exit AV and BV auction. Estimations with each exit price as a separate dummy yieldsqualitatively similar results.
23The exit option dummy is interacted with sunk bids, auction bids placed, round, active bidders,leading lag and no. exited.
24The more typical LR Test is forgone since clustering of standard errors does not produce truelikelihoods. A more conservative Bonferroni adjusted test also produces results consistent with theWald test - Korn and Graubard (1990).
45
Table 6: Probit Marginal Effects of Decision to Bid(B1) (B2) (B3) (B4)
Sunk Bids 0.0572*** 0.0759*** 0.1736*** 0.1867***(0.0038) (0.0087) (0.0226) (0.0233)
Sunk Bids ^2 -0.0073*** -0.0089***(0.0020) 0.0022
Sunk Bids * Threshold -0.0235*** -0.0406** -0.0581**(0.0079) (0.0215) (0.0221)
Sunk Bids ^2 * Threshold 0.0040** 0.0055**(0.0020) (0.0022)
Round -0.0107*** -0.0101*** -0.0219*** -0.0209***(0.0007) (0.0007) (0.0013) (0.0014)
Round ^2 0.0001*** 0.0001***(0.0000) (0.0000)
Match -0.0113* -0.0091* -0.0066 -0.0061(0.0060) (0.0048) (0.0042) (0.0039)
Exit Option 0.0557** 0.0229 0.0263 0.0263(0.0260) (.0255) (0.0243) (0.0243)
Exit Price Gap -0.0066 -0.0118** -0.0095 -0.0136*(0.0049) (0.0056) (0.0066) (0.0071)
Bid Lag -0.0871*** -0.0757*** -0.0830***(0.0119) (0.0112) (0.0108)
Active Bidders 0.0524*** 0.0404*** 0.0399***(0.0066) (0.0055) (0.0062)
Leading Lag 0.1732*** 0.1086*** 0.1074***(0.0234) (0.0191) (0.0286)
No. Exited 0.0479** 0.0554** 0.0361(0.0205) (0.0274) (0.0308)
Demographics yes yes yes yesExit Interactions no no no yes
Pseudo R2 0.1518 0.1772 0.2303 0.2325Wald test p-value 0.0000 0.0000 0.0450
Note: * p-value < 0.10, ** p-value < 0.05, *** p-value < 0.01
Note: p-value for variables with quadratic terms pertains to joint significance
Although the exit option is statistically significant at the 5% level and positively
signed as expected (with the exit option increasing the probability of bidding) in
model (B1), the effect becomes insignificant as we account for additional auction lag
and polynomial terms. Where significant, the exit option only increases probability
of bidding by 5.72% (at the mean value of the sample of 67 sunk bids). This suggests
that bidding behaviour as a whole is not strongly influenced by the mere presence
of the exit option alone when additional drivers of behaviour (both treatment and
46
non-treatment related) are accounted for.25
The exit price gap variable captures the relative effects of the exit price, between
auctions with differing exit prices, on bidding behaviour on top of what is captured
by the exit option dummy. Auctions with a positive exit price gap (Exit AV auctions)
are correlated with a lower bidding probability relatively to negative exit price gap
(Exit BV auctions).26
By far the most important effect of the exit option on the decision to bid is
captured by the Valuation Exit Threshold variable. As discussed in Section 4, the
Valuation Exit Threshold marks the point at which bidders have sunk bids greater
than or equal to the difference between their valuation and the exit price. The
threshold variable captures whether the bidder has entered this state of bounded
losses where costless bidding is possible and the dominant strategy is to bid. We in-
teract this threshold variable with sunk bids to obtain the additional effect, reaching
this Valuation Exit Threshold, has on bidding behaviour as bids sunk increases. We
find significant additional effects of this threshold variable in addition to the already
significant sunk bids variable. Given the difficulty of analysing marginal effects at
means of continuously variables, we compute average predicted probabilities of the
probability to place a bid over our sample and report the results in Figure 14 on the
next page.27
For the No Exit auction case, as sunk bids increase from 0 to 10, the probability
of bidding rises from 20% to 80%. The effect is however diminishing with bidding
probability falling back to 20% at 20 bids placed and 0% past 25 bids placed. The
positive correlation between sunk bids and bidding probability lends some evidence
to the argument that bidding behaviour in pay-per-bid auctions is due to suscept-
ibility to the naive sunk cost fallacy. To the extent that auction, bidder state and25A joint test of significance for the exit option and all exit option interaction terms returns a
p-value of 0.0702 for model (B4).26The formulation of the exit price gap variable necessarily forces an opposing relationship
between Exit AV and Exit BV auctions. A check against estimation of the probit regressionwith categorical variables for each Exit auction price supported this relationship.
27The average predicted probabilities are calculated by using our estimation results to predictback into the sample and graphing the corresponding average predicted probabilities over sunkbids.
47
Figure 14: Average Predicted Probabilities of Decision to Bid
strategic variables have been properly controlled for within our probit regression,
participants become more likely to place bids when they have incurred irrecoverable
costs.28 This potential escalation of commitment appears to diminish as parti-
cipants approach the maximum bids that can be placed by participants (given $15
endowment and $0.50 bidding fees). This is possibly consistent with a desire by
participants to conserve their remaining endowment for strategic bidding and loss
mitigating purposes.
In comparison, the Exit with Threshold average predicted probabilities capture
the effect of sunk bids on bidding probabilities for participants who have reached the
Valuation Exit Threshold.29 The average predicted probability profile is markedly
different, with an increasing propensity to bid which exceeds a 95% chance of bidding
past 10 sunk bids. The key effect of reaching the threshold is that the decreasing
probability of bidding in higher values of sunk bids seen in the No Exit case, which
may be attributed to loss mitigation and strategic conservation, is nullified by the
presence of the exit option. This sustained high probability of bidding drives the
lower hazard rates observed in exit option auctions.28Even if participants were to anchor their bidding strategy to spending some fraction of their
endowment, such behaviour should not lead to an increasing bidding probability over sunk bids.29Although statistically not significant, the Exit auction case, where we only consider the static
effect on bidding probability of participating within an auction with an exit option, does predicthigher bidding probability over all values for sunk bids.
48
Bidding behaviour in both the Exit and No Exit auctions are affected positively
by an increase in sunk bids, but it is worth noting the differing effects at work.
Increase in bidding under the No Exit auction may be attributed to naive sunk cost
fallacy, but an increase in bidding under Exit auction may be reasonable considering
the zero marginal costs of doing so once the Valuation Exit Threshold is reached.
In terms of the current round, which captures the effects of the current auction
price and the total bids placed by all auction participants, the probability of bidding
falls at a decreasing rate as auction bids placed increases which reflects the increasing
cost of winning the item.
The control for auctions rounds, which investigates whether any order or learning
effects occur across successive auctions, results in both statistically and economically
insignificant effects once additional factors are controlled for.
The effect of the bid lag variable (8.30% less likely to bid), which captures the
probability of bidding if the participant had bid in the previous stage but was not
selected as the winner, suggests that there is no persistence in the bidding process if
no costs are incurred. The leading lag estimate, which measures the effect of bidding
after being outbid in the previous stage, returns a significant 10.74% increase in
bidding probability. With the bid lag controlled for, this suggests a competitive
effect of the auction, where being “dethroned” by another bidder, increases bidding
propensity.
We find no significant effects of bidding behaviour attributable to the level of
risk aversion as measured by the Holt and Laury (2002) risk instrument.30 From a
sample mean of 5.1, an additional unit of risk aversion results in participants being
0.44% less likely to place a bid.
No significant effects were found when controlling for demographics (gender,
student locality and field of studies).30The instrument was not financially incentivised and was carried out as part of the post ex-
perimental questionnaire. The assessment was given to and completed by subjects in half oursessions (n=115). A separate (B4) model with the subsample of subjects was estimated with therisk measure. No significant changes in coefficients and standard errors were observed.
49
6.5 Auctioneer Profits
In addition to understanding how drivers of individual bidding behaviour change
across Exit and No Exit auctions, it is worth stepping back to examine how such
behaviour translates to outcomes for the auctioneer/operator. The distribution of
auctioneer profits per auction arising from our experiments are displayed in Figure
15. Auctioneer profits are calculated by assuming the item is acquired at costs
equal to the valuation of bidders and represents the net profit after deductions for
additional items sold and bidding fees refunded where applicable.
Figure 15: Auctioneer Profits
Similar to Ockenfels et al. (2010), our auctioneer experienced slight negative
profits (mean of -$3.24) in the No Exit auctions.31 This is contrasted with mean
losses of -$11.66 and -$34.56 respectively for the Exit AV and Exit BV auctions.
Although the auctioneer profits cannot be interpreted directly as they are a product
of our experimental setup, we can analyse the difference in profits between the
auction types to get a sense of how profits may change from the implementation
of an exit option. It appears that the implementation of an exit option does not
attract a sufficiently greater amount of bids to compensate for loss in bidding revenue
(refund of bids) arising from the exercising of the exit option.
However, the above method of comparing profits ignores a key difference between31The lack of over bidding within our experiment is contrary to experiments involving standard
all-pay auctions, e.g. Gneezy and Smorodinsky (2004). However, this may simply be a consequenceof our chosen experimental parameters.
50
Exit and No Exit pay-per-bid auctions. In the case of No Exit auctions, profits are
wholly a function of bid revenue, ending price and cost of goods. For Exit auctions,
bid revenue, exit price, refund of bidding fees, cost of goods sold and number of
items sold all factor into the profit equation. Since more than one item may be sold
for every Exit auction, the relative impact of the cost of the goods is greater with
respects to Exit auctions. Figure 16 depicts mean profits as cost of goods sold is
varied away from the valuation.
Figure 16: Auctioneer Profits and Cost of Goods
As the profit margin of the good sold increases, Exit AV profits rise steadily
as gains from goods sold cover diminished bidding revenue. Although it would be
difficult to extrapolate such an analysis to the real world to pin point a particular
break-even point between offering of the exit option and a standard pay-per-bid auc-
tion, this does suggest that there is a reason why the exit option could be beneficial
to auctioneers.
It is also worth noting that the standard deviation of Exit AV auction profits
(10.31) are significantly lower (p-value of 0.0000) than that of the No Exit Auction
across cost of goods sold. A tighter variation in profits, for an auction platform that
is characterised by volatile earnings, represents a significant reduction of risks which
further increases the potential attractiveness of offering such an option.
51
6.6 Bidder Gains
Just as pertinent as auctioneer outcomes are bidder outcomes. Figure 17 illustrates
the distribution of bidder gains per auction in our experiment. We calculate bidder
gains as the value of the voucher, minus the bidding fees, minus auction or exit price
paid.
Figure 17: Bidder Gains
Average bidder gains from the Exit AV auctions (2.13) significantly outweigh
gains for those who participated in the No Exit auctions (-1.01).32 Although the
standard deviation of Exit AV auctions (6.08) are larger than that of No Exit auc-
tions, the added variation appears to be right skewed towards higher bidder gains.32Exit BV auctions unsurprisingly result in far higher bidder gains of 6.99.
52
7 Discussion
With respect to our empirical analysis of the exit option, we find giftcard auc-
tions last longer while non-giftcard auctions actually end earlier than their No Exit
counterparts. Given the pure common value nature of giftcard auctions and the un-
observed, potentially very heterogeneous private valuations in non-giftcard auctions,
it would be reasonable to put more weight on the results of the giftcard auctions. In
this sense, the exit option appears to increase bidding aggressiveness, particularly
in the early rounds of the auction.
The discrepancy between the two groups of auctions may however be attributed
to an exponential effect of sunk cost fallacy in response to higher valued items in
non-giftcard auctions. If bidders are more susceptible to the naive sunk cost fallacy
when the value of items are higher, then it is possible this effect, which is eliminated
in an exit option, that would cause No Exit auctions to last relatively longer. This
would suggest some level of interplay between sunk costs effects and exit option
effects of reaching the Valuation Exit Threshold.
Likewise, due to the higher value of items, bidders may be more reluctant or
have a smaller probability of reaching the Valuation Exit Threshold and this would
reduce the effect of the exit option. Another reason for the difference may be the
interplay of strategies between bidders with asymmetric valuations. If one bidder
reaches the threshold significantly sooner than the rest, their bidding aggression may
deter bids from other participants. This is in contrast to giftcard auctions where
multiple bidders are likely to reach the threshold with relatively fewer bids placed
and within a few rounds of each other.
However, the lack of data on the use of the exit option limits our empirical ana-
lysis of No Exit auction outcomes for auctioneers and bidders. Crucially, while the
Exit auction increases bidding, auctioneer profits and bidder gains hinges on the
frequency and point at which the exit option is used. This serves to motivate our
experimental investigation which not only allows for a cleaner test of the theoret-
ical formulation of the auction, but also resembles the giftcard auctions analysed
53
empirically.
Our experimental evidence suggest that the introduction of an exit option to the
pay-per-bid auction affects behaviour in the ways predicted within the theoretical
section. Overall bidding increases with the exit option and this leads to longer
lasting auctions and higher gross bid revenue for the auctioneer. Participants take
advantage of the exit option as expected, increasing bidding aggressiveness once they
reach the Valuation Exit Threshold leading to higher welfare gains for bidders. We
observe that the effect of entering the threshold on bidding dominates the sunk cost
effect of the No Exit auction. However, the implications for the auction operators
are less clear and are highly contingent on profit margins on the goods auctioned.
While the Exit auction results in reduced bidding revenue, additional quantities
of the goods sold per auction compared to the No Exit auction could potentially
make up for the difference. To the extent that profit margins on goods are sufficient
there appears to be potential welfare gains for both auction operators and auction
participants, with a more equitable split of efficiency gains from trade.
Such a result has important implications for the regulatory treatment of pay-per-
bid auctions. Since the exit option appears to eliminate the reliance on exploiting
bidders’ naive sunk cost fallacy, caps potential losses and improves bidder welfare,
there may be a lessened need to intervene in the online pay-per-bid auction market.
Duration analysis undertaken shows that qualitatively, our experimental data
matches our empirical observations with respects to giftcard auctions, which gives
strength to the external validity of our experimental conclusions. Particularly, ter-
mination rates for Exit auctions in early rounds are lower than No Exit auctions
in both experimental and empirical analyses. However, given budget limitations,
questions remain regarding the effect of the exit option on items of high value as in
non-giftcard auctions. Further research could therefore be undertaken to isolate how
the effects of Exit and No Exit auctions change as the value of the item is scaled.
It is worth noting that a key abstraction made in our experimental setup is
the fixing of bidder numbers within our auction. Although bidding is not enforced
54
within our experiment and bidders still earn a payoff if they choose not to bid, it
stands to reason that the potential pool of bidders are different between platforms
offering the exit option and those which do not offer it. When faced with a choice of
auction platforms, the exit option might attract a greater number of participants,
as basic evidence in our empirical analysis suggests. The experimental setup may
therefore be understating bid revenues for the Exit auctions.
A possible extension of the experimental design would be to allow participants
to choose between the two auction platforms or even multiple Exit and No Exit
auctions. This may lead to further differences between the Exit and No Exit auction
outcomes and may provide an alternative explanation to why certain operators have
decided to offer the Exit auction. From an institutional choice perspective, if one
auction platform offers the exit option, it may force all other operators to do the
same to retain their bidders.
55
8 Conclusion
In this paper, we investigate how bidding behaviour and auction outcomes change
in pay-per-bid auctions when an exit option is made available to bidders. The exit
option, which allows bidders to purchase the item at the recommended retail price
less sunk costs, is predicted to increase bidding propensity when the bidder has
reached the Valuation Exit Threshold. Empirically, we find evidence that supports
our conjectures in pure common value giftcard auctions. Giftcard auctions with the
exit option are found to last longer than their No Exit counterparts. Experimental
data which isolates individual bidder behaviour corrobates the effect of the exit
option. The increase in bidding in the Exit treatment displaces and dominates
the potential sunk cost effects in the No Exit auctions. This results in unambiguous
welfare gains for bidders. Auctioneer profits are however heavily dependent on profit
margins as the exit auction trades-off bid revenues for more items sold per auction.
Provided profit margins are sufficiently large, the Exit auction has the potential to
be pareto improving.
56
References
[1] N. Augenblick. (2009) Consumer and Producer Behavior in the Market for
Penny Auctions: A Theoretical Empirical Analysis. Unpublished
[2] M.R. Bayes, D. Kovenock and C.G. de Vries. (1996) The All-Pay Auction with
Complete Information. Economic Theory. 8(2): 291-305
[3] D.T. Bishop and C. Cannings (1978) A Generalized War of Attrition. Journal
of Theoretical Biology. 70(1): 85-124
[4] J. Bulow and P. Klemperer. (1999) The Generalized War of Attrition. The
American Economic Review. 89(1): 175-189
[5] J.W. Byers, M. Mitzenmacher, and G. Zervas. (2010) Information Asymmetries
in Pay-Per-Bid Auctions: How Swoopo Makes Bank. Unpublished
[6] U. Fischbacher. (2007) z-Tree: Zurich Toolbox for Ready-made Economic Ex-
periments. Experimental Economics. 10(2): 171-178
[7] U. Gneezy and R. Smorodinsky. (2004) All-pay auction - an experimental study.
61: 255-275
[8] B. Greiner. (2003) An Online Recruitment System for Economic Experiments.
Forschung und wissenschaftliches Rechnen. 63: 79-93.
[9] T. Hinnosaar. (2009) Penny Auctions. Unpublished
[10] C.A. Holt and S.K. Laury. (2002) Risk Aversion and Incentive Effect. American
Economic Review. 92(5): 1644-1655
[11] E.L. Korn and B.I. Graubard. (1990) Simultaneous testing of regression coeffi-
cients with complex survey data: use of Bonferroni t statistics. The American
Statistician. 44: 270-276
[12] V. Krishna and J. Morgan. (1997) An analysis of the war of attrition and the
all-pay auction. Journal of Economic Theory. 72(2): 343-362
57
[13] C. MacDonald. (2011) The Economics of Online Penny Auctions. Unpublished
[14] S. Mittal. (2010) Equilibrium Analysis of Generalized Penny Auctions. Unpub-
lished
[15] J. Nanney. (2010) “Entertainment Shopping” An Analysis of Profit and Strategy
in a New Auction Format. Unpublished
[16] A. Ockenfels, P. Tillmann, and A. Wambach. (2010) English All-Pay Auctions:
An Empirical Investigation. Unpublished
[17] B. Platt, J. Price, and H. Tappen. (2010) Pay-to-bid Auctions. NBER Working
Paper. Working Paper 15695
[18] M. Shubik. (1971) The Dollar auction game: A paradox in noncooperative
behavior and escalation. Journal of Conflict Resolution. 15(1): 109
58
A Empirical Regression Outputs
Table 7: OLS on Unique Bidders in an Auction: CoefficientsGiftcard Non-giftcard
(GU1) (GU2) (NU1) (NU2)RRP 0.0715*** 0.0687*** 0.0542*** 0.0566***
(0.0025) (0.0026) (0.0017) (0.0024)No BidAgent 0.0677
(0.1397)Exit 2.1996*** 2.2943*** -25.8453*** -26.4918
(0.1323) (0.1448) (3.0190) (20.9921)Exit and No BidAgent 0.1417*** -3.0949
(0.3207) (22.6313)Bonus -2.9784
(21.0969)Bonus and Exit -28.8138
(20.9502)Beginner -1.3380
(0.2304)Beginner, Bonus and Exit -1.1067 -27.5901
(2.0433) (21.3185)Constant 2.2508*** 2.3896*** 29.3757*** 29.9491
(0.0865) (0.0896) (3.1803) (21.0134)
Adjusted R2 0.2477 0.2557 0.2477 0.4281N 3664 3664 2107 2107
Note: * p-value < 0.10, ** p-value < 0.05, *** p-value < 0.01
59
B Experimental Regression Outputs
B.1 Probability of Bidding
Table 8: Probit Regression on Bidding Decision: Coefficients(B1) (B2) (B3) (B4)
Sunk Bids 0.1464*** 0.1948*** 0.4440*** 0.4777***(0.0097) (0.0223) (0.0581) (0.0600)
Sunk Bids ^2 -0.0186*** -0.0226***(0.0052) (0.0057)
Sunk Bids * Threshold -0.0604*** -0.1038* -0.1486***(0.0202) (0.0551) (0.0567)
Sunk Bids ^2 * Threshold 0.0102** 0.0141**(0.0050) (0.0056)
Round -0.0274*** -0.0259*** -0.0559*** -0.0534***(0.0018) (0.0017) (0.0031) (0.0036)
Round ^2 -0.0003*** 0.0003***(0.0000) (0.0000)
Match -0.0289* -0.0234* -0.0168 -0.0155(0.0153) (0.0124) (0.0108) (0.0100)
Exit Option 0.1428** 0.0587 0.0672 0.0393(0.0667) (.0655) (0.0622) (0.1223)
Exit Price Gap -0.0169 -0.0302** -0.0243 -0.0347*(0.0127) (0.0144) (0.0169) (0.0182)
Bid Lag -0.2236*** -0.1936*** -0.2123***(0.0308) (0.0313) (0.0276)
Active Bidders 0.1345*** 0.1034*** 0.1020***(0.0167) (0.0140) (0.0159)
Leading Lag 0.4624*** 0.2832*** 0.2801***(0.0639) (0.0505) (0.0759)
No. Exited 0.1229** 0.1418** 0.0924(0.0527) (0.0703) (0.0789)
Constant 0.2269*** -0.0542 -0.0488 -0.0277(0.0786) (0.0814) (0.0648) (0.0842)
Demographics yes yes yesExit Interactions no no no
Pseudo R2 0.1518 0.1772 0.2303 0.2325N 25279
* p-value < 0.10, ** p-value < 0.05, *** p-value < 0.01
60
B.2 Probability of Exiting
Table 9: Probit Regression on Exit Decision: Coefficients(E1) (E2) (E3)
Sunk Bids 0.0214*** 0.0206*** -0.0123**(0.0054) (0.0056) (0.0170)
Sunk Bids ^2 0.0012**(0.0005)
Round -0.0003 -0.0003 -0.0079*(0.0016) (0.0016) (0.0047)
Round ^2 0.0001**(0.0000)
Match -0.0123 -0.0172 -0.0194(0.0213) (0.0208) (0.0204)
Exit Price Gap -0.0455*** -0.0498*** -0.0534***(0.0065) (.0073) (0.0077)
Bid Lag 0.0624 0.0531(0.0660) (0.0663)
Active Bidders -0.2059*** -0.1773***(0.0324) (0.0302)
Leading Lag 0.0324 0.0762(0.0829) (0.0850)
No. Exited -0.0622 -0.0610(0.0384) (0.0407)
Constant -2.4515*** -2.0647*** -1.8942***(0.0726) (0.0985) (0.0929)
Demographics yes yes yes
Pseudo R2 0.0319 0.0537 0.0670N 25279
* p-value < 0.10, ** p-value < 0.05, *** p-value < 0.01
61
C Experimental Documentation
C.1 Experimental Script
If you are here for the experiment, could you please form a line and ready your
student id cards so that we can record your attendance and get you seated.
[Wait for participants to be seated]
You should find three pieces of paper in front of you, a participant information
sheet which you should read, an informed consent form which you should sign and
a payment receipt form which you will need at the end of the experiment
[Wait for participants to complete forms]
We will now collect the consent forms before we distribute the experiment in-
structions.
[Wait for collection of consent forms]
We will now distribute the instructions for the experiment. Please read and
understand the rules carefully as this will help you make the decision that is best
for you. Before the experiment begins, there will also be a short comprehension
test to assist your understanding of the rules of the experiment. If anyone has any
questions please just raise your hand and we will come over to assist you.
[Wait for participants to finish reading instructions]
We will now start the comprehension test, if you have any difficulties with an-
swering the question, please just raise your hand and we will come over to assist
you.
[Wait for participants to complete comprehension test]
The experiment will begin shortly. There will be a welcome screen, once all
participants have clicked the OK button, the auction will begin immediately. Please
note that there may be a waiting period between each auction as the program waits
for all groups to finish an auction before continuing. You can raise your hand at any
point in the experiment if you need help, we will come out to assist you.
[Wait for participants to complete experiment]
62
The experiment is now over, we now need to randomly select the auction that is
paid out. I have an eight-sided die here which I will need a volunteer to roll. The
number facing up will be the auction number that is paid out for real.
[Dice roll]
While we prepare the payments, we will distribute some questionnaires for you
to complete. If you have any thoughts about the experiment, you can write them
on the back of the questionnaire.
[Prepare payments]
Note that the seat number is the number on the wall of your station, not your
bidder id.
Please also fill out your receipt form with your details, you can leave the amount
paid blank for now, we will fill that in when we call your seat number up for payment.
63
C.2 Experimental Instructions
64
65
66
C.3 Comprehension Test
Questions were administered via the computer terminals.
Q1) When more than one bidder decides to place a bid in a stage, how is the leading
bidders in the next stage decided? Only one of the following is true:
• All bidders who placed a bid become leading bidders.
• The bidder who places their bid first during the countdown becomes the leading
bidder.
• The leading bidder is selected randomly from the bidders who decided to place
a bid during the countdown in the previous stage.
• The bidder who placed their bid last during the countdown in the previous
stage becomes the leading bidder.
Q2) When is the bidding fee of $0.50 deducted from your endowment? Only one of
the following is true.
• Immediately after you decide to place a bid, before the stage ends.
• After the stage ends whether you are the leading bidder or not.
• After the stage ends whether you decide to place a bid or not.
• After the stage ends, if you place a bid AND is randomly selected as the leading
bidder.
Q3) Who will win the auction and obtain the voucher at the current price if no bids
are placed in a stage? Only one of the following is true.
• All the bidders who decided to place a bid in the previous stage.
• The bidder in the last stage who decided to place a bid and was selected as
the leading bidder.
• Any bidder who was selected as a leading bidder at any stage of the auction.
67
• A randomly selected bidder out of all the bidders.
Q4) How do you calculate your balance at the end of an auction? Only one of the
following is true.
• $20 if you have a voucher. $0 if you have no voucher.
• $20 if you have a voucher, less the price you paid for your voucher. $0 if you
have no voucher.
• $20 if you have a voucher, less the price you paid for the voucher, plus your
remaining endowment.
• Just your remaining endowment.
Q5) How much will you be paid at the end of the entire experiment which is made
up of multiple auctions? Only one of the following is true.
• The sum of all your balances at the end of each auction.
• Your balance in one randomly selected auction.
• The average of all your balances in all the auctions.
• Only if you have won a voucher.
Q6) At the start of a new auction, only one of the following is true.
• You will start the new auction with your remaining endowment from the pre-
vious auction.
• You will start the new auction with a voucher if you won one previously and
will be able to accumulate vouchers.
• You will not be able to bid if you previously obtained a voucher.
• You will start the new auction with a a separate new endowment and none of
the actions from the previous auction will carry forward to this new auction.
68
Q7) What happens when you choose to buy the voucher now? Only one of the
following is true.
• The auction will end for everyone and only you will get the voucher at the buy
now price.
• You will obtain the voucher at the buy now price and you will no longer be
able to place a bid in any new auctions in the future.
• You become the winning bidder and pay the current price to obtain the
voucher.
• You will obtain the voucher at the buy now price less your bidding fees paid.
Q8) If you have $1.00 of remaining endowment and the buy now price after discount-
ing the bidding fees paid is $6.00, will you be able to buy the $20 voucher now?
Q9) You have an initial endowment of $15, the bidding fee is $0.50 and the voucher
is worth $20. You decided to place 5 bids, was successful in placing a bid (selected
as the leading bidder) 2 times and you won the voucher at $1.50. What would your
balance be at the end of the auction?
Q10) You have an initial endowment of $15, the bidding fee is $0.50, the voucher is
worth $20 and the initial buy now price is $22. You decide to place 10 bids and was
successful in placing a bid (selected as the leading bidder) 8 times but did not win
the auction. What would your balance be at the end of the auction if you chose to
buy now at the end of the auction?
69
C.4 Post Experimental Questionnaire
70
71