Académique Documents
Professionnel Documents
Culture Documents
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Management Science.
http://www.jstor.org
MANAGEMENT
SCIENCE
int!?
Vol.
51, No.
12, December
2005,
pp.
1753-1762
DOI i0.l287/mnsc.l050.0433
of
Better, Faster, Cheaper: An Experimental
Analysis
a Multiattribute
Reverse Auction Mechanism
with
Restricted
Information
Terry P. Harrison,
J. Thomas
Douglas
Chen-Ritzo,
Ching-Hua
M.
Kwasnica,
Anthony
Smeal
of Business,
College
The
(chenritzo@psu.edu,
Feedback
State University,
Park, Pennsylvania
University
Pennsylvania
tharrison@psu.edu,
kwasnica@psu.edu,
dthomas@psu.edu}
use
a
auctions
of reverse
for procurement
majority
single-attribute
on
straints
attributes
and
lead time. Alternatively,
such as quality
The
nonprice
a multiattribute
can
a
auction
both
and
where
bidders
levels
price
specify
be
realized
increased
mechanism
supply
the
given
theoretical utility
chain
procurement
utility
problem
function.
a
price-only
words:
1.
we
setting,
provides
was
with
In this
auction.
the
is known
consider
economics;
experimental
by Sunil Chopra,
the
authors
3 months
operations
while
could
of nonprice
and
we
research,
con
providing
to conduct
choose
attributes.
While
lead
restricted
present
an
such
improvement
auction
ascending
and dependent
quality
format
buyer
the performance
of this multiattribute
experimentally,
with
mechanism.
the price-only
auction,
Compared
auctions;
History: Accepted
paper
function
effective
Key
explore,
auction
of
complexity
to the buyer,
(price)
a
16802
time
for
feedback
find
two
attributes
regarding
auction
we
the
mechanism
that
the buyer's
as
our mechanism
by a
in addition
utility
compared
design
to
is
(supplier) profits.
supply
chain
management
for 2 revisions.
Introduction
are auctions
in which
Reverse
auctions
the auction
on
behalf of a buyer, solicits bids from a group
eer,
a promising
of potential
sellers. They provide
alter
native to buying goods or services at a fixed price or
and can be easily implemented
through negotiation
General Electric
Covisint,
FreeMarkets,
electronically.
(Kwasnica and Thomas 2002), and Mars
(Hohner et al.
are
to
of firms that have attempted
2003)
examples
use auctions
are
an
auctions
for this purpose. While
excellent
form of price discovery,
there are usually
in addition
to price that can affect the
other aspects
value of the outcome.
In such cases, if the auction
on price only, the
focuses
process
competition
buyer
is forced to ignore these other aspects or to handle
them outside of the auction by subsequent
contractual
arrangements.
number
of
researchers
have
considered
dimen
sions
in the procurement
other
than price
setting
from a purely theoretical perspective.
Che (1993) and
Branco
(1997) study the optimal
(buyer utility-maxi
auction when
there are only two attributes,
mizing)
In their models,
and quality.
bid
however,
price
can still be represented
der private
information
by
a one-dimensional
allows
type. This simplification
them to apply powerful mechanism
tools first
design
discussed
there
has
been
(1981). While
by Myerson
some recent promising
on multidimensional
work
mechanism
2000, Williams
1999),
design
(Armstrong
the extension
to
of one-dimensional
optimal auctions
is not straightforward.
dimensions
higher
Recently,
Parkes and Kalagnanam
(2002), Beil and Wein
(2003),
and Asker and Cantillon
(2004) have studied auctions
with
the potential
for many
attributes.
and
Parkes
for
look
efficient
(2002)
(total
Kalagnanam
surplus
auctions
when
there may
be many
maximizing)
1753
Chen-Ritzo
et al.: A Multiattribute
general
of
this
1
Cramton
the
three characteristics
not fully reveal
the
we compare,
at both
following
First, we do
function. Second,
bidder
and individual
level, observed
in
behavior
the theoretically
expected
and a three-attribute
auction. Third,
highlight
research.
buyer's utility
the aggregate
behavior with
both a price-only
Auction Mechanism
Science
1754 Management
with Restricted
Information
51(12), pp. 1753-1762, ?2005
Feedback
INFORMS
our comparison
of the single-attribute
English auction
to the three-attribute
auction
is motivated
English
by
a
chain
procurement
specific supply
setting.
A differentiating
feature of our auction design
is
the type of information we provide
the bidders about
the buyer's
function.
the
Instead of revealing
utility
entire utility function, we only provide
infor
buyer's
on the marginal
mation
utilities
related to submitted
can be used to
bids. This information
guide bidders
toward their optimal bid. We made
this design choice
it may
for two reasons. First, in many
applications
not be practical
for a buyer to reveal their entire util
either for strategic
reasons, or because
ity function,
are
to
unable
it.
the informa
Second,
express
they
to the bidders
to be simple
tion provided
is designed
and intuitive. Previous
studies
of com
experimental
that simple
have demonstrated
infor
plex auctions
can often be effective
in generating
mation
efficient
et al. (2005) found that properly
Kwasnica
outcomes;
auc
in a combinatorial
prices
designed
single-item
in
tion can improve
the
fact
that,
efficiency
despite
not
these
do
all
the
necessary
prices
theory,
provide
It is possible
that the limited
informa
information.
tion we provide may
auction
per
actually
improve
the buyer's utility
formance
relative to fully revealing
the situation
function by simplifying
for the bidder.3
One of the objectives
of this research is to investigate
a reasonably
auc
whether
successful multiattribute
with
tion can be designed
information
that is both
for the buyer and simple for the bid
readily available
ders to understand.
one of the first to use experi
Bichler
(2000) was
ments
to study multiattribute
auctions. He examined
auctions
human
two-attribute
subjects experi
using
in general, have
ments
and found that the auctions,
It is pos
modest
auctions.
gains over single-attribute
of
the auc
sible that, due to the added
complexity
two
the
observed
with
attributes
tion,
gains
only
are considered.
may disappear when more attributes
in
Therefore,
our
experiments,
we
an
consider
auc
three-attribute
experiments.
We
Reverse
of using
an
English
auction
2
albeit somewhat
and Roth
See Kagel
outdated,
(1995) for a good,
on
literature
auctions.
review of the large experimental
simple
auction
where
experiments,
English
the buyer's
scoring function was not at all revealed
in the other. In
in one treatment
and fully revealed
a multiat
we
contrast to both these studies,
examine
tribute
auction
buyer's
scoring
that provides
function.
partial
revelation
3
if we published
For example,
the buyer's
utility
to arrive
to perform
calculations
ders would
need
attribute
bid levels. Assuming
the bidders
recognize
to do
correct
this,
it is still
calculations.
questionable
whether
of the
bid
function,
at the optimal
that they need
they would
make
the
Chen-Ritzo
Management
with
Restricted
which
auction,
they did not compare
auction case.
the
single-attribute
against
to
the bidders
the
information
Moreover,
presented
in both cases is actually very limited with
respect to
We
elabo
toward
their
bid.
optimal
guiding bidders
on
in
rate
this point
?3.1.
as follows:
In ?2, we
is organized
This paper
forms for buyer
the functional
describe and motivate
In
functions.
and
bidder
?3, we present
profit
utility
we
In ?4,
the exper
describe
the auctions examined.
are presented
in ?5. We con
imental design. Results
in ?6.
future directions
clude our work and discuss
of
the multiattribute
results
these
2.
The
4
We
the
=
q, I) v(q,
sourcing
auction when
suppliers
l)-p
+
aQOqb& aL0(l
is linear
the two
I. In our
contract
l**) -p,
(1)
to
related
be interesting
questions
at potentially
different
multiple
suppliers
multi
and Kalagnanam
levels. Bichler
(2005) consider
of the
auctions. We focus on the design
for multiattribute
recognize
incentive
attribute
> 0 and
are quality-related
constants
where
aQ0
bQ0
constants.
and aL0 > 0 and bL0 are lead-time-related
The buyer always prefers higher quality. In our exper
value of
iments, we chose bQ0 such that the marginal
<
<
is
(0 bQ0 1).
decreasing
quality
i for an
The profit function
for bidder
(supplier)
contract
is linear in price, with
awarded
c{(q, I) rep
i's cost as a function of quality and
resenting bidder
have
In our experiments,
bidders
lead-time
levels.
function:
cost functions
that yield the following
profit
=
ir,-(P, <7,1) V
cAq, 0
*Q,VQi Ml
lhu)>
(2)
> 0 and
are quality-related
constants
where
aQi
bQi
constants.
and aLi > 0 and bLi are lead-time-related
in
cost increases with
Each bidder's
improvements
cost of
and lead time. The marginal
both quality
>
is increasing
1). If the bidder
higher
quality
(bQi
she will have a profit of
the auction,
does not win
zero.
Define
between
the
a contract
created
total welfare
by
i at quality and lead
and bidder
the buyer
Environment
a single indi
is procuring
is a single buyer who
visible object or contract from a set of three potential
attributes
suppliers.4 The contract has three measured
level q, and lead time Z.
of interest: price p, quality
as the amount
of
Lead time is typically
thought of
order and
time that the buyer must wait between
we use the subscripts
0
receipt of goods. Throughout,
and the sub
and i for the bidders
for the buyer
related to qual
scripts Q and L to denote parameters
There
U(p,
Feedback
Information
that
there may
to select
one
select at most
the buyer must
different
levels of attributes.
can offer
supplier
and
3.
The Auctions
the
5
In the event
of
tie bids,
earliest
always
wins.
Chen-Ritzo
et al.: A Multiattribute
if no new standing
ended
bids were placed
for k seconds. The purpose
of extending
the auction
in this way was to eliminate
the advantage
that might
a bid
to bidders by sniping
be provided
(i.e., placing
in the last few seconds of the auction).6 In the PO auc
two minutes
and
tion, the initial time period was
/c= 45 seconds.
In the MA
the initial time
auction,
and k = 90 seconds.
period was five minutes
auction
3.1.
The Multiattribute
Auction
were
In the MA
to iden
bidders
auction,
required
were
not
multiattribute
that
bids
tify
only profitable,
but that would
also improve
the buyer's utility with
respect to the current standing bid. A critical element
of
the MA
bidders.
auction
we
Because
where
the buyer
with
certainty, we
the buyer's
utility
is the feedback
to the
provided
in mind
future applications
function may not be known
utility
have
(2002)
for more
the extreme
on
the strategy
condition
where
of sniping.
all bidders
In that case,
attributes.
nonprice
no more
about potential
information
ing would
provide
ments
than a simple price-only
auction.
the rank
improve
with Restricted
Information
51(12), pp. 1753-1762, ?2005
Feedback
INFORMS
our experi
that issue necessitates
Wfhile the above feedback
is pro
vided
for every bid submitted
in the MA
auction,
the auction,
throughout
only bids that beat the cur
rent best bid are accepted.
When
bidders
can determine
the optimal
quality
and lead-time
combinations
(q*,l*), it is straightfor
to show that it is a dominant
ward
strategy for them
to bid at those levels. Therefore,
consistent with
the
standard English
it is also a dominant
strat
auction,
the price offered until they
egy to continue
lowering
make zero profits. If all bidders
follow their dominant
strategy, the auction will be efficient and the expected
to the second
utility of the buyer will be equivalent
offer of the bidders.
See Chen-Ritzo
highest welfare
et al. (2005) for a formal discussion
of these properties
of the MA auction.
If the buyer had information
about bidder
costs,
then he may
find it in his interest to misrepresent
his
true utility
in the MA
function
auction. Che
(1993)
in a simpler multiattribute
shows that this is optimal
bidder
setting where
types are still one-dimensional.
Of
will
result in lost
any misrepresentation
not
We
did
examine
efficiency.
buyer misrepresenta
tion because our focus is on examining
bidder behav
ior in these auctions.
the
focus of this
Specifically,
on
an
can
is
whether
auction
be
that
paper
designed
to
sufficient
decision
allow
the
bid
provides
support
ders to bid as theory predicts
in a complex, multi
attribute
The
in our
bidders
setting.
human-subject
or
knew
that
the
func
experiment
utility
scoring
tion used by the auctioneer was programmed
by the
issues
of
Therefore,
experimenter.
buyer misrepresen
tation
course,
were
not
a concern.
Chen-Ritzo
Management
with
Restricted
4. Design
of the Experiment
in our exper
treatments
four different
considered
bidders
iment: (1) The PO auction with
inexperienced
bidders
the
PO
auction
with
(PO), (2)
experienced
bid
with
MA
auction
the
(3)
(P02),
inexperienced
ders (MA), and (4) the MA auction with experienced
variables
of interest
bidders
(MA2). The response
realized
the
were, primarily,
buyer utility and seller
We
profits.
for a total of 24 experi
We
recruited
subjects
in
each session we exposed
where
mental
sessions,8
one of the four treatments
to
previously
subjects
of several indepen
Each session consisted
described.
or periods.
The same three bidders
dent auctions,
were matched with each other in all periods,
and each
in
PO auc
the
bidders
either
of
group
participated
for all periods. We decided
tion or the MA auction
in separate sessions
to study the auction
institutions
the
rather than using a within-subjects
design where
same subjects would
in both the PO
have participated
us to isolate
the
auctions.
This allowed
and MA
a
we
to
for
observe
effects
that
learning
expected
effects
from
order
auction
(i.e.,
type
potential
given
variation
due to the order in which
subjects would
in the different
auction
have participated
types in a
our
because
treatment).
Additionally,
within-subject
reason
was relatively homogeneous,
it
is
subject pool
that any effects relating to differences
able to assume
be small.
between
the subject groups would
In every auction (or period), subjects were provided
in the cur
with
about their bid history
information
rent auction,
their costs for delivering
the asset at all
the time
lead-time and quality combinations,
possible
current
bid
in
the
and
the
best
auction,
remaining
or not they were
the
their bid status (i.e., whether
session
group
refers
of subjects
to a continuous
is exposed
of
time where
period
to a single experimental
a fixed
treatment.
Feedback
across
sessions.
Whenever
a case
was
repeated
a session,
to bid
of subjects
the matching
within
no subject was
so
the
that
der types was
changed
same type in both replicates of each case. This elim
be associated with
inated learning effects that might
more
same bidder
than once. The
the
type
playing
were
constant
type matchings
kept
subject-bidder
across all sessions. The exact parameters
used in each
screen shots, and other
case, experiment
instructions,
are available
in the online
information
supplement
(http://mansci.pubs.informs.org/ecompanion.html).
at the
conducted
All experimental
sessions were
for
Economic
of Business's
Smeal College
Laboratory
and Auctions
(LEMA), The Pennsylva
Management
2003 and April
nia State University,
between March
or graduate
student who
2003. Any undergraduate
in this experiment
had not previously
participated
was
that
to be eligible
for treatments
considered
students
those
bidders. Only
required
inexperienced
in the PO auc
had previously
who
participated
in the
to
tion (MA auction) were
participate
eligible
treatments
that
PO auction
(MA auction)
required
bidders. The auctions were
implemented
experienced
(Fischbacher
1999).
using Z-Tree software
In the PO auction
lasted
each period
sessions,
auc
in
MA
5-10
and
the
minutes,
approximately
10-15
tion sessions, each period
lasted approximately
minutes.
took around 1-1.5
The PO auction sessions
to complete,
sessions
while
the MA
auction
hours
took around 1.5-2 hours to complete, on average. The
rate used in all treatments was 0.1 U.S. dol
exchange
fee
dollar. A $7 show-up
lars for every experimental
was
to the total earnings
of each subject at
added
incentive
cash
besides
the end of the session. No
was provided
to subjects. Subjects earned, on aver
age, $22 and $16 in the MA and PO auction sessions,
respectively.
5.
8
A
Information
Results
et al.: A Multiattribute
Chen-Ritzo
Reverse
1758
Auction
Mechanism
Science
Management
Table 1
with Restricted
Information
51(12), pp. 1753-1762, ?2005
Feedback
INFORMS
Case 2
CaseCase
3 4
ExpectedMA (PO)outcome
Quality(reserve)
Leadtime (reserve)
Buyer
utility
Seller
profit
Improvementinbuyerutility
MA (PO)observed
Averageprice
% Expectedquality
% Expected leadtime
Averagedistance
% Expectedwinner
Average buyerutility
Average
seller profit
Numberof observations
MA2 (P02) observed
Average price
% Expectedquality
% Expected leadtime
Averagedistance
% Expectedwinner
Average buyerutility
Average seller profit
Numberof observations
Price
130(550)
4(4)
10(5)
Winner
3(2)
285 (250)
62(10)
14%
646.6 (559.5)
44% (N/A)
25% (N/A)
4.33 (N/A)
25% (80%)
240.5 (241.5)
44.1
(-128.4)
between
94.5(299.8)
33% (N/A)
100% (N/A)
0.67 (N/A)
100% (100%)
94.2 (83.2)
46.8(4.0)
6(6)
6(6) 6(6)
buyer
utility.
In Table 1 we
report the expected
improvement
as
in buyer utility
in the MA
to
auction
compared
9
The
number
a technical
data
of observations
difficultly
related
in the MA
treatment
to
due
vary
to the recording
to a
of the results
file.
10
non
test is a powerful
The Wilcoxon-Mann-Whitney
rank-sum
to the standard
f-test when
substitute
data has at least
parametric
ordinal
data
measurement
from
generated
the data do not meet
(Siegel
human
and Castellan
subjects,
the assumptions
2.6
(-64.6)
126.7(630.2)
100% (N/A)
100% (N/A)
0.00 (N/A)
100% (83%)
288.3 (170.8)
58.1(87.6)
6(6)
treatments.10
auction
-6.1
1988). When
examining
to assume
it is typical
that
for a f-test.
required
2,399(2,364)
1,986 (1,870)
577.1 (381.9)
2,035.5 (2,350.6)
38% (N/A)
50%
(N/A)
44% (N/A)
36% (N/A)
2.52 (N/A)2.45
(N/A)
75% (85%)
50% (45%)
281.3 (249.1)
1,887.1 (1,884.4)
15 (20) 16
We begin by examining
the utility
that the buyer
at the end of the auction.
achieves
In theory, because
bidders
in their bidding
choice
(sellers) have more
under the MA auction, they should provide
the buyer
with higher utility. However,
the MA auction may keep
izing these potential
gains.
1. The MA
Conclusion
190.7 (297.2)
38% (N/A)
56% (N/A)
2.45 (N/A)
63% (70%)
98.5 (85.9)
16 (20)
In Table 1, we present
without
previous
experience.
outcome
the expected
the
bid, win
(i.e.,
winning
and
seller
bidder,
ning
buyer utility,
profit) and the
outcome
observed
for
both
auction
institu
average
studied. The
tions, in each of the cases and treatments
number
of observations
collected
for each treatment
in the different
institutions
is also included
in this
the results of the rank-sum
table.9 Table 2 presents
tests that we use for statistically
certain
comparing
outcomes
52(299)
561(294)
3(2)3(4) 10(9)
10(6)
3(7) 5(3)
3(2)
2(3)
1(1)
98 (84)
333 (237)
46(4)
48(6)
88(49)
41%
6% 17%
(-25.2)
115.4
(-580.5)
14
(20)
(20)
595.3(393.8)
2,090.0(2,355.2)
67% (N/A)
67% (N/A)
67% (N/A)
33% (N/A)
0.47 (N/A)1.69
(N/A)
100% (100%)
67% (50%)
300.7 (237.2)
1,863.7 (1,879.8)
62.0(5.8)
149.0(39.6)
Wilcoxon-Mann-WhitneyRank-SumTest Results
Test(fy,)
Buyer utility
MA>P0
Buyer utility
MA2>P02
MA^PO
Seller profit
MA2^P02
distance
M/4>M/12
Note. The
Casel
Case 2
w = 325.0
w = 343.0
p = 0.1802
w = 57.0
p = 0.0023
n/ = 262.0
Seller profit
Euclidean
w = 53.5
p = 0.0079
w = 246.5
Case 4
tv= 368.0
w = 272.0
p = 0.0112
iv= 51.0
p = 0.0311
w = 355.0
p = 0.1764
w = 39.0
p = 0.5000
w = 385.0
p = 0.2858
p = 0.4414
p = 0.0616
p = 0.0000
iv= 33.0
iv= 51.0
u/ = 52.0
w = 28.5
p = 0.3751
p = 0.0557
p = 0.0435
p = 0.1025
iv= 210.0
= 0.0264
p
w = 220.0
p = 0.0038
left-hand column
native hypothesis
p is the p-value.
p = 0.0682
Case 3
w = 202.0
= 0.0883
p
iv= 157.0
= 0.2139
Chen-Rltzo
Management
et al.: A Multiattribute
Reverse Auction Mechanism
INFORMS 1759
Science 51(12), pp. 1753-1762, ?2005
the experienced
has only a small
and PO auctions
MA
it
auction,
able to conclude
the MA auction.
with
treatment.
Case
is greater under
that buyer utility
the
The PO auction does not provide
case.
in
any
greater utility
we have seen that the MA auction
can
buyer with
Although
to
improve buyer utility, it is worthwhile
significantly
examine how well
the auctions performed
relative to
the expected outcomes. Again,
referring to Table 1,we
on
that
MA
find
auction achieved
91%
the
average,
and 95% of the expected buyer utility for the inexpe
rienced and experienced
The
treatments,
respectively.
PO auction
tends to yield slightly higher utility than
On average,
the PO auction yields
101%
predicted.
of the buyer utility predicted.
This difference
disap
treatment, and the P02 auc
pears in the experienced
tion only yields
92% of the predicted
buyer utility.
The auctions' performances
relative to the theoretical
benchmarks
shed light on the fact that we did
may
not always observe a significant
increase in buyer util
ity in the MA auction relative to the PO auction. At
least for the inexperienced
the MA auction
treatments,
full expected
fails to achieve
buyer utility, whereas
to overshoot
the PO
auction
tends
predictions
slightly.
The winning
bidder, quality, and lead time are fre
for the MA auction format. As an
quently as predicted
additional measure
of the closeness
of the observed
bid to the expected
multiattribute
bid, we
report
distance
(Table 1) the Euclidean
levels from
ity and lead-time
MA auction. With the exception
of the observed
qual
in the
those expected
of Case 1 for the MA
this distance
is usually
small.
auction also offers potential
gains for the
In all four cases, the theoretical
bidders.
seller profits
are higher under the MA auction than the PO auction.
seem to contradict
this may
the previous
Although
of the MA auction often
result, the added flexibility
benefits
both parties. We examine
seller profitabil
in Conclu
auction institutions
ity under the different
sions 2 and 3.
2. The MA auction does not degrade,
Conclusion
and occasionally
increases, seller profits.
treatment,
The MA
Restricted
Information
Feedback
1 shows, average
seller profits are often
This is because,
sellers placed
negative.
occasionally,
not individually
In 11 out
bids that were
rational.
of 80 (14%) PO observations
and 5 out of 62 (8%)
MA observations,
the winning
bidder ended up los
it is impossible
to determine
While
the
ing money.
source of this behavior,
were
not in any way
bidders
forced to bid below
their cost (at a loss) and were
As
Table
to help avoid
checks
given
the most
likely determinant
this behavior. We
of these losses was
had to understand
feel
the
in
the bidders
complex
setting
a time-constrained,
which
environment,
competitive
was our primary motivation
for studying
these auc
tions experimentally.
Conclusion
3. The frequency
of seller losses are
not
Despite
der losses
in the observed
difference
of losses under
frequency
the two institutions
is actually
somewhat
surprising.
A priori we expected
that bidder errors would
be a
in the MA auction.
In principal,
the
greater problem
PO auction was no different
from a standard reverse
where previous
experiments
rarely observed
it appears
losses. However,
that the added
of the multiattribute
cost tables, despite
information
in the bidding,
its irrelevancy
have affected
might
bidder
behavior.
with
Consistent
other experimen
tal studies of auctions, we have
identified behavior
that differs dramatically
from the theory. The level of
auction,
bidder
losses actually
tends to be greater in the PO auction
as well;
losses averaged
dollars
1,641 experimental
in the PO auction versus 217 experimental
in
dollars
a rank-sum
the MA auction. Using
test, we find that
losses under the PO auction are stochastically
larger
= 57.0 and
(w
p < 0.0200). These high losses also affect
the buyer utility results presented
earlier because bid
as increased
der losses tend to manifest
themselves
for the buyer. When
with
observations
losses
utility
are removed, we find that the PO auction
buyer utility
as a percentage
of expected drops from 101% to 93%.
The decline
in the MA auction
is more modest
from
91% to 88%. All buyer rank-sum utility comparisons
earlier in support of Conclusion
1 still hold
presented
if observations
with
losses are dropped.
never made
bidders
losses in either
Experienced
treatments. Therefore,
the P02 or MA2
it appears that
losses might be more related to learning the interface
and environment
rather than some systematic
prob
our
lem with
the auction
setup. This also suggests
next
conclusion.
et al.: A Multiattribute
Chen-Ritzo
Reverse
Auction
Mechanism
Science
1760 Management
4. Performance
auction
of the MA
Conclusion
with
institution
improves
experience.
While
buyer utility and seller profits rarely change
to the experi
from the inexperienced
significantly
treatments
Table
the
auctions
enced
(see
2),
appear
to perform more consistently
with expectations
when
the
bidders. When
there are experienced
comparing
MA
and MA2
the predicted
to 92%. The
the proportion
treatments,
increases
winner
is observed
of times
from 53%
of correctly predicted
proportion
qual
from 42% to 67% and
ity and lead times increases
in
These differences
from 40% to 75%, respectively.
a
are
at
the
10%
chi
all significant
level;
proportions
test
these proportions
yielded
squared test comparing
and
6.993
of
9.485
3.303
statistics
(winner),
(quality),
distance
the Euclidean
(lead time). If we measure
combination
and
lead-time
of the winning
quality
from the expected
outcome, we find that it decreases
In three of four cases,
from 2.96 to 0.71 on average.
to
the distance drops significantly
from inexperienced
a
as
treatments
As
Table
result,
(see
2).
experienced
mentioned
earlier, the observed buyer utility as a per
from 91% to 95%
increases
centage of the predicted
with experience.
less
of the PO auction
The performance
improves
are
no
and
observed
losses
longer
dramatically. While
and
bidders
the reservation
quality
always provide
times
that
the
lead time,11 the proportion
of
predicted
bidder wins
from 70% to 83%, which
only increases
test statis
is not statistically
significant
(chi-squared
tic of 1.059). In addition,
buyer utility actually goes
as a
relative to predicted
down
utility; buyer utility
to
93%.
of predicted
percentage
drops from 101%
the aggre
results have highlighted
The previous
of the MA and PO auctions. We can
gate performance
and how it
individual
behavior
also examine
bidding
to the theoretical
Bidders
should
predictions.
at
and
lead
the
levels
of
quality
optimal
place
auction. Only
9% and 11%
time (??*,/*) in the MA
in the MA
and MA2
of all bids placed
treatments,
are at the predicted
levels. This is not
respectively,
once one considers
the fact
surprising
particularly
relates
bids
quality
the
treatment,
inexperienced
lead time in excess
and/or
three
of
bidders
the reservation
either
provided
amounts.
Feedback
INFORMS
is often different
than the the
treatment,
experienced
are
it
be
that
bidders
ory predicts,
may
only mak
errors
in
small
their
Given
the struc
ing
bidding.
ture of the cost and utility
functions
there
utilized,
movements
to
small
from
be
the
effi
may
gains
only
cient quality and lead-time
levels. By measuring
the
Euclidean
distance
of each bidder's
best bid from
the predicted
levels, we have a
quality and lead-time
measure
of distance
from the predicted
strategy. In the
MA and MA2
the distance
2.40
treatments,
averaged
is
and 2.01, respectively.
Even more
that
only
telling
one
3 out of 33 total bidders
bids within
averaged
unit distance
of the optimal
strategy.
are not always bidding
Even though bidders
the
correct levels of quality and lead time, they may
still
in utility
be placing
bids close,
terms, to the opti
mal bid. In theory, losing bidders
should offer a bid
at the efficient quality
all the surplus,
that provides
and lead time, to the buyer. By examining
the highest
we can examine
the
utility offers of the losing bidders,
accuracy of this benchmark.12 We can examine when
leave "money on the table," or they do
losing bidders
not bid despite
the fact that there was a profitable bid
bid. In the
beat the final winning
for them that would
51%
of
bidder
obser
treatment,
inexperienced
losing
are
two
vations
auction
observations
for
each
(there
a profitable
failed to bid when
bid was avail
period)
able. The proportion
falls to a more
but
reasonable,
treatment. A chi
still large, 35% in the experienced
in these proportions
yields
squared test for difference
a test statistic of 2.70, which
at the 10%
is significant
the previous
level. This suggests,
result,
along with
are often failing to realize the optimal
that bidders
to the relatively
strategy. This is in contrast
bidding
on an aggre
MA
auction
of
the
strong performance
must
it
these
be that
level.
To
reconcile
results,
gate
is close to
the actual utility offered by these bidders
the theoretical prediction. We can measure
this by cal
the utility of each losing bidder's
best bid
culating
the buyer with
the greatest
(i.e., the bid that provided
util
of the maximal
possible
utility) as a percentage
that
bidder
could
offer
of
(i.e., provision
greater
ity
result in bidder
losses). Figure 1
buyer utility would
the fraction of the maximum
possible
buyer
was offered by each losing bidder's
best
that
utility
treat
In the MA and MA2
bid in the MA auctions.13
offer 57% and
the losing bidders
ments,
respectively,
shows
12
The bids
bidders
of the winning
when
others
stop bidding
should
bid
13
to their zero
up
profit
cannot
be examined
stop bidding.
Thus,
because
they will
they
not
point.
to note
that this measure
is biased
downward.
It is important
is following
if a losing bidder
the optimal
she may
Even
strategy,
to be offering
This
is
less than 100% under
this measure.
appear
because
offer
11
In
with Restricted
Information
51(12), pp. 1753-1762, ?2005
essary.
pliance
over
bidder may
the maximal
the winning
utility
jump
of such a bid unnec
the losing bidder, making
placement
to use this measure
it gives noncom
because
We decided
of
with
the theory
the benefit
of
the doubt.
Chen-Ritzo
Management
et al.: A Multiattribute
Reverse Auction Mechanism
Science 51(12), pp. 1753-1762, ?2005
INFORMS 1761
Figure 1
with
Restricted
Information
Table 3
Rank-Sum
Wilcoxon-Mann-Whitney
Test
p-value
Mann-Whitney
0.8
Test
?0.6
Xxxx
x\
CO
X MA
0.2
(HA)
xX
***?**.
x
xxx
Results
1.0
1,075.5
PO^MA
684.0
PO^MA2
214.0
P02^MA
104.0
P02^MA2
0.0000
0.0049
0.0144
0.1120
MA2
Note. The
Feedback
X
12
XX
)X?3?
3
i-Xa
i-r^X-rX-Xr
4
5
left-most column
of the statistical
represents
the alternative
hypothesis
test.
Period
individual
is not exactly as
bidding behavior
across
in
it
consistent
is
predicted
theory,
relatively
the two auction mechanisms.
Individual behavior
also
to
the
closer
behavior
with
gets
predicted
experience
in the MA
that the added
auction.14 This suggests
of the auction environment
does not lead
complexity
to systematically
different
behavior.
Variations
from
the theory are expected
in both cases, and highlights
the need to study auction design
from experimental
a
in
to
addition
theoretical
devel
perspective
purely
cases
seem
in
The
fact
that
to
both
bidders
opment.
While
in auctions,
sur
it should not be entirely
for
the
theoretical
prising
specific
predictions
a
MA
auction
do not do particularly
well.
Rather,
more
to
would
be
the
indi
appropriate
comparison
vidual
In the
behavior
in the PO auction.
bidding
behavior
that
earlier
for the MA
auction, we
to the MA
treatments,
compared
the
PO
is significantly
smaller than
only
proportion
the MA proportion
test statistic of 7.68).
(chi-squared
are
In both cases, these proportions
indistinguishable
from the MA2 proportion
of bidders
failing to make
we
If
bids.
calculate
the
profitable
utility of the los
as
a
bidders'
best
offer
of the maximal
ing
percentage
for that bidder, we find that, on
offered 92% and 90% of the maxi
for the PO and P02 auctions,
respec
offer available
utility
average, bidders
mal utility offer
tively. Ifwe compare
the percentages
of the maximum
in the PO auc
offered
bids
by losing
utility
tions with
those in the MA
find that,
auctions, we
at the 5% level, only the P02 and MA2
are
auctions
not significantly
different. Therefore, with experience,
available
the percentage
of the maximum
available buyer
offered
bids
is
consistent
between
ity
by losing
two auction
The
results
of
the
rank-sum
types.
are provided
for these comparisons
in Table 3.
Conclusion
ual
bidding
auctions.
6. At
behavior
least with
is similar
experience,
in the MA
util
the
tests
behave
reconcile
the superior
aggre
similarly helps
auction
of
the
MA
to the
gate performance
(compared
PO auction) despite obvious
deviations
from the the
oretical benchmark.
6.
Conclusions
The
results
here
presented
of MA
auctions
are promising
for the
for procurement
when
matter.
bid
Providing
application
several nonprice
components
ders with
information
the
only marginal
regarding
buyer's utility, we are able to construct a simple auc
tion mechanism
that clearly betters,
experimentally,
the PO auction
format where
and lead time
quality
are set in advance. The MA auction benefits both the
the expected dif
buyer and the sellers. It is only when
ferences are small that we do not observe
significant
from the experimental
MA auction.
improvements
The experimental
real
results, by showing where
from
differs
ity
theory, provide
potentially
important
Bidder
insights into the MA auction design problem.
was not perfect.
behavior
It took experienced
sub
an
on
to
MA
that
auction
jects
get
closely matched,
an aggregate
the
dominant
level,
strategy predictions.
Even then, individual
behavior was not par
bidding
The level
ticularly close to the theoretical predictions.
of bidder losses was also not predicted. Despite
these
the
MA
auction
still
differences,
outper
consistently
forms the PO auction
in the lab. However,
future
more
or
for
realistic
designs
complex
settings might
individ
and PO
14
to note
It is interesting
in the PO auction.
true
trend does
not
appear
to be
et al.: A Multiattribute
Chen-Ritzo
in an attempt
consider
these deviations
better MA auction.
to design
Reverse
Auction Mechanism
Science
1762 Management
P. J., C. R. Plott.
(BICAP) mechanism
to use
railroad
right
Brewer,
with Restricted
Information
51(12), pp. 1753-1762, ?2005
1996. A
for
the
Feedback
INFORMS
conflict
binary
ascending
decentralized
allocation
tracks.
Internat.
J. Indust.
price
of the
Organ.
14(6)
857-886.
Y.-K.
Che,
1993. Design
competition
J. Econom. 24(4)
tions.
RAND
auc
multidimensional
through
668-680.
A. M. Kwasnica,
T. P. Harrison,
D. J. Thomas.
C.-H.,
A multi-attribute
chain
Better,
faster, cheaper:
supply
auction mechanism.
Technical
of Busi
report, Smeal College
ness Center
for Supply Chain Research,
The Pennsylvania
State
Park, PA.
University,
University
Chen-Ritzo,
2005.
V., V. L. Smith,
Coppinger,
in English,
Dutch
and
J. Titus.
sealed-bid
1980.
Incentives
auctions.
and
Econom.
behavior
Inquiry
43
1-22.
P. 1998. Ascending
auctions.
Eur. Econom. Rev. 42 745-756.
U. 1999. Z-Tree: A toolbox
for readymade
economic
no. 21,
of Zurich,
paper
experiments.
Working
University
Switzerland.
Zurich,
Cramton,
Fischbacher,
concerns
H. S. Lee,
Hohner,
G., J. Rich, E. Ng, A. Davenport,
J. Kalagnanam,
C. An. 2003. Combinatorial
and quantity-discount
procurement
auctions
and its suppliers.
benefit Mars,
Incorporated
Interfaces
23-35.
33(1)
allocation
rules in auctions
with
values:
affiliated
private
55 1275-1304.
laboratory
study. Econometrica
O. R, E. van Heck.
2003. Information
architecture
and
Koppius,
in multidimensional
electronic
market
auctions.
performance
and
Acknowledgments
The authors gratefully acknowledge
support from the Cen
ter for Supply Research (CSCR), the Institute for the Study
of Business Markets
(ISBM), and the Laboratory for Eco
nomics Management
and Auctions
(LEMA) at the Smeal
State University
College of Business at The Pennsylvania
an
thank
also
They
referees.
anonymous
thoughtful
anonymous
This
article
associate
has
editor
and
from
benefitted
two
input.
M.
67(3)
2000.
Optimal
multi-object
auctions.
Rev.
2002. Case
trans
electric
study: General
D. Thomas,
eds. eBusiness
and
Grenoble,
Processes.
The Supply
Chain Council,
Pittsburgh,
M.
Rothkopf,
auctions
auction
Manage
experimental
Support
Systems
Siegel,
auc
of multi-attribute
analysis
29 249-268.
2005. Configurable
Bichler, M.,
J. Kalagnanam.
ner determination
in multi-attribute
auctions.
and
Eur. ]. Oper.
win
Res.
160(2) 380-394.
F. 1997. The design
Branco,
Econom. 28(1) 63-81.
of multidimensional
auctions.
RAND
auction
Math.
design.
Oper.
J.
are
E. P. Kahn.
1990. Why
H., T. J. Teisberg,
Econom. 98 94-109.
rare? /. Political
1988. Nonparametric
S., N. J. Castellan.
ioral Sciences. McGraw-Hill,
New
York.
Strecker,
offers
1981. Optimal
Res.
2004.
2003. An inverse-optimization-based
Beil, D. R., L. R. Wein.
a multiattribute
to support
mechanism
RFQ process.
Bichler,
tions. Decision
The Nether
Rotterdam,
University,
A., D. Thomas.
systems. W
portation
R. B.
Myerson,
58-73.
Econom.
455-481.
J., E. Cantillon.
lished manuscript,
M.
Kwasnica,
Erasmus
A. M.,
D. Porter,
C. DeMartini.
2005.
Kwasnica,
J. O. Ledyard,
A new and improved
iterative
for multi-object
auctions.
design
Sei. 51(3) 419-434.
Management
of scoring
auctions.
Properties
Unpub
Harvard
MA.
University,
Cambridge,
uncer
D. P. Porter.
1989. Allocating
Banks,
J. S., J. O. Ledyard,
resources:
An experimental
tain and unresponsive
approach.
RAND
J. Econom. 20(1) 1-25.
Asker,
paper,
Working
lands.
Supply Chain
PA, 40-44.
their
References
Armstrong,
Stud.
S. 2003.
Preference
revelation
Statistic
for
in multi-attribute
Vickrey
the Behav
reverse
auctions: A laboratory
english
study. Internat. Conf Inform. Sys
tems (ICIS). Association
for Information
Seattle, WA,
Systems,
271-282.
Williams,
S. 1999. A
compatible
characterization
mechanisms.
Econom.
of efficient,
Theory
incentive
Bayesian
155-180.
14(1)