Vol. 24, No. 6, June 2015, pp. 1012–1027 DOI 10.1111/poms.

12301
ISSN 1059-1478|EISSN 1937-5956|15|2406|1012 © 2014 Production and Operations Management Society

Split-Award Procurement Auctions—Can Bayesian

Equilibrium Strategies Predict Human Bidding
Behavior in Multi-Object Auctions?
Martin Bichler
Decision Sciences & Systems, TU M€
unchen, Boltzmannstr. 3, 85748, Garching, Germany, bichler@[Link]

Kemal Guler
Hewlett Packard Labs, 1501 Page Mill Rd, Palo Alto, California 94304, USA, [Link]@[Link]

Stefan Mayer
Decision Sciences & Systems, TU M€
unchen, Boltzmannstr. 3, 85748, Garching, Germany, [Link]@[Link]

e analyze if and when symmetric Bayes Nash equilibrium predictions can explain human bidding behavior in
W multi-object auctions. We focus on two sealed-bid split-award auctions with ex ante split decisions as they can be
regularly found in procurement practice. These auction formats are straightforward multi-object extensions of the first-
price sealed-bid auction. We derive the risk-neutral symmetric Bayes Nash equilibrium strategies and find that, although
the two auction mechanisms yield the same expected costs to the buyer, other aspects of the two models, including the
equilibrium bidding strategies, differ significantly. The strategic considerations in these auction formats are more involved
than in single-lot first-price sealed-bid auctions, and it is questionable whether expected utility maximization can explain
human bidding behavior in such multi-object auctions. Therefore, we analyzed the predictive accuracy of our equilibrium
strategies in the laboratory. In human subject experiments we found underbidding, which is in line with earlier experi-
ments on single-lot first-price sealed-bid auctions. To control for regret, we organize experiments against computerized
bidders, who play the equilibrium strategy. In computerized experiments where bid functions are only used in a single
auction, we found significant underbidding on low-cost draws. In experiments where the bid function is reused in 100
auctions, we could also control effectively for risk aversion, and there is no significant difference of the average bidding
behavior and the risk-neutral Bayes Nash equilibrium bid function. The results suggest that strategic complexity does not
serve as an explanation for underbidding in split-award procurement auctions, but risk aversion does have a significant
impact.
Key words: procurement; multi-object auction; Bayes Nash equilibrium; risk-aversion; experiment
History: Received: December 2012; Accepted: August 2014 by Haresh Gurnani, after 3 revisions.

hundreds of millions of dollars using different types

1. Introduction of multiple sourcing auctions (Elmaghraby 2000,
A growing number of papers over recent years have Tunca and Wu 2009). These firms usually want to
focused on the design of auctions for multiple non- have more than one supplier for risk considerations.
identical objects. The simultaneous multi-round auc- In global business often a new supplier with cheaper
tion, for example, and, more recently, combinatorial but possibly unreliable technology enters the market-
auction designs have been used for selling spectrum place to win orders from firms by beating the price of
licenses worldwide (Cramton 2013). Similar auction their reliable but more expensive competitors (Gur-
design problems regularly arise in industrial procure- nani et al. 2012). Split-award auctions can be used to
ment and logistics where multiple heterogeneous make sure the new entrant does not win the entire
goods or services need to be purchased (Cramton quantity and the second source with the required pro-
et al. 2006). duction facilities is available in case the new entrant
In this study, we focus on split-award auctions, a sim- defaults.
ple but widespread type of multi-object auction often We focus on equilibrium bidding strategies in two
used for multi-sourcing, which is a typical require- forms of sealed-bid auction formats that we fre-
ment in procurement negotiations. Companies such quently encounter in procurement practice for dual
as Sun and HP, for example, procure products worth sourcing. The lots in such auctions could be the 30%
1012
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1013

and the 70% share of the demand for a particular raw and Baranov 2010, Goeree and Lien 2009, Sano 2012).
material, and the buyers choose the split ex ante and Actually, so far there is no Bayes Nash equilibrium
restrict the suppliers to submitting either one single characterization for first-price sealed-bid combinato-
bid or two bids on the two lots. More specifically, we rial auctions. Some very recent research uses numeri-
analyze two first-price sealed-bid procurement auc- cal methods to derive equilibrium predictions, which
tions for auctioning the two lots of a product with dif- also shows the difficulties of Bayesian models for
ferent sizes at one go. The first auction model is the multi-object markets. Given the strategic complexity
so-called Yankee auction, in which each bidder just of these multi-object auctions, it is not clear that
submits one bid for the unit price of the product for RNBNE predictions explain human behavior well.
which they bid. Then, the bidder with the lowest bid The bidders’ decision is single-dimensional only con-
wins the large lot and the bidder with the second low- cerning the level of bid-shading for first-price sealed-
est bid wins the small lot. The second split-award bid auctions of a single object. Bidders in multi-object
sealed-bid auction we study is an extension called the first-price sealed-bid auctions also need to decide
parallel auction, in which each bidder submits one bid which objects they want to bid on and how much they
for the unit price of the product for each lot and the want to shade their bids in each of the lots. Bid shad-
bidder with the lowest bid on each lot wins that lot. ing is defined by the difference between the bid and
Each bidder can win at most one lot. Such auction for- the production costs for a lot. In parallel split-award
mats with an ex ante split are easy to implement for auctions, for example, not only the number of bidders
procurement managers, and it is important for them and the prior distribution, but also the split parameter
to understand the bidding strategies in such auctions. determines the level of bid shading. In the Yankee
Corey (1978), Woodside and Vyas (1987), and Sesh- auction, the bidders also need to take into account the
adri et al. (1991) have already discussed such split- risk of winning the small lot rather than the large lot
award contracts with pre-defined splits in a variety of with a certain bid price.
industries, and such auctions are common in e-sourc- The predictive accuracy of RNBNE predictions for
ing nowadays. Interestingly, we are neither aware of multi-object auctions in the laboratory is largely
a game-theoretical analysis nor of experimental work unexplored. However, there is growing literature on
on these auction formats. Note that the auction for- first-price sealed-bid auctions of single objects which
mats analyzed in this study are different from the shows that bidding behavior in the laboratory devi-
split-award auctions analyzed in Anton and Yao ates substantially from the RNBNE prediction and
(1992) Anton and Yao (1989) and Anton et al. (2010), overbidding is a common phenomenon in sales auc-
where suppliers submit bids on each possible split of tions. In our procurement auction context this means
a contract and the entire quantity and the split choice underbidding below the RNBNE equilibrium predic-
is endogenous. Our study is also different from Perry tion, and we will refer to underbidding and the
and Sakovics (2003), who study a setting in which the underbidding phenomenon throughout unless we
buyer commits to a split ex ante, but the items are then want to distinguish between the results of sales and
sold in a sequential second-price auction. Recently, procurement auctions in the laboratory. The first-
Gong et al. (2012) assume a single-bid second-price price auction controversy and the discussions about
split-award auction with an ex ante split similar to the underbidding already played out in a controversy
Yankee auction in our study, but their model focuses among experimental economists in the December
on incentives of suppliers to invest. We contribute the 1992 issue of the American Economic Review. It
theoretical analysis and Bayes Nash characterization raised the question, how to establish sufficient experi-
of these split-award auctions with a predefined split, mental control to establish empirical regularities in
which is different from the one with endogenous the laboratory, and how to modify theory in light of
determination of market structure, and analyze to countervailing empirical evidence. Ever since, under-
which extent such models have predictive power in bidding has been source of substantial research in the
laboratory experiments. experimental auction literature (see section 3.1). Risk
Bayes Nash equilibrium analysis is the standard aversion, regret, and uncertainty about the rationality
approach to model sealed-bid auctions and a lot of of others have served as explanations for underbid-
recent research has tried to extend this type of ding (Engelbrecht-Wiggans and Katok 2009). A num-
analysis to multi-object auctions (Krishna 2009). The ber of authors have challenged the overall approach
risk-neutral Bayes Nash equilibrium (RNBNE) of of models based on rational choice and expected util-
multi-object auctions is technically much more chal- ity maximization (Bourdieu 2005, Nell 2007).
lenging than that of single-lot auctions and, as of yet, Even if bidders are able to mimic their RNBNE
there are only a small number of papers deriving strategy in a single-lot auction, it is far from obvi-
RNBNE strategies for specific combinatorial or non- ous that RNBNE models would still be a good
combinatorial multi-object auction formats (Ausubel predictor for multi-object auctions. In single-lot
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1014 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

auctions, bidders might estimate the right level of earlier experiments on single-lot first-price sealed-
bid shading. As described above, split-award auc- bid auctions. Although the impact of risk aversion
tions are strategically more complex, and it is should be reduced with many repeated auctions,
interesting to understand if bidders are able to residual risk aversion, wrong expectations about
master the strategic complexity and mimic their other bidders or regret can all serve as explana-
RNBNE strategy at least in a controlled environ- tions for this underbidding. Interestingly, we do
ment. If this is the case, the model can serve as a not find significant underbidding in our computer-
baseline starting point from which one can add ized experiments, where bidders submit bids only
other known phenomena such as risk aversion or once, but their bids are reused in 100 computerized
regret. However, if the RNBNE strategy does not auctions. There is no significant difference between
explain bidding behavior even in a controlled the average empirical bid functions and the
environment, there is little hope that such models RNBNE bid function in the single-lot and in both
would explain bidding behavior in more complex split-award auctions when subjects played against
multi-object auctions such as combinatorial auc- computerized bidders. This provides evidence that
tions. In summary, we try to understand if, in strategic complexity or wrong expectations cannot
spite of the increased strategic complexity of split-award serve as an explanation for underbidding. Comput-
auctions, bidders are able to bid according to the erized experiments where the bid function of a bid-
RNBNE bid function in a controlled experiment. Our der is only used once and not in 100 auctions
contribution is two-fold: exhibit significant underbidding, which indicates
Firstly, we derive closed-form increasing Bayes that risk aversion has considerable impact on the
Nash bidding strategies for the Yankee and the bidding strategies. Andreoni et al. (2007) write that
parallel auction in the symmetric independent pri- “one element of the theory that cannot be repli-
vate values model and also compare and contrast cated in an experiment is the risk neutrality of bid-
the total cost implications. This Bayes Nash char- ders, for the risk attitudes of the subjects cannot be
acterization has been missing from the growing lit- controlled.” The computerized experiments where
erature of multi-object auctions and is particularly the bid function is reused in 100 auctions effec-
relevant for procurement. It is interesting that until tively achieve this goal.
now there is no closed-form expression of Bayes The results provide evidence that bidders in the
Nash equilibrium bidding strategies for first-price laboratory behave as expected utility maximizers in
combinatorial auctions. We find that, although the our auctions and they are able to mimic the compli-
parallel and the Yankee auction mechanisms yield cated equilibrium strategies with surprisingly high
the same expected costs to the buyer, other precision in the computerized experiments, if we con-
aspects of the two models, including the equilib- trol for regret and risk aversion. In other words, the
rium bidding strategies as well as winning bid- RNBNE model describes the basic strategic consider-
ders’ ex post profits, differ significantly. Most of ations of bidders in the laboratory well in the parallel
the previous studies in this area (Anton and Yao but also in the more complex Yankee auction, and that
1992, Armstrong 2000) focus on the comparison of this is independent of the split parameter. We also
auction mechanisms in terms of the expected reve- show that risk aversion leads to significant underbid-
nue, while here, we also compare the different ding, once we do not control for risk aversion. In the
mechanisms in terms of other measures, such as field wrong expectations about others, risk aversion,
the equilibrium bidding strategies and winning and regret can all influence bidder behavior. The level
bidders’ ex post profits, which are important con- of underbidding will depend on the frequency of such
siderations in real-world procurement practice. auctions and bidder idiosyncrasies. Still, we argue
Secondly, we report on laboratory experiments that bidders in split-award auctions are able to under-
with split-award procurement auctions testing stand the strategic situation and that the RNBNE
RNBNE predictions. We designed laboratory exper- model can serve as a useful baseline model for practi-
iments with different levels of control where tioners and future research.
human bidders compete either against other human The study is organized as follows. The two compet-
bidders or against computerized bidders. The latter ing procurement auction models are introduced in
are designed to mitigate the impact of behavioral section 2, where we also discuss equilibrium bidding
influences, such as risk aversion, regret and incon- strategies and their implications. In section 3, the
sistent expectations. In experiments with human experimental design is introduced and the results are
subjects, bidders compete in repeated auctions. presented in section 4. Finally, section 5 concludes
These experiments are modeled after procurement with a discussion on further extensions of the models.
auctions as they can be found in the field. We Technical proofs and the experimental instructions
observe underbidding for low-cost draws similar to are included in the Appendices.
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1015

price for both lots of the products. The large lot is then
2. Theoretical Models awarded to the bidder with the lowest bid and the
2.1. The Auctions small lot is awarded to the bidder with the second
A buyer is to procure a given volume of a product lowest bid. The winning bidders are again compen-
using an auction. For exogenous reasons, the buyer sated according to the bids they make. In this auction
divides the total required volume, normalized to be model, the expected payoff for a generic bidder
1 unit, into two lots with q units in lot 1 and 1
q whose private per-unit production cost is c and whose
units in lot 2. We further assume that q > 0.5, and bid is b is
hence, lot 1 is the large lot. There are N risk-neutral
bidders competing for the two lots. Bidder i’s pri- pðb; cÞ ¼ Prðbidder i wins Lot 1Þ  ½ðb
cÞq
K
vate constant marginal production cost ci is identi- þ Prðbidder i loses Lot 1 and wins Lot 2Þ ð2Þ
cally and independently distributed according to  ½ðb
cÞð1
qÞ
K:
the distribution F() with support ½c; c and density
f(). In addition to the per-unit marginal production Since in this auction, bidders only submit one
cost ci, each supplier must also incur a fixed cost K bid, there are only two possible outcomes other
to complete the production. Finally, we assume that than losing. The first case is that bidder i submits
there is no reserve price. the lowest bid, and hence, wins the large lot. The
other possibility is that his bid is the second lowest
2.1.1. The Parallel Auction. The rules of the par- and he is awarded the small lot. Conditional on the
allel auction are as follows: After observing his other bidders’ behavior, bidder i chooses the bid-
private per-unit constant marginal production cost ding strategies by maximizing the expected payoff
for the product, ci, each bidder submits two bids, b1i Equation (2).
and b2i , as the per-unit price for each lot of the prod-
ucts. The lots are awarded to the lowest bid on each
lot and the winning bidder gets the contract with 2.2. Characteristics of The Equilibria
the payment that he bids. If one bidder has the low- We will discuss the Bayes Nash equilibrium strategies
est bids on both lots, he will be awarded only the in what follows. All proofs can be found in Appendix
large lot. S1.
We will denote the expected payoff to a generic bid-
der whose private per-unit production cost is c and 2.2.1. The Parallel Auction. We restrict our atten-
whose bids are b1and b2 by p(b1, b2, c). Then tion to symmetric Bayes Nash equilibrium strategies
in which bids are continuous, strictly increasing and,
almost everywhere, differentiable functions of costs.
pðb1 ; b2 ;cÞ ¼ Prðbidder i wins both LotsÞ
Let (b1(ci), b2(ci)) denote the equilibrium pair of bid-
 ½ðb1
cÞq
K ding strategies for the two lots.
þ Prðbidder i wins Lot 1 and loses Lot 2Þ
ð1Þ
 ½ðb1
cÞq
K PROPOSITION 1. In the parallel sealed-bid first-price pro-
curement auction model with two lots and N > 2 risk-
þ Prðbidder i loses Lot 1 and wins Lot 2Þ
neutral bidders, the unique symmetric Bayes Nash
 ½ðb2
cÞð1
qÞ
K: equilibrium bidding strategies are given by
The three terms in Equation (1) reflect the three K
possible outcomes of the auction game other than los- b1 ðcÞ ¼ c þ
q
ing. The first possibility is that a bidder submits the Z c 
lowest bid on each lot, and hence, could win both lots. 1 1
FðxÞ N
2
þ fq½1
FðxÞ:
However, according the auction rule, he is only q½1
FðcÞ c 1
FðcÞ
awarded the large lot, that is, lot 1. The second possi- þð1
qÞðN
1Þ½FðxÞ
Fðci Þgdx
bility is that he only wins the large lot 1, and the last Z c 
K 1
FðxÞ N
2
term is the case where he only wins the small lot 2. b2 ðcÞ ¼ c þ þ dx;
Conditional on the other bidders’ behavior, bidder i 1
q c 1
FðcÞ
chooses the bidding strategies by maximizing the
with the boundary conditions (1) b1 ðcÞ ¼ c þ Kq and (2)
expected payoff Equation (1).
b2 ðcÞ ¼ c þ 1
q
K
:
2.1.2 The Yankee Auction. The Yankee auction Furthermore, b1(c)q > b2(c)(1
q)and b2(c) > b1(c) if
works as follows: After observing his private per-unit the entry cost K is greater than a threshold K0 where
 N
2 h i
constant marginal production cost for the product, ci, ð1
qÞ R c ½Nq
ðN
1Þ½FðxÞ
FðcÞ
K0 :¼
ð2q
1Þ c 1
FðxÞ ½1
FðcÞ dx:
each bidder i submits only one bid bi as the per-unit 1
FðcÞ
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1016 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

Proposition 1 calls for a few comments. Firstly, COROLLARY 1. In the Yankee sealed-bid first-price pro-
the equilibrium bidding function for the small lot is curement auction model, ex post,
the same as the bidding strategy in a standard
sealed-bid first-price auction with N
1 bidders (i) the winner of the large lot always makes a positive
competing for the small lot. Secondly, the bidders profit; and
bid less aggressively than the equilibrium bid they (ii) the profit for the winner of the small lot can be
submit in a standard sealed-bid first-price auction either positive or negative.
with N bidders competing for the large lot, because
Corollary 1 reveals an important difference between
they can still win the small lot. This means that bid-
the two models. In the parallel auction mechanism, bid-
ders need to take into account the possibility of los-
ders bid in such a way that, ex post, they always make
ing the large lot and winning the small lot when
positive profits no matter which lot is awarded. In con-
they submit bids for the large lot. The strategy is
trast, in the Yankee mechanism, ex post, the bidder who
similar to a sequential auction where the large lot is
wins the small lot will earn either a positive or negative
sold first. In addition, ex post, the winning bidders
profit depending on his private production costs.
make positive profits no matter which lot they win.
Lebrun (2006) shows that with standard assump-
PROPOSITION 4. In the Yankee sealed-bid first-price procure-
tions on the type distributions such an equilibrium
ment auction model with two lots and N risk-neutral bidders,
is unique and it is characterized as the solution to
the system of differential equations corresponding (i) As N increases, bidding becomes more (less) aggres-
to the first-order conditions. sive if the fixed cost K is greater than (smaller than)
a threshold K1:
PROPOSITION 2. In the parallel sealed-bid first-price pro-
curement auction model with two lots and Nrisk-neutral @bðcÞ
bidders, [ 0ð\0Þ if K [ ð\ÞK1 ðN; c; FÞ
@N
where
i. As the number of bidders increases, bidding becomes Rc
more aggressive for both lots: ci fA þ B þ Dgdx
K1 ðN; c; FÞ :¼

½1
FðcÞN FðcÞð2q
1Þ
@b1 ðcÞ @b2 ðcÞ
\0 and \0: A ¼ ½FðcÞ
FðxÞ½1
FðxÞN
1 ½1
FðcÞq2
@N @N
B ¼ ½FðcÞ
FðxÞ½1
FðxÞN
2 FðxÞFðcÞð1
qÞ2
ii. As q increases, bidding becomes more (less) aggres-
sive for the large lot 1 (the small lot 2): D ¼ ½FðcÞ
FðxÞ½NFðcÞ
2ðN
1ÞFðcÞFðxÞ
þNFðxÞ
1½1
FðxÞN
2 qð1
qÞ:
@b1 ðcÞ @b2 ðcÞ
\0 and [ 0:
@q @q (ii) As qincreases, bidding becomes more (less) aggres-
sive if the fixed cost K is greater than (smaller than)
2.2.2. The Yankee Auction. As above, we restrict a threshold K2:
our attention to strictly increasing differentiable sym-
@bðcÞ
metric Bayes Nash equilibrium strategies. At such [ 0ð\0Þ if K [ ð\ÞK2 ðN; c; FÞ
equilibrium, bidder i chooses his bid bi = b(ci) by @q
maximizing his expected payoff. where
1 N
1
PROPOSITION 3. In the Yankee sealed-bid first-price pro- K2 ðN; c; FÞ :¼ :
1 þ ðN
2ÞFðcÞ NFðci Þ
1
curement auction model with two lots and N > 2 risk- Z c  
neutral bidders, the unique symmetric Bayesian Nash 1
FðxÞ N
2
½FðxÞ
FðcÞdx:
equilibrium bidding strategies are given by c 1
FðcÞ

Rc n o
c ½1
FðxÞN
1 q þ ðN
1ÞFðxÞ½1
FðxÞN
2 ð1
qÞ dx
bðcÞ ¼ c þ n o
½1
FðcÞN
1 q þ ðN
1ÞFðcÞ½1
FðcÞN
2 ð1
qÞ
f1 þ ðN
2ÞFðcÞg
þK ;
f½1
FðcÞq þ ðN
1ÞFðcÞð1
qÞg

with the boundary conditions bðcÞ ¼ c þ 1
q

K
.
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1017

Comparative statics analysis of the equilibrium bid- under the two procedures showed that bidders were
ding strategy for the Yankee model is more involved. risk averse in the first-price auction but risk-neutral,
Partial derivatives are not of uniform sign throughout or moderately risk-loving, under the BDM procedure.
the parameter space, but regions where the partial Overall, risk attitudes not only differed across assess-
effects are positive and negative can be characterized ment methods, but also varied within the same
in terms of thresholds on the magnitude of the fixed method used (Krahnen et al. 1997, MacCrimmon and
cost K. Wehrung 1990, Payne et al. 1980, Schoemaker 1993).
Paired lottery choices, as introduced by Holt and Lau-
2.3. Procurement Cost Comparisons ry (2005), have become more popular recently, but
Having derived the symmetric equilibrium bidding risk aversion is still recognized as a complex- and con-
strategies in both the parallel and the Yankee auc- text-specific phenomenon, which is difficult to mea-
tions, we can now compare the implications of the sure (Dohmen et al. 2005).
two mechanisms on expected cost to the buyer. The Engelbrecht-Wiggans (1989) introduced post-auc-
proof for Proposition 5 can be found in Appendix S1 tion regret as another explanation for the bidding
and it also follows from Engelbrecht-Wiggans (1988) behavior in first-price sealed-bid auctions. Engelbr-
and Krishna (2009, p. 206). echt-Wiggans and Katok (2009) found support for the
regret model in experiments with information feed-
PROPOSITION 5. Suppose that the private production back about the highest or second-highest bid after the
costs are independently and identically distributed and all auction. Filiz-Ozbay and Ozbay (2007) also found
bidders are risk-neutral. Then the sealed-bid first-price support for regret. Deviations from the RNBNE bid
parallel procurement auction and the sealed-bid first-price function could also be due to wrong expectations of the
Yankee procurement auction yield the same expected cost bidders about the bids of others (Stahl and Wilson
to the buyer. 1995). Goeree et al. (2002) demonstrate that misper-
ceived probabilities of winning the auction would
The equivalence result is a useful piece of informa- explain underbidding as well as risk aversion.
tion for business decision-makers, because in theory Problems in computing a best response are yet
they can now look into other dimension of the differ- another conjecture why bidders might not be able to
ences between the two formats and recommend follow the RNBNE strategy. The mathematical deriva-
which format to adopt in practice without worrying tion of the RNBNE of split-award auctions in sec-
about the expected costs by using different auctions. tion 2 is more complex than in single-lot auctions and
it is far from obvious that human bidders in the labo-
ratory would be able to mimic these derivations.
3. Experimental Setup Actually, there are contradictory claims as to whether
Predicting strategies of bidders in sealed-bid auctions people reason according to Bayesian inference (Gige-
turned out to be a challenge, as factors such as the renzer and Hoffrage 1995). We will refer to the prob-
uncertainty about the “rationality” of other bidders, lem of deriving the RNBNE bid function based on
risk aversion, and regret can all play a role. In fact, the given prior distributions about valuations as the stra-
literature on single-lot first-price sealed-bid auctions tegic complexity of the auction.
shows a consistent and significant level of underbid- The literature on underbidding is extensive and
ding compared to the Bayes Nash equilibrium predic- beyond what we can discuss in this section. Some
tion (Roth and Kagel 1995). We will firstly summarize authors mention spite and joy of winning as
lessons learned from the research on underbidding in potential reasons for deviations from equilibrium
single-lot first-price sealed-bid auctions before we predictions in second-price auctions (Cooper and
introduce our experimental design. Fang 2008). However, risk aversion, regret, wrong
expectations, and strategic complexity are the most
3.1. The Underbidding Puzzle in Sealed-bid natural conjectures for deviations in our experi-
Auctions ments.
A number of authors have used risk aversion to explain
bids above the RNBNE (Andreoni et al. 2007, Chen 3.2. Experimental Design and Hypotheses
and Plott 1998, Cox et al. 1988, Kirchkamp and Reiss In our experiments, we wanted to test our theoretical
2006). However, measuring the risk aversion of labo- results and used four designs which mirror the deci-
ratory subjects turned out to be a challenge. Isaac and sion situation in split-award auctions with different
James (2000) compare estimates of risk preferences levels of control for the conjectures discussed in the
from first-price sealed-bid auctions to the Becker– previous section. We organized computerized experi-
DeGroot–Marshak (BDM) procedure for comparably ments where subjects competed against computer
risky choices. Aggregate measures of risk preferences agents to test hypotheses for underbidding, and
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1018 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

human subject experiments to analyze environments their expected payoff. Bidders did not learn about
which mirror real-world environments. other bids in the auction, just whether they won or
In all treatments, we used the strategy method in a lost an auction, which should minimize the impact of
way similar to Selten and Buchta (1999), G€ uth et al. regret as it was shown in Engelbrecht-Wiggans and
(2003), and Kirchkamp and Reiss (2011). The strat- Katok (2009). Of course, risk aversion and wrong
egy method elicits bid functions rather than bids for expectations can still be a driver for deviations from
individual cost draws and it allows us to observe the RNBNE.
bidding functions in much more detail. Other exper-
iments by Kirchkamp and Reiss (2006) show that HYPOTHESIS 2. Bidding against computerized agents
bidding behavior that is observed with the strategy without information about the bids of others after the
method is very similar to the behavior observed auction eliminates underbidding.
with alternative methods. In contrast to this earlier
work, in our computerized experiments, we reuse a The third group of treatment combinations (C100) is
bid function 100 times in auctions against computer identical to C1, but the bid function of a user is reused
bidders to eliminate risk aversion. We will see that in 100 auctions, which should mitigate risk aversion.
this procedure is very effective. Let us briefly moti- For each new auction, we drew a cost value randomly
vate the three groups of treatment combinations in and determined the bid based on the bid function of a
our experiments. bidder to participate in an auction against computer-
The first group of treatment combinations (H) is mod- ized bidders. The subject is then paid the average of
eled after real-world procurement practices. These his winnings in the 100 auctions. The impact of regret
experiments model an environment where four bid- should also be minimal, because bidders did not learn
ders compete in the same type of auction against a about the outcome of individual auctions or the bids
pool of unknown bidders, but the competitors change of others. The difference between C100 and C1 pro-
over time. Four bidders received cost draws and com- vides an estimate for the impact of risk aversion in
peted against each other. Bidders submitted bids in 16 these auctions.1
subsequent auctions and they are re-matched ran-
domly after each auction. These experiments allow HYPOTHESIS 3. Bidding against computerized agents
for learning about the auction format, and the 16 repe- without information about the bids of others, where bid
titions mitigate risk aversion to some extent, which functions are reused in 100 auctions, eliminates
can be motivated by real-world tenders. After each underbidding.
auction, bidders had a possibility to revise their bid
functions and new variable costs are drawn for each The fourth group of treatment combinations (C100+)
bidder independently. After each auction, the bids of uses the same experimental design as C100, but we
all competitors are revealed, as is typically the case, also provide explicit information about the RNBNE
for example, in construction or in public sector auc- function of the computerized agents. Since the bid-
tions. This is often done to combat collusion or brib- ders are all ex ante symmetric, the information in
ery (Thomas 1996). In contrast to the computerized C100+ tells the subjects implicitly what their equi-
experiments, which are designed to understand the librium bidding function would be. Bidders should
potential impact of risk aversion or wrong expecta- just replicate the RNBNE strategy of others to
tions on underbidding, the results of the treatment determine their best response. Here we control
combination H should have external validity as they for wrong expectations about the computerized
are close to real-world practices. Our main hypothesis bidders, which might be different from wrong
for this set of baseline treatment combinations is expectations that bidders have in human subject
therefore: experiments. Still, it is valuable to understand
which impact explicit information about the bid-
HYPOTHESIS 1. Bidders in human subject experiments ding strategies of others has on bidders compared
will underbid below the RNBNE bid function. to a treatment where this information is not avail-
able in C100. We consider deviations from the
As we found underbidding in line with earlier RNBNE in C100+ as ground noise or irrationality,
experiments on single-lot auctions, we introduced which provides a baseline for other experiments.
additional treatments to control for different conjec- Still, bidders might not understand that their best
tures why bidders underbid. A second group of treat- response is to mirror the RNBNE in this treatment
ment combinations (C1) had human subjects compete and strategic complexity can still be a reason for
against computerized agents, which played their deviations. However, we cannot expect subjects in
RNBNE strategy. Bidders are told that their oppo- experiments with less control to be closer to the
nents are rational computer agents, who maximized RNBNE prediction.
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1019

HYPOTHESIS 4. Bidding against computerized agents support of [0, 10.0]. The fixed cost K is 1 for all bid-
without information about the bids of others, where bid ders. The split parameter in the experiments reported
functions are reused in 100 auctions, and bidders see the in the following with treatment combinations H,
equilibrium bid functions of their computerized C100, and C100+ is q = 0.7.
opponents, eliminates underbidding. We also ran additional computerized experiments
of C100+ and C100 with a split parameter of q = 0.9 to
Table 1 provides an overview of how we control make sure that the high predictive accuracy of the
for different hypotheses for deviations from the RNBNE function that we found for the split of q = 0.7
RNBNE in the four different treatment combinations. is robust against changes of the split parameter. This
As discussed earlier, in treatment C100+ strategic could be confirmed. We provide the results of the split
complexity can still be an explanation for possible of q = 0.9 only in Appendix S4, to limit the number of
deviations from the RNBNE. Bidders in C100 only get treatment combinations in the main part of the study.
the prior cost distributions, and therefore wrong In addition, we performed experiments with single-
expectations and strategic complexity can both lot auctions (q = 1) to understand how the results
explain deviations. The difference between C100 and compare with traditional reverse auctions. This is nec-
C1 is only the number of times in which the bid func- essary, because we are not aware of similar experi-
tion is reused. Therefore, this difference can be ments with a reverse first-price sealed-bid auction.
explained by risk aversion to a large extent. Regret Overbidding on high-cost draws in sales auctions
should have little impact, because bidders did not might just be different from underbidding for low-
learn about the bids of others in an auction. Treatment cost draws in reverse auctions, and these experiments
combination H allows for all explanations, although provide us with a baseline.
the 16 repetitions should mitigate risk aversion to
some extent. 3.3. Experimental Procedures
The individual treatment combinations are All experiments were conducted from November
described in Table 2. Overall, 209 subjects were 2011 to November 2013 with students in computer
involved in the experiments. In all treatment combi- science, mathematics, physics, and mechanical engi-
nations, variable costs per unit, ci, were i.i.d. random neering. The subjects were recruited via e-mail lists
variables drawn from a uniform distribution with a and experiments were conducted in a computer labo-
ratory at our university. At the beginning of the
experiment, participants were randomly assigned to
Table 1 Control for Reasons of Deviations from the RNBNE in seats in the laboratory. The participants obtained
Different Treatment Combinations written instructions. These instructions varied slightly
depending on the treatment (see Appendix S2). All
H C1 C100 C100+
the instructions were read aloud and participants had
Strategic complexity + + + + to participate in a test about the economic environ-
Wrong expectations + + +

Risk aversion + +


ment and the auction rules. In addition, we conducted
Regret +


a test auction to make sure that the subjects were
familiar with the auction design and the user inter-
The + sign indicates a possible reason for deviation, while
indicates face. Bidders could take as much time as they wanted
that this reason is unlikely in this treatment.
to write down their bid functions. The average

Table 2 Overview of Treatment Combinations in the Experiments Describing the Number of Auctions in Which a Bid Function is Used, the Types of
Bidders Against Which a Subject Competed (Human or Computer Bidder), the Information Available to Bidders Before an Auction, the
Auction Format (Parallel or Yankee Auction), and the Number of Bidders Involved in Auctions with This Treatment Combination

Treatment Bid function reused Opponents Information Split Auction format No. of subjects
H.S 1 Human Prior distribution, bids of past auctions 1.0 Single object 16
H.P 1 Human Prior distribution, bids of past auctions 0.7 Parallel 16
H.Y 1 Human Prior distribution, bids of past auctions 0.7 Yankee 16
C1.S 1 Computer Prior distribution 1.0 Single object 12
C1.P 1 Computer Prior distribution 0.7 Parallel 12
C1.Y 1 Computer Prior distribution 0.7 Yankee 11
C100.S 100 Computer Prior distribution 1.0 Single object 11
C100.P 100 Computer Prior distribution 0.7 Parallel 13
C100.Y 100 Computer Prior distribution 0.7 Yankee 13
C100+.S 100 Computer Prior & RNBNE bid function 1.0 Single object 10
C100+.P 100 Computer Prior & RNBNE bid function 0.7 Parallel 11
C100+.Y 100 Computer Prior & RNBNE bid function 0.7 Yankee 11
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1020 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

auction duration for all sealed-bid formats was 4.1.1. Results on Single-Lot Procurement
around 10 minutes. The repeated auctions in H took Auctions.
less time, as bidders usually finished their update S1. There is underbidding in treatment H, but the level
after 5–10 minutes. The experiments C1, C100, and of underbidding decreased after a few rounds. We fail to
C100+ took 45 minutes overall on average, while reject Hypothesis 1.
experiment H took 2 hours and 30 minutes on
average. S2. The RNBNE strategy describes the empirical data
Each participant received a show up fee of in C100 and C100+ in the single-lot auction well. This
10 euro for H, which took more time, and 5 Euro suggests that wrong expectations and the strategic
for the computerized experiments. Losses could complexity have little impact on the bidding behavior. We
reduce the fee. Payoffs from auctions were trans- fail to reject Hypothesis 3 and 4.
lated from Franc, the experimental currency, into
Euro through a fixed exchange rate. If participants S3. We found underbidding on low-cost draws in C1
had colluded, they would have been excluded from and reject Hypothesis 2. The difference to C100 indicates
the experiment without any payment. However, we that risk aversion has substantial impact on bidder
did not find evidence of collusion. If subjects had behavior.
made a loss, which was not covered by the show up
and the payoff of all auctions, they would also have Support: To analyze the empirical bid functions,
been excluded from the experiment and the session we analyze the outcome of linear regression models
would have been cancelled. Actually, there were no in the different treatments and compare it with the
losses which were not covered by the show up fee. linear RNBNE bid functions for the small and the
Subjects only participated once in one session. Over- large lot.2 We use a fixed effects model with a dummy
all, 103 students participated in the experiments variable ui to estimate the unobserved heterogeneity
with split-award auctions where the split was of bidders i.
q = 0.7, and 57 students in experiments with a split yit ¼ a þ bcit þ rit þ dwit þ ui þ rit :
parameter of q = 0.9 (see Appendix S4). An addi-
tional 49 students participated in the experiments The dependent variable yit describes the bid t sub-
with single-lot reverse auctions. mitted by bidder i. The unit costs cit are used as the
From the 103 students participating in split-award main independent variable. The coefficients ui for the
auctions (q = 0.7), 22 were in C100+ (average payoff bidder ID of all the bidders control for bidder idiosyn-
13.38 €), 26 in C100 (average payoff 13.82 €), 23 in crasies, but they are omitted from the table to focus
C1 (average payoff 8.27 €), and 32 in H (average on the main variables. Variable rit controls the number
payoff 46.76 €). From the 49 students in single-lot of the auction (or round) and is only used in repeated
reverse auctions, 10 participated in C100+ (average experiments in treatment combination H. c is the
payoff 13.15 €), 11 in C100 (average payoff 13.49 €), coefficient for the number of an auction in an experi-
12 in C1 (average payoff 7.25 €), and 16 in H (average ment. Variable wit-1 describes whether a bidder won
payoff 53.45 €). in the previous auction and d describes the impact
of winning in the last round. Table 3 summarizes
the main parameters: the intercept a, the regression
4. Experimental Results coefficient b for the unit costs cit, and the multiple R²
We will now describe bidder behavior in the of the linear regression.
laboratory. We want to test the theoretical predictions We also compute the mean squared error (MSE) of
and understand how well the RNBNE bid function the RNBNE function to understand how well the
explains the empirical observations in the different model explains the data in the different treatment
treatments. combinations. This metric is lowest in C100+, indicat-
ing that the variance around the RNBNE bid function
4.1. Single-Lot Procurement Auction is low. Plots of the empirical bid functions can be
Before we look at split-award auctions, we analyze sin- found at the end of Appendix S3. We compare the
gle-lot procurement auctions in which the entire quan- MSE of the linear RNBNE function against the MSE of
tity goes to one supplier. This will provide us with a a LOESS estimation of the data (Cleveland and Devlin
baseline against which we can compare bidding 1988). LOESS is also known as locally weighted poly-
behavior in split-award procurement auctions. We nomial regression, which can be considered a best
will discuss the four hypotheses from the previous sec- case model for the empirical data. At each point in the
tion and organize them in three results because we can dataset a low-degree polynomial is fitted to a subset
aggregate C100 and C100+. Then we provide statistics of the data. The value of the regression function
supporting the results. for the point is obtained by evaluating the local
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1021

Table 3 Regression Coefficients for the Empirical Bid Functions (w/o Bidder ID) of the Single-Lot Auction

Single-lot

a b (unit cost) SE b (p-value) Mult. R2 MSE RNBNE MSE LOESS No. bids/bidders
RNBNE 2.53 0.748
H.S 2.11 0.821 0.005 (0.000) 0.923 0.854 0.676 2560/16
H.S (# 1) 1.51 0.869 0.028 (0.000) 0.897 2.127 1.936 160/16
H.S (# 7) 1.87 0.793 0.016 (0.000) 0.952 0.587 0.387 160/16
H.S (# 16) 1.70 0.829 0.014 (0.000) 0.966 0.699 0.353 160/16
C1.S 1.92 0.767 0.039 (0.000) 0.862 4.828 (0.598) 4.740 (0.382) 120/12
C100.S 2.53 0.727 0.033 (0.000) 0.922 0.415 0.392 110/11
C100+.S 2.08 0.797 0.008 (0.000) 0.991 0.104 0.083 100/10

polynomial using the explanatory variable values for ments C100+ (p = 0.999) and C100 (p = 0.991). This
that data point. means that, even without information about the bid
The lines H.S (#1) to H.S (#16) in Table 3 describe functions of computerized agents, the empirical bid
the results of the regression for the empirical bid func- functions in C100, and C100+ are very close to the
tions in individual auctions (numbers 1, 7, and 16) RNBNE bid function, which is also illustrated by the
in H. The low intercept a together with a higher b low MSE. Note that the difference in the intercept a
compared to the RNBNE function indicates that there between C100 and C100+ is caused by an outlier (see
is underbidding on average on low-cost draws in H Figure S15), but that the MSE is lower in C100+ as
compared to the RNBNE bid function. In auction #7, expected. We will also find no significant difference
for example, there is an underbidding of 18.7% at a between C100+, C100, and the RNBNE bid function in
unit cost of 1 compared to the RNBNE bid function, the split-award auctions.
whereas there is underbidding of 2.75% for high-cost Most, but not all, empirical bid functions in the
draws of 9 Francs. single-lot auction increased monotonously, but there
The value of intercept a, which can be used as an are also some spikes. These small non-monotonici-
estimator for underbidding on low value draws, is at ties which we found across all treatments might be
a mean value of 1.74. The a decreased slightly in the due to errors that bidders make when typing in the
last six rounds from a value of 1.81 to a value of 1.70. data in spite of the graphical display of their bid
This can be explained by some bidders who became function.
more aggressive on low-cost draws to become win- Discussion: As already mentioned, overbidding
ners before the experiment was over. More aggressive on large valuations in experiments on first-price
bidding across rounds overall is also illustrated by a sealed-bid sales auctions is a consistent pattern. Kir-
significantly negative, but low, coefficient c =
0.04. chkamp and Reiss (2011) report median overbidding
We found a small but significantly negative impact of of up to 30% over the RNBNE on high-value draws,
winning in the last round (d =
0.06), which cannot but even modest underbidding for low-value draws.
be explained by regret. This small negative impact Pezanis-Christou and Sadrieh (2003) report average
can also be observed in the different split-award auc- relative overbidding over the RNBNE prediction of
tions. Note that in this study, we want to analyze 34–37% for their experiments with symmetric bid-
when the RNBNE can explain bidding behavior in ders. These authors also use the strategy method.
first-price auctions. This allows us to rule out explana- However, since they test sales auctions and the
tions such as strategic complexity as reasons for number of auctions and competitors is different, the
underbidding. The question, whether risk-aversion or level of overbidding in their experiments cannot
rather regret determine the underbidding in our easily be used as an estimate for underbidding in
experiments may be a fruitful exercise to look at in our reverse auctions. Kirchkamp and Reiss (2011),
the future. for example, used 12 iterations and two bidders in
The high MSE (RNBNE) in treatment combination experiments, while we had four bidders in 16
C1.S is due to a single bidder who bid substantially rounds of a reverse auction. Underbidding for low
above the RNBNE bid function. Without this bidder, unit costs of 1 Franc in our experiments is on aver-
the MSE is 0.589. The average underbidding at a unit age 18–19% below the RNBNE bid function in treat-
cost of 1 is 18.02% below the RNBNE bid function. ments C1 and H, as described above. As we will
The Chow test allows testing whether the regression see, the level of underbidding in treatments C1 and
coefficients of two linear regressions are significantly H increases in split-award auctions, while C100 and
different from the RNBNE. There is no significant C100+ are not significantly different from the
difference between the RNBNE prediction and treat- RNBNE prediction as well.
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1022 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

4.2. Parallel Auction used afterward. The line H.P in Table 4 describes the
We will now discuss the results of the parallel auc- relevant regression coefficients of all human subject
tion, which requires bidders to think about their bids experiments where we control for bidder idiosyncra-
for the small and the large lot. The results are orga- sies and the number of auctions. The subsequent lines
nized similar to those of the single-lot auction. describe the results of the regression for the empirical
bid functions in individual auctions (numbers 1, 7,
4.2.1. Results on for the Parallel Procurement and 16) in treatment combination H. The low intercept
Auctions. a together with a higher b compared to the RNBNE
P1. There is underbidding in treatment H, but the level bid function indicates that there is underbidding on
of underbidding decreased after a few rounds. We fail to average on low-cost draws in H compared to the com-
reject Hypothesis 1. puterized treatments in C100 and C100+. However,
there is even more underbidding in C1. Note that in
P2. The RNBNE strategy describes the empirical data in
the initial sealed-bid treatments with C100 and C100+,
C100 and C100+ in the parallel auction well. This suggests
we have elicited the bid function for 20 unit costs from
that wrong expectations and the strategic complexity of the
0.5 to 10 Francs, while for the human subject experi-
parallel auction have little impact on the bidding behavior.
ments where students had to submit their bid func-
We fail to reject Hypothesis 3 and 4.
tion multiple times; we reduced this to 10 parameters.
P3. We found underbidding on low-cost draws in C1 and In test experiments with treatment combination C1,
reject Hypothesis 2. The difference to C100 indicates that we did not find that this had any impact on the shape
risk aversion has substantial impact on bidder behavior. of the bid function in the experiments.
In Table 5, we have provided the MSE of all three
Support: Table 4 follows the format of Table 3, but models for the parallel auction. The MSE is again
describes the regression coefficients for the large and lowest in C100+ and in C100 (marked in bold in
the small lot. The regression line of C1 has a much Table 5). Actually, C100 has even a lower MSE than
lower intercept compared to the RNBNE bid C100+ for the small lot. The MSE of C1 is also much
function, which shows underbidding on low-cost higher than that of C100 and C100+ and higher than
draws. This can be explained by the impact of risk H, where some of the risk aversion might be elimi-
aversion, because the only difference from C100 is nated due to the 16 repetitions of the experiment.
the number of auctions in which the bid function is We have added additional statistics to compare the

Table 4 Regression Coefficients for The Empirical Bid Functions of the Parallel Auction

Large lot Small lot

Α b (unit cost) Mult. R2 SE b (p-value) a b (unit cost) Mult. R2 SE b (p-value) No. bids/bidders
RNBNE 3.59 0.643 3.37 0.67
H.P 1.70 0.778 0.951 0.004 (0.000) 2.68 0.735 0.929 0.004 (0.000) 2560/16
H.P (# 1) 1.45 0.777 0.937 0.017 (0.000) 2.07 0.765 0.950 0.015 (0.000) 160/16
H.P (# 7) 1.74 0.772 0.961 0.013 (0.000) 2.79 0.722 0.940 0.016 (0.000) 160/16
H.P (# 16) 1.28 0.785 0.963 0.013 (0.000) 1.85 0.731 0.950 0.015 (0.000) 160/16
C1.P 1.48 0.854 0.938 0.022 (0.000) 1.81 0.805 0.947 0.019 (0.000) 120/12
C100.P 3.70 0.697 0.966 0.008 (0.000) 3.67 0.680 0.977 0.007 (0.000) 260/13
C100+.P 3.49 0.626 0.963 0.008 (0.000) 3.39 0.621 0.957 0.009 (0.000) 220/11

Table 5 MSE of The RNBNE in the Split-Award Auction, the RNBNE of a Single-Item Auction, and the MSE of a Constant Profit Margin Model

Large lot Small lot

MSE MSE MSE single-item MSE constant MSE MSE MSE single-item MSE constant
LOESS RNBNE RNBNE factor LOESS RNBNE RNBNE factor
H.P 0.329 1.160 0.449 1.124 0.403 0.607 0.484 1.089
H.P (# 1) 0.578 1.107 0.629 1.165 0.568 0.708 0.690 1.059
H.P (# 7) 0.269 1.037 0.386 1.059 0.319 0.496 0.424 1.050
H.P (# 16) 0.319 1.340 0.508 1.223 0.453 0.750 0.539 1.234
C1.P 0.827 1.668 0.948 1.234 0.682 1.057 0.730 1.129
C100.P 0.269 0.328 0.406 1.037 0.187 0.192 0.443 1.055
C100+.P 0.141 0.175 0.464 1.337 0.206 0.260 0.542 1.441

The MSE of the LOESS estimate serves as a baseline to compare against.

Bold numbers highlight the lowest MSE values among the MSE RNBNE, the MSE of the single-item RNBNE, and the MSE of the constant factor model.
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1023

RNBNE against the predictive power of a model 40.63% for treatment H (auction #7) and 44.79% for
with a constant profit margin and the RNBNE of a C1. On the small lot, we observed underbidding of
single-item first-price sealed-bid auction. This should 13.07% for treatment H (auction #7) and 35.27% for
help understand how sensitive the predictions are. C1. By comparison, in the single-lot reverse auction,
The model assuming bidders had a constant profit we observed around 18–19% for both treatments.
margin had the worst MSE in all treatments. For this Risk aversion can serve as a natural explanation for
model, we used the average markup of the RNBNE the underbidding in C1.P. In H.P, the residual risk
function across all draws as the profit margin. For aversion in spite of the 16 repetitions, but also other
the treatment combinations C1 and H the RNBNE of conjectures such as regret, can be potential reasons for
the single-item auction has a lower MSE than the underbidding on low-cost draws. However, given
RNBNE of the split-award auction. This can easily our experimental design where bidders do not learn
be explained by the underbidding observed in these about their opponents, we conjecture that risk aver-
treatments. In a single-item auction with the same sion serves as the most likely explanation for under-
number of bidders the competition is higher, which bidding in treatment H.
brings down the bid prices in equilibrium below that The Chow test shows equivalence between the
of the RNBNE in the split-award auction. Again, the RNBNE bid function and those in C100+ for the large
MSE for the treatment C1 is highest, which can be (p = 0.7399) and the small lot (p = 0.934). There is also
explained by risk aversion and the differences in no significant difference to the average bid functions
how bidders respond to risk aversion. A few bidders in C100 for the large (p = 0.2698) and the small lot
deviated substantially from the RNBNE prediction, (p = 0.031). The test shows that the bid functions in
which led to a high MSE (see Figure S3). C1 and H are both significantly different from the
As an example, Figure 1 describes the bid functions RNBNE bid function for both lots (p = 0.000).
in the first and seventh auction of H for the large lot.
A thick solid line describes the average bid function P5. The correlation between the markups of bidders in
based on a linear regression, while a thick dashed line the large and in the small lot is high, and the markup on
shows the RNBNE bid function. In both plots, we can average in the small lot is significantly higher than in the
see the underbidding in the lower values compared to large lot, as theory predicts.
the RNBNE bid function. Figures S1–S7 in Appendix
S3 show the bid functions for all other treatments in Support: We found the markups between the large
the parallel auction. and the small lot to be highly correlated (H
The intercepts of both C1.P and H.P are much lower [q = 0.981], C1 [q = 0.968], C100 [q = 0.977], and
than those of the RNBNE bid function for the large C100+ [q = 0.973]). In other words, bidders with a
and the small lot. Underbidding below the RNBNE high markup on the large lot also have a high markup
for low costs of 1 Franc on the large lot is on average on the small lot. The differences in the markup

Figure 1 Scatter Plot of Bids and The Optimal Bid Functions for H on The Large Lot for the First (Left) and the 7th Auction (Right) for the Parallel
Auction (q = 0.7) and a Fixed Cost of K = 1
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1024 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

between the large and the small lot in all treatments 4.3.1. Results for the Yankee Auction.
are significant throughout (paired t-test, a = 0.01). Y1. There is underbidding in treatment H, but the level
This suggests that bidders on average followed the of underbidding decreased after a few rounds. We fail to
same strategy in both the large and the small lot and reject Hypothesis 1.
on average they bid lower on the small lot where
there is also less competition, which is in line with the Y2. The RNBNE strategy describes the empirical data in
equilibrium prediction. C100 and C100+ in the Yankee auction well. This
Discussion: Overall, the results from the computer- suggests that wrong expectations and the strategic
ized experiments C100+.P and C100.P confirm complexity of the Yankee auction have little impact on the
rational bidding behavior according to the RNBNE bidding behavior. We fail to reject Hypothesis 3 and 4.
model. Wrong expectations or strategic complexity do
impact bidding behavior significantly. Underbidding Y3. We found underbidding on low-cost draws in C1
in C1.P on low-cost draws can again be explained by and reject Hypothesis 2. The difference to C100 indicates
risk aversion. This underbidding is higher than in the that risk aversion has substantial impact on bidder
single-lot auction, in particular on the large lot. We behavior.
conjecture that bidders tried to win the large lot with
low prices, because it promised a higher total payoff Support: We provide the same statistics as for the
with 70 units. Risk aversion can also serve as one of parallel auction in Table 6. The MSE values are com-
the reasons for underbidding in H.P. Due to the 16 parable to the parallel auction. The lowest MSE values
repetitions risk aversion is mitigated, but not elimi- are again achieved for C100+ and C100. C1.Y has also
nated. Regret and spite might also play a role, but due a significantly lower intercept that can be attributed
to the experimental design, where bidders did not to risk aversion. In C1.Y, there is a clear outlier, a bid-
know their opponents, we assume these behavioral der who submitted very high bid functions leading to
conjectures to be less important. This is supported by a high MSE RNBNE of 3.372. Without this bidder the
comments of participants after the experiments MSE RNBNE for C1.Y is 1.024.
explaining their bidding strategy. Also in treatment The line H.Y in Table 6 describes the regression
H.P, bidders are aggressive on the large lot, and they coefficients of all the bid functions with the number of
started with a higher bid on the small lot in case of the auction as an additional covariate. As in the paral-
low cost draws. However, the bidding in the last lel auction, we find a low intercept a together with a
rounds also became aggressive on the small lot with higher b compared to the RNBNE function in the
lower bids on the low-cost draws. analysis of auctions 1, 7, and 16 in H. This means that,
also in the Yankee auction, bidders in H underbid on
4.3. The Yankee Auction low-cost draws compared to the RNBNE function,
In addition to the parallel auction, we also analyzed which we also observe in C100 and C100+.
how well equilibrium bidding strategies explain We have included the MSE for the single-item
bids in the different treatment combinations in the RNBNE and that of a constant profit factor model in
Yankee auction. The strategic complexity is higher Table 6. In the Yankee auction, the split-award
than in the parallel auction, because bidders do not RNBNE model had the lowest MSE in treatments
know if an aggressive bid will actually win the C100+.Y, and C100.Y. For H.Y and C.1 (w/o the out-
large lot, and if they do not win the large lot, they lier) the single-item RNBNE had a lower MSE, which
might win the small lot with a very low payoff. can again be explained by the fact that the single-item
Again, we provide results in Table 6 and scatter RNBNE model leads to higher competition with the
plots (Appendix S3). same number of bidders and lower equilibrium bid

Table 6 Regression Coefficients for the Empirical Bid Functions (w/o Bidder ID) of the Yankee Auction with q = 0.7

b (unit Std. error b MSE MSE MSE single-item MSE constant Number of
Α cost) (p-value) Mult. R² LOESS RNBNE RNBNE factor bids/bidder
RNBNE 3.53 0.647
H.Y 2.49 0.756 0.004 (0.000) 0.941 0.475 0.766 0.532 1.062 2550/16
H.Y (# 1) 2.68 0.724 0.018 (0.000) 0.930 0.907 0.989 1.297 1.623 160/16
H.Y (# 7) 2.19 0.743 0.012 (0.000) 0.963 0.317 0.573 0.360 0.941 160/16
H.Y (# 16) 1.92 0.800 0.011 (0.000) 0.975 0.355 0.914 0.418 0.892 160/16
C1.Y 2.05 0.783 0.028 (0.000) 0.860 3.191 3.372 3.531 3.647 110/11
C100.Y 3.21 0.684 0.007 (0.000) 0.973 0.192 0.213 0.396 1.039 260/13
C100+.Y 3.66 0.682 0.008 (0.000) 0.972 0.148 0.166 0.342 0.995 220/11

The numbers in brackets for C1.Y describe MSE values w/o one outlier.
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1025

price, which better fits the average behavior of risk- Support: First, the results S1-3, P1-3, and Y1-3 are
averse bidders. in line. In addition, to analyze the predictive accuracy
As an example, Figure S8 describes the bid func- of the RNBNE model across auction formats and split
tions in the first and seventh auctions of H. We parameters, we have pooled all observations (Parallel
provide the bid functions for other treatment combi- and Yankee auction) within each of the four treatment
nations in the Yankee auction in Figures S8, S9–S15 in combinations, as well as the observations for different
Appendix S3. split parameters (0.7 and 0.9). We have then used the
The Chow test shows that there is no significant dif- RNBNE prediction as right-hand side variable for the
ference between the RNBNE bid functions in C100+ bids in a regression. A coefficient on the RNBNE pre-
(p = 0.847) and C100 (p = 0.907), but there is a differ- diction close to 1 is strong evidence for the RNBNE
ence to H and C1 (p = 0.000). Underbidding below model, and it demonstrates that subjects understand
the RNBNE for low-cost draws of 1 Franc is on aver- the strategic differences across the auction sessions.
age 29.78% for treatment H (auction #7) and 32.18% Again, for C100+ and C100 we find evidence for the
for C1. This is less than in the parallel auction. RNBNE model, while the results in Table 7 indicate
Discussion: Bidders in the Yankee auction do not underbidding in the treatments C1 and H. Apart from
know a priori if they will win the large lot or the small the standard t-test for b = 0, we have also tested the
lot with their single-bid price. The high predictive null hypothesis of b = 1. The difference of the coeffi-
accuracy of the RNBNE function in C100 and C100+ cient b to a value of 1 is significant in all cases,
is, therefore, an interesting result. In particular, there although the b is already very close to 1 for C100 and
is no significant difference between C100 and C100+ C100+.
and no significant underbidding, indicating that stra-
tegic complexity had little impact even in the Yankee 4.5. Procurement Cost Comparisons
auction. In line with what we see in the single-lot and The final result of our theoretical analysis in sec-
in the parallel auction, we find significant underbid- tion 2.3 is that the expected costs of the parallel and
ding below the RNBNE bid function in C1, which can the Yankee auction are the same. In this subsection,
be explained by risk aversion. In the treatment combi- we report on allocative efficiency and a cost ratio
nation H, the level of underbidding on low-cost which normalizes the actual procurement costs by the
draws increases slightly across the 16 auctions in a costs of the bidders in the optimal solution. This
session, which can be explained by losing bidders in allows for comparison across different cost draws in
the initial rounds who bid more aggressively in later the auctions, because average procurement costs can
rounds. differ significantly due to the cost draws of the bid-
ders in individual auctions.
4.4. Predictive Accuracy across Auction Formats Allocative efficiency is computed as E ¼ ðc1 þ
Let us now summarize the results across all auction c þ 2KÞ=ðC1i þ c2j þ 2KÞ, where c1 and c2 are the vari-
2

formats, the single-lot, the parallel, and the Yankee able costs in the efficient allocation for the large and
auction. the small lot, and c1i and c2 are the costs of those
bidders i and j who won the auction. Cost ratio is
4.4.1. Results for The All Three Auctions. defined as C ¼ ðc1 þ c2 þ 2KÞ=ðb1j þ b2j , where b1i and
A1. There is underbidding in treatment H. We fail to b2j describe the winning bids by bidders i and j on the
reject Hypothesis 1. large and the small lot, respectively. If bid prices in
the winning allocation decrease, this ratio increases,
A2. The RNBNE strategy describes the empirical data
that is, a higher number is better for the buyer.
in C100 and C100+ in the single-lot auction well. We
fail to reject Hypothesis 3 and 4.
4.5.1. Result on Procurement Costs Across
Auctions.
A3. We found underbidding on low-cost draws in C1
C1. The parallel auction and the Yankee auction exhibit
and reject Hypothesis 2. The difference to C100 indicates
no significant differences in efficiency and cost ratio
that risk aversion has substantial impact on bidder
behavior. within the same treatment, as predicted by Proposition 5.

Table 7 Regression Coefficients of the RNBNE Prediction Across Auction Formats

Treatment a b (RNBNE) SE (b) p-value (b = 0) p-value (b = 1)

H
1.867 1.188 0.005 <2e-16 0.000
C1
2.151 1.268 0.052 <2e-16 0.000
C100
0.460 1.060 0.009 <2e-16 0.000
C100+ 0.139 0.980 0.006 <2e-16 0.004
 Bichler, Guler, and Mayer: Split-Award Procurement Auctions
1026 Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society

Table 8 Efficiency and Auctioneer’s Cost Ratio computerized experiments where bid functions are
Efficiency E (%) Cost ratio C (%)
reused in 100 auctions. This is different to earlier
experiments of first-price sealed-bid auctions, and
H.P 96.17 83.63
C1.P 96.62 71.32
we attribute the result to our experimental design.
C100.P 99.01 51.91 The ability to control risk aversion in the experiments
C100+.P 99.39 51.63 allows us to analyze whether bidders are able to
H.Y 96.49 79.62 cope with the strategic complexity in these markets.
C1.Y 99.24 67.48 The results provide evidence that bidders are well
C100.Y 98.50 51.95
C100+.Y 98.78 51.60
able to mimic their RNBNE bid function even in stra-
tegically complex Yankee split-award auctions. Over-
all, there is no evidence that the strategic complexity
or wrong expectations of these auctions explain
Support: Overall, efficiency is high in all experi- underbidding in human subjects experiments. How-
mental treatments. We did not find a significant dif- ever, the results show that risk aversion has substan-
ference in efficiency E or cost ratio C within the same tial impact on the bidder behavior and we conjecture
treatment combination between the parallel and the it is a major driver for the underbidding that we see
Yankee auction using a Wilcoxon rank sum test in human subject experiments. We find underbid-
(a = 0.01). ding for low-cost draws compared to the RNBNE bid
There are significant differences between the treat- function, which typically increases in the latter auc-
ment combination H and C100+ for both split parame- tion rounds.
ters (a = 0.01). H has lower efficiency and a higher Regret and risk aversion are two possible exten-
cost ratio, which means a lower cost, in both the paral- sions of our models and also the experimental work
lel and the Yankee auction. The lower cost in H can be in this area. It would also be interesting to analyze
attributed to the underbidding that we described ear- markets with very high fixed costs in the laboratory
lier in human subject experiments. So, in spite of dif- or explore more complex cost functions. However, we
ferences in the bid function and the underbidding in firstly wanted to find evidence that human subjects
the laboratory the outcomes of the auction are in line are able to mimic their equilibrium bid functions even
with the theoretical prediction. without this added complexity.

5. Conclusions Acknowledgments
In this study, we analyze two sealed-bid split-award
The authors gratefully acknowledge support of the Deut-
auctions which are regularly used in procurement sche Forschungsgemeinschaft (DFG) (BI 1057/7-1).
practice. We obtain closed-form symmetric Bayes
Nash equilibrium bidding strategies and several
interesting model implications. We prove that if sup- Notes
pliers are risk-neutral and the production cost infor- 1
We recently learned about a working paper by Kirchk-
mation is private, then the two auction mechanisms amp et al. (2014), who used an experimental design for
are equivalent, in the sense that they yield the same single-object auctions where the bid function was reused
expected costs to the buyer. Experimental work on in 50 auctions, but they could not eliminate, but only
the first-price sealed-bid auction has shown a consis- reduce overbidding. The experiments differed in a number
tent pattern of underbidding. This can be due to risk of ways including the feedback to bidders.
2
aversion, regret, and wrong expectations, but also Seemingly unrelated regression (SUR) is one possibility
due to the complexity of deriving the RNBNE bid to deal with these two sources of data. However, because
each equation contains exactly the same set of regressors,
function. It is not obvious that RNBNE strategies
the estimators of a SUR are numerically identical to ordin-
could be a good predictor for split-award procure-
ary least squares estimators, which follows from Kruskal’s
ment auctions, where bidders are exposed to theorem (Davidson and MacKinnon 1993).
increased strategic complexity compared to single-lot
reverse auctions.
We provide the results of laboratory experiments References
with different levels of control. The experiments Andreoni, J., Y.-K. Che, J. Kim. 2007. Asymmetric information
about rivals’ types in standard auctions: An experiment. Game
against computerized bidders are meant to limit the Econ. Behav. 59(2): 240–259.
influence of risk aversion and regret as far as possi- Anton, J. J., D. A. Yao. 1989. Split awards, procurement, and inno-
ble. Interestingly, there is no significant difference vation. RAND J. Econom. 20: 538–552.
between the average bid function of bidders in the Anton, J. J., D. A. Yao. 1992. Coordination in split award auctions.
laboratory and the RNBNE bid function in our Q. J. Econ. 107(2): 681–707.
Bichler, Guler, and Mayer: Split-Award Procurement Auctions
Production and Operations Management 24(6), pp. 1012–1027, © 2014 Production and Operations Management Society 1027

Anton, J. J., S. Brusco, G. Lopomo. 2010. Split-award procurement Isaac, R. M., D. James. 2000. Just who are you calling risk averse?
auctions with uncertain scale economies: Theory and data. J. Risk Uncertainty 20(2): 177–187.
Games Econ. Behav. 69(1): 24–41. Kirchkamp, O., J. P. Reiss. 2006. Another Explanation for Overbid-
Armstrong, M. 2000. Optimal multi-object auctions. Rev. Econ. ding and another Bias for Underbidding in First-Price Auctions.
Stud. 67(3): 455–481. University of St Andrews, Scotland.
Ausubel, L. M., O. V. Baranov. 2010. Core-selecting auctions with Kirchkamp, O., J. P. Reiss. 2011. Out-of equilibrium bids in auc-
incomplete information. University of Maryland. tions: Wrong expectations or wrong bids. Econ. J. 121(557):
Bourdieu, P. 2005. The Social Structures of The Economy. Polity, 1361–1397.
Cambridge, UK. Kirchkamp, O., J. P. Reiss, K. Sadrieh. 2014. A pure variation of
Chen, K.-Y., C. R. Plott. 1998. Nonlinear behavior in sealed bid risk in private-value auctions. Working paper, Karlsruhe
first price auctions. Game Econ. Behav. 25(1): 34–78. Institute of Technology. Downloaded in June 2014.
Cleveland, W., S. Devlin. 1988. Locally-weighted regression: An Krahnen, J. P., C. Rieck, E. Theissen. 1997. Inferring risk attitudes
approach to regression analysis by local fitting. J. Am. Stat. from certainty equivalents: Some lessons from an experimen-
Assoc. 83(403): 596–610. tal study. J. Econ. Psychol. 18(5): 469–486.
Cooper, D. J., H. Fang. 2008. Understanding overbidding in sec- Krishna, V. 2009. Auction Theory. Academic Press, San Diego, CA.
ond price auctions: An experimental study*. Econ. J. 118(532): Lebrun, B. 2006. Uniqueness of the equilibrium in first-price auc-
1572–1595. tions. Games Econ. Behav. 55: 131–151.
Corey, E. R. 1978. Procurement Management: Strategy, Organization, MacCrimmon, K. R., D. A. Wehrung. 1990. Characteristics of risk
and Decision-Making. CBI Publishing Company, Boston, MA. taking executives. Manage. Sci. 36(4): 422–435.
Cox, J., V. Smith, J. Walker. 1988. Theory and individual behavior Nell, E. J. 2007. Rational Economic Man. Cambridge University
of first-price auctions. J. Risk Uncertainty 1(1): 61–99. Press, Cambridge, UK.
Cramton, P. 2013. Spectrum auction design. Rev. Ind. Organ. 42(2): Payne, J. W., D. J. Laughhunn, R. Crum. 1980. Translation of gam-
161–190. bles and aspiration level effects in risky choice behavior. Man-
Cramton, P., Y. Shoham, R. Steinberg. 2006. Combinatorial auc- age. Sci. 26(10): 1039–1060.
tions. MIT Press. [Link] Perry, M. K., J. Sákovics. 2003. Auctions for split-award contracts.
auctions J. Indust. Econom. 51(2): 215–242.
Davidson, R., J. G. MacKinnon. 1993. Estimation and inference in Pezanis-Christou, P., A. Sadrieh. 2003. Elicited bid functions in (a)
econometrics. OUP Catalogue. symmetric first-price auctions. CentER Discussion paper, Til-
Dohmen, T., A. Falk, D. Huffman, U. Sunde, J. Schupp, G. Wagner. 2005. burg University.
Individual risk attitudes: New evidence from a large, representative, Roth, A. E., J. H. Kagel. 1995. The Handbook of Experimental Eco-
experimentally-validated survey. No. 1730. IZA Discussion Papers. nomics. Princeton University Press, Princeton, NJ.
Elmaghraby, W. J. 2000. Supply contract competition and sourcing Sano, R. 2012. Non-bidding equilibrium in an ascending core-
policies. Manuf. Serv. Oper. Manag. 2(4): 350–371. selecting auction. Games Econ. Behav. 74(2): 637–650.
Engelbrecht-Wiggans, R. 1988. Revenue equivalence in multi- Schoemaker, P. J. 1993. Determinants of risk-taking: Behavioral
object auctions. Econ. Lett. 26(1): 15–19. and economic views. J. Risk Uncertainty 6(1): 49–73.
Engelbrecht-Wiggans, R. 1989. The effect of regret on optimal bid- Selten, R., J. Buchta. 1999. Experimental sealed bid first price auc-
ding in auctions. Manage. Sci. 35(6): 685–692. tions with directly observed bid functions. Games and human
Engelbrecht-Wiggans, R., E. Katok. 2009. A direct test of risk aver- behavior: Essays in honor of Amnon Rapoport, 101–116.
sion and regret in first price sealed-bid auctions. Decis. Anal. Seshadri, S., K. Chatterjee, G. L. Lilien. 1991. Multiple source pro-
6(2): 75–86. curement competitions. Mark. Sci. 10(3): 246–263.
Filiz-Ozbay, E., E. Y. Ozbay. 2007. Auctions with anticipated regret: Stahl, D. O., P. W. Wilson. 1995. On players0 models of other players: The-
Theory and experiment. Am. Econ. Rev. 97(4): 1407–1418. ory and experimental evidence. Games Econ. Behav. 10(1): 218–254.
Gigerenzer, G., U. Hoffrage. 1995. How to improve Bayesian reasoning Thomas, C. J. 1996. Market Structure and the Flow of Information in
without instruction: Frequency formats. Psychol. Rev. 102(4): 684. Repeated Auctions. Bureau of Economics, Federal Trade Com-
Goeree, J. K., Y. Lien. 2009. On The Impossibility of Core-Selecting mission, Washington, DC.
Auctions. Institute for Empirical Research in Economics, Uni- Tunca, T. I., Q. Wu. 2009. Multiple sourcing and procurement pro-
versity of Zurich, Zurich. cess selection with bidding events. Manage. Sci. 55(5): 763–780.
Goeree, J. K., C. A. Holt, T. R. Palfrey. 2002. Quantal response Woodside, A. G., N. Vyas. 1987. Industrial Purchasing Strategies:
equilibrium and overbidding in private-value auctions. J. Recommendations for Purchasing and Marketing Managers. Lex-
Econ. Theory 104(1): 247–272. ington books, Lanham, Maryland.
Gong, J., J. Li, R. P. McAfee. 2012. Split-award contracts with
investment. J. Public Econ. 96(1): 188–197.
Supporting Information
Gurnani, H., M. Gumus, S. Ray, T. Ray. 2012. Optimal procurement strat-
Additional Supporting Information may be found in the
egy under supply risk. Asia Pacific J. Oper. Res. 29(1): 132–165.
online version of this article:
G€
uth, W., R. Ivanova-Stenzel, M. K€ onigstein, M. Strobel. 2003.
Learning to bid–an experimental study of bid function adjust-
ments in auctions and fair division games*. Econ. J. 113(487): Appendix S1: Proofs.
477–494. Appendix S2: Experimental Instructions.
Holt, C. A., S. K. Laury. 2005. Risk aversion and incentive effects: Appendix S3: Scatter Plots.
New data without order effects. Am. Econ. Rev. 95(3): 902–904. Appendix S4: Experiments with a Split Parameter of 0.9.