Estimating the First-Price Auction Model with Entry: A Parametric Approach * Shinichi Sakata; Pai Xu † April 5, 2010 Preliminary. Please do not cite. Abstract This paper studies the structural estimation of first-price auctions with endogenous entry. Most of the empirical auction literature to date is based on the theoretical work of Levin and Smith (1994, LS hereafter) in which the potential bidders make their entry decisions before knowing their own valuations. This paper, instead, assumes that a potential bidder already knows his own valuation when deciding to pay an entry cost and submit a bid. Such an assump- tion reflects the reality in certain important auctions. We propose estimating the distribution of bidders’ private valuations in the entry model by using the simulated method of moments. We implement the proposed estimator on a dataset of highway procurement auctions in Oklahoma. * We are grateful to Han Hong, Quang Vuong, Harry Paarsch, Vadim Marmer, Brian Krauth and Unjy Song for useful comments, as well as the seminar participants at the University of Hong Kong. We also thank Kun Li for his excellent research assistance. The second author gratefully acknowledges the financial support of HKU under the seed funding scheme for basic research. † Sakata: Department of Economics, University of British Columbia; [email protected]. Xu: School of Economics and Finance, University of Hong Kong, [email protected].
33
Embed
Estimating the First-Price Auction Model with Entry: A ...stevexu/papers/estimate_Sentry_Sakata_Xu.pdf · Estimating the First-Price Auction Model with Entry: A Parametric Approach
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Estimating the First-Price Auction Model with Entry:
A Parametric Approach∗
Shinichi Sakata; Pai Xu†
April 5, 2010
Preliminary. Please do not cite.
Abstract
This paper studies the structural estimation of first-price auctions with endogenous entry.
Most of the empirical auction literature to date is based on the theoretical work of Levin and
Smith (1994, LS hereafter) in which the potential bidders make their entry decisions before
knowing their own valuations. This paper, instead, assumes that a potential bidder already
knows his own valuation when deciding to pay an entry cost and submit a bid. Such an assump-
tion reflects the reality in certain important auctions.
We propose estimating the distribution of bidders’ private valuations in the entry model by
using the simulated method of moments. We implement the proposed estimator on a dataset of
highway procurement auctions in Oklahoma.
∗We are grateful to Han Hong, Quang Vuong, Harry Paarsch, Vadim Marmer, Brian Krauth and Unjy Song for
useful comments, as well as the seminar participants at the University of Hong Kong. We also thank Kun Li for his
excellent research assistance. The second author gratefully acknowledges the financial support of HKU under the
seed funding scheme for basic research.†Sakata: Department of Economics, University of British Columbia; [email protected]. Xu: School of
where the third equality follows by the fact that ζ(vρ) = 0; and Y(1) is the highest of N − 1 random
draws from F . The above equality can be further rewritten as
F (v)N−1 (E[maxY(1), vρ |Y(1) < v]−B(v)) = κ. (2.5)
(2.5) provides an intuitive interpretation of equilibrium bidding strategy in the S entry model. It
says that the bidder with value v chooses the bid level at which his best “expected” competitor’s
profit would be nonpositive, if the competitor outbids him. (2.5) is the counterpart of the well-
known Revenue Equivalence Theorem for the auctions in a standard setup in which there are no
endogenous entries associated with participation costs of any format.
Though a potential bidder with v < vρ does not bid, we artificially assign r to his bid value for
convenience in our econometric analysis, so that B(v) = r whenever v < vρ. By rewriting (2.3) and
5
using r as the bids of non-bidders, we obtain that
B(v) =1(v ≥ vρ)F (v)−(N−1)
(∫1(y ≤ v) maxy, vρ (N − 1)F (y)N−2 F (dy)− κ
)+ 1(v < vρ) r
=m(v)
≡∫ (
1(v ≥ vρ)F (v)N−1
(1(y ≤ v) maxy, vρ(N − 1)F (y)N−2 − κ
)+ 1(v < vρ) r
)F (dy). (2.6)
The jth moment of the bid is therefore equal to∫m(v)j F (dv). Such construction provides a basis
and convenience for our design of estimators in the entry context.
2.2 Identification
In general, the observables in our current model are the bids, the number of potential bidders
and the number of actual bidders, while the private values and their distributions are not observed.
The identification problem here amounts to a discussion on whether the observed variables can
uniquely determine both the latent distribution F and the participation cost κ.1
Following GPV (2000)[7], the inverse equilibrium bidding function can be written as
v = ε(b∗) ≡ b∗ +1− p+ pG∗(b∗)(N − 1)pg∗(b∗)
(2.7)
where b∗ is observed bids. G∗ and g∗ are their CDF and pdf, respectively. And the rate of
participation p = 1− F (vρ). (2.7) implies that if both p and G∗ (therefore, g∗) are identified, then
so is ε for v ≥ vρ.
Furthermore, as F (vρ) = 1−p and vρ is identified (from 2.7), then (2.1) implies that κ is identified.
It is clear from above argument that the identification hinges on the facts that p and vρ are both
identified. However, p is always identified, as we assume econometricians can observe both the
number of potential bidders and the number of actual bidders.1We provide a heuristic argument on the identification of the S model. More thorough discussion can be found in
Xu (2009)[20], Marmer, Shneyerov and Xu (2010)[15].
6
3 Estimation of the Auction Model with Entry Costs
3.1 An Econometric Model
This paper considers a parametric framework for the power of counterfactual analysis as outlined
in the introduction. We therefore assume that a parametric specification for the value distribution is
given. We also allow for heterogeneity of auctions through the dependence of the value distribution
and the entry cost on a k × 1 vector consisting of exogenous variables. In a formal format, the
following assumption summarizes the elements of an econometric model in the auction context.
Assumption 1 The parameter space Θ is a nonempty compact subset of Rp, and F : R×Rk×Θ→
R is a function measurable B ⊗ Bk ⊗ B(Θ)/B that satisfies that for each z ∈ Rk and θ ∈ Θ,
F (·, z, θ) : R → R is a cumulative distribution function with a pdf f with respect to the Lebesgue
measure, and for each (v, z) ∈ R × Rk, F (v, z, ·) : Θ → R and f(v, z, ·) : Θ → R are continuous.
Also, κ : Rk×Θ→ [0,∞) is a function measurable Bk⊗B(Θ)/B([0,∞)) such that for each z ∈ Rk,
κ(z, ·) : Θ→ R is continuous.
The value distribution of an auction with the exogenous variables z is modeled as F (·, z, θ), where
θ ranges over Θ.
In terms of data generating process, we assume that the number of potential bidders (N), all of
the bids (B), the exogenous variables (Z), and the reserved price (R) for each auction, where the
bids are given through the mapping m of the unobserved values (V ) in (2.6) suitably modified to
allow for the heterogeneity of auctions.
Assumption 2 (i) On a probability space (Ω,F, P ), (Nt, Rt, Zt, Vt) : Ω → N × R × Rk ×
R∞t∈N is an independently and identically distributed (i.i.d.) process such that there uniquely
exists θ0 ∈ Θ such that each element of v1 are independently and identically distributed with
F (·, Z1, θ0) conditionally given N1, R1, and Z1.
(ii) For each θ ∈ Θ, F (R1, Z1, θ) > 0 a.s.-P .
Remarks.
(i) In Assumption 2, the conditions imposed on the first observation (indicated by subscript 1)
applies to all observations, because (Nt, Rt, Zt, Vt)t∈N is i.i.d.
7
(ii) Despite that in each auction t, we only observe Nt bids, which are mapped from vt1, . . . , vtNt ,
the random vector vt consists of infinitely many elements in Assumption 2. This is merely
for technical convenience, and it imposes no restrictions on the applicability of our approach.
The special feature of endogenous entry in our auction model introduces a threshold (vρ) on the
valuation distribution F . We define it in the econometric model accordingly. vρ : R×Rk×N×Θ→ R
such that
vρ(r, z, n, θ) ≡ infv ∈ R : (v − r)F (v, z, θ)n−1 − κ(z, θ) ≥ 0
Because v 7→ (v − r)F (v, z, θ)n−1 − κ(z, θ) is nondecreasing and continuous, and its range con-
tains [0,∞), vρ(r, z, n, θ) is the minimum zero of v 7→ (v − r)F (v, z, θ)n−1 − κ(z, θ). Under As-
sumption 2(ii), vρ(R1, Z1, N1, θ) is a singleton a.s.-P for each θ ∈ Θ, so that for each θ ∈ Θ,
vρ(R1, Z1, N1, θ) is a unique zero of v 7→ (v −R1)F (v, Z1, θ)n−1 − κ(Z1, θ) a.s.-P .
For purpose of structural analysis, the observed bids are assumed to be generated following the
symmetric equilibrium strategy. Define m : R× R× Rk × N×Θ→ R by
As suggested by the second expectation in (3.13), there requires no bias correction by using simu-
lated moments in the objective function, as long as the simulation error is orthogonal to h1(Zt).
In general, h1 needs to be chosen as the gradient of l1 to attain the efficiency for GMM estimation.
(3.13) further implies that, regarding the choice (of estimation) of h1(Zt), the simulation method
may also be applied, provided that the simulation errors from simulating moments of bids and
from simulating h1 are uncorrelated. However, in this paper, we have to avoid the choice of such
an efficient estimator, as l1 is non-smooth in parameters, therefore may be non-differentiable.
3.3 Large-Sample Properties
To ease the exhibition, we first define the population counterpart of GS,T as G(θ) = J(θ)′WJ(θ),
where
J(θ) = E
Bti − l1(Zt, θ)h1(Zt)
B2ti − l2(Zt, θ)h2(Zt)
· · ·
Bkti − lk(Zt, θ)hk(Zt)
1(Bti = r)− F (vρ(Zt; θ))hρ(Zt)
.
To ensure the desired asymptotic properties of our simulated GMM estimator, we assume the
following regularity conditions:
Assumption 4 (i) W is positive semi-definite and WJ(θ) 6= 0 for θ 6= θ0.
(ii) v vρ(Z, θ) v, ∀Z, θ.
(iii) fz the marginal density of Z is bounded.
11
(iv) FzZ : F−1(., Z, θ)− vρ(Z, θ) = 0 = 0
(v) S1 ... Sk are fixed natural numbers no less than 2 and independent of t.
As G is continuous on the compact parameter set, assumption 4(i) requires there be only one
minimizer of G. This by and large is a typical identification condition for the GMM type estimators.
4(ii) states that the truncation introduced by the entry cost is bounded away from the supports
of the valuation distribution. This assumption is generally assumed in the auction setup with
endogenous entry, as otherwise the entry issue will become trivial (either full entry or no entry).
Assumption 4 (iii) and (iv) are not restrictive in the sense that the boundedness and continuity of
covariates are usually assumed in the auction theory. The last item in assumption 4 reinforces that
our asymptotic analysis is for a fixed number of simulations as T →∞.
The following theorem states the consistency property of θ.
Theorem 1 Given Assumptions 1 to 4, θ converges in probability to θ0 as T →∞.
We further impose the following assumption for the asymptotic distribution.
Assumption 5 (i) θ0 is an interior point of Θ.
(ii) ∀Z, ∂F (., Z, θ)/∂θ exsits and is bounded ∀θ ∈ Θ.
(iii) ∀Z, ∂f(., Z, θ)/∂θ exists and is bounded ∀θ ∈ Θ.
(iv) (Zt, θ) has a density that is bounded above uniformly over θ ∈ Θ.
Assumption 5 (ii) and (iii) impose conditions on f and F over the parameter set Θ. These
restrictions can be satisfied if f is chosen from the commonly used distributions. Especially, in our
empirical applications and all of other empirical auction works with parametric approaches, the
private values are usually specified to follow lognormal distribution. Then these assumptions hold.
The last item of assumption 5 asks the boundedness of joint density of covaiates and parameters.
This is not generally restrictive in applications.
12
We now present the following theorem, which states the asymptotic normality for θ.
Theorem 2 Given Assumptions 1 to 5,√T (θ−θ0)→ N(0,Σ), where Σ = (′W)−1′WΩW(′W)−1
and = E[∇θJ(Z, θ0)], and Ω = E[J(Z, θ0)J(Z, θ0)′].
To use the asymptotic normality of θ to conduct inference on θ0, it is necessary to estimate
the asymptotic covariance matrix Σ. This matrix depends on , Ω and W . Following the same
reasoning and spirit, we propose consistent estimators of and Ω. We postpone the discussion of
estimating W later for the efficiency of simulated GMM estimation. For now, we simply assume
we have a consistent estimator of W , W .
Our proposed estimator that relies on simulating first order derivatives and that is also consistent
for a fixed number of simulations as the number of auctions increases to infinity. We define Y [ι]t =
∂X[ι]t (θ)/∂θ and Fθ = ∂F (v;θ)
∂θ . Then,
(θ) = − 1T
T∑t=1
Nt∑i=1
Y
[1]t (θ)h1(Zt)
· · ·
Y[k]t (θ)hk(Zt)
Fθ(θ))hρ(Zt)
.
and
Ω(θ) =1T
T∑t=1
Jt(θ)Jt(θ)′
where
Jt(θ) =
Bti − X [1]
t (u[1]t,1, ..., u
[1]t,S1
, Zt, θ)h1(Zt)
· · ·
Bkti − X
[k]t (u[k]
t,1, ..., u[k]t,Sk
, Zt, θ)hk(Zt)
1(Bti = r)− F (vρ(Zt; θ))hρ(Zt)
.
Therefore, Σ = (′W )−1′W ΩW (′W )−1. The following theorem states that Σ is a consistent
estimator of Σ.
Theorem 3 Σ converges to Σ in probability as T →∞.
13
A practical advantage of the proposed Σ is that it does not require the number of simulations to
be large, as otherwise the computation burden may become incredibly heavy due to the complex
of bidding strategy. Moreover, Ω can typically be obtained from the last iteration of minimization
routines with which θ was obtained.
3.4 Comparison with LOV (1995)
LOV (1995) [11] considers a first-price auction model without endogenous entry. In there, the
value distribution is truncated by the presence of a binding reserve price. Conditional on partici-
pation, the Revenue Equivalence Theorem implies:
b = ε(v,N, F ) = E[max(v(2), r)|v(1) = v,N, F ] (3.14)
where v(1) and v(2) are the largest and second largest order statistic among v1, v2, ..., vN . Considering
only winning bids and extending to R, (3.14) gives
bw = E[max(v(2), r)|v(1) = v,N, F ]1[v ≥ vρ] + r · 1[v < vρ]
= E[max(v(2), r)|v(1) = v,N, F ]. (3.15)
Then, taking the expectation with respect to v, (3.15) gives
E[bw|N,F ] = E[max(v(2), r)|N,F ]
which can be written as an integral with respect to f :
E(bw|N,F ) =∫ v
vmaxu(2), rf(u1)...f(uN )du1...duN . (3.16)
LOV propose a simulated non-linear least squares estimator for the parameters of F based on the
moment condition (3.16). Though our work is related to LOV, there are subtle differences between
the two:
(i) Bidding strategies. (2.4) vs. (3.14).
There is no endogenous entry issue in LOV’s model. They assume that the bidders know
the number of competitors they are actually facing when submitting the bids. Then, r
just serves as the lower bound for the truncated value distribution. Therefore, the revenue
14
equivalence theorem takes the regular form as (3.14). However, in our model, the revenue
equivalence theorem implies a different form as (2.4), which reflects the free entry condition
in the equilibrium of the model.
(ii) Asymptotics. The objective function maximized by LOV’s estimator, which is based on
(3.16), is smooth, as long as F is smooth in parameters. In our model, on the other hand,
the truncation point corresponding to r in (3.16) is vρ, which is one of the parameters to
be estimated. Our estimation objective function is therefore non-smooth in the parameters.
This feature makes the asymptotic analysis of our estimator more involved.
(iii) Estimation strategy. The model primitive in LOV is the parameters that characterize the
distribution of valuations. In their specification of lognormal distribution, let µ and σ2 be the
mean and variance, respectively. LOV suggest a SNLLS estimator to estimate the parameters.
However, in essence, their LS estimator only matches the first moment of observed (winning)
bids with data. µ and σ are not jointly identified. Therefore, in practice, LOV use the market
data to fix σ first, then apply their SNLLS estimator to recover µ.
Besides the distributional parameters µ and σ, we also have a truncation parameter κ (or,
equivalently, vρ) to estimate in our entry context. We propose a simulated-GMM estimator
for estimating both the distributional and truncation parameters, instead of extending LOV’s
SNLLS. By doing so, we are able to recover both µ and σ jointly.
(iv) Simulation method. LOV proposed using the importance sampling to simulate the first
moment of bids, while we provide a general approach to simulate higher moments by extending
LOV’s method.
An alternative way to estimate the model is to extend LOV’s SNLLS. Focusing on the first
moment and suppressing the covariates, a SNLLS estimator θ, for any fixed number of simulations
S, minimizes
Q1S,T (θ) =
1T
T∑t=1
Nt∑i=1
[(Bti − Xt(θ)
)2 − 1S(S − 1)
S∑s=1
(Xt(θ)− Xt(θ)
)2].
We next compare our simulated GMM method with simulated SNLLS by LOV up to front. Such
comparison allows us to see more clearly the advantage of using our approach in practice.
15
(i) Bias correction. The second term in bracket is the bias correction term, by incorporating
which enables Q1S,T converges in probability to Q1
T . To see this point, we notice that
E[Bti − Xt(θ)2]
= E[Bti − l1t(θ) + l1t(θ)− Xt(θ)2]
= E[Bti − l1t(θ)2] + E[l1t(θ)− Xt(θ)2] (3.17)
E[l1t(θ) − Xt(θ)2] in (3.17) corresponds to the bias correction term in Q1S,T . The cross
product term disappears due to the fact that B is always uncorrelated with the simulation
errors.
The bias correction is needed as SNLLS takes the expectation of squares. Therefore, the bias
can always be formulated as the variance of simulation errors. Instead, the objective function
of Simulated GMM takes the form of squares of the moments. Then no bias is occurred in
the objective function as long as the moments are unbiasedly simulated.
(ii) Computation. If we consider higher moments of bids, the simulation and bias correction
required in NLLS approach are demanding. The computation in NLLS is more involved than
GMM.
(iii) the objective function approximated in LOV is unbiased but bears a huge variance when F
puts a large probability mass in the area where G puts a low probability mass.
4 Empirical Application
Our dataset consists of 277 auctions for surface paving and grading contracts let by Oklahoma
Department of Transportation (ODoT) during the period of January 2002 to December 2005.2 The
available data items include all bids, the engineer’s estimate, the time length of the contract (in
days), the number of items in the proposal and the length of the road. The ODoT implements a
policy under with all bids over 7% of the engineer’s estimate are typically rejected, so there is a2Our choice of surface paving and grading contracts is motivated by the fact that Hong and Shum (2002)[8],
in their study of New Jersey highway procurement auctions, find little support for common values for this type of
contracts. This is important because in this paper, we assume the independent private value framework.
16
binding reserve price. In reality, we do observe bids above the reserve price (although extremely
few winning bids were above the reserve price). We treat these bids as non-serious.
Importantly, we observe the list of eligible bidders (planholders) for each auction. In the vo-
cabulary of this paper, these eligible bidders are the potential bidders. the list of planholders is
published on the ODoT website prior to bidding. A firm becomes a planholder through the follow-
ing process. All projects to be auctioned are advertised by the ODoT 4 to 10 weeks prior to the
letting date. These advertisements include the engineer’s estimate, a brief summary of the project,
location of the work and the type of the work involved.
Interested firms can then submit a request for plans and bidding proposals, the documents that
contain the specifics of the project (in particular, the items schedule). An important feature of
the qualification process is that only eligible firms are allowed the access to these documents. A
firm is deemed eligible if it satisfies certain qualification requirements. The goal of the qualification
process is to ensure that the winning firm will have sufficient expertise and capacity to undertake
the project. While the expertise part is typically determined at the pre-qualification stage, the
capacity part is project-specific. An important requirement is that the prospective bidder is not
qualified for the aggregate amount of work that exceeds 2.5 times its current working capital. Given
that the bidders know the sizes of all projects to be let but not the project specifics, it is plausible
that the decision of a firm to request the plan for a particular project is primarily determined by
the project size as well as the sizes of other projects for which it is pre-qualified, in relation to the
available capacity of the firm. The capacity may be determined by a number of factors, such as for
example the amount of resources committed to other projects, including but not limited to those
previously contracted with ODoT.
Before turning to our parametric specification, we investigate the importance of various observ-
able covariates on bid levels and the decisions to submit a bid with the help of usual OLS and logit
regression. The description of the variables used in the regressions and their summary statistics
are reported in Table 1. The results of the OLS bidding regressions and the entry logit regression
are presented in Tables 2 and 3.
17
In the OLS regression, the dependent variable is logarithm of bid, where bid is in millions of
dollars. The size of the project (EngEst) has a strong and positive effect on the bids. Clearly, it
is the most important variable in the OLS regression, so the impact of the other variables is much
smaller. In the order of importance, the next variable is the number of business days needed for
the completion of project (Ndays). Though small, its impact is statistically significant to explain
the variation in bids. We also mention that the project size has a negative (but not statistically
significant) effect on the probability of submitting a bid.
The complexity of the project is captured by the number of items in the construction plan
(Nitems). Table 1 shows that there is a substantial variation in Nitems, with a mean of 49 and
standard deviation of 58 items. One may expect that the cost of preparing a bid is an increasing
function of Nitems. The conjectured effect of Nitems is therefore to reduce the probability of
submitting a bid. The estimate of the Nitems coefficient in the logit regression confirms this
conjecture. Increasing Nitems by 1% at mean level, reduces the odds of bidding by 55%.
A discussion of the variable NPotential, the number of potential bidder, is in demand. Its effect
is statistically significant in both regressions. Having more potential bidders reduces the odds of
submitting a bid, and also results in lower bids. Whether such observed effects are consistent
with the auction model considered in the paper is yet beyond the scope of the paper.3 Instead,
this paper focuses on the participation and bidding behaviors for auctions with same number of
potential bidders. This effectively treats NPotential as an exogenous parameter which simply
defines the bidding environment as in the common knowledge of bidders. By doing so, we avoid the
strong assumption of the exogenous variation in NPotential. Table 4 shows the empirical frequencies
and sample sizes (observed bids) for each N. As a practical matter of implementing our estimator,
it is desirable that our estimation results and interpretation on counterfactual experiments are
comparable across the auctions with different number of potential bidders. Therefore, we choose
to focus our data analysis on the auctions with N = 3, 4, 5, 6.3Interested readers may refer to Li and Zheng (2009) [13] and Marmer, Shneyerov and Xu (2010) [15] for the
comparative studies on the equilibrium bidding with respect to the number of potential bidder.
18
We restrict our empirical analysis in Log-Normal distribution family, that is, F ∼ LogN(µ, σ).
Based on the previous preliminary data analysis, we assume the following functional form specifi-
cations:
µ = θ1 + θ2 · Log(EngEst) + θ3 · Log(Ndays)
σ = θ4 + θ5 · Log(EngEst) + θ6 · Log(Ndays)
κ = θ7 + θ8 · Log(EngEst) + θ9 · Log(Nitems)
The estimation results are presented in Table 5. The EngEst is significant in determining all the
parameters. It has positive impacts on (µ, σ). In other words, the larger the project, the higher
value of moments for bids. However, the larger project reduces the entry cost κ, which may suggest
an economy of scale in the bidding preparation process. Unlike the results in preliminary analysis,
we found that Ndays has no statistically significant impact on the bids. The effect of Nitems on
the bids is interesting. Only for auctions with 6 potential bidders, we found a significant impact of
Nitems. We suspect that for auctions with few potential participants, say 2 or 3, they are large and
complicated in nature already. Therefore, the impact of adding more bidding items does not change
the entry behavior of bidders. Instead, for auctions with certain smaller scale, which could attract
a few bidders in the first place, getting more bidding items into the plan significantly increases the
bidding costs.
5 Potential counterfactual (policy) analysis
There is a small literature that examines the cost of government intervention in the procurement
markets. In particular, these works have been focusing on the small business programs, in which
governments favor the firms with certain characteristics in the procurement process. Looking at
federal dredging contracts, Denes (1997) [6] estimates the effects on the winning bids when some
contracts are set aside for small businesses. Marion (2007) [14] studies the cost effect of the
bid preference program for small businesses in California highway procurement auctions. Both
of these works find that ”assessments of intervention to the programs for disadvantaged firms
(minorities or women) in contracting deserve special attention.” Discrimination by contractors, or
the vulnerability of these programs to fraud through the formation of nonminority front companies,
could weaken or reverse the estimated effect on the procurement costs. For example, Blanchflower
19
and Wainwright (2005) [4] suggest the existence of front companies as a possible reason, when they
find little empirical evidence supporting the impact of the preference programs on minority self-
employment. These messages suggest a great difficulty in studying the preference programs through
a reduced-form approach, when directly analyzing the data observations under the program.
Structural assessment of the bid preferential program has been studied by Krasnokutskaya and
Seim (2006) [10]. They directly investigate the effect of the preference program in the highway
procurement auctions in California. A critical problem they have encountered in their study is
multiplicity of the equilibrium. When bidders are treated asymmetrically, the participation and
bidding equilibrium in the auction model is not unique. Through simulation experiments, they
statistically find one equilibrium which might generate the data.
This paper provides an alternative approach to look at the bid preference program. In a sub-
sample of auctions, the involved bidding firms in Oklahoma can be regarded as symmetric bidders.
Then the unique entry and bidding equilibrium allows one to use the proposed estimation method-
ology to recover bidders’ valuation distribution and entry costs. At the end, one can use these
estimated model elements to conduct policy experiments. In particular, it will make it possible
to predict how the bid preference program may affect the bidding outcome, i.e., the market price
(winning bids). In the policy experiments, the preferential treatment is reflected by the differ-
ent truncation parameters for bidders. By doing so, one no longer face the problem of multiple
equilibriums, which is a fundamental and difficult issue to handle in structural analysis.
The counterfactual exercise in this case amounts to focusing on how to implement asymmetric
auctions towards the optimal outcome. The possible improvement of asymmetric equilibrium on
the auction outcomes has been documented by Celik and Yilankaya (2009) [5] in economic theory.
The researcher can compute the actual improvement a state could have made if the state had been
effectively imposing entry barriers for different groups of bidders.
20
6 Conclusion and extensions
A seminal work by LOV provides a new parametric estimation strategy for analyzing auction data.
They propose a simulated NLLS estimator to approximate the bidders private-value distributions.
Their work greatly broaden the class of distributions that empirical researchers can handle when
analyzing auction data sets.
This paper can be viewed as an extension to LOV’s work and complement to the methodological
literature on the auction data analysis. We propose a simulated-GMM estimation method to a first-
price auction model with endogenous entry. The particular entry pattern in such models makes
LOV’s objective function no longer twice continuously differentiable. The loss of smoothness on the
statistical objective demands another investigation on the asymptotic properties for simulated-based
estimators. We derive a set of sufficient conditions for the consistency and asymptotic normality
to remain valid. We also apply our estimation method to Oklahoma highway procurement auction
data to estimate the model parameters.
The proposed estimation method involves simulating the moments of the observed bids. It is a
vital role that the number of random draws S plays in the simulation. Specifically, S appears in
the variance-covariance matrix. Therefore, it is not a trivial question how the choice of S is going
to affect the estimation accuracy in the finite sample. We leave this for the future research.
Our estimation framework is quite general and the empirical findings are by and large intuitive.
But there is also an important limitation that the future research should address. In auction
datasets, one typically finds that the variation in bids is only partially explained by their variation
within auctions. The between-auction variation is typically present. It is also observed in our
dataset. We deal with this issue by pooling auctions with same number of potential bidders together
and controlling the observed auction heterogeneity. Though our appraoch reasonable and intuitive,
the observed auction heterogeneity may also present and be important in explaining the bids
variation, within IPV paradigm. Recently, krasnokutskaya (2010) [9] has developed a structural
estimation method that can be applied even in the presence of unobserved auction heterogeneity.
21
Another extension would be to allow bidder asymmetries. The obvious difficulty here would be
the necessity to deal with multiple equilibria. Bajari, Hong and Ryan (2009) [3] obtain a number
of identification results in this direction and estimate a parametric model with multiple equilibria
for highway procurement auctions. Finally, incorporating dynamic feature as in Pesendorfer and
Jofre-Bonet (2003) [18] is also left for future research.
22
References
[1] Andrews D.W.K. Consistency in Nonlinear Econometric Models: a Generic Uniform Law of
Large Numbers. Econometrica, 55:1465–1471, 1987.
[2] Andrews D.W.K. Empirical Process Methods in Econometrics. Handbook of Econometrics,
IV, 1994.
[3] Bajari, P. and H. Hong and S. Ryan. Identification and Estimation of a Discrete Game of