WPS4855 P olicy R eseaRch W oRking P aPeR 4855 Bidders' Entry and Auctioneer's Rejection Applying a Double Selection Model to Road Procurement Auctions Antonio Estache Atsushi Iimi The World Bank Finance, Economics and Urban Development Department Economics Team March 2009 Policy ReseaRch WoRking PaPeR 4855 Abstract Limited competition has been a serious concern in any point in the selection process. This paper applies infrastructure procurement. Importantly, however, there the double selection model to procurement data from are normally a number of potential bidders initially road development projects in developing countries and showing interest in proposed projects. This paper examines why competition ends up restricted. It shows focuses on tackling the question why these initially that bidders are self-selective and auctioneers also tend interested bidders fade out. An empirical problem is to limit participation depending on the size of contracts. that no bids of fading-out firms are observable. They Therefore, limited competition would likely lead to high could decide not to enter the process at the beginning infrastructure procurement costs. of the tendering or may be technically disqualified at This paper--a product of the Economics Team, Finance, Economics and Urban Development Department--is part of a larger effort in the department to understand and enhance efficiency and effectiveness in public procurement for infrastructure. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at aiimi@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team BIDDERS' ENTRY AND AUCTIONEER'S REJECTION: APPLYING A DOUBLE SELECTION MODEL TO ROAD PROCUREMENT AUCTIONS Antonio Estache Université Libre de Bruxelles Atsushi Iimi § World Bank (FEU) Key words: Public procurement; auction theory; endogenous bidder entry; infrastructure development. JEL classification: D44, H54, H57, C21, D21. § We express our special thanks to Japan International Cooperation Agency (former Japan Bank for International Cooperation (JBIC) ODA Operations) for their collaboration in data collection and their insightful comments on an earlier version of this paper. We are also grateful to Decio Coviello, Marisela Montoliu Munoz, Tomás Serebrisky and Tina Søreide for their various suggestions and comments. We also thank many seminar participants in the Chaire Bernard Van Ommeslaghe Conference at the Université Libre de Bruxelles. -2- I. INTRODUCTION Infrastructure procurement is still a challenging task for developing country governments. The limited degree of competition remains among the major concerns (Foster, 2005; NOA, 2007; Estache and Iimi, 2008a). A fundamental problem for auctioneers, namely executing agencies, is how to contain government procurement costs while ensuring the good quality of public work would be delivered. Even though competitive bidding is now commonly required in public procurement systems, participants are often limited because infrastructure projects tend to be highly valuable, complex and customized. Traditional auction theory suggests that increased competition would lower procurement costs under the fixed-n approach. Although how the degree of competition would affect the auction outcome will depend on model, this has been confirmed in many empirical auctions (e.g., Brannman et al., 1987; Paarsch, 1992; Gupta, 2002; Estache and Iimi, 2008a). In the independent private value paradigm, by which our traditional ODA projects could be characterized, the winning bid should tend to approach the lowest possible procurement price, as the number of participants in an auction increases. 1 Therefore, competition is in principle considered welcome for auctioneers, and the issue examined in the current paper is closely related to the results derived by Bulow and Klemperer (1996), discussing the value of an extra bidder in an auction in terms of the buyer's expected traded prices. Importantly, however, bidders' participation is endogenous. On one hand, potential bidders themselves will decide whether or not to enter the tendering, depending on their endowments, 1 How to determination the paradigm must of necessity defer to individual empirical works (e.g., Paarsch, 1992; Bajari and Hortaçsu, 2003). The ODA-related infrastructure project procurement auctions may be more likely to be characterized by the independent private value paradigm, because auction-specific asymmetric uncertainty among bidders plays a more important role to determine the individual bid prices than symmetric uncertainty does. Typical are labor costs of individual firms. Even though the same amount of inputs is required to implement a project, unit costs (e.g., wages and equipment prices) are different across firms (Bajari et al., 2006). Also those private factors remain different even after the contract is awarded. By contrast, political instability and regulatory credibility are considered as a commonly uncertain component. These often affect the public- private partnership (PPP) infrastructure projects. However, in our traditional ODA projects to procure only specific construction works or equipment, these are less important than firm individual cost factors. Furthermore, ODA contracts are not supposed to be resold once awarded. -3- rivals' behavior and the size and nature of projects being auctioned. On the other hand, auctioneers also desire to limit competition in the sense that they might disqualify some of the initial applicants prior to taking their price bids into account. Project owners often want to exclude from the selection process those who would likely fail to fulfill the agreed contract. Without doubt, the number of bidders observed at the final price evaluation stage results from a sequence of choices made by prospective firms and an auctioneer. For this reason, partial observability must emerge as a potential empirical problem. One can observe bids only if firms participate in the tender and are qualified through prerequisite conditions. Auction theory has discussed the endogeneity issue associated with bidders' entry (e.g., McAfee and McMillan, 1987; Levin and Smith, 1994; Menezes and Monteiro, 2000). Basically, it shows that high entry costs would reduce the optimal number of bidders, and bidders would enter until their expected profits are driven to the entry cost. Unfortunately, however, the selection is rarely singleton in practice. Auctioneers can be selective as well. Moreover, even qualified bidders may desire to withdraw their proposal in the middle of the process. The current paper collects data on potential bidders from public procurement auctions for road development projects in developing countries and aims at tracking their sequential decisions toward the final decision about who would be the awardee. In general, it is empirically uneasy to identify potential bidders, simply because they are not observable. 2 The current paper pays attention to those who purchased prequalification or bidding documents as a good set of serious prospective bidders. They may or may not participate in the tendering, as will be described later. The paper, using the sequential-response model (Amemiya, 1975; Maddala, 1983), examines how auctioneers and individual firms decide whether or not to proceed to the next step toward the final price bid evaluation stage. Then, 2 In the existing endogenous entry auction literature, this difficulty tends to be avoided by relating the realized (not potential) number of bidders to explanatory variables in a Poisson or negative binomial regression model (e.g., Li and Perrigne, 2003; Li and Zheng, 2006; Ohashi, 2008; Estache and Iimi, 2008b). Otherwise, one can assume that those who participated in a series of auctions at least once in the past would constitute a set of potential bidders. -4- the equilibrium bid function is estimated by the Heckman (1979) and Lee's (1978) two-step estimation method with a double selection process. The remaining paper is organized as follows. Section II describes the traditional public procurement procedure and overviews the behavior of potential and actual bidders in our road procurement data. Section III develops the empirical methods and summarizes our used data. Section IV presents the main results and discusses some policy implications. II. AN OVERVIEW OF PARTICIPATION DECISIONS BY BIDDERS AND AUCTIONEERS IN ROAD PROCUREMENT Public road projects are traditionally procured through the first-price sealed-bid competitive tenders. Notably, however, there are a variety of procurement systems involving different institutional elements. For instance, some auctions invite only domestic companies, while others accept international enterprises as well. But the most important difference in terms of auction design may be that some auctions adopt the strict lowest price criteria where only the price bids are compared, while others account for "quality," which refers to anything the auctioneer cares about other than prices. 3 A typical example is prequalification. In large- scale infrastructure projects, the technical evaluation prior to comparing prices is normally required in order to ensure the quality of the object or public work, the contract cost agreed, and the period designated. From the auctioneer's point of view, apparently, there is a tradeoff between price and quality if the latter is costly for firms to produce and attain; higher quality, higher prices. In addition, if quality is also costly to prepare before the auction, the optimal number of bidders that are allowed to enter would decline in the presence of increased entry costs, whence limiting 3 There are a certain amount of theoretical works on multidimensional auctions. They show that the two-stage bid evaluation system can implement the optimal mechanism maximizing auctioneer's expected profits (Che, 1993; Cripps and Ireland, 1994). However, it is still far from applicable to practice. -5- competition and raising the equilibrium bid. Moreover, if there are a limited number of companies that could meet highly specific qualification requirements, then the competition would also become narrow. 4 Too exclusive conditions might even be considered corruption. 5 From the bidder perspective, they must be of necessity self-selective for a number of reasons. The fundamental reason is that potential contractors are under resource constraints to a larger or lesser extent. They may not be technologically competent enough to apply for a complex contract. 6 Even though they are potentially able to implement that project, they may not tentatively have available resources for new contracts because they are already devoted to other development projects elsewhere. In addition, if the entry cost is too high, bidders may decide not to enter the tendering process, though having shown interest. Finally, the bidders' strategy would also be influenced by rivals' entry and bidding behavior. In theory, it could be affected if bidders are asymmetric, 7 or if bidders have private signals but do not ex ante know the true common value of the object being auctioned. 8 Hence, bidders and auctioneers have various different reasons for not participating or not allowing them to participate in the tendering process. Our data uncover the typical auction process for road construction and rehabilitation works, during which the applicants gradually shrank from 13 to 4 on average (Figure 1). At the very beginning of the process, auctioneers usually publish tender notices. About 13 companies or consortia per contract showed interest 4 See Estache and Iimi (2008b) for more discussion. 5 A crucial shortcoming of multidimensional auctions is that the award process would be less transparent and more vulnerable to corruption; authorities can easily exploit their excessive discretion (Klein, 1998; Ware et al., 2007; Estache et al., 2008). 6 One of the solutions to this resource constraint problem is joint bidding. Bidders can overcome the possible prerequisites by pooling their financial and managerial resources with each other (e.g., Moody and Kruvant, 1988; Hendricks and Porter, 1992; Iimi, 2004). 7 Asymmetric auction theory predicts that a weaker (fringe) bidder tends to bid more aggressively in the presence of a strong (incumbent) bidder. Maskin and Riley (2002) shows that if a weak bidder faces a strong bidder rather than another weak bidder, he responds with a more aggressive bid distribution in the sense of stochastic dominance. The empirical evidence is supportive of this (e.g., De Silva et al., 2002, 2003; Estache and Iimi, 2008c). 8 Under the common value paradigm, competition may increase the equilibrium bid due to the winner's curse effect (e.g., Milgrom and Weber, 1982; Paarsch, 1992; Klemperer, 1998). -6- and purchased the relevant prequalification or bidding documents. If prequalification is adopted, about 10 potential bidders would apply for it. Then, about two-thirds are qualified. At this stage, only 6­7 bidders remain. About one bidder may decide not to submit a price bid for some reason, though prequalified. After the staged process involving the detailed technical evaluation prior to opening the price bids, only four price bids are evaluated and compared at the final stage. In our sample, the majority of auctions adopted the two-envelope procedure, in which all (qualified) bidders are requested to submit both price and technical bids simultaneously and the auctioneer opens the price bid only if the corresponding technical bid is substantially responsive to the bidding documents and other technical requirements. As the result, the degree of competition tends to be rather limited at the very end of the auction process. Figure 1. Numbers of Potential and Actual Bidders in Road Procurement Auctions 14 Average number of firms/consortia 12 (­3.1) 10 (­3.6) 8 (­0.7) 13.1 6 (­1.6) 9.9 4 6.4 5.6 2 4.0 0 Purchasing Applying for Prequalified Submitting Price PQ PQ if PQ is if PQ is bids evaluated documents adopted adopted after technical qualification Source: Author's calculation. Because the intensity of competition changes along the bidding procedure, the relationship between the bid strategy and competition also varies depending on the stage of the selection process. Figure 2 illustrates the competition effect when it is measured at the beginning of the tender process, i.e., the number of firms that purchased prequalification or bidding documents. Figure 3 depicts the competition effect at the final price comparison stage. Of course, both ignore partial observability; one cannot observe bids from firms that not participated, though showed interest. But the comparison between them indicates that Figure 3 ignores many unrealized bids that could have been scattered where the horizontal axis is -7- greater than seven if they were observed. It suggests that without controlling for the sample selection bias, the empirical competition effect would be biased. Figure 2. Potential Competition and Bids Figure 3. Final Competition and Bids 10 10 Evaluated bid / Engineering cost estimate Evaluated bid / Engineering cost estimate 9 9 8 8 7 7 6 6 5 5 4 4 3 3 2 2 1 1 0 0 0 5 10 15 20 25 30 0 5 10 15 20 25 30 Number of firms/consortia who purchased PQ documents Number of firms/consortia whose price bids are evaluated after all Source: Author's calculation. Source: Author's calculation. III. THE EMPIRICAL MODELS The following sequential decisions made by an auctioneer and potential bidders are considered. Suppose that L firms purchased prequalification and/or bidding documents. These constitute a set of potential bidders, out of which M firms applied for prequalification if the prequalification process is introduced, or just entered into the price competition if prequalification is omitted. This is the first selection made by each bidder of , , L, 1 denoted by d1. Then, the auctioneer qualifies N firms from actual applicants , , M due to 1 technical reasons. Disqualification occurs at either the prequalification or technical evaluation level. This is the second selection, denoted by d2. Even though qualified, bidders can still choose not to proceed to the final price comparison stage. Some qualified bidders in fact cease participating in the auction, possibly because of unexpected financial and equipment constraints. But such cases are relatively rare, as shown in the previous section. Hence, the model focuses on only two major decision nodes: (i) whether firms apply for the bidding process, and (ii) whether bidders are qualified and advance to the price competition. -8- The selection rules are as follows: d1 1 if d1* Z ' 1 1 0 (1) d1 0 otherwise d 2 1 d1 1 if d 2* Z ' 2 2 0 (2) d 2 0 d1 1 otherwise where d1* and d2* are latent variables but have dichotomous observable realizations, d1 and d2, respectively. Z is a vector of bidder- and auction-specific characteristics determining the selection mechanisms. The two error terms are assumed to follow a multivariate normal distribution with zero means and variances equal to unity. Potentially, the error terms can be correlated with one another. Denote this by Cov 1 , 2 2 . One might expect that an auctioneer disqualifies firms that have certain common unobserved characteristics, which could also systematically encourage bidders to enter into the auction. In this case, the correlation will be negative. Given the two selections, each bidder of , , N submits the following conventional 1 symmetric equilibrium bid (e.g., Porter and Zona, 1993; Gupta, 2002; Iimi, 2006): ln BID X ' u (3) where X is composed of bidder- and auction-specific variables that influence firms' underlying cost parameters and control for heterogeneity among projects to be auctioned. BID is the evaluated bid price. This is observed only for firms that entered and were qualified. Consequently, the ordinary least squares (OLS) estimation would lead to the sample selection bias if Equations (1) and (2) are significant (Lee, 1978; Heckman, 1979). -9- Following the existing literature focusing on the double or generalized selection process (e.g., Poirier, 1980; Lee, 1983; Tunali, 1986; Mohanty, 2001), Equation (3) can be consistently estimated by correcting the two selectivity biases: E ln BID d1 1, d 2 1 X ' 1u 1 2u 2 (4) Z ' j Z ' i where iu Cov ( i , u ) and i Z ' i B Z ' , Z ' , for 1 2 i j i, j 1,2 and j i . is the standard normal density function and is the cumulative normal distribution function. B is the bivariate standard normal distribution function. To estimate Equation (4), the two-step method is adopted. The first stage is estimated by a censored bivariate probit model because of partial observability, because the second selection rule d2 is observed only if bidders apply for the prequalification process, i.e., d1 1 . The corresponding log likelihood function is: N L 1 d1 ln 1 ( Z ' 1 ) d11 d 2 ln B( Z ' 1 , Z ' 2 , ) d 2 ln BZ ' 1 , Z ' 2 , The first term is associated with the probability that a bidder decides not to apply for the process. The second represents the case that a bidder determines to enter into the auction but does not proceed to the price comparison stage. The last expression is the probability that a bidder is qualified and its bid is evaluated after all. Given estimated parameters, 1 and 2 are calculated and used for the second stage. Then, ^ ^ Equation (4) can be estimated by OLS. This two-step estimator will be consistent, and the standard errors are corrected with the possible bias into account, which is mainly caused by the additional variance of 1 and 2 added to remove the selection bias. The asymptotically consistent mean square error is: - 10 - v 1 u2 d1 1, d 2 1 N 2 i 1 i u i N where vi is the residual from the OLS regression of Equation (4), and the additional error is b( zi ' ^1 , zi ' ^2 , ) u2 i u12 zi ' ^11i u 2 2 zi ' ^2 2i u11i u 2 2i 2 2 u1 u 2 ( u21 u22 ) ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ B( zi ' ^1 , zi ' ^2 , ) ^ . ki ( zi ' ^k ) ( zi ' ^h zi ' ^k ) / 1 2 ^ ^ for k , h 1,2 k h . 9 B ( zi ' ^k , zi ' ^h , ) ^ To specify the equilibrium bid function in Equation (3), X is composed of four types of project- and bidder-specific characteristics: (i) the number of bidders participating in an auction, which is often referred to as N, (ii) project-specific observables, such as length of roads (LENG), number of lanes (LANE), engineering cost estimates (COST) and expected contract duration (MONTH), (iii) country-specific fixed effects, and (iv) observable bidder attributes. For the last, a set of dummy variables representing bidder nationalities are used to control potential heterogeneity across bidders. As one can expect, whether bidders are foreign or domestic is among the most important determinants of the bidding equilibrium strategy in large-scale development projects. Regarding the first variable N, it must be noted that we choose the number of bidders that were qualified if prequalification was applied and stepped in the price competition stage. If prequalification is not adopted, N is simply the number of firms submitting bids. N is expected to capture the competition effect particularly in the independent private value paradigm. Notably, this is comparative statics in the conventional fixed-n approach; more competition would lead to lower procurement costs. 10 In auction theory, the endogeneity of N 9 See Tunali (1986) for a full description. 10 It is still theoretically possible to characterize our auctions as the common value paradigm, in which the equilibrium bid could increase with the number of bidders. However, there is little evidence supportive of this in our data, regardless of functional forms. A significant negative relationship between the bids and the number of bidders was found even in a partially nonparametric specification (Estache and Iimi, 2008a). In the current - 11 - is another important question for considering competitiveness (e.g., McAfee and McMillan, 1987; Levin and Smith, 1994; Bajari et al, forthcoming). At least from the empirical point of view, however, a necessary question may be whether N is common knowledge among bidders, whatever it represents. Our defined N is the best proxy in this regard. In the practical ODA procurement circumstances, serious tenderers are likely to share a good sense of how many rivals are remaining for the final price competition, either through the formal prequalification result publication and informal business relationship. Therefore, the number of bidders that were prequalified and submitted price bids is considered the best information of N known by tenderers. 11 Following the existing literature (e.g., Porter and Zona, 1993; Bajari et al., 2006; and Price, 2008), two instruments are used for a vector of selection determinants, Z. 12 One measure is the total amount of contracts--in our sample--awarded to each firm in the three-year period before a particular contract is auctioned (BKLG). If a bidder forms a joint venture, the average backlog among consortium members is used. Another variable to complement this is the amount of total transport ODA disbursed to each country in the three-year period prior to the auction (CAID). 13 The basic idea behind these variables is that if prospective firms are already devoted to other development projects elsewhere, fewer bidders would be willing to bid on further new works because they might be too busy. An implicit assumption is that firms are resource-constrained, which will in fact be verified in the following analysis. paper, the competition effect is specified in logarithmic form, because it is known that the mapping between winning bids and the number of bidders is never linear (Rezende, 2005) and the logarithmic models often better fit actual auction data (e.g., Porter and Zona, 1993; Gupta, 2001, 2002). 11 Technically, our N is different from the sum of d2a and the sum of d2b as well. It represents the number of prequalified bidders bud may include those who are disqualified technically in the pre-stage of price bid evaluation. This is the case where the two-envelope procedure is adopted. In this system, all potential bidders are requested to submit both price and technical proposal and an auctioneer opens the price bids only if the submitted technical proposal meets the required standards. It will play the similar role as prequalification, but the order is difference and the scope of the examination may be more technical (see Estache and Iimi (2008b) for more discussion on this). 12 X is also part of Z except for N, because at the preselection stage, participants may not be sure how many firms would finally decide to apply for it. 13 Our sample merely represents several percent of total ODA allocated to the transport sector all over the developing world. Therefore, CAID is expected to capture the broader market conditions than BKLG. The used aid data come from the OECD Creditor Reporting System database. - 12 - Table 1 shows the summary statistics. Our sample covers about 325 potential bidders that are interested in one of the 31 road procurement auctions under 11 projects in nine developing countries. 14 70 percent of them applied for the tendering process. 184 firms are identified as those who were technically evaluated, and about 64 percent passed the technical examination. There were 117 bidders that were technically qualified, but not all of them participated in the price competition; only 84 percent did. Therefore, we observe 98 winning and losing bids, of which the average is about 23 million U.S. dollars. Note that both winning and losing bids are equally informative to estimate the equilibrium bid function, because the auction format normally used for infrastructure project procurement is the standard first-price sealed-bid auction. The contracts differ considerably in financial and technical terms. While lower bids are less than 0.4 million U.S. dollars, high bids may exceed 100 million U.S. dollars. The average length of roads is about 40 km, but it ranges from 1 km to 280 km. The number of lanes also varies from two to six. The average project duration is estimated at 31 months. The average of total contract amounts obtained by a firm in the three-year period prior to the auction is about 8 million U.S. dollars. But again, there is huge variation; some firms were awarded several similar projects in recent years, and others not. In addition, each recipient country seems to receive a considerable amount of aid in the transport sector alone. Note that the last is a country-specific but time-variant variable. 14 Our sample originally covers about 450 potential bidders but is reduced to 325 due to missing relevant data. - 13 - Table 1. Summary Statistics Variable Abbr. Obs Mean Std. Dev. Min Max Evaluated bid 1/ BID 98 22.70 23.90 0.37 115.00 Dummy for applicants to the process d1 325 0.70 0.46 0 1 Dummy for the technically qualified d 2a 184 0.64 0.48 0 1 Dummy for bidders of which bids are compared d 2b 117 0.84 0.37 0 1 Number of bidders proceeding to the price bidding stage N 325 5.74 1.92 2.00 11.00 Length of roads (km) LENG 325 39.59 50.50 0.90 278.55 Number of lanes LANE 325 2.95 1.29 2.00 6.00 Engineering cost estimate 1/ COST 325 33.12 48.83 0.39 176.74 Estimated contract duration (months) MONTH 325 31.12 13.81 9.00 45.00 Firm's backlog in the past three years 1/ BKLG 325 7.78 24.40 0.00 189.00 Total transport aid received by the project country 1/ CAID 325 410.89 318.82 5.73 1,000.97 1/ In millions of constant 2000 U.S. dollars. Source: Author's calculation. IV. ESTIMATION RESULTS AND IMPLICATIONS Before presenting the equilibrium bid equation estimated by the two-step double selection technique, the sequential response model is performed to examine how prospective firms determine to participate in each step of the auction and how the auctioneer rejects some of the applicants. For this, the second decision d2 is separated into two stages. First, the auctioneer can determine to technically qualify each applicant prior to the price comparison; this process is denoted by d2a. Then, qualified bidders can decide whether to proceed to the price comparison. This is denoted by d2b. It must be noted that d2a is the decision making on the auctioneer side, while d2b as well as d1 are primarily determined by bidders. Assuming the normal distribution, the probabilities can be written as: Pr d1 0 Z ' 1 Pr d1 1, d 2 a 0 1 Z ' 1 Z ' 2 a (5) Pr d1 1, d 2 a 1, d 2b 1 1 Z ' 1 1 Z ' 2 a Z ' 2b Since these are sequential decisions, the parameters can be estimated by performing the probit model with only relevant subsample in each case (Amemiya, 1975; Maddala, 1983). An important assumption for these estimators is obviously that the random factors influencing responses at individual stages are independent. - 14 - In our data, 325 firms constitute a set of potential bidders, 15 out of which 228 decided to apply for the selection process. Then, prior to the following price competition, at least 117 and 67 bidders were technically qualified and disqualified, respectively. 16 Out of the qualified bidders, 98 firms actually decided to proceed to the final tendering process and thus 98 price bids are observable in our data. Table 2 presents the probit results. 17 There is a noticeable contrast between the auctioneer's and bidders' behavior related to auction participation. Bidders are more likely to enter in the tendering when their potential backlogs, which are measured by either BKLG or CAID, are considered moderate. This is consistent with the previous literature. It implies that bidders have certain capacity constraints. On the other hand, auctioneers would be more likely to qualify bidders that have more backlogs. This must reflect the tendency toward more experienced contractors. This is not surprising, because the prequalification process normally requires prospective bidders to have experienced similar development projects in the past. And intuitively, auctioneers may normally be comfortable contracting with incumbents. These results highlight an important challenge in procuring infrastructure projects: Potential contractors are resource-constrained, but auctioneers may prefer to contract with those who have undertaken many other development projects. As the result, the market competition at the final stage will tend to be limited. But then the expected contract amount will increase due to lack of competition among contenders, and market concentration may also be of potential concern. 15 Only 252 observations are used for the first probit model, because one of the explanatory variables takes the same value for 73 cases. These are perfectly predicted in our specification. 16 44 observations are not used for the second probit model, because the prequalification results are unavailable in these cases. 17 Note that many country- and nationality-specific fixed effects are omitted because there are many cases where the fitted probability that one of the country or bidder nationality dummy variables equals unity is exactly either one or zero. - 15 - Other than the backlog variables, it is found that larger road projects would attract fewer bidders. It is because firms that can undertake large-scale projects are limited due to their inevitable resource boundaries; the coefficients of LENG, LANE and COST are broadly negative and significant for the bidder's entry decisions d1 and d2b. It is also because auctioneers have a propensity to put more emphasis on bidders' technical capabilities and past experiences when contracting out a large project. Thus, the probability that an auctioneer disqualifies an applicant would increase with the size of contract holding everything else constant. In the model for d2a, again, LENG, LANE and COST have negative coefficients. Table 2. Binary Response Model for Bidders' Participation Dependent variable d1 d1 d 2a d 2a d 2b d 2b ** ** ** *** *** ln LENG ­0.046 ­0.371 ­0.254 ­0.560 ­0.856 ­0.713 (0.094) (0.152) (0.109) (0.287) (0.321) (0.274) ** ** * ** ln LANE ­1.019 ­0.194 ­0.498 ­1.613 ­1.627 ­1.910 (0.402) (0.538) (0.450) (0.790) (0.903) (0.850) ** * ln COST ­0.068 ­0.323 ­0.192 ­0.187 ­0.115 ­0.329 (0.105) (0.133) (0.119) (0.149) (0.185) (0.302) *** ln MONTH ­0.308 1.317 0.163 0.850 ­1.448 3.220 (0.299) (0.456) (0.358) (0.763) (1.118) (3.684) *** ** ln BKLG 0.024 0.019 0.005 0.013 (0.007) (0.008) (0.012) (0.012) ** * ln CAID ­0.093 0.059 ­0.104 0.005 ­0.781 ­3.144 (0.081) (0.138) (0.104) (0.506) (0.317) (1.673) *** ** * *** *** Constant 3.113 ­0.291 2.225 3.119 15.508 13.678 (0.844) (1.279) (1.025) (1.846) (4.676) (4.300) Obs. 252 252 184 184 117 117 *** *** *** *** *** Wald-chi2 68.23 145.38 27.48 57.98 21.46 42.03 No. of country dummies (0) (3) (0) (4) (0) (2) No. of bidder nationality dummies (0) (0) (0) (7) (0) (5) Chi2 test statistics: H0: Coef. of country and bidder 61.89 *** 43.24 *** 8.38 nationality dummies= 0 Note: The dependent variable is the bidder's entry selection at each stage. The robust standard errors are shown in parentheses. *, ** and *** indicate the 10%, 5% and 1% significance levels, respectively. Source: Author's calculation. Turning back to the formal selection model we analyze, two main decision nodes are focused on: d1 and d2. d2a and d2b can be considered to have been merged. Notably, however, d2 is supposed to represent mainly the former decision by auctioneers (d2a), because the number of bidders that were qualified but did not participate in the price competition is much small in our data. It is also noteworthy that when estimating a censored bivariate probit model for d1 and d2 as the first stage regression, the two error terms can be correlated with one another unlike the above sequential response model. Table 3 shows the results, which are broadly - 16 - consistent with the sequential response models estimated above; potential bidders are faced with resource constraints and often decide not to enter the tender especially when the contract amount is large. The coefficient of lnCOST is negative, and the coefficient of lnCAID is also negative in the second column model. 18 On the other hand, auctioneers would consistently prefer to preselect enterprises that have more past experiences, even though they may not be available due to their resource commitment to something else. The correlation between the two binary responses seems subtle. It is negative but insignificant. In the following analysis, the second column model is used for inferring 1 and 2 for Equation (4), because the country and nationality fixed effects are found unignorable; the chi square test statistic is estimated at 315.36. If this is the case, it mans that there are some unobservables that would facilitate bidders' entry but discourage auctioneers from qualifying them. Some characteristics shared by new growing companies may be one possibility. While new entrants are eager to enter the market, they may not be trustworthy enough given their past experiences. 18 All four measures of project heterogeneity are related to the scale or size of the work. Unsurprisingly, therefore, the estimated coefficients of these variables may be sensitive to specifications. One of the most important independent variables controlling for project heterogeneity may be engineering cost estimates. If the other three variables are omitted from the models, it can be found that the coefficient of the project value is always negative for both equations. The larger a project, the fewer contractors there are who could apply for it. And the larger a project, the fewer applicants there are who are prequalified. - 17 - Table 3. Censored Bivariate Probit Model for Bidders' Entry and Auctioneer's Rejection Coef. Std.Err. Coef. Std.Err. Equation d 1 ln LENG ­0.039 (0.092) 0.823 (1.085) ln LANE ­0.886 (0.427) ** 6.679 (2.950) ** ln COST ­0.145 (0.128) ­5.741 (1.815) *** ln MONTH 0.026 (0.312) 27.871 (10.268) *** ln BKLG 0.176 (0.005) *** 0.010 (0.012) ln CAID ­0.084 (0.084) ­26.505 (10.843) ** Constant 4.805 (0.876) *** 40.519 (19.833) ** Equation d 2 ln LENG ­0.125 (0.106) ­0.782 (0.256) *** ln LANE 0.296 (0.464) ­1.169 (0.669) * ln COST ­0.188 (0.126) ­0.041 (0.133) ln MONTH ­0.250 (0.339) 0.001 (0.601) ln BKLG 0.015 (0.014) 0.018 (0.008) ** ln CAID 0.074 (0.088) 1.275 (0.333) *** Constant 1.140 (1.054) 0.060 (1.773) ­0.310 (0.476) ­0.109 (0.374) Obs. 285 285 Wald-chi2 1482.66 *** 1390.33 *** No. of country dummies (0) (4) No. of bidder nationality dummies (0) (5) Chi2 test statistics: H0: Coef. of country and bidder 315.36 *** nationality dummies= 0 Note: The dependent variables are the bidder's entry decisions at individual stages. The robust standard errors are shown in parentheses. *, ** and *** indicate the 10%, 5% and 1% significance levels, respectively. Source: Author's calculation. Finally, the consistent equilibrium bid function is presented in Table 4. The first column model is the simple OLS estimation, which may not be consistent because of uncontrolled selectivity bias. As usually expected, nonetheless, the competition effect is negative, meaning that the equilibrium bid, whence the government procurement cost would decrease with the number of participating bidders. 19 It is also reasonable that the submitted bids would be higher for more valuable contracts, longer highways, and wider roads. When correcting the possible double selection bias, it is found that the competition effect is still statistically significant but slightly larger (in absolute terms) at ­0.707. 20 It means that unrealized bids for some reason would be pro-competitive if they were observable. This can be interpreted as the potential benefits of unrealized competition. The other explanatory variables have the similar coefficients to the OLS regression. 19 For more discussion on this static competition effect, see Estache and Iimi (2008a). 20 Note that the standard errors were corrected by accommodating the additional disturbance caused by the two bias correction terms; but the coefficient still remains significant. - 18 - The hypothesis that the difference in coefficients between the two models is not systematic cannot be rejected by the standard chi-square test. The test statistic is 2.11 and well below the conventional critical values. Therefore, it can be concluded as follows: There are many prospective bidders, of which bids are not observed, and various factors would likely influence bidders' entry decision and auctioneer's qualification decision. Nonetheless, the equilibrium bid function may be able to be estimated effectively through the OLS regression with only observed bids. The competition effect may be slightly underestimated, but the bias will be small. The estimated bid function reveals another important challenge in procuring infrastructure projects, which is the tradeoff between competition and economies of scale. The bid function exhibits economies of scale; For instance, longer roads are unsurprisingly more expensive, but the unit cost would decrease with the length of roads in a package. The coefficients are well below one (i.e., 0.368 in the OLS case and 0.447 in the double selection model), meaning that a one percent increase in road length would raise the equilibrium bid by only about 0.4­0.5 percent, holding everything else constant. Therefore, auctioneers could save public spending, when they design relatively large procurement packages. However, it will contradict the competition objective; as shown above, a smaller number of applicants could be qualified for longer road projects, and more valuable contracts could also attract fewer bidders. To highlight the above dilemma between competitiveness and economies of scale, a comparative static analysis is considered. The baseline scenario is evaluated at sample means for explanatory variables. Suppose that the auctioneer downscales the size of contract by 10 percent in terms of engineering cost estimate, leaving everything else unchanged. 21 Then, the probability that a prospective firm decides to apply for the process would increase from 69 to 74 percent (Table 5). This is because more enterprises would be able to accommodate 21 Apparently, it is more natural that some attributes of the project would be changed if the engineering cost is changed. For simplicity, however, the analysis focuses on a simple shock in one single variable. - 19 - smaller works. The probability of an applicant being qualified would also increase 34 to 36 percent, because the auctioneer might be less likely to disqualify bidders for smaller projects. If the predicted probabilities are simply translated into our competitiveness variable N, the degree of competition would increase by 5.2 percent, which would result in 3.7 percent lower bids according to our estimated bid function. This is the benefit from the intensified competition. At the same time, a 10 percent reduction in engineering cost estimates will reduce the bid by 2.3 percent. This is the direct effect of changing the specification of the project. However, recall that the cost estimate was reduced by 10 percent in our scenario. Despite that, the auctioneer is expecting only a 6 percent reduction in predicted bids. As in this scenario, hence, the simple unbundling approach, which aims at dividing a project to a number of small packages, may not always be the solution. The policy implication is straightforward but challenging. Essentially because potential bidders are resource-constrained, the market competition must of necessity become limited throughout the selection process. If auctioneers divide a project into several small contracts in order to loosen bidders' resource constraints, the benefit from economies of scale would be sacrificed. These contradictory difficulties could not be solved unless more skilled and competent contractors would emerge and enter in the public procurement markets. In this regard, new entry and local business development, though not achievable in the short term, are of particular importance. Although it is beyond the scope of this paper, the above discussion may provoke governments to use an alternative approach to procure public contracts, i.e., through negotiation. Theoretically, Bajari and Tadelis (2001) show that more complex transactions would be likely by plagued by ex post adaptations and are better accommodated by a cost- plus contract, which is opposite to a fixed price contract typically assumed in the auction- based procurement. Bajari et al. (2006) also provide the supportive evidence of this in the private sector building construction industry. Obviously, however, some important advantages of auctions may have to be sacrificed, e.g., transparency and corruption - 20 - prevention, in addition to competitiveness. These factors may be unignorable in particular in the context of public procurement. Table 4. Two-Step Estimation for Equilibrium Bid Strategy with Double Selection Process Double OLS selection model ** ln NUM ­0.687 ­0.707 * (0.318) (0.368) *** ln LENG 0.368 0.447 *** (0.093) (0.159) *** ln LANE 1.517 1.621 *** (0.275) (0.393) *** ln COST 0.240 0.226 ** (0.088) (0.110) ln MONTH ­0.075 ­0.108 (0.504) (0.580) 1 0.341 (0.274) 2 ­0.314 (0.388) Constant 15.054 *** 15.283 *** (1.916) (2.206) Obs. 98 98 R-squared 0.939 0.940 F -statistics 62.77 *** 56.97 *** No. of country dummies (6) (6) No. of bidder nationality dummies (8) (8) F test statistics: H0: Coef. of country and bidder 11.85 *** 9.41 *** nationality dummies= 0 Note: The dependent variable is the bidder's entry selection at each stage. The standard errors are shown in parentheses. *, ** and *** indicate the 10%, 5% and 1% significance levels, respectively. Source: Author's estimation. Table 5. Scenario Analysis­An Example Scenario: 10 percent reduction in Baseline engineering cost estimates Avg. predicted probability (d 1= 1) 68.8 74.1 Avg. predicted probability (d 2= 1 given d 1= 1) 34.4 36.2 Predicted bid changes associated with N -3.7 Predicted bid changes associated with COST -2.3 Source: Author's estimation. V. CONCLUSION The limited degree of competition is one of the major concerns in infrastructure procurement. The degree of competition that one can observe at the very end of the auction process tends - 21 - to be limited perhaps to about 4­5 bidders. Although one can still expect the conventional competition effect in a static sense, it is important to recall that the entry decision of potential bidders is potentially endogenous. There are a number of prospective bidders, who are interested but may decide not to be involved into the selection process. Hence, the observed distribution of bids will be much different from the distribution of bids that would have been submitted if those potential bidders have participated. Because of this partial observability, the paper applied the double selection model to procurement data from road development projects in developing countries and examined why some companies did not participate or were not allowed to participate in the bidding process. The estimation results highlight several challenges in infrastructure procurement. Bidders are resource-constrained and self-selective. A few companies can apply for larger road projects. At the same time, auctioneers also tend to limit participation as the contract size increases. They prefer to contract with only experienced contractors. This is the reason the market competition ends up being rather restricted, resulting in high infrastructure procurement costs. The estimated equilibrium bid strategy suggests possible economies of scale; government procurement costs could be saved by increasing the contract size. But it would risk jeopardizing the expected competition effect because, again, larger projects could attract fewer bidders. - 22 - REFERENCE Amemiya, Takeshi. 1975. Qualitative models. Annals of Economic and Social Measurement, Vol. 4, pp. 363-372. Bajari, Patrick, Ali Hortaçsu. 2003. Winner's curse, reserve prices and endogenous entry: empirical insights from eBay auctions. The RAND Journal of Economics, Vol. 34(2), pp. 329-355. Bajari, Patrick, Steven Tadelis. 2001. Incentives versus transaction costs: a theory of procurement contracts. The RAND Journal of Economics, Vol. 32, pp. 387-407. Bajari, Patrick, Robert McMillan, Steven Tadelis. 2006. Auctions versus negotiations in procurement: an empirical analysis, Journal of Law, Economics and Organization, forthcoming. Brannman, Eric, Douglass Klein, Leonard Weiss. 1987. The price effect of increased competition in auction markets. The Review of Economics and Statistics, Vol. 69, pp. 24-32. Bulow, Jeremy, Paul Klemperer. 1996. Auctions versus negotiations, The American Economic Review, Vol. 86(1), pp. 180-194. Che, Yeon-Koo. 1993. Design competition through multidimensional auctions. Rand Journal of Economics, Vol. 24, pp. 668-680. Cripps, Martin, Norman Ireland. 1994. The design of auctions and tenders with quality thresholds: the symmetric case. The Economic Journal, Vol. 104, pp. 316-326. De Silva, Dakshina, Timothy Dunne, Georgia Kosmopoulou. 2002. Sequential bidding in auctions of construction contracts. Economics Letters, Vol. 76, pp. 239-244. De Silva, Dakshina, Timothy Dunne, Georgia Kosmopoulou. 2003. An empirical analysis of entrant and incumbent bidding in road construction auctions. The Journal of Industrial Economics, Vol. 51, pp. 295-316. Estache, Antonio, Jose-Luis Guasch, Atsushi Iimi, Lourdes Trujillo. 2008. Multidimensionality and renegotiation: evidence from transport-sector public-private- partnership transactions in Latin America. Policy Research Working Paper No. 4665, the World Bank. Estache, Antonio, Atsushi Iimi. 2008a. Procurement efficiency for infrastructure development and financial needs reassessed. Policy Research Working Paper No. 4662, the World Bank. Estache, Antonio, Atsushi Iimi. 2008b. Auctions with endogenous participation and quality thresholds: evidence from ODA infrastructure procurement. Mimeograph. Estache, Antonio, Atsushi Iimi. 2008c. Bidder asymmetry in infrastructure procurement: are there any fringe bidders? Policy Research Working Paper No. 4660, the World Bank. - 23 - Foster, Vivien. 2005. Ten years of water service reforms in Latin America: toward an Anglo- French model. Water Supply and Sanitation Sector Board Discussion Paper Series No. 3, The World Bank. Gupta, Srabana. 2001. The effect of bid rigging on prices: a study of the highway construction industry. Review of Industrial Organization, Vol. 19, pp. 453-467. Gupta, Srabana. 2002. Competition and collusion in a government procurement auction market. Atlantic Economic Journal, Vol. 30, pp. 13-25. Heckman, James. 1979. Sample selection bias as a specification error. Econometrica, Vol. 47, pp. 153-161. Hendricks, Kenneth, Robert Porter. 1992. Joint bidding in federal OCS auctions. The American Economic Review, Vol. 82, pp. 506-511. Iimi, Atsushi. 2004. (Anti-)Competitive effect of joint bidding: evidence from ODA procurement auctions. Journal of The Japanese and International Economies, Vol. 18, pp. 416-439. Iimi, Atsushi. 2006. Auction reforms for effective official development assistance. Review of Industrial Organization, Vol. 28, pp. 109-128. Klein, Michael. 1998. Bidding for concessions. Policy Research Working Paper Series No. 1957, The World Bank. Klemperer, Paul. 2002. How (not) to run auctions: the European 3G telecom auctions. European Economic Review, Vol. 46, pp. 829-845. Lee, Lung-Fei. 1978. Unionism and wage rates: a simultaneous equation model with qualitative and limited dependent variables. International Economics Review, Vol. 19, pp. 415-433. Lee, Lung-Fei. 1983. Generalized econometric models with selectivity. Econometrica, Vol. 51, pp. 507-512. Levin, Dan, James Smith. 1994. Equilibrium in auctions with entry. The American Economic Review, Vol. 84, pp. 585-599. Li, Tong, Isabelle Perrigne. 2003. Timber sale auctions with random reserve prices. The Review of Economics and Statistics, Vol. 85, pp. 189-200. Li, Tong, Xiaoyong Zheng. 2006. Entry and competition effects in first-price auctions: theory and evidence from procurement auctions. Cemmap working paper CWP13/06, University College London. Maddala, G.S. 1983. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge University Press. Maskin, Eric, and John Riley. 2000. Asymmetric auctions. The Review of Economic Studies, Vol. 67, pp. 413-438. McAfee, Preston, John McMillan. 1987. Auctions with entry. Economics Letters, Vol. 23, pp. 343-347. - 24 - Menezes, Flavio, Paulo Monteiro. 2000. Auction with endogenous participation. Review of Economic Design, Vol. 5, pp. 71-89. Milgrom, Paul, Robert Weber. 1982. A theory of auctions and competitive bidding. Econometrica, Vol. 50, pp. 1089-1122. Mohanty, Madhu. 2001. Testing for the specification of the wage equation: double selection approach or single selection approach. Applied Economics Letter, Vol. 8, pp. 525-529. Moody, C.E., W.J. Kruvant. 1988. Joint bidding, entry and the price of OCS leases. Rand Journal of Economics, Vol. 19, pp. 276-284. NOA. 2007. Improving the PFI Tendering Process. The National Audit Office. Ohashi, Hiroshi. 2008. Effects of transparency in procurement practices on government expenditure: a case study municipal public works. CIRJE Discussion Paper No. 548, University of Tokyo. Paarsch, Harry. 1992. Deciding between the common and private value paradigms in empirical models of auctions. Journal of Econometrics, Vol. 51, pp. 191-215. Poirier, Dale. 1980. Partial observability in bivariate probit models. Journal of Econometrics, Vol. 12, pp. 209-217. Porter, Robert, Douglas Zona. 1993. Detection of bid rigging in procurement auctions. Journal of Political Economy, Vol. 101, pp. 518-538. Price, Michael. 2008. Using the spatial distribution of bidders to detect collusion in the marketplace: evidence from timber auctions, Journal of Regional Science, Vol. 48, pp. 399-417. Rezende, L. 2005. Auction econometrics with least squares. Mimeograph, University of Illinois, Urbana-Champlain, IL. Tunali, Insan. 1986. A general structure for models of double selection and an application to a joint migration/earning process with remigration. Research in Labor Economics, Vol. 8, Part B, pp. 235-282. Ware, Glenn, Shaun Moss, Edgardo Campos, Gregory Noone. 2007. Corruption in public procurement: a perennial challenge, in The Many Faces of Corruption: Tracking Vulnerabilities at the Sector Level, eds. by Edgardo Campos and Sanjay Pradhan, pp. 295-334. The World Bank.