Cheap Signals, Costly Proof¶
Award-Layer Triage for Cartel Enforcement
Insper Institute of Education and Research, Sao Paulo, Brazil
83% Reduction in the bid-microdata pool the forensic stage must work on, while still recovering 131 of 193 adjudicated cobidders
Abstract¶
Can cheap data triage cartel investigations before costly forensics? Most antitrust enforcers see who won a public contract and at what price---not what every bidder offered. This thinner data is enough to flag which firms warrant costly forensic investigation. Our architecture sequences an award-layer screening stage before a costly bid-layer forensic stage, instantiated with a frequent-loser flag on São Paulo's BEC (2009--2019). The flag cuts the forensic stage's bid-microdata pool by 83% while still flagging two-thirds of adjudicated cartel cobidders. On the same data, it matches the seven-feature Imhof--Wallimann bid-distribution pipeline that requires bid microdata, adding non-redundant signal in combination. The discrimination is out-of-sample: built on 2009--2016 participation only, the screen prospectively flags adjudicated cobidders in 2017--2019. The screen triages; it does not adjudicate cartel membership. A simple separating-equilibrium argument motivates endogenous loser-side participation as the ranking primitive. Wherever award records are routinely available while bid microdata are forensic-recoverable, screening should be sequenced before forensics.
JEL Classification: D44 D73 H57 K21 L41
Keywords: screening under incomplete observability cover bidding cartel adjacency award-layer enforcement separating equilibrium
Key Findings¶
3.6--7.7% conditional price association
FL-present tenders exhibit 3.6--7.7% higher conditional prices across four estimation approaches: cross-fit (3.6%), IPW (5.5%), OLS with item + year + PBU FE (6.4%), and CEM matching (7.7%). The estimates cluster rather than scatter, indicating a stable association across designs.
AUC = 0.94 against CADE convictions
The FL screen achieves AUC = 0.94 against competition-authority convictions, versus 0.79 for an Imhof-style bid-level proxy. FL firms co-participate with convicted cartelists at 3.5 times the baseline rate (\(p < 0.001\)). The two screens capture largely non-overlapping information (correlation 0.06); naively combining them into a single score degrades detection to AUC = 0.61, confirming that they operate on different dimensions and are best deployed sequentially.
Coordinated cover bidding (Regime 2)
BIC strongly favors coordinated cover bidding (\(\Delta\)BIC = \(-91{,}473\)): FL bids are 28% less dispersed than non-FL bids (\(\sigma_c / \sigma_g = 0.72\)). The price association concentrates in competitive markets (0.126) and vanishes in concentrated ones (\(-0.018\)), consistent with strategic complementarity (\(\hat{\gamma} = 0.69 > 0\)).
Three-stage enforcement pathway
The FL screen requires only participation and outcome data (no bid values). We propose a screen--triage--investigate pathway that exploits a 12.5x gradient in the FL--price association across PBU size quartiles to direct investigative resources where oversight is weakest.