Competition or Collusion? Bid-Level Evidence from Medical Gas Procurement

We examine suspected collusive behavior in Brazil's medical and industrial gases sector, a major case investigated by the Brazilian competition authority (CADE). Using public procurement auction data from 2006–2019, we combine an econometric screening model with an unsupervised machine-learning clustering approach. The econometric analysis indicates higher bids by firms accused of cartel formation during the collusive phase and relatively lower bid levels in the post-cartel subperiods. The machine-learning analysis complements this benchmark by identifying heterogeneous bidding regimes based on bid-intensity measures and distributional screens. In the cartel phase (2006–2010), nearly 80% of auctions fall into a cluster associated with patterns typical of coordinated bidding. Weaker post-cartel screening signals should not be interpreted as evidence of cartel absence, since firms may adapt their strategies to reduce detectable statistical traces. More generally, our results show that even when traditional price-based screening indicators weaken after enforcement, structured statistical patterns consistent with coordination can still be detected using flexible, data-driven methods. Together, these findings illustrate how integrating econometric screens with unsupervised learning can support procurement monitoring and competition policy.