Public entities procure goods, services and construction projects to a value of about 12% of the global GDP (Bosio et al., 2022) and approximately 13 percent of GDP within OECD countries (OECD, 2023). A competitive bidding process on public contracts makes it possible for public authorities to get the best offer based on a combination of price and quality and hence an efficient use of public funds. Therefore, the public procurement process is in many countries highly regulated to ensure open competition, a transparent and fair bidding process and award method. Still, firms engage in unlawful collusive bidding behavior on public contracts to increase profits. Comprehensive studies have found that, on average, cartels increase prices by 15 to 30% (Connor and Bolotova, 2006; Boyer and Kotchoni, 2015; Bolotova, 2009; Froeb et al., 1993). Hence, bidding rings potentially impose a significant cost on taxpayers.
Collusive biding may take many forms and submitted bids could be designed to be consistent with competitive bidding. Those who engage in collusive bidding on public contracts are often innovative and those who survive are perhaps cleverer than those chasing them. This makes detecting bid rigging from bid data extremely difficult and previous studies estimate that the probability of a cartel being caught and convicted is only around 10 to 20% per year (Harrington and Wei, 2017; Combe et al., 2008; Bryant and Eckard, 1991).
Hence, cartels are problematic as they are hard to detect, reduce competition and increase prices. This has driven the development of statistical methods for detecting cartels to support further investigations. As the collection of complementary information (besides the separate bids) from procurements is costly, time-consuming and in some cases almost impossible, an ideal method would enable to test for collusive bidding behavior based on the bids alone.
Aapo Aaltio, Riku Buri,Antto Jokelainen and Johan Lundberg, in their recent paper, focus on cartel detection tests designed for auctions that require only bid data from the auctions. This contrasts with other methods in the literature that often rely on firms’ cost data, which is much harder to obtain. Since most public procurement databases report bids, methods based on bid distributions are well-suited for automated, large-scale cartel screening. The different tests analyzed in this paper are applied on the bidding behavior of two convicted cartels that operated in the Finnish and Swedish asphalt markets during the 1990s and early 2000s.
The paper has two objectives. First, by comparing bidding patterns before and after then cartel investigations, we estimate how the distribution of bids changed after the collapse of the cartel. We find that during the cartel, a large share of bids are within 10% of the winning bid. This clustering of bids is particularly prevalent in the Finnish market. We also observe that during the cartel period, winning bids are isolated, with losing bids typically being at least 1% higher than the winning bid. Together, the clustering of bids and isolated winning bids result in a bimodal distribution of bids during the cartel period. Both of these features, the clustered bids and the missing mass of nearly tied bids, have been proposed as markers of collusion in previous literature (Chassang et al., 2022; Clark et al., forthcoming; Imhof et al., 2018). After the start of cartel investigations, both in Finland and Sweden, the distribution of bids becomes unimodal and the share of bids within 10% of the winning bid decreases. The change is considerably larger in Finland. To support the causal interpretation of our results, we conduct several robustness checks and also conduct a difference-in-differences analysis using data from a control market.
Second, we test the performance of two cartel detection methods, which can be implemented using only information on the distribution of bids. The first method is a distributional regression approach suggested by Clark et al. (2025). This method is based on the observation that while a cartel might find it optimal to leave a gap between the winning bid and the second lowest bid, it does not have similar incentives to manipulate the difference between the other non-winning bids. The method works by comparing two sets of bid differences, where the bid difference is defined as the difference between a bid and the lowest rival bid. The first set of bid differences is calculated from a sample that includes all the bids, whereas the second set is calculated from a sample where the winning bid is excluded. The null hypothesis is that, with competitive bidding, the two distributions should be similar for bid differences close to zero. In both Finland and Sweden, the null hypothesis is rejected for the cartel period. In both cases, consistent with the intuition of the test, we find that during the cartel period, the full set of bid differences has a much lower density close to zero, indicating that the cartel firms avoided leaving bids very close to the winning bid. The null hypothesis is not rejected for the post-investigation period for either of the countries. However, the results for the post-investigation period in Sweden are not as conclusive as they are for Finland.
The second method, developed by Huber and Imhof (2019), uses machine learning to classify tenders as competitive or collusive. As predictors, the machine learning model uses different statistical screens calculated from the distribution of bids. These include, for example, the standard deviation of the bids and the difference between the winner and the runner-up. When the predictive model is trained using the data from the same country, the model correctly classifies about 90% of the tenders with the Finnish dataset and 74% with the Swedish dataset. When the model is trained using data from another country, the prediction rates decrease for both Finland and Sweden. We also find substantial variation in the prediction rates depending on the model specification and the data used to train the model.
To summarize, our study highlights two key findings. First, the Finnish and Swedish cartels caused significant distortions in bid distributions, reinforcing the idea that such patterns can indicate collusion. Second, the effectiveness of detection methods depends on the bidding behavior of the cartel. The Finnish cartel exhibited stronger collusive patterns, making it easier to detect, whereas the Swedish cartel sometimes resembled a competitive market, reducing detection rates. Together, these results underscore that the effectiveness of statistical methods to detect collusive bidding depends on the specific bidding behavior of the cartel.
The full paper is published in IJIO Volume 98, January 2025.