Simulating Collusion: Challenging Conventional Estimation Methods

Abstract

The empirical literature in industrial economics relies on hazard rate models to estimate the probability of death and survival as well as to explain the duration of collusion. Estimations are based on detected and convicted offenses. Detected cartels are, however, a non-random sample of their population of collusive activity. We question whether hazard rate and linear estimation methods derive consistent unbiased estimators explaining collusion. We simulate collusive behavior of industries with different number of firms based on three classical models of collusion, additionally varying four variables of antitrust enforcement. It is the first easily amenable and amendable simulation tool for collusion. The simulation provides a ground-truth data set of undetected and detected cartels; a population and its sample. Applying hazard rate and linear models on the sample fails to deliver consistent unbiased estimates for the population. Controlling for sample and feature selection on the population of all potential offenders does not improve prediction. The use of average treatment effects and average duration bias shows to quantify the magnitude of any bias well; a solution for future research relying on detected cartel cases. JEL Classification: C13 , C63 , D43 , 43 , L41 , L44