Partial Identification and Mergers

Inference on partially identified parameters has now become a standard tool kit for econometricians. Econometric models and properties of available data sets are such that it is often impossible to point-identify parameters of interest even if the sample size is infinity and the exact data generating process is available to econometricians without sampling noise. Numerous papers have been written on the issue, and Manski (2003) is probably a good starting point to have an idea about the fundamental issues there. Manski (2011) extends the analysis, and discusses policy choice when a planner faces ambiguity in policy outcomes. In this note, we extend Manski’s insight to the problem of merger simulation by incorporating uncertainty about post-merger conduct into a simple simulation model. At a broad level, merger review by antitrust agencies requires the prospective evaluation of two types of likely merger effects. Unilateral effects refer to the incentives for the merged firms to unilaterally raise price in the absence of disciplining competition from the other firm. Coordinated effects refer to how the merger is expected to change the conduct, for example the opportunity and incentive to engage in tacit or explicit collusion, of the post-merger industry overall. In its typical implementation, merger simulation maps pre-merger estimates of demand and firm costs into forecasted, post-merger price changes via assumptions about post-merger costs and firm conduct. Since the assumption about conduct is almost always Nash in prices or quantities, this implementation of the simulation methodology is only useful, at best, for assessing unilateral effects.1 Coordinated effects must then be considered separately and, necessarily, in a non-integrated way (Kovacic et al. 2006, 2009). Our focus of attention in this note is on the problem of uncertainty regarding overall postmerger conduct, i.e. integrated unilateral effects and coordinated effects, and the complications it entails for policy choice. In order to isolate this aspect of the problem, we work with a simple "conduct parameters" model in the spirit of Bresnahan (1982) and Lau (1982). This model allows integration of non-Nash conduct into the simulation but does not, under the functional restrictions imposed by Bresnahan (1982) and Lau (1982), face issues of partial identification of