Hierarchical Bayesian ML models for recovering advertiser valuations from observed bid data in online ad auctions — an attribution problem under partial observability and selection. The recovered valuations are then plugged into an auction-theoretic model to study counterfactual auction designs.