Fair Incrementality Learning and Conservative Policy Selection without Persistent Identifiers: Calibrated Counterfactual Evaluation for Ads and Job Ranking

Chen Yang (Northwestern University)
Derek Peterson (Brown University)
Lin Feng (University of California, Irvine)
Rachel Adams (University of Colorado Boulder)

Journal of Information Systems and Business Technology
Vol 2 No 4 (2026): Journal of Information Systems and Business Technology

Identifier loss complicates advertising measurement, while job recommenders must balance utility and opportunity. We linked two separate tracks through conservative selection: randomized Criteo Uplift v2.1 for intention-to-treat estimation and FairJob for identifier-free ranking and proxy-group audit. S-HGB achieved test Qini 4.689×10⁻³, top-decile uplift 6.746 points, and top-20% gain 0.974 points; a lower-confidence-bound rule treated 30% and gained 1.007 points. FairJob's best PR-AUC was 0.00833. Proxy penalization raised MRR from 0.574 to 0.585 but widened group disparity from 0.023 to 0.071, so deployment was rejected. Evidence cards quantified local faithfulness without interpreting anonymized fields. Calibration and gating supported identifier-free decisions, but predictive parity did not ensure fair ranking.