Research map / Learning theory and algorithms / Prediction and decision losses

From Individual Experience to Collective Evidence: A Reporting-Based Framework for Identifying Systemic Harms

ICML 2025 · Learning theory and algorithms › Prediction and decision losses · approach: Adaptive testing and experiments · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a reporting-database framework that treats individual harm reports as a sequential hypothesis test for whether subgroups are overrepresented among reports, with a Bonferroni correction and two test variants (sequential Z-test, betting-style e-value test), evaluated on COVID vaccine myocarditis reports and simulated HMDA mortgage denials.

Problem

Individuals reporting adverse experiences with a system have no way to know whether their experience reflects a broader pattern of disproportionate harm to subgroups; existing fairness audits are batch or post-hoc and do not model reporting behavior.

Approach

Test, for every subgroup defined by feature combinations, whether its share of reports exceeds beta times its base preponderance in the reference population, using anytime-valid sequential tests (sequential Z-test or betting-style test with Online Newton Step) plus Bonferroni correction; under reporting-rate assumptions, this lower-bounds relative risk or incidence rate.

Key results

• On real VAERS myocarditis reports, all three tests first flag (M, 18-29) before (M, 12-17); e.g., at beta=2.0 the betting-style test flags at 61 reports (Mar. 23) versus 241 (May 8) for the 12-17 group.
• On simulated HMDA reporting at beta=1.6, the asymptotic Z-test stops fastest (886 reports under Correlated) while the betting-style test identifies groups with higher true relative risk (e.g., 1.82 vs 1.68 for the asymptotic Z-test under Correlated).
• Stopping time rises sharply with beta; at beta=1.8 some algorithm/reporting-model combinations fail to stop within 40,000 steps in some trials.

Measured effects reported in the paper

Surprising findings

• The asymptotically valid Z-test stops much faster but identifies less severely harmed groups in the mortgage simulation under the anti-correlated model (relative risk 1.05 at beta=1.6).
• Stopping time varies by orders of magnitude across reporting models for small changes in beta.

Limitations

• Assumes i.i.d. reports and known base preponderances; relaxation discussed but not analyzed in experiments.
• Flagging depends on reporting-rate assumptions via beta; true incidence cannot be determined from report data alone.
• Mortgage experiment uses simulated reporting models rather than real reports.

Experimental setup

Datasets: VAERS reports (COVID-19 vaccine, myocarditis keyword) CDC VaxView vaccination coverage HMDA 2019 mortgage data (conventional loans, three large lenders)

Benchmarks: VAERS myocarditis reports HMDA mortgage denials (simulated reporting)

Metrics: stopping time (number of reports to first alarm) true relative risk of first-identified group time to identification

Baselines: Asymptotic Z-test Finite-sample Z-test Betting-style test

Builds on: Ville's inequality Online Newton Step betting-style sequential tests

Closest papers by meaning

• Assessing Disparate Impact of Personalized Interventions: Identifiability and Bounds (NeurIPS 2019 · similarity 0.84)
• Learning Triggers for Heterogeneous Treatment Effects (AAAI 2019 · similarity 0.84)
• Participatory ML for Social Harm Should be Constructed, Validated and Reasoned Through First-Person Accounts (NeurIPS 2026 · similarity 0.84)
• Doubly robust identification of treatment effects from multiple environments (ICLR 2025 · similarity 0.83)
• Data Fusion for Partial Identification of Causal Effects (NeurIPS 2025 · similarity 0.83)
• Controlling Counterfactual Harm in Decision Support Systems Based on Prediction Sets (NeurIPS 2024 · similarity 0.82)