Single-barrier fixes yield under 0.02% gain in algorithmic recourse model

When an algorithm denies someone healthcare, credit, employment or housing, the path to overturning that decision runs through a chain of obstacles: integrating scattered data, correcting inaccurate records, and navigating institutional channels. A new formal analysis models this recourse process as a series system of 11 empirically parameterized stages spanning three layers—data integration, data accuracy and institutional access—and proves that single-barrier interventions are bounded by the baseline system success, so removing any one barrier produces almost no improvement in overall outcomes.

Under baseline parameters drawn from federal datasets and peer-reviewed algorithmic audit studies, the end-to-end probability of successful recourse is 0.0018%. Within the model, eliminating any single stage yields a gain of less than 0.02%. The result is structural rather than a statement about any particular organization: as long as many low-probability stages remain in the series, dramatic improvement at one bottleneck barely moves overall success.

Factorial decomposition of the model reveals where the leverage actually lies. The three-way cross-layer interaction among data integration, data accuracy and institutional access accounts for 87.6% of the achievable improvement. The author supports this attribution with Shapley value analysis, Sobol sensitivity analysis and bootstrap resampling with 1,000 iterations across prespecified perturbation ranges, all pointing to the same comparative conclusion: coordinated reform across layers, not isolated fixes, drives large modeled gains.

The analysis also probes robustness. Introducing dependence among stage pass-events with a Gaussian-copula analysis, using latent correlations up to ρ = 0.5 while preserving each marginal pass probability, leaves the comparative conclusion intact—single-layer reform still yields far less than coordinated reform. A repeated-attempt topology, however, attenuates the gap, delimiting the model’s domain of validity.

The author is careful about scope. These findings are properties of the modeled system rather than empirical measurements of real recourse pathways, and they yield testable hypotheses for digital-health governance—chiefly that single-layer fairness interventions are structurally unlikely to produce large absolute gains—pending validation against observed recourse data.

Subject of Research: Structural limits of single-barrier reform in algorithmic recourse: A formal series-system model with implications for digital health

Article Title: Structural limits of single-barrier reform in algorithmic recourse: A formal series-system model with implications for digital health

Article References: Demidont, A. C. (2026). Structural limits of single-barrier reform in algorithmic recourse: A formal series-system model with implications for digital health. PLOS Digital Health, 5(10), e0001784. https://doi.org/10.1371/journal.pdig.0001784

Image Credits: AI Generated

DOI: 10.1371/journal.pdig.0001784

Keywords: algorithmic recourse, series-system model, digital health governance, algorithmic auditing, Shapley attribution, Sobol sensitivity analysis, Gaussian copula, factorial decomposition, fairness interventions, decision systems, Structural, limits