This section documents the methodological literature that informs the Open Road Risk pipeline. It is not a comprehensive survey of road safety research. It is a targeted review focused on one question: what does the peer-reviewed evidence say about building an exposure-adjusted road collision risk model at national scale from open UK data?
Purpose and scope
Open Road Risk estimates crash risk for approximately 2.17 million OS Open Roads links across England, using STATS19 injury collision data, estimated AADT from sparse DfT AADF counts, and road/network/contextual features derived from open data. The pipeline has three stages:
- Stage 1a — machine-learning AADT estimation from ~8,000 DfT AADF count sites
- Stage 1b — time-zone traffic profile estimation from WebTRIS National Highways sensors
- Stage 2 — Poisson GLM and XGBoost collision risk model at link × year grain
The literature review was designed to answer questions that arise directly from these design choices:
- Which count model family is appropriate for zero-heavy, low-mean integer outcomes?
- How should traffic exposure be handled when most AADT values are estimated rather than observed?
- What does the evidence say about temporal disaggregation of exposure?
- How do spatial autocorrelation and cross-validation design interact?
- What is the evidence base for junction-level risk mechanisms that the link-year model cannot capture?
- When and how should severity be modelled separately from frequency?
- What validation metrics are appropriate for a sparse count ranking model?
- Which methods from the international literature are actually transferable to UK open data?
Papers were selected for extraction if they spoke directly to these questions, used methods comparable to the pipeline’s approach, or provided cautionary evidence about failure modes.
Review process
Each paper was extracted using a structured prompt schema that records:
- response variable, spatial unit, temporal resolution
- exposure treatment and traffic data source
- model family and validation design
- transferability to Open Road Risk’s data stack
- repo-actionable implications
High-priority papers received two independent AI extractions, a cross-audit, and a reconciliation to produce a final record. Lower-priority papers received a single extraction and lightweight sanity check.
The review process is documented in full in literature/prompts/README_literature_extraction.md.
What this review is not
This review does not:
- Cover crash causation or behavioural risk factors (driver behaviour, alcohol, vehicle condition). These are documented as a structural explanatory ceiling in the Validation and Metrics page.
- Survey pedestrian or cyclist safety literature, except where it provides methodological evidence relevant to the collision frequency model.
- Cover the Highway Safety Manual or US AASHTO design guidance as primary sources; specific findings from US studies are included where they transfer to UK open data.
- Constitute a systematic review in the epidemiological sense. Coverage is selective and driven by pipeline questions, not exhaustive search and inclusion criteria.
Literature pages
Each page below synthesises findings across multiple papers on a connected topic. The pages document what papers found; they do not describe the current state of the pipeline. For the pipeline-state view, see the alignment page.
| Page | What it covers |
|---|---|
| Crash Frequency Models | Poisson, NB, zero-inflation, EB shrinkage, overdispersion, serial correlation |
| Exposure and Traffic Volume | AADT elasticity, exposure offset structure, estimated vs observed AADT, temporal exposure disaggregation, argument-averaging bias |
| Spatial Methods and Network Risk | Spatial autocorrelation in crash models, spatial CV design, network point processes, MAUP |
| Junctions and Conflict Structure | Junction risk mechanisms, why the link-year model cannot capture them, derivable proxy features |
| Severity Modelling | Frequency vs severity as separate estimands, ordered response models, joint modelling, post-event variable leakage, STATS19 underreporting |
| Validation and Metrics | Metric taxonomy, balanced accuracy, pseudo-R², CURE plots, spatial CV, temporal holdout, structural explanatory ceiling |
| Transferability and Open Data Limits | Per-paper transferability assessment, UK open-data availability matrix, negative-transfer examples |
Pipeline alignment
The Literature–Pipeline Alignment page consolidates the implications from all seven pages into one place, organised by pipeline stage. It maps each literature recommendation to the current pipeline state and recommended action. This is the page that changes when the pipeline changes; the literature pages themselves are stable.
Literature register
The Evidence Register tracks all extracted papers, their extraction status, active reconciliation work, and the candidate Quarto pages they feed. It is the primary provenance record for the literature review.
Coverage gaps
Several topic areas are not yet well-covered in the reviewed set:
Temporal disaggregation at the model level. Mensah & Hauer (1998), Qin et al. (2006), and Dutta & Fontaine (2020) establish the theoretical and empirical case for temporal exposure conditioning. The pipeline infrastructure (Stage 1b time-zone profiles) exists, but whether conditioning Stage 2 on time-zone fractions materially improves held-out performance has not been formally validated. An ad-hoc diagnostic using core_overnight_ratio showed a modest improvement (~+0.004 R²) under a single seed; this should be confirmed with the 5-seed harness before production adoption.
Spatial validation. No paper in the reviewed set was designed to validate a link-level crash model on a network as large or diverse as Open Road Risk’s. Mahoney et al. (2023) provide the best available methodology guidance for spatial CV, but their simulation used a regular grid with a single crash type. The empirical spatial autocorrelation range for Open Road Risk’s mixed national network has not been measured.
Severity at link-year scale. The severity literature (Boulieri 2016, Gilardi 2022, Michalaki 2015, Quddus 2010) is mostly at aggregate spatial units or event level. The sparse KSI count problem at individual OS Open Roads link-year grain — where most links have zero KSI events in any given year — is not directly addressed in the reviewed set.
Minor road exposure estimation. The Stage 1a AADT estimation problem for unclassified and classified unnumbered roads (approximately 60% of the network) is documented as a gap, but no reviewed paper directly addresses ML-based AADT estimation at this spatial density and road-type composition.