Literature–Pipeline Alignment

Where the evidence base meets the current pipeline

Consolidated mapping of literature evidence to the current Open Road Risk pipeline: what is implemented, what is pending, and what the literature recommends for each stage.

This page consolidates the pipeline-state implications from all seven literature review pages into one place. It is the page that changes when the pipeline changes. The individual literature pages document what papers found; this page documents where that evidence leaves the current implementation.

The structure follows pipeline stages, with a section for cross-cutting concerns.

How to read this page

Each table has four columns:

  • Requirement — what the literature collectively recommends
  • Literature basis — which page(s) and paper(s) support the recommendation
  • Current pipeline — the actual current state
  • Gap / action — what remains to be done, with effort indication

Actions are graded: documentation note (lowest disruption) → diagnostic → small pilot → candidate feature → production change (highest disruption). The literature rarely justifies production changes from a single paper; most recommendations are diagnostic first.

Stage 1a — AADT Estimation

Requirement Literature basis Current pipeline Gap / action
AADT coverage: observed counts as ground truth, ML fills the gap Exposure: Roll 2026 three-tier hierarchy; Jayasinghe 2019 DfT AADF (~0.4% of links) → Stage 1a ML estimate for all ~2.17M links Documentation: document that ~96% of AADT is estimated, not observed
Low-AADT links hardest to estimate; geometry dominates below 1000 vpd Exposure: Jayasinghe 2019 (RMSE 193–412% for lowest AADT band); Huda 2024 (R² drop 0.009 when AADT removed at low-volume) Stage 1a trained on AADF without low-AADT stratification Diagnostic: report Stage 1a CV error separately by road class and AADT band
Application sanity checks on full-network predictions Exposure: Roll 2026 (XGBoost produced negative AADT; NB produced implausible maxima; CV metrics did not reveal this) CV metrics reported; full-network distribution not checked by road class Diagnostic: compare distribution of predicted AADT by road class and rural/urban against AADF observations
Learning-curve diagnostic for sparse-count validation Exposure: Jayasinghe 2019 (~40 calibration points → RMSE < 30%) Not run Diagnostic (low effort): plot Stage 1a CV error vs number of AADF sites in each road class
Centrality features support AADT estimation Transferability: Jayasinghe 2019; Gilardi 2022 Betweenness centrality already in Stage 1a feature set No gap — document as confirmed by literature

Stage 1b — Time-Zone Profiles

Requirement Literature basis Current pipeline Gap / action
Temporal disaggregation improves crash model performance over AADT-only Exposure: Dutta & Fontaine 2020 (20–38% MSPE improvement average-hourly vs AADT on Virginia freeways); Sung 2024; Mensah & Hauer 1998 (argument-averaging bias theory) Stage 1b produces time-zone fractions per link; these are not currently used in Stage 2 Candidate feature: join core_overnight_ratio from timezone_profiles.parquet to Stage 2 training data and test as a feature
Average-hourly profiles outperform raw hourly (noise in raw data degrades performance) Exposure: Dutta & Fontaine 2020 (23% of raw hourly observations failed quality checks) Stage 1b already builds smoothed time-zone profiles rather than using raw hourly data No gap — current approach is consistent with this finding; document
Argument-averaging bias: AADT underestimates SPF by ~5–8% for typical β Exposure: Mensah & Hauer 1998 correction factor w Not quantified Diagnostic: estimate correction factor w using free-elasticity diagnostic β and Stage 1b CV(q) per road class
Function-averaging: combining daytime/nighttime in one SPF loses information Exposure: Mensah & Hauer 1998; Qin 2006 Single annual model; no time-of-day stratification Documentation: note as known limitation; Stage 1b profiles are the infrastructure for future temporal conditioning
Dutta improvement magnitude is upper bound for Open Road Risk Exposure — Documentation: note that estimated profiles (Stage 1b) will produce smaller gains than observed sensor profiles

Stage 2 — Collision Risk Model

Model family

Requirement Literature basis Current pipeline Gap / action
Run posterior predictive zero check before deciding on model family Crash frequency: Pew 2020 Not yet run Diagnostic (low effort): sample from fitted Poisson GLM; compare predicted zero rate to observed
NB GLM is priority before ZINB; π ≈ 0 means overdispersion dominates Crash frequency: Pew 2020 Poisson GLM Diagnostic → candidate: fit NB GLM; compare held-out pseudo-R² to Poisson baseline
Equate random effect structures when comparing model families Crash frequency: Pew 2020 N/A Apply when NB vs ZINB comparison is run
Facility stratification: per-family models reduce overdispersion and enable per-family EB weights Crash frequency: Chengye 2013 (MSPE −24% from ramp split); Al-Omari 2021 Diagnostic v1 (risk_scores_family.parquet) exists Validation: run grouped or temporal holdout on per-family models before production
Single-vehicle and multi-vehicle crashes have opposing flow relationships; combining inflates function-averaging bias Crash frequency: Qin 2006; Mensah & Hauer 1998 Total injury collisions combined Documentation: note as known limitation; SV/MV split diagnostic is a candidate action

Exposure offset

Requirement Literature basis Current pipeline Gap / action
Log-offset of AADT × length is supported for most road classes Exposure: Gilardi 2022; Hauer 2001; National Highways 2022 Fixed offset log(AADT × link_length_km × 365 / 1e6) No gap — document as literature-supported
Test AADT elasticity as free covariate; sub-linear likely for some classes Exposure: Aguero-Valverde 2008 (0.63–0.71); Wang 2009 (1.2–1.9 motorway); Al-Omari 2021 (0.39–0.63 dense urban) Elasticity constrained to 1.0 via offset Diagnostic: fit Stage 2 GLM with log(AADT) and log(length) as free covariates; report estimated elasticities by road class
Exposure uncertainty not propagated into Stage 2 rankings Exposure: estimated vs observed AADT gap Stage 2 treats estimated AADT as observed Documentation: document as first-class limitation; EB shrinkage partially absorbs it for sparse links
Gao 2024 no-exposure model is a cautionary negative example Exposure; Transferability Exposure offset implemented Documentation: cite as documented cautionary contrast

Empirical Bayes shrinkage

Requirement Literature basis Current pipeline Gap / action
Per-family overdispersion parameter φ from NB regression for EB weights Crash frequency: Hauer 2001; Chengye 2013 Global method-of-moments k Candidate: per-family NB φ for v2 EB weights
Full EB procedure: sum year-specific μ_t across years Crash frequency: Hauer 2001 equation 7 Year-specific AADT available Candidate: implement full EB summing annual SPF predictions
Crude KSI ranking unreliable without shrinkage Severity: Boulieri 2016 (smoothing reorders high-severity rankings substantially) EB shrinkage for total counts; no severity-split Pilot: extend EB shrinkage to KSI sub-band

Features

Requirement Literature basis Current pipeline Gap / action
core_overnight_ratio from Stage 1b: ad-hoc diagnostic shows ~+0.004 R², correct sign Exposure Not yet added to production Candidate feature: add core_overnight_ratio join from timezone_profiles.parquet; confirm with 5-seed harness
late_evening_frac shows unexpected sign; collinearity with road class suspected Exposure Not in pipeline Do not add until collinearity with road class is resolved
Junction density (nodes degree ≥ 3 per km) is a consistently significant predictor Junctions: Al-Omari 2021; Wang 2015 Not currently in pipeline Candidate feature: count junction nodes per link length from OS Open Roads topology
Junction-proximity (distance to nearest junction node) Junctions: Baddeley 2021; Ziakopoulos 2020 Not in pipeline Candidate feature: distance from link midpoint to nearest OS Open Roads junction node
Betweenness centrality: test whether it adds value over road type + AADT Junctions: Wang 2015 supports it; Gilardi 2022 finds it insignificant after road type controlled Candidate feature Diagnostic: collinearity check against road class and AADT before adding to production
Speed limit is a road-type proxy, not a direct safety predictor Junctions: Al-Omari 2021 negative coefficient is confound OSM speed limit in pipeline Documentation: note that negative speed-limit coefficient proxies for low junction density
HGV proportion supports inclusion Severity: Michalaki 2015 (strong severity predictor) Candidate feature (AADF HGV proportion) Documentation: confirm it is a road-level proxy, not crash-level variable

Spatial structure

Requirement Literature basis Current pipeline Gap / action
Spatial autocorrelation in residuals is present and biases coefficient SEs Spatial: Aguero-Valverde 2008; Gilardi 2022; Wang 2009 (UK motorways) Not modelled Diagnostic: Moran’s I on Stage 2 GLM residuals (sampled 10k–50k links, first-order adjacency)
CAR spatial model is computationally infeasible at 2.17M links Spatial Not attempted Documentation: note as known limitation; Moran’s I is the feasible alternative
Geographic residual mapping to identify persistent high-residual corridors Spatial: Aguero-Valverde 2008 Not yet done Diagnostic: choropleth map of Stage 2 GLM residuals on OS Open Roads geometry for a pilot area
Do not use planar KDE or Euclidean Moran’s I on crash point locations Spatial: Baddeley 2021 Not currently used Documentation: note constraint for any future crash-point visualisation
Junction-segment spatial correlations are the strongest spatial dependencies Spatial: Ziakopoulos 2020 Not explicitly modelled Documentation / candidate feature: junction-proximity feature addresses this indirectly

Severity

Requirement Literature basis Current pipeline Gap / action
Frequency and severity are different estimands; separate models warranted Severity: Quddus 2010; Michalaki 2015; Ma 2019; Savolainen 2011 Single count model (all injury combined) Documentation: document as known design choice; plan severity layer as future work
KSI and slight crashes have different predictor sets Severity: Wang et al. 2011 (lanes significant for slight only; grade significant for both) Not modelled separately Pilot: separate KSI and slight diagnostic models
Joint slight/KSI model substantially improves KSI estimation Severity: Boulieri 2016 (ρ ≈ 0.74); Gilardi 2022 (ρ_φ ≈ 0.83–0.90) Not implemented Future work: joint Bayesian model after EB shrinkage pilot
Severity-weighted composite (Gao 2024 weights 1/2/3) conflates frequency and severity Severity Not used Documentation: note as design approach to avoid
STATS19 underreporting: slight injuries ~75% under-reported Severity: Savolainen 2011 citing Elvik & Myssen 1999 Inherits reporting limitation Documentation: document as known limitation of the outcome variable
Post-event STATS19 variables must not enter Stage 2 Severity: full leakage catalogue Collision-derived variables excluded per repo dossier Documentation: link the full leakage catalogue explicitly
Congestion index is insignificant for crash frequency (M25 null result) Severity: Quddus 2010 Not in production Documentation: note null result as caution against prioritising congestion features

Validation

Requirement Literature basis Current pipeline Gap / action
Grouped-link CV controls within-link temporal leakage but not spatial autocorrelation Validation Grouped-link CV implemented Documentation: record distinction explicitly in validation documentation
Temporal holdout (hold out 2023–2024; train on 2015–2022) Validation: Quddus 2007; Chengye 2013 Not yet implemented Diagnostic (straightforward): add temporal holdout as a second validation split
Spatial CV with exclusion buffer matching residual autocorrelation range Validation: Mahoney 2023 (V-fold only 2% reliable; spatial CV ~60%) Not implemented Diagnostic → pilot: variogram first, then police-force holdout
Police force area holdout as practical spatial CV approximation Spatial; Validation: Mahoney 2023 Not implemented Pilot: hold out one force area; evaluate Stage 2 performance
Balanced accuracy (pool confusion matrices across folds, do not average) Validation: Brodersen 2010; Gilardi 2022 Not yet implemented Diagnostic: implement after choosing classification threshold
AccHR@k ranking quality metric Validation: Gao 2024 Not yet implemented Diagnostic: compute for top-1% and top-5% predicted links
CURE plots by AADT quantile and link-length quantile Validation: Roll 2026; Dutta 2020 Not yet implemented Diagnostic: 50-quantile bins; in-sample only
Posterior predictive zero check Validation; Crash frequency: Pew 2020 Not yet run Diagnostic (low effort): should precede NB vs ZINB decision
Exposure-only baseline comparison Validation: Roll 2026 Not yet run Diagnostic: compare full feature model against exposure-only NB/Poisson
Cluster-robust standard errors grouped by link_id Validation: Quddus 2007; Savolainen 2011 Not implemented Diagnostic → documentation: compute ACF on high-crash links first; if lag-1 ACF > 0.15, add cluster SEs
Serial correlation ACF diagnostic on high-crash links Validation: Quddus 2007 Not run Diagnostic (low effort): sample 500–1000 highest-crash links
Structural explanatory ceiling: road-environment models cannot explain behavioural factors Validation: Roshandel 2015 (~93% of crash causation is behavioural) Not documented Documentation: contextualise held-out R² of ~0.32 as consistent with the ceiling

Data and Transferability

Issue Literature basis Current pipeline Gap / action
UK geography ≠ UK data availability Transferability — Documentation: Gao 2024 and Balawi 2024 cited as cautionary negative-transfer examples
Lane width, shoulder width, driveway density are unavailable nationally Transferability: Huda 2024; Chengye 2013 Not in pipeline Documentation: note as unavailable; OS MasterMap or HE data would be out of scope
CART threshold derivation should use UK data, not US thresholds Transferability: Huda 2024 US-derived thresholds not used Documentation: if CART or tree-based thresholds are introduced, derive from Open Road Risk data
STATS19 CF/RSF 2024 structural break Transferability; Severity: DfT 2024/2025 guidance Collision-derived fields excluded from Stage 2 Documentation: note break and its implications for trend analysis
MAUP: OS Open Roads link is a defensible unit; network-lattice MAUP less severe than zone MAUP Spatial: Gilardi 2022 MAUP sensitivity test OS Open Roads links used throughout Documentation: present risk percentile as one view at one spatial resolution
Hotspot rankings sensitive to model choices, time periods, road-user types Spatial: Ziakopoulos 2020 Not documented explicitly Documentation: add caveat to production risk percentile description

Summary: Priority Actions

The table below lists all actions in priority order, combining effort and information value.

Priority Action Type Stage Effort
1 Posterior predictive zero check on Stage 2 Poisson GLM Diagnostic S2 Low
2 Free-elasticity diagnostic: log(AADT) and log(length) as free covariates Diagnostic S2 Low
3 Temporal holdout: hold out 2023–2024; evaluate Stage 2 on unseen years Diagnostic S2 / validation Low
4 Moran’s I on Stage 2 GLM residuals (sampled ~10k–50k links) Diagnostic S2 / spatial Low
5 Stage 1a CV error by road class and AADT band Diagnostic S1a Low
6 Stage 1a full-network sanity checks by road class Diagnostic S1a Low
7 ACF diagnostic on high-crash links; add cluster-robust SEs if ACF > 0.15 Diagnostic S2 Low
8 core_overnight_ratio feature addition: 5-seed harness confirmation Candidate feature S2 Low–medium
9 NB GLM diagnostic: compare held-out pseudo-R² to Poisson baseline Diagnostic S2 Medium
10 CURE plots by AADT quantile and link-length quantile Diagnostic S2 / validation Medium
11 Exposure-only baseline comparison Diagnostic S2 / validation Medium
12 Junction density feature (nodes degree ≥ 3 per km) Candidate feature S2 Medium
13 Junction-proximity distance feature Candidate feature S2 Medium
14 Police force holdout as practical spatial CV Pilot S2 / spatial Medium
15 EB shrinkage extended to KSI sub-band Pilot S2 / severity Medium
16 Per-family NB overdispersion parameter for EB weights Candidate S2 / EB Medium
17 Balanced accuracy and AccHR@k implementation Diagnostic Validation Medium
18 Argument-averaging correction factor w computation Diagnostic S2 / S1b Medium
19 Temporal holdout validation for per-family models Validation S2 Medium
20 Separate KSI and slight diagnostic models Pilot S2 / severity High