Junctions and Conflict Structure

Why junction risk requires different units, data, and exposure structures from the current link-year model

Open Road Risk models crash frequency at the road link level. This is a practical and defensible design choice at national scale with open data, but it creates a known structural gap: the mechanisms that drive crash risk at junctions are fundamentally different from those that drive risk on mid-link road sections, and most of those mechanisms are not captured by link-level features or a link-level AADT exposure offset.

This page documents what the reviewed literature says about junction risk, identifies which junction-related ideas are implementable with UK open data, and establishes what should be documented as a known limitation of the current pipeline.

Why junctions are different

Road crashes concentrate at junctions. Multiple streams of traffic interact at a single point; conflicts arise from turning movements, opposing through traffic, pedestrian crossings, and sight-line restrictions that have no counterpart on a uninterrupted mid-link section. The spatial unit and the exposure structure both need to change to model this correctly.

Poch and Mannering (1996) make this explicit. Fitting negative binomial models to 63 signalised intersections in Bellevue, Washington — 1,385 intersection approach-year observations — they find that the strongest predictors of total accident frequency are total opposing approach volume (elasticity 2.95) and left-turn volume (elasticity 2.28). These are movement-specific traffic volumes, not segment AADT. A 1% increase in total opposing approach volume is associated with approximately 3% more accidents; the equivalent elasticity on link-level AADT in segment models is typically 0.6–1.0. Junction conflict exposure scales super-linearly with the volume of opposing and turning movements in ways that a simple link AADT cannot capture.

Poch and Mannering (1996) also estimate separate models for rear-end, angle, and approach-turn accidents. Each accident type has a different predictor set. Rear-end accidents respond to lane configuration and sight-distance restrictions; angle accidents respond to signal phase configuration and opposing volumes; approach-turn accidents depend heavily on protected versus permitted left-turn movements. A single total-crashes count model on a road link conflates all of these mechanisms into one coefficient on one exposure variable.

Ziakopoulos and Yannis (2020) review the broader spatial road safety literature and document a finding from joint segment-intersection models: spatial correlations between intersections and their connected segments are consistently more significant than those between intersections alone or between segments alone. The mechanism proposed is common unobserved factors — approach speed, signal timing, road surface condition — that are shared between an intersection and the immediately adjacent road sections. This means that even in a link-level model, links adjacent to junctions will have correlated residuals that are driven by the junction’s unmodelled risk, not by the link itself.

What junction risk depends on: the data picture

The table below maps the key junction-risk mechanisms identified across the reviewed papers to their data availability in the UK open-data stack.

Mechanism Paper evidence Required data UK open-data availability
Opposing approach volume Poch & Mannering 1996 — elasticity 2.95 Per-approach AADT at intersection Not available nationally. DfT AADF counts per link; no turning-movement counts.
Left-turn and right-turn volume Poch & Mannering 1996 — elasticities 2.28 / 0.92 Directional turning counts Not available nationally.
Signal phasing (2-phase vs 8-phase; protected left turn) Poch & Mannering 1996 — significant predictors Signal control inventory with phase detail Not available. OSM records signal presence but not phase/control type.
Signalised intersection density (per km) Al-Omari 2021 — most consistently significant predictor across all 8 Florida road classes; Wang 2015 — +37% per signal/km Count of signalised junctions per segment length Partially available. OSM highway=traffic_signals provides signal locations; coverage is uneven nationally but better in urban areas. Junction count derivable from OS Open Roads topology.
Access point density (per km) Al-Omari 2021 — consistently significant; Wang 2015 — +11% per access/km Count of access points (driveways, side roads) per segment length Partially available. Minor road connections to a link derivable from OS Open Roads network topology. True driveway accesses are not in OS Open Roads.
Sight-distance restriction Poch & Mannering 1996 — significant positive predictor Site survey / design plan data Not available nationally. Curvature and grade proxies are imperfect substitutes.
Road network pattern (grid vs tree structure) Wang 2015 — lollipop networks ~100% more crashes than grid Network topology / betweenness centrality Derivable from OS Open Roads using betweenness centrality. Already present as a candidate feature.
Junction leg count / complexity Roll 2026; Ziakopoulos 2020 Node degree from road network graph Derivable from OS Open Roads topology.
Traffic control type (signal / roundabout / priority) Ziakopoulos 2020; Roll 2026 Junction control classification Partially available from OSM junction tags; form of way in OS Open Roads captures roundabouts.
Pedestrian crossing presence Roll 2026 — marked vs unmarked crossing as model stratifier Crossing inventory Partially available from OSM; coverage sparse and inconsistent nationally.

The gap is stark. The data that drive junction-specific crash risk — turning movement volumes, signal phasing, sight-distance surveys, approach-level geometry — are not available in any UK national open-data source. They require either proprietary traffic engineering data (Highways England network management systems) or intersection-level field surveys neither of which is in scope for Open Road Risk.

What is derivable: proxy features

Three categories of junction-related proxy features are derivable from OS Open Roads and OSM at national scale, albeit imperfectly.

Network topology: betweenness centrality

Wang et al. (2015) study Shanghai suburban arterials at TAZ level and find that road-network topology matters for crash frequency. Links embedded in tree-structure (lollipop) networks, which force all local traffic onto a single arterial with no parallel collector routes, have roughly twice the crash frequency of equivalent links in grid networks. The mechanism is captured by betweenness centrality: a link that lies on nearly every shortest path through its neighbourhood carries more exposure than its AADT alone reflects, because it serves as the only conduit for a catchment area.

Betweenness centrality is already a candidate feature in Open Road Risk Stage 1a and Stage 2. Wang et al. (2015) provide conceptual support for it as a junction-conflict proxy beyond its use as an AADT estimator. However, note that Gilardi et al. (2022) find betweenness centrality is not significant in their Leeds network-lattice model after controlling for road type and exposure — suggesting it may be absorbed by other features in a well-specified model. Its contribution deserves a diagnostic test rather than assumption.

The Wang 2015 results should be treated cautiously for other reasons: they are from a single-year in-sample TAZ-level model in Shanghai with no holdout validation. The directional findings (junction density and network tree structure increase crash frequency) are consistent with physical expectations and the broader literature, but the coefficient magnitudes are not transferable.

Junction-proximity as a binary or continuous feature

Baddeley et al. (2021) observe from network point-process theory that crashes concentrate at network vertices (junctions), and that standard continuous-network models must be modified to handle this — essentially because junctions create discrete spatial concentrations of crash risk that cannot be represented by a smooth intensity function along link edges.

For Open Road Risk’s link-level model, the practical analogue is a junction-proximity feature: the distance from a link’s midpoint to the nearest OS Open Roads junction node (degree ≥ 3), or equivalently a binary flag for links that are themselves junction connectors (slip roads, roundabout circulating carriageways). Links adjacent to junctions absorb some of the crash risk from the junction itself. A distance-to-nearest-junction feature could help the model distinguish genuinely high-risk mid-link sections from links that appear risky because they are close to a junction.

The exposure problem at junctions

Open Road Risk uses log(AADT × link_length_km × 365 / 1e6) as its exposure offset. At a mid-link road section, this is a reasonable approximation: the expected number of crashes is roughly proportional to the number of vehicle-kilometres of travel on that section.

At a junction, this breaks down in two ways.

First, the relevant exposure is not vehicle-kilometres on the approach link but the product of conflicting traffic volumes. A T-junction with 1,000 vehicles per day on the major road and 100 on the minor road has a conflict exposure that depends on both volumes simultaneously. The link-level offset captures only the major-road AADT and misses the minor-road conflict contribution entirely.

Second, junctions are points, not lengths. The AADT × length offset normalises by segment length — a longer link accumulates more vehicle-kilometres of exposure. At a junction, length is not the relevant denominator; the conflict happens at a single location. A very short junction connector link will have a very small AADT × length exposure and therefore appear to have an anomalously high crash rate per vehicle-kilometre, not because it is especially dangerous but because the exposure denominator is wrong for a point-like conflict location.

Hauer et al. (2001) make this explicit in the EB tutorial: intersections are treated as separate entities with different SPF forms (accidents per year, not accidents per km-year), and the overdispersion parameter φ is estimated per intersection-type, not per-unit-length. The link-year model cannot directly handle this — but it does imply that the highest-ranked links in the current risk percentile may include many junction-connector links whose apparent high risk is partly an artefact of the length-denominator mismatch.

Scale and feasibility constraints

Roll et al. (2026) develop a pedestrian SPF for Oregon urban intersections. The data engineering alone — classifying 64,000 intersections by leg count, control type, and marked-crossing presence; contracting complex divided-road intersections to single nodes; integrating signal inventories from five data sources; estimating pedestrian AADPT via a random-forest data-fusion model — required a substantial research effort for a single US state. Replicating this for the full national OS Open Roads network would be a major undertaking well outside Open Road Risk’s current scope.

The Roll 2026 paper is most useful for two things. First, it confirms the general principle that junction risk needs different spatial units, different exposure variables, and different model structures from the current link-year model. Second, it provides a template for the data-engineering challenges involved if a separate junction layer is ever added: intersection contraction, control-type classification, and exposure estimation are each non-trivial steps.

Note

Roll et al. (2026) also find that their final pedestrian SPFs use only vehicle AADT and pedestrian AADPT as predictors — the complex contextual feature set added little beyond exposure. This is consistent with Al-Omari (2021)’s finding of no significant difference between simple and full SPF in-sample performance. For Open Road Risk, this supports running an exposure-only baseline comparison before investing in complex junction feature engineering.

Open Road Risk alignment

Junction issue Literature evidence Current pipeline Gap / recommended action
Turning/opposing volumes Poch & Mannering 1996 — elasticity 2.95–2.28 Not available; AADT only Document as a fundamental limitation of link-level exposure.
Signal phase / control type Poch & Mannering 1996 Not available Document; OSM signal presence is the maximum achievable.
Signalised junction density Al-Omari 2021 — most significant predictor; Wang 2015 Not currently in pipeline Candidate feature: count junctions (degree ≥ 3) per link length from OS Open Roads; optionally signalised variant from OSM.
Access point density Al-Omari 2021 — consistently significant Not currently in pipeline Candidate feature (partial): minor-road connections per km derivable from OS Open Roads topology; true driveway accesses not available.
Network tree structure / betweenness Wang 2015 Betweenness centrality already in pipeline as Stage 1a / Stage 2 candidate Diagnostic test of whether betweenness adds value over road type + AADT in Stage 2 GLM.
Junction proximity Baddeley 2021; Ziakopoulos 2020 — segment-intersection correlation Not currently in pipeline Candidate feature: distance from link midpoint to nearest OS Open Roads junction node.
Exposure denominator mismatch at junctions Hauer 2001 — intersections need different SPF form Link-level offset treats all links identically Document; add flag for junction-connector links; consider sensitivity analysis removing or up-weighting junction links.
Speed limit as junction-density confound Al-Omari 2021 — negative coefficient is confound, not causal OSM speed limit in pipeline Documentation note: negative speed-limit coefficient proxies for low junction density; do not interpret causally.
Pedestrian/cyclist junction risk Roll 2026 Not modelled Future work: separate junction layer with pedestrian exposure if data becomes available.

What not to do

Two failure modes are worth calling out explicitly.

Do not import coefficient values from these papers. Poch and Mannering’s elasticities are from 63 selected operationally deficient intersections in 1990s Bellevue. Al-Omari’s SID coefficients are from Florida suburban roads. Wang’s signal-density coefficient is from Shanghai TAZ-level data with no holdout validation. The directions of these effects are plausible and transfer as qualitative evidence; the magnitudes do not.

Do not add junction-proximity features without a diagnostic test first. Junction density and betweenness centrality are correlated with road classification and AADT. In Gilardi et al. (2022), betweenness centrality is not significant after road type and exposure are included. Any junction feature added to Stage 2 should be tested for collinearity with existing features (road classification, AADT, link length) before inclusion, and evaluated on a held-out validation set rather than in-sample fit alone.

References

ID Citation
LIT-003 Al-Omari, M.M.A. (2021). Crash analysis and development of safety performance functions for Florida roads in the framework of the context classification system. MSc thesis, University of Central Florida. Available: stars.library.ucf.edu/etd2020/633
LIT-004 Baddeley, A., Nair, G., Rakshit, S., McSwiggan, G. & Davies, T.M. (2021). Analysing point patterns on networks — a review. Spatial Statistics, 42, 100435. DOI: 10.1016/j.spasta.2020.100435
LIT-015 Hauer, E., Harwood, D.W., Council, F.M. & Griffith, M.S. (2001). Estimating safety by the empirical Bayes method: a tutorial. National SPF Summit, Chicago.
LIT-012/013/014 Gilardi, A., Mateu, J., Borgoni, R. & Lovelace, R. (2022). Multivariate hierarchical analysis of car crashes data considering a spatial network lattice. JRSS-A, 185(3), 1150–1177. DOI: 10.1111/rssa.12823
LIT-044 Poch, M. & Mannering, F. (1996). Negative binomial analysis of intersection-accident frequencies. Journal of Transportation Engineering, 122(2), 105–113.
LIT-028/045 Roll, J., Anderson, J. & McNeil, N. (2026). Developing a pedestrian safety performance function for Oregon. FHWA-OR-RD-26-06.
LIT-030 Wang, X., Yuan, J., Schultz, G.G. & Meng, W. (2015). Investigating safety impacts of roadway network features of suburban arterials in Shanghai, China. TRB Annual Meeting.
LIT-031/041 Ziakopoulos, A. & Yannis, G. (2020). A review of spatial approaches in road safety. [Year and DOI not confirmed — verify before formal citation; active reconciliation pending.]