Research reviewed November 3, 2026 · v1.2
Michigan-left prevalence: no defensible statewide percentage
The available sources do not establish the share of all Michigan intersections that accommodate left turns using Michigan-left geometry. A reliable estimate requires a current census of complete intersection systems and a matching denominator of left-turn-capable intersections, across state and local roads.
FHWA’s 2007 synthesis reported over 700 directional crossover openings on the state highway system. Openings are not complete intersections. The 2015 MDOT urban-trunkline project explicitly merged multiple boulevard nodes and removed median-turnaround nodes before counting 10,621 intersections. Neither inventory provides the statewide numerator and denominator needed here. Dividing the two totals would be invalid. FHWA synthesis · MDOT RC-1628, pp. 23–25
Historical signalized comparison
The lab starts with FHWA Table 3 intersection-related rates: 1.644 conventional and 1.388 MUT crashes per million entering vehicles. Table 4 supplies the following all-crash shares. Its counts are expected for five years at 60,000 AADT, not raw observations. These inputs are a legacy scenario, not current statewide probabilities. Source: FHWA-HRT-07-033, Tables 3–4.
| Crash category | Standard share | Michigan share |
|---|---|---|
| Rear-end | 46.94% | 68.29% |
| Angle, straight | 24.73% | 12.12% |
| Head-on / left turn | 11.87% | 2.66% |
| Angle, turning | 5.62% | 6.84% |
| Rear-end / left turn | 3.27% | 0.88% |
| Rear-end / right turn | 1.02% | 1.19% |
| Sideswipe, same direction | 0.42% | 0.31% |
| Sideswipe, opposite | 0.25% | 0.20% |
| Head-on | 0.56% | 0.35% |
| Fixed object | 4.36% | 6.13% |
| Overturn | 0.95% | 1.03% |
Do not interpret a share as a probability per trip or as a percent reduction. The sampler normalizes rounding (standard sums to 99.99%). The published severity totals do not fully reconcile; no severity probabilities are inferred. Precise crossover inclusion is unclear, limiting transfer to our full-system boundary.
Recent Michigan U-turn segments: a different crash mix
A 2025 NCDOT-sponsored study examined 7,523 crashes during 2019–2023 at 264 crossover segments belonging to 132 selected Michigan MUT sites. It excluded the central intersection. Crossover controls included 162 signals, 74 stops and 28 yields. This is neither a statewide census nor a matched conventional-versus-MUT comparison. Source: Chakraborty et al., Tables 4.4–4.5 and §4.1.3.
| Published crossover category | Count | Share of 7,523 (calculated) |
|---|---|---|
| Rear-end | 3,778 | 50.22% |
| Sideswipe | 1,746 | 23.21% |
| Angle | 986 | 13.11% |
| Backing | 44 | 0.58% |
| Head-on | 29 | 0.39% |
| Other | 249 | 3.31% |
| Unaccounted for by listed type rows | 691 | 9.19% |
The listed type counts sum to 6,832, leaving 691 unspecified in that table. We retain this residual rather than inventing a category or renormalizing it away. The segment mix is contextual evidence for including merge, queue and access conflicts; it is not merged into the historical full-junction input distribution.
Newer comparisons do not imply one universal reduction
The abstract for Kay et al.’s 2023 signalized study describes 82 MUT and 85 reference intersections. For conversion from undivided approaches it reports crash modification factors (CMFs) of 0.656 for fatal/injury and 0.684 for property-damage-only crashes. For already-divided approaches, the reported factors are 1.423 and 1.668. These are adjusted comparison ratios, not shares of crash types; they depend on the comparison design. The full methods and comparable-site diagnostics were not available for verification, so these values are not implemented as presets. TRB 2023 synthesis, paper 23-01580, p. 175.
Kay et al.’s published 2022 study covers unsignalized intersections. Its injury and property-damage CMFs are not substituted into this lab’s signalized RYG comparison. Primary paper.
“T-bone” and “caught in the turn” need distinct codes
Michigan’s UD-10 coding uses pre-impact travel directions. Opposing approaches with a left turn are coded head-on–left-turn; perpendicular approaches with an approximately right-angle side impact are angle. Side damage alone does not determine the code. Rear-end–left-turn and rear-end–right-turn are separate from ordinary rear-end. Only the first crash type is recorded for a multiple-impact crash. Thus “caught in the turn” is a scenario description, not an extra mutually exclusive crash category. MSP UD-10 Manual, pp. 14–19.
What the implemented processes mean
| Scenario | Placement assumption | Closure assumption |
|---|---|---|
| Rear-end | Time-weighted queued nodes, including return and U-turn queues | Affected queue resource |
| Opposing direct-left | Standard direct-left location; Michigan residual is an external access/violation event | Main junction |
| Angle / turning conflict | Matching through or turning nodes where available | Main junction or U-turn queue |
| Sideswipe / merge | Michigan crossover or approach proxy | Affected queue resource |
| Head-on, object, overturn | External incident at an approach/node | Junction for head-on; affected resource otherwise |
These are explicit scenario mappings, not empirical causal assignments. An incident is not linked to a physically colliding pair of cars. If the relevant movement is absent, the event is flagged as external rather than silently removed. The selected clearance blocks service; a second run propagates the queue, time and fuel effects. Vehicles remain in the cohort. Placement uses the undisturbed traffic trace and stays fixed during the disrupted run; feedback into crash risk and secondary crashes are excluded.
Counts use independent seeded Poisson draws with mean λ = rate × unique entering trips / 1,000,000. One Michigan-left trip counts once. All measured entrants, including trips finished in the drain period, provide exposure. Warmup/tail incident generation is excluded. The 1,000× demonstration changes event frequency and therefore all traffic results; it is prominently labeled. Rate and share inputs remain fixed when traffic controls change, so the model cannot discover an independent safety benefit from a chosen rate advantage.
A calibrated next study would geocode complete intersection systems, distinguish permitted turn movements and signal types, apply common crash influence boundaries, retain all UD-10 categories, collect turning volumes and AADT, and fit exposure-adjusted type/severity models with comparable reference sites. Individual crash narratives would be needed to attribute cause or distinguish queue, U-turn, weaving and return-signal mechanisms reliably.