Methodology
Data, formulas, sources. Open.
No black box. Everything you see on the page can be independently verified — from public sources or by replicating the calculation.
01. Data and sources
Public sources, each with its own license.
| Source | Content | Range |
|---|---|---|
| Czech Police (Policie ČR) | Accident dataset — location, severity, type, cause, damage | 2015 – present |
| Road and Motorway Directorate (ŘSD) | Road network and traffic-intensity data (traffic counts) | Network current, counts from 2020 and 2025 |
| Prague road administration (TSK hl. m. Prahy) | Traffic counts on segments and junctions of the monitored Prague network | counts 2016 – 2025 |
| Prague Institute of Planning and Development (IPR Praha) | Street centrelines — geometry of TSK segments and junctions (CC BY 4.0) | current |
| Transport Research Centre (CDV) | Unit losses per person killed or injured, updated annually | 2015 – 2024 |
| Weather data | Historical weather from the ERA5-Land reanalysis (Copernicus); forecast from the ICON-EU model (DWD) via Open-Meteo | 2015 – present, 48 h forecast |
| Czech Office for Surveying, Mapping and Cadastre (ČÚZK) | DMR 5G digital terrain model — segment elevation profile in the risk forecast | current |
02. Spatial matching
How an accident finds its segment.
Each accident from the Czech Police is a point on the map. A 15 m band is drawn around the accident, and the accident is assigned directly to the ŘSD traffic census segment the band intersects. If the band intersects more than one, the nearest wins. If it intersects none, the accident has no indicators — there is no search at a greater distance.
The segment length L in the formulas is the length of the ŘSD traffic census segment (DELKA_SU). For motorways and dual carriageways, the length is not counted for both carriageways.
The result is a dataset where every assigned accident is bound to a specific traffic census segment — a precondition for computing all indicators at the segment level.
Prague (TSK network): an accident within 15 m of a TSK network node belongs to the junction. Other accidents are assigned to TSK segments by the same rule as in the rest of the country. Junction intensity is the sum of entering vehicles, so the junction indicator is relative to one million entries. TSK reports intensity for an average working day; it is converted to the annual average daily traffic with a coefficient from TP 189 for local roads.
03. Indicator formulas
Three indicators, each answering a different question.
Unit: accidents per million vehicle kilometres per year. Answers the question "What is the accident rate on a segment relative to vehicle exposure?"
Fatal accidents weighted 130×, seriously injured 70× — the index respects severity, not just count.
Specific values AF, AS, AM for 2015–2024 are in the table below.
CDV unit costs 2015 – 2024 (CZK)
Loss per person as published by CDV: AF = killed, AS = seriously injured, AM = lightly injured. In the H indicator the rate is multiplied by the number of accidents by the most severe outcome. Accidents from 2025–2026 use the 2024 rates. Methodological discontinuity in 2020 – 2021 (introduction of WTP survey).
| Year | AF (killed) | AS (seriously injured) | AM (lightly injured) |
|---|---|---|---|
| 2015 | 20,790,000 | 5,033,600 | 649,800 |
| 2016 | 19,411,000 | 5,094,200 | 668,500 |
| 2017 | 19,784,000 | 5,097,500 | 716,700 |
| 2018 | 22,534,000 | 5,983,000 | 739,700 |
| 2019 | 25,041,000 | 5,567,000 | 809,000 |
| 2020 | 35,021,000 | 5,800,000 | 362,600 |
| 2021 | 58,235,000 | 12,211,000 | 575,600 |
| 2022 | 66,763,000 | 13,847,000 | 655,000 |
| 2023 | 75,000,000 | 16,575,000 | 1,544,000 |
| 2024 | 78,184,600 | 16,002,300 | 749,000 |
04. Relationships between metrics
Spearman rank correlation between RSI, H, RAR, and accident count.
If all indicators expressed the same thing, one would suffice. Each describes a different aspect of a segment, and the degree of mutual similarity changes with the applied filter. The correlation view in the application respects the active filter and lets you verify relationships across different network slices.
- RSI and H are methodologically related — severity weights resemble the CDV coefficients. A very strong correlation is therefore expected and serves as a data-consistency check.
- RAR vs. accident count — the weaker relationship shows that segments with the highest accident counts are not necessarily the riskiest relative to traffic exposure (depends on traffic intensity).
- The By severity switch separates correlations for fatal, serious, light, and property-only accidents.
- The strong RSI–H link is also the reason a TOPSIS composite index is built on top of the indicators — it combines both indicators into a single severity indicator so the same information is not counted twice.
05. Composite index
TOPSIS — one ranking from three independent criteria.
The RSI and H indicators essentially measure the same thing — how severe the outcomes of accidents on a segment are. On the reference sample (Central Bohemian Region, 2023, 416 segments with an injury accident) their rank (Spearman) correlation is ρ = 0.98. Adding them into a single score counts severity twice — and the ranking then floats up segments that stand out in that one property alone.
The TOPSIS method (Technique for Order of Preference by Similarity to Ideal Solution) avoids that: the pair that measures the same thing is first combined into a single severity indicator, and segments are then ranked on three criteria, each describing a different aspect of risk. The result is a single number between 0 and 1 and the ranking that follows from it, where rank 1 goes to the most hazardous segment in the selected set. In the application the index is available as a sortable column in the segment table, as a map colouring option, and as a value in the segment detail.
Three independent criteria
The number of injury accidents relative to how many vehicles pass through the segment and how long the segment is. A busy motorway is therefore not automatically more hazardous than a short road with lighter traffic.
A single severity indicator derived from the RSI and H pair. On the reference sample it stands for 95.3 % of the information both metrics carry jointly — so severity enters the ranking once, not twice.
How many injury accidents occurred on the segment at all, regardless of traffic intensity. Covers accidents with fatal, serious, and light injuries.
- The absolute count of all accidents is not a criterion. It is not related to traffic intensity and is negatively correlated with the RAR indicator (ρ = −0.28 on the reference sample) — including it would shift the ranking toward high-traffic rather than high-risk segments. The count of injury accidents is used instead — the same quantity RAR is built on.
- The index is relative to the active filter. TOPSIS is recomputed over the currently selected set of segments (year, region, severity, road class). Both the rank and the index value change with the filter — it is a ranking within the chosen slice of the network, not an absolute scale. Values from two different filters are not comparable.
- Which segments have an index: those with at least one injury accident and a known traffic intensity from the ŘSD count. The remaining segments have no index. In Prague, TOPSIS is computed separately for TSK segments and for TSK junctions.
The procedure follows our previously published work: doi.org/10.1016/j.ets.2026.100075. The reference figures quoted above (ρ, share of joint variance) come from its analysis of the Central Bohemian Region for 2023.
06. Limitations and caveats
What the page cannot do and what to watch out for.
- Traffic counts: average daily traffic intensity comes from the 2020 and 2025 ŘSD traffic censuses. For each year it is interpolated between the two censuses and held constant outside their range. A zero in the census means “not measured”, not zero traffic — such a census is skipped in the interpolation, and a segment with no measured intensity has no indicators. In years with markedly different traffic (2020 COVID), the value can be skewed.
- Roads without traffic counts: they appear on the map as a grey layer with the accident count, but without RAR, RSI, H, or the TOPSIS index — these cannot be computed without traffic intensity.
- Accident-location accuracy: Czech Police coordinates have an accuracy on the order of single metres. The 15 m band covers most cases, but segments running close to each other can in rare cases be confused.
- Matching in 2D: the matching does not take elevation into account, so at grade-separated junctions an accident can be assigned to a segment on a different level. For about 1 % of accidents, the distance difference between candidate segments is below 1 m.
- Observed period: the period length t ends on the last day for which accident data is available, not at the end of the calendar year. A segment with traffic counts where nothing happened in the selected period has indicators equal to zero.
- Prague and grade-separated locations: matching in the TSK network is also 2D. Accidents at possible grade-separated nodes and crossings are flagged in the application, not resolved. Indicators of TSK segments and junctions are not directly comparable with ŘSD segments, so they have their own rankings.
- CDV methodological discontinuity: the jump in AF between 2020 and 2021 is not inflation but a change in methodology (introduction of the WTP survey). Comparing annual H values across this boundary is invalid — for cross-year comparison use the relative view between segments within the same year.
- Weather data: used for weather-based risk analysis, with spatial resolution on the order of single to tens of kilometres — local phenomena (valley fog, bridge ice) are not faithfully captured. It serves as a proxy for segment-level analysis, not a substitute for a meteorological station.
- Risk forecast: based on the ICON-EU model with a resolution of about 7 km. The weather effect on a segment is estimated from its accident history and, for segments with few accidents, leans on roads of the same class. The forecast shows the change in risk for an average vehicle versus the segment average — it is not an accident forecast.
- Accident clusters (KDE+): the method of Bíl, Andrášik and Janoška (2013, doi.org/10.1016/j.aap.2013.03.003) finds places within a single traffic census segment where accidents concentrate more than chance would explain. It does not change RAR, RSI, or H. Do not compare results (number of clusters, their length and strength) across periods of different length. A cluster close to the end of a segment is labelled “at node” — a segment end is typically a junction.
- Places for inspection: ranks accident clusters by the road geometry at the spot and by the segment indicator. The geometry is derived from the road centreline, not from design documentation, and motorways are left out by default. The result is a list of candidates for an on-site inspection, not a conclusion about accident causes or the need for construction work.
Questions
Get in touch if you'd like to discuss the methodology
Comments, suggestions for extensions, or alternative data sources are welcome.