Official Statistics

Linking police and health data on road collisions: hospital admissions data - methodology annex

Updated 24 September 2026

Applies to England

Overview

This methodology note accompanies the publication of a feasibility study exploring the linkage of police recorded road casualty data from the STATS19 dataset with hospital admission records, providing further detail of the linkage methodology used.

The linkage identifies records that are likely to refer to the same road traffic casualty in two independent administrative datasets:

• the Department for Transport’s STATS19 road casualty dataset
• NHS England’s Hospital Episode Statistics Admitted Patient Care (HES APC) dataset

The datasets do not contain a common person-level identifier. It was therefore not possible to link records deterministically using a unique identifier. Instead, a probabilistic linkage method was developed.

As a probabilistic linkage, it was not possible to achieve a definitive match; some quality assurance was done and this is outlined below, together with a summary of limitations of the approach. Further work is required to assess the robustness of, and where possible, further improve the linkage achieved. Nonetheless, this work is considered to establish the feasibility of linkage of the two datasets, for population-level analyses.

Data used for linkage

STATS19

The STATS19 linkage file contains one record for each casualty. The variables used for linkage are:

• age
• sex
• casualty severity
• casualty type
• collision date and time
• collision location coordinates
• casualty postcode, where available

Additional variables were derived from the STATS19 coding to improve comparability with HES APC. These include harmonised casualty-type groups and an indicator of whether another vehicle was involved in the collision.

HES APC

The HES APC linkage file is restricted to admissions likely to relate to road traffic collisions.

Records are retained where:

• the primary external cause code indicates a transport accident, using ICD-10 V-codes (V01 to V89, excluding V81 which relates to railway trains)
• the admission method indicates an emergency admission
• the record represents the first episode of care within the admission

The linkage file includes age, sex, admission date and postcode, together with transport-related classifications derived from ICD-10 external cause codes including casualty type and collision type (whether another vehicle was involved).

To reduce duplicate observations, admissions occurring within 14 days of a previous transport-related admission for the same individual are excluded. This is intended to minimise multiple admissions relating to the same collision event.

Matching process

Following the preparation of the data files for linkage, the linkage method has 3 main stages:

• generating candidate record pairs
• comparing candidate pairs by calculating probabilistic linkage scores
• selecting the final links, based on the candidate pair with the highest score

Generating candidate record pairs

A blocking approach reduces the number of record pairs requiring detailed assessment. A HES APC admission and a STATS19 casualty are retained as a candidate pair only where all the following conditions are met:

Variable	Matching rule
Sex	Exact agreement required
Age	HES APC age within 2 years of STATS19 age
Date	Admission date is the collision date or the following day

This stage retains plausible matches where there may be a small difference in recorded age or a delay between injury and hospital admission, but ensures the number of candidates is computationally manageable with available processing resources.

Comparing candidate pairs

Each candidate pair is compared using variables available in both datasets.

Demographic characteristics

Sex is used as an exact-match blocking variable. Age is assessed during both candidate generation and probabilistic scoring.

Date of collision and admission

The interval between the collision date and admission date is calculated. An admission on the collision date provides stronger evidence of a match than an admission on the following day.

Casualty and collision characteristics

The comparison includes:

• broad casualty class (driver, passenger or pedestrian)
• casualty or road-user type
• whether another vehicle was involved

Agreement across these variables provides additional evidence that the records relate to the same incident.

Injury severity

The reported STATS19 severity category is used to prioritise casualties coded as seriously injured within STATS19 (for example, it is assumed that where there are multiple casualties with similar characteristics within the same collision, it is more likely that one recorded as serious will have been admitted to hospital compared with one recorded as slight).

Postcode agreement

Where postcode is available in both datasets, spaces are removed and the postcode strings are compared. Exact and near agreement are identified using Levenshtein edit distance.

Geographic proximity of collision to home

A National Statistics Postcode Lookup is used to obtain coordinates for the HES residential postcode. The distance between the residential postcode centroid and the STATS19 collision location is then calculated. A shorter distance provides stronger evidence of a match as within STATS19 data, most collisions are shown to occur close to the residence of the casualty. This is used for records where a valid postcode is not recorded in STATS19 (as postcode provides the strongest evidence to determine a match, and this is considered the next best alternative of the variables available).

Calculating the probabilistic linkage score

Each candidate pair receives a weighted score that reflects agreement and disagreement across the matching variables. Positive weights are assigned where records agree on key characteristics, while negative weights are assigned where they disagree. The score includes contributions from:

• age agreement
• date agreement
• casualty class agreement
• casualty type agreement
• other-vehicle involvement
• injury severity (whether STATS19 record is coded serious)
• postcode agreement
• geographic proximity

Variables with greater discriminatory value, particularly postcode agreement and geographic proximity, receive larger positive weights than variables with lower discriminatory value.

The overall linkage score is calculated as:

Overall linkage score = age score + date score + class score + type score + severity score + other-vehicle score + postcode score + distance score

A higher score indicates a greater likelihood that the HES APC admission and STATS19 casualty refer to the same individual and collision event.

Linkage weights

The linkage weights were derived from the data and previous linkage work (linking STATS19 data to ambulance records). The previous work was used to flag ‘true’ matches within the link candidates dataset, and then scores were based on the probability of agreement given a record was deemed to be matched, or otherwise. There is a degree of judgement involved in assigning the weights in this method, which is not a formal Fellegi-Sunter approach (but uses a similar logic).

The following provides an indication of which variables have the highest power to distinguish between candidate records as to which is the most likely match.

Age

Condition	Weight
Exact age agreement	+1.29
Otherwise	-4.91

Date

Date difference	Weight
Same day	+0.84
Admission one day later	+0.20
Any other difference	-5.04

Geographic distance

Distance is calculated between the STATS19 collision coordinates and the HES home postcode centroid coordinates.

Distance band	Weight
2 miles or less	+8.45
More than 2 miles and up to 10 miles	+6.13
More than 10 miles and up to 50 miles	+1.74
More than 50 miles and up to 100 miles	-2.14
More than 100 miles	-3.84
Location unavailable	0.00

Other vehicle involvement

Condition	Weight
HES variable unavailable or unknown	0.00
Agreement	+0.71
Disagreement	-2.17

Casualty class

For casualty class, pedestrians receive a 0 score as pedestrian is also a casualty type.

Condition	Weight
Agreement	+0.53
Disagreement	-2.77
Pedestrian records, ICD-10 V0x codes	0.00

Casualty type

Condition	Weight
Agreement	+1.93
Disagreement	-3.91

Injury severity

Here, agreement means STATS19 record is flagged as serious.

Condition	Weight
Agreement	+2.17
Disagreement	-3.00

Postcode similarity

Postcode similarity is assessed using Levenshtein edit distance after spaces have been removed. This broadly counts the number of changes needed to transform the STATS19 postcode to match the HES one. ‘Distance 1’ means a Levenshtein edit distance of 1.

Levenshtein distance	Weight
Exact agreement, distance 0	+9.64
Distance 1	+7.97
Distance 2	+7.97
Distance greater than 2	-2.74
Missing postcode	0.00

Selecting final links

A hospital admission can potentially match more than one STATS19 casualty. A best-link approach is therefore used. For each HES APC admission:

• the candidate pair with the highest overall score is identified
• tied highest-scoring pairs are removed through de-duplication
• the remaining highest-scoring pair is assessed against a minimum acceptance threshold

Different thresholds are used depending on whether the STATS19 casualty postcode is available.

STATS19 postcode availability	Acceptance threshold
Postcode available	Score greater than 7
Postcode unavailable	Score greater than 6

This approach reflects the additional confidence provided by postcode agreement while allowing high-quality links to be retained where postcode is missing, and was developed using an iterative approach. Within the linked records, a further ‘confidence score’ is assigned to allow distinction to be made between matches with a higher score, allowing different levels of confidence to be used for different analyses.

Quality assurance

Calibration using the SWAST linkage

The linkage weights and acceptance thresholds were calibrated using records from the earlier South Western Ambulance Service NHS Foundation Trust (SWAST) linkage exercise. The SWAST exercise provided an independently linked reference dataset against which alternative scoring approaches could be compared on the same dataset (a small subset of the full HES APC data used) and is a valuable comparator as the ambulance data contains collision location which makes it much easier to determine a correct match.

Candidate linkage variables, weighting structures and score thresholds were assessed by comparing whether the HES linkage and SWAST linkage identified a match for the same record. The selected method sought to retain valid links, minimise false-positive links and maintain consistency with the independently linked records, bearing in mind that the SWAST linkage itself was probabilistic in nature and therefore subject to some potential error.

Comparison with the SWAST linkage

The comparison covered 556 records available to both linkage exercises. The 2 approaches agreed on whether a record was linked for 507 records, equivalent to 91% agreement. Both methods identified a link for 295 records and neither identified a link for 212 records.

The HES linkage identified 21 additional links not present in the SWAST linkage, while SWAST identified 28 links not present in the HES linkage. Some, but not all, of the 21 additional HES linkages are likely to represent false positives; conversely some, but not all, of the 28 additional SWAST linkages represent missed links in the HES data.

In the latter case, many of the linkages were to records with non-matching postcodes in the two datasets. These records would be hard to link in any probabilistic approach between HES and STATS19, without introducing risk of further false positives.

HES linkage result	SWAST not linked	SWAST linked	Total
Not linked	212	28	240
Linked	21	295	316
Total	233	323	556

Of the 295 records linked by both methods, 292 were linked to the same STATS19 casualty and 3 were linked to a different casualty. This means that 99% of the records linked by both methods agreed on the selected casualty.

These comparisons provide reassurance that the scoring framework generally identifies the same links as the independent linkage, though with some errors and missed matches. However, the SWAST comparison is a validation check rather than a complete clerical review and cannot establish that every accepted link is correct.

Overall, the linkage quality was deemed suitable for this feasibility study, and for population-level analysis. A further step was to add linkage confidence bands to provide an indication of the likelihood of incorrect matches.

Linkage confidence bands

Each accepted link was assigned to 1 of 5 confidence bands: worst, low, moderate, high or best. The bands use the overall linkage score and the difference between the highest and next-highest candidate scores. A higher overall score indicates stronger agreement across the linkage variables. A larger difference from the next-best candidate indicates that the selected link is more clearly distinguished from alternative candidates.

The confidence bands are relative indicators of linkage quality rather than estimated probabilities that an individual link is correct. They allow users to apply different inclusion thresholds depending on the analytical purpose. The main analyses treat moderate, high and best matches as linked. Lower-confidence links can be retained for sensitivity analysis or excluded where greater confidence in individual links is required.

For HES records classified as traffic collisions, 371,307 links were assigned a confidence band. Of these, 221,717, or 60%, were in the best band. Overall, 94% of the linked traffic-collision records were in the moderate, high or best bands.

Linkage confidence band	Number of linked traffic-collision records	Percentage of links
Worst	7,850	2%
Low	12,970	3%
Moderate	52,962	14%
High	75,808	20%
Best	221,717	60%
Total	371,307	100%

Internal diagnostic checks

After linkage, diagnostic checks were undertaken to assess the stability and plausibility of the results. These included examining:

• linkage rates and confidence bands by year
• overall linkage scores and the separation between the selected and next-best candidate
• agreement on postcode, casualty class, casualty type and injury severity
• age differences and differences between collision and admission dates
• geographic distance between the home postcode centroid and collision location
• linkage rates by road-user type, collision type, police force area, demographic characteristics and clinical characteristics

These checks do not determine whether each individual link is correct. They are intended to identify systematic anomalies, changes over time or differences between groups that may indicate linkage bias, data-quality issues or instability in the matching method.

Patterns in the linkage results

For HES inpatient records classified as traffic collisions, 350,487 of 735,219 records, or 48%, were linked when moderate, high and best matches were treated as links.

Linkage rates varied across road-user and collision types. For example, the linkage rate was 61% for pedestrians, 55% for car occupants, 52% for motorcyclists and 23% for pedal cyclists. The rate for pedal cyclists was 64% where the HES classification indicated a collision with a motor vehicle, but 4% where no collision counterpart was recorded. These differences are consistent with the scope of STATS19, which excludes some non-collision and non-road incidents, and show why the linked records should not be treated as fully representative of all HES transport admissions.

Linkage rates were higher for records with indicators of more substantial injury and treatment. The proportion linked rose from 36% for hospital spells shorter than 1 day to 65% for spells of 28 days or more. It also rose from 35% for records with no recorded injury diagnosis to 75% for records with 10 or more injury diagnoses. These patterns are plausible because more severe road traffic casualties are more likely to come to police attention, but they also show that the probability of linkage is not uniform across the HES cohort.

A further analysis of the linkage rates is presented in the separate analysis report.

Limitations

The linkage is probabilistic rather than deterministic. It is therefore subject to both false-positive and false-negative links.

Potential sources of linkage error include:

• inaccuracies or differences in recorded age, date, postcode, road-user type or injury severity
• incomplete postcode information
• a delay between the collision and hospital admission (of more than the one day allowed by the method)
• more than one plausible STATS19 candidate for a HES admission, for example several casualties with similar demographic characteristics in the same collision
• more than one hospital admission relating to the same collision

Linked records represent the most probable match identified by the linkage algorithm. They should not be interpreted as definitive confirmation that the records relate to the same person and collision.

The use of residential postcode to assess geographic proximity also has limitations. A casualty may be injured a long distance from home, and a postcode centroid does not represent an exact residential location. However, it is considered that use of the distance between postcode and collision improves the linkage overall.

Scope differences between the datasets

HES transport admissions and STATS19 do not cover identical populations. STATS19 records personal-injury collisions reported to the police on a public highway. HES can contain admissions arising from non-traffic incidents, incidents not reported to the police, and boarding or alighting incidents. Some HES records classified as transport accidents are therefore outside the intended scope of STATS19 and cannot be expected to link. It is not possible to establish with certainly how reliable the coding of cause of injury is within the HES dataset; if this is inaccurate then some ‘traffic’ admissions might fall outside the scope of the STATS19 data.

Variation in linkage rates

Linkage rates vary by road user type, collision circumstances, geography, demographic characteristics and indicators of injury severity. In particular, admissions involving pedal cyclists without a recorded motor-vehicle counterpart have a substantially lower linkage rate than admissions involving a motor vehicle. Analyses based only on linked records may therefore under-represent some casualty groups and incident types.

Because linkage is more likely for longer hospital spells and records containing more injury diagnoses, the linked cohort may also over-represent more seriously injured casualties relative to the full set of HES transport admissions. Linkage rates should not be used as a direct measure of STATS19 completeness without considering these scope and selection effects.

Changes over time

Changes in hospital coding, police recording, data availability or the composition of the HES cohort may affect linkage rates and confidence bands over time. Apparent trends in linked data should therefore be considered alongside changes in the source datasets and recording practices. In particular the overall number of transport-related admissions in HES fell during the pandemic and has not returned to the previous level, despite an increase in casualties in STATS19 over this period. This pattern requires further investigation in the next phase of this linkage work.

Interpretation of confidence bands

The confidence bands rank accepted links according to the strength and distinctiveness of the match. They are not calibrated probabilities of correctness. The SWAST comparison provides useful validation evidence, but it relates to the records available in that exercise and is not a full record-by-record clerical review of the national linkage. Further work to establish the accuracy of the linkage is required for a better understanding of its quality.

Next steps

The linkage method will continue to be reviewed as new data become available. This will include:

• monitoring linkage rates and confidence-band distributions over time
• reviewing results by population group, road-user type, collision type and police force area
• investigating changes that may indicate a shift in data quality or recording practice
• undertaking sensitivity analysis using different confidence-band thresholds
• considering targeted clerical review of selected records, including lower-confidence and ambiguous links
• reassessing linkage weights and acceptance thresholds where validation evidence suggests that recalibration is needed
• documenting material changes so that the method remains transparent and reproducible

Instructions for printing and saving

Depending on which browser you use and the type of device you use (such as a mobile or laptop) these instructions may vary.

Tablets and mobile devices normally have the option to “find in text” and “print or save” in their sharing or quick options menu of their browser, but this will vary by device model.

How to search

Select Ctrl and F on a Windows laptop or Command and F on a Mac

This will open a search box in the top right-hand corner of the page. Type the word you are looking for in the search bar and press enter.

Your browser will highlight the word, usually in yellow, wherever it appears on the page. Press enter to move to the next place it appears.

Contact details

Road safety statistics

Email roadacc.stats@dft.gov.uk

To hear more about DfT statistical publications as they are released, follow us on X (formerly known as Twitter) at DfTstats.