Under-reporting of road traffic casualties in police records has been well documented,12 as has the substantial international variation in police recording levels of hospital admissions.3 Relying solely on police data on road traffic crashes (RTCs) and its consequences is thus potentially misleading.4^–6 A specific cause for concern is the reporting of casualty severity by police forces.7 A recent report from England8 found that the fall in the number of serious traffic injuries over time observed in police statistics was not matched by a similar decrease in hospital admissions arising from road crashes. The authors hypothesized that this was probably due to a reduction in the completeness of reporting of serious injuries by the police rather than a real reduction in incidence. An investigation by the UK Department for Transport1 comparing hospital and police road casualty data (STATS19) claimed that “the number of serious casualties in STATS19 could be under-reported and/or be under-recorded by as much as a half, and it is possible that this has risen over recent years”. The same report also pointed to the difficulty in interpreting only one source of data on serious injuries, as the severity of some casualties may be misclassified. We therefore decided to investigate the extent of possible under-reporting of casualties due to RTCs in the West of Scotland.

METHODS

The analysis was part of a larger study designed to assess the health impact of (speed and red light) cameras in the Strathclyde police region of Scotland. This mixed urban–rural region is home to more than 2.3 million residents (around two-fifths of the population of Scotland) and contains concentrations of severe social deprivation, a factor that is correlated with the risk of RTCs (particularly pedestrian casualties) and perhaps also their likelihood of being reported. Our approach differed from other studies in that we used an extensive linked database of hospital and police records covering a comparatively large geographical area over a relatively long time period.1910

We used data from two sources: police road casualty data (STATS19) from Strathclyde Police Information Resources Department, and NHS hospitalization data (SMR01) from the Information Services Division (ISD) of NHS National Services Scotland.

The UK Department for Transport STATS19 data are derived from the recording of personal injuries occurring on public roads as a result of RTCs that become known to the police. The data include pedal cyclists injuring themselves or a pedestrian, but exclude injuries occurring in car parks and similar locations. The reporting form incorporates information on RTC circumstances, vehicles, and casualties in great detail. The police officer recording an injury selects a severity grading from three levels: fatal, serious, and slight.11

• Fatal *injury: immediate death or within 30 days of the crash.

• Serious injury: requiring hospital treatment either immediately or later (includes injuries to casualties who die more than 30 days after the crash).

• Slight injury: injuries not necessarily requiring medical treatment.

Although not all fatal or serious injuries recorded by the police require hospital admission, all hospital admissions should by definition be recorded as either fatal or serious casualties on police records.

SMR01 inpatient data comprise demographic, episode management, and coded clinical information from the routine hospital discharge returns completed by Scottish (non-obstetric, non-psychiatric) hospitals for all patients. The admissions we selected were those to which ICD-9 and ICD-10 (International classification of diseases, 9th and 10th revisions, respectively) diagnostic codes for RTC injury ICD codes (ICD-9: E810–E819; ICD-10: V01–V82, V87, V89.2) were attached. ICD-10 was used for all hospital admissions, while death registration switched from ICD-9 to ICD-10 in January 2000. The SMR01 data were internally linked with all hospital episodes belonging to an individual and grouped together allowing episodes constituting a single continuous inpatient stay (including intra-hospital and inter-hospital transfers) or re-admissions to be identified. Death records with mention of RTC were obtained from General Register Office for Scotland (GROS) and attached to the SMR01 data or directly to a STATS19 record (if death was instant and no hospitalization was required).

The hospitalization data for initial admissions were restricted to hospitals located in health boards within or bordering Strathclyde police force area (to allow for RTC casualties admitted to neighboring hospital care): Greater Glasgow, Argyll & Clyde, Ayrshire & Arran, Forth Valley, Lanarkshire, Lothian, and Dumfries & Galloway Board areas. Subsequent admissions included the whole of Scotland but not the remainder of the UK. All casualties were included regardless of their place of residence, as the location of the RTC was the relevant variable.

We created a linked database comprising both police and hospital in-patient data. Linkage between STATS19 and SMR01 was performed by ISD. The linkage methods involved a combination of deterministic matching of postcodes, age, sex and date (of RTC and admission1213) and probabilistic matching on these variables in addition to casualty type and severity.1415 Of the linked records, 57% were matched by postcodes, 36% had postcodes missing but an exact match on age, sex, and date, and 7% were linked solely on the basis of a high probability match (table S1 online).

The Scottish Medical Record Linkage system has in the past been shown to have 3% false positive and 3% false negative rates,16 but we recognize that, owing to the very limited identifying information available, our RTC linkage may not have achieved such high accuracy. As we did not have access to names or unique common person identifiers, it was impossible to estimate the true positive and false negative linkage rates with any degree of accuracy. Rigorous clerical checking of a large sample of best matching pairs at a wide range of probability weights was, however, carried out. Different probability thresholds were set depending on whether a postcode was available or not. A crude estimate of the false negative rate was 2.6%—that is, 2.6% (1802) of the unlinked STATS19 records were missed links (ie, should have been linked). A crude estimate of the false positive rate was 15.6%—that is, 15.6% (1607) of the linked STATS19 records were false positives (ie, should not have been linked). The number of missed linked records was therefore nearly compensated for by the number of erroneously linked records. We acknowledge, however, that this does not avoid possible misclassification bias when characteristics of linked data are considered

The data were grouped into three categories: casualties with only STATS19 records, casualties with only SMR01 records, and casualties with both SMR01 and STATS19 records (linked data). Contrasting the number of SMR01-identified road casualties that were also recorded (“linked”) in STATS19 records with those that were not (“unlinked”) gives an indication of the extent of under-reporting of hospitalized road casualties to the police. The unlinked SMR01 data excluded all non-traffic RTC injuries (as STATS19 does not record non-traffic RTCs, defined by ICD-10 as “any vehicle accident that occurs entirely in any place other than a public highway”17). These constituted ∼28% of the total unlinked SMR01 RTC casualties at the start of the process.

The study period was 1997–2005 inclusive. However, the time period was restricted to 1997–2004 in the analysis of road users, as the definitions for the road user categories in the STATS19 data changed slightly in 2005. Casualty reduction over time was assessed by comparing the first 3 years with the middle and last 3 years, and time trends were assessed using linear regression (using all years). Three-year groups were used to iron out any random fluctuations in the data (especially where fatalities are included).

To assess which factors influenced whether or not a RTC hospital admission was recorded by the police, we tested, using Pearson’s χ² statistic, for any association of the following variables with linked and unlinked SMR01: age (0–14, 15–24, 25–34, 35–44, 45–54, 55–64, and 65+), sex of casualty, road user (car user, pedal cyclist, pedestrian, moped/motorcycle rider, and other vehicle), length of stay in hospital (day case, 1, 2–3, 4–7, and >7 nights stay), third-party involvement, Scottish Index of Multiple Deprivation (SIMD18), year, month, and day of week of crash. Using binary logistic regression, univariable odds ratios and 95% confidence intervals were calculated for each variable of interest. To assess the effect of the variables simultaneously on recording, we performed multivariable binary logistic regression analysis, which produced adjusted odds ratios and confidence intervals.

RESULTS

Comparing STATS19 killed and seriously injured (KSI) rates with SMR01 admission rates shows both that STATS19 had fewer casualties than SMR01 in this category and that the decline over time was steeper (fig 1). Comparing the first 3 years with the last 3, the reduction in SMR01 RTC admissions rates was 21%, compared with the STATS19 reduction in KSI rates of 38%.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1 Casualties by SMR01 road traffic crash admissions and STATS19 killed and seriously injured (KSI) records in Strathclyde 1997–2005 per 100 000 population.

Distribution of STATS19 data over time and severity

Casualties coded as fatal in Strathclyde as recorded by STATS19 comprised just over 1% of the total in all years (online table S2). The relative proportions of serious and slight casualties changed over time. Injuries coded as serious decreased from 18% to 12% of the total, whereas injuries coded as slight increased from 81% to 87% over the study period.

There was a significant downward trend (trend gradient −12.7) in the overall numbers of road casualties in STATS19 (−18%, from 431 per 100 000 in the first 3 years to 355 in the last 3 years). There were large differences across the three severity categories: fatalities declined by 7% (a non-significant reduction), serious casualties by 40%, and slight injuries by 13% (table 1).

Distribution of the linked data over time

To exclude artifactual errors due to the linkage process itself, we compared the proportion of linked STATS19 casualties in the 3-year groups over the study period. This shows that there is little difference in the distribution, ie, in all periods, 12.5–14.7% of STATS 19 records were linked.

More than half of all casualties hospitalized following an RTC (excluding those not occurring on public roads) were identified in STATS19 records. The proportions did not change greatly over time (varying between 50% and 60%). Of the SMR01 casualty records, 55% could be linked over the whole study period, suggesting an under-reporting rate for hospitalized casualties of 45%.

Distribution of linked and unlinked data over time and severity

There was a significant decline in RTCs in all linked and unlinked SMR01 and STATS19 categories (table 2 and online fig S1). The decline in unlinked SMR01 data, however, appears to have reached a plateau over the last 6 years of the study period. At the same time, there appears to be a higher proportion of SMR01 casualties with a STATS19 record during the middle time period (2000–2002).

The STATS19 slight casualties category linked to the SMR01 data is the only (of linked and unlinked casualties) that showed an increase in numbers over the study period (+5% comparing first and last 3 years; table 3), whereas the numbers of both linked and unlinked STATS19 KSI (combined fatal and serious casualties) decreased substantially (−27% and −47%, respectively). There were increases over the study period in the proportions of both total KSI (47% in 1997–9 and 55% in 2003–5) and total slight casualties (5% in 1997–9 and 6% in 2003–5, not shown in table 3) that could be linked to SMR01.

There was a relatively high proportion of the unlinked SMR01 casualties in the first 3-year period (3512 out of 8701, ie, 40%). This may reflect the general time trend in RTC casualty reduction, rather than a linkage problem in earlier years (as a similar proportion of SMR01 were linked to STATS19 over time; see online fig S1).

In summary, about one-third (33%) of linked SMR01 records were recorded as slight by the police and a little over half of KSI STATS19 records linked to an SMR01 record.

Linked and unlinked casualties by road user

Both STATS19 and SMR01 data include information on road user (casualty class/vehicle type in STATS19 and external cause in ICD-10). Comparing road user categories shows which types of casualties were more likely to be missed by police records. Table 4 includes both the STATS19 and the equivalent ICD-10 definitions (using the STATS19 definition for unlinked STATS19 and linked STATS19/SMR01, and the ICD definition for unlinked SMR01).

Car occupants comprised nearly two-thirds (62%) of the total injured casualties. Of hospitalized car occupants (SMR01), 38% did not have a STATS19 record and 10% of all STATS19 injured car occupants could be linked to SMR01.The proportions of unlinked SMR01 were higher than linked in the following categories: pedal cyclists, motorcycles/mopeds, and other vehicle.

Pedal cyclists appear to be the road user type most likely to be missed by police recording: 22% of total unlinked traffic-related hospital admissions, while only comprising 6% of total casualties; whereas 47% of all recorded injured cyclists were admitted to hospital (2145 of total 4531), 82% of these had no STATS19 record. In STATS19, there were 2769 pedal cycle casualties, of which 14% linked to an SMR01 record.

Pedestrians made up a substantial proportion (19%) of the unlinked SMR01 records, which is a similar proportion to the total recorded pedestrian casualties (20%). However, 30% of hospitalized pedestrians had no STATS19 record, and, of all pedestrians recorded injured in STATS19, 24% were linked to SMR01.

Characteristics of unlinked and linked SMR01 records

In the univariable analysis, the following categories of admitted RTC casualties were associated with a significantly increased risk of being unrecorded in STATS19 (online table S4): no third-party involvement, age 0–14, earlier years (in this dataset), road user (other than car), length of stay in hospital, day of week and month of crash. The SIMD score and sex of casualty were not significantly associated with a risk of being unrecorded in STATS19.

All variables were tested using multivariable logistic regression (table 5). The significant factors associated with hospitalized casualties not being recorded by the police were: no third-party involvement, age (progressively higher risk with older age), year (hospitalized casualties in the earlier years appeared to be slightly less likely to be recorded by the police), vehicle occupant (pedal cyclists were over eight times as likely to be missed by police compared with a car occupant), length of stay (day cases were less likely to be recorded by police), and sex (females were less likely to be recorded by the police).

DISCUSSION

Our results suggest a consistent under-reporting in police road casualty data of ∼45%. Notwithstanding the potential pitfalls of data linkage, to which we have alluded above, this is a similar figure to that reported in previous studies linking accident and emergency or hospital data with STATS19 data.1910

There are several possible reasons why a proportion of total casualties are unrecorded by the police in STATS19. There is no legal requirement to report a road crash when only the driver is injured and the only damage caused is to their vehicle. Under-reporting, however, particularly seems to apply to RTCs involving motorcycles or pedal cycles when no other vehicles are involved, perhaps because these road users regard police reports as being necessary only when insurance claims arise for vehicle damage. This hypothesis is consistent both with our findings and those of other studies.2^–41920 We also found, that age and length of stay in hospital (in common with a previous study5) as well as sex and year of RTC were significantly associated with under-reporting. Road users driving stolen vehicles, driving while impaired by alcohol or drugs, or driving with no insurance, driving license, or MOT (roadworthiness certificate) are also less likely to have their crash brought to the attention of the police, especially if no other vehicle is involved. Casualties are not recorded if they absent themselves from the scene without giving their details to the other parties in the crash. Police forces should, in theory, update the severities on the reports, but in practice do not always do so, and there seems to be no adequate system for doctors to update police officers about changes in severities—that is, the police may not be advised that the casualty has subsequently been admitted to a ward and hence automatically become a “serious” casualty. Finally, patients admitted to hospital may fraudulently claim to have been involved in an RTC or may exaggerate the injury symptoms, perhaps to obtain compensation or to conceal the real reason for their injuries.

As we also included matching links of casualties from neighboring health boards (4%, ie, 444 out of 10 005 hospital admissions—excluding immediate deaths), we recognized that there would be a proportion of unlinked SMR01 records that had a matching STATS19 record in a neighboring police force too. Consequently, if an equal number of SMR01 records in Strathclyde matched a STATS19 elsewhere, the actual proportion of RTC SMR01 records that were not recorded by the police would be 42% rather than 45%.

Regarding severity, our findings are in agreement with those of Gill et al,8 suggesting that there has been an increase over time in the under-reporting by the police of serious casualties. Seriously injured STATS19 casualties declined in frequency more than the SMR01 would indicate. There are several potential explanations (some of which were also highlighted by the Department for Transport21). There may have occurred a true, disproportionate fall in the number of serious casualties caused by RTCs. As the reduction was not reflected in the number of fatal injuries or total injuries in SMR01, this seems improbable. Alternatively, public reporting of serious RTCs may have declined, as might the proportion of less severe casualties being treated in accident and emergency departments, although there is little evidence in favor of either phenomenon. Hospital admission policies, such as a rise in hospital admissions of less severe road casualties, may have occurred, but that also seems unlikely. Hospital reporting of casualties may have changed over time, but the same reporting standards have been used in Scotland for decades, and the completeness of SMR data are audited regularly in terms of coding and accuracy with consistent results over time.22 We believe that the most likely explanation is that some types of casualties, formerly recorded by the police as severe, were in the later years recorded as slight (ie, a shift in police judgment of casualty severity).

If our findings are generalisable across the UK, then this combination of under-recording and misclassification by the police has major implications, not least for the British Government’s national casualty reduction targets for 2010, which are based solely on police STATS19 data.23 Even more important are the thousands of injuries potentially missing from the databases of the police, local authorities, and trunk-route operators who rely on STATS19 information to target dangerous sites for remedial measures such as new road layouts or traffic calming. Likewise, the national safety camera program also relies heavily on the completeness and accuracy of police data—especially fatal and serious RTCs—to locate correctly its cameras at casualty “blackspots.” Furthermore, if certain categories of road users are particularly likely to under-report their injuries, this could have implications for the accuracy and targeting of road safety education by public agencies and road safety groups. Finally, the Government’s estimates of the total cost to society of not preventing RTCs will need to be radically revised.2324

We conclude that a general decline in the completeness of STATS19 is unlikely to have occurred, but that there may have been an increasing tendency over time for police officers to report injuries as slight rather than serious. These findings have major implications for road safety policy in Scotland, and perhaps the UK as a whole, and should inform the future collection, interpretation, and utilization of road casualty statistics. To improve the quality of this information, routine linkage of road casualty data derived from police and hospitalization databases should be considered.