WOl-l575/93 $6.00 + .OO 0 1993Pergamon Press Ltd.
Accid. Anal. and Prev. Vol. 25, No. 6, pp. 659465, 1993 Printed in the U.S.A.
THE CONSTRUCTION OF A ROAD INJURY DATABASE ANNA M. FERRANTE”, DIANA L. ROSMAN+, and MATTHEW W. KNUIMAN Department
of Public Health, University (Received
of Western Australia,
3 June 1991; in revised form
Nedlands,
22 January
Perth, W.A. 6009, Australia
1993)
Abstract-In order to effectively examine possible causes and determinants of road trauma, reliable information on the participants, circumstances, and resultant injuries and deaths must be available. Characteristics of participants (persons and vehicles) and the circumstances of road accidents are routinely collected by police and road authorities, whereas details of the injuries and medical care provided to casualties are collected by hospital and ambulance services. A road injury database, linking data collected by the Health, Police, and Main Roads Departments of the Government of Western Australia with records of the St. John Ambulance Association and the Death Register, has been established. This paper describes the procedures used to link the various sources of data and discusses the design, construction, and quality of the resultant
relational database. and the level of confidence that the linked records referred to the same person. They found that a linkage rate of 50% would require acceptance of a 10% error rate. With the advantage of access to named data, the linkage project described below is the first attempt to link accident data from four sources over an extended period (15 months) for an entire state. Construction of the database required the collaboration of a number of authorities, each of which maintained its own computerized records relating to traffic accidents and casualties. Subject to certain security and confidentality provisions, name-identified data were obtained from police, hospital, ambulance, and death records and linked using ‘GIRLS’ (Generalized Iterative Record Linkage System) (Hill 1981; Hill and Pring-Mill 1986). Name, age, gender, and data of accident were used to link information from the various sources. This article describes the methods and procedures used to construct the database, highlights particular problems encountered in the process, presents results on the success of the linkage process and the quality of the linked data, and describes the design of the resultant database. The design chosen readily allows the extraction of datasets suitable for analyses in which the accident, the vehicle, or the casualty is the unit of analysis. The information in the database has already been utilised in a number of studies including a comparison of hospital and police-reported casualties (Rosman and Knuiman 1992), an investigation of conspicuity as an issue in accidents involving motor
INTRODUCTION A database was established consisting of all records relating to all road crashes occurring in Western Australia during a E-month study period (October 1987 to December 1988) in which at least one person was injured or killed. The purpose was to create a comprehensive road injury information resource that would be used in the study of the epidemiology of road injury, the selection of special road user groups for further investigation, and the evaluation of cost effective methods of reducing road trauma. In addition, the linkage of hospital and police data has also provided a means of estimating the extent of underreporting of casualty crashes resulting in hospital admission. Linked databases are not new to the field of road accident research, particularly to the study of fatal accidents. In the United States, databases have been created that link crash circumstances with cause of death for fatalities (Fife 1989). Police and medical sources have also been linked for noncrash events (Agran and Dunkle 1985). Within Australia, a pilot study, carried out by Gordon et al. (1986) in the Hunter region of the state of New South Wales, measured the feasibility of linking various data sources without using names or other unique identifiers. This pilot study concluded that there was a trade-off between the proportion of records linked *Currently with the Crime Research Centre, University Western Australia. +A11 correspondence should be directed to this author.
of
659
A. M. FERRANTEet al.
660
cycles (Cercarelli et al. 1992), pedestrian crash risk (Arnold , Rosman, and Thornett 1992), and the hospital costs of road trauma (Hendrie, Rosman, and Harris 1992). METHOD Sources of data Hospital admission records. The Health Department of Western Australia operates a comprehensive Inpatient Morbidity System of all hospital admissions for the state. All available information for inpatient episodes occurring during the study period as a result of a traffic accident was extracted from this Morbidity System. A traffic accident was defined according to the International Classification of Diseases 9th Revision Clinical Modification (ICD9-CM) code for external cause (E810.0 to E819.9) (National Center for Health Statistics, 1988). Due to a specific interest in pedal cycle accidents, admissions with external cause E826.1 were included. Late effects of a traffic accident (E929.0) were also included for completeness. The resultant dataset contained a total of 5,697 hospital episodes (4,633 traffic accident casualties-E810.0 to E819.9; 642 pedal cycle casualties-E826.1; 20 other nontraffic (E826 and E829) casualties; and 402 casualties from late effects of a traffic accident-E929.0). Information retrieved from the Inpatient Morbidity System for each hospital discharge included name and address, age, race, country of birth, occupation, gender, hospital type, admission date, separation date, separation type, as well as several diagnoses, operations, and complications, all coded according to the ICD-9-CM formula. Accident details. Detailed information about the accident site and about the vehicles and persons involved in road accidents was obtained from a Road Accident Database (ROTARS) jointly maintained by the Police and Main Roads Departments of Western Australia. Records of accidents occurring between October 1987 and December 1988, where one or more persons were killed or injured, were extracted from this database for linkage to other sources. Details of all vehicles and drivers involved in these “casualty” accidents were extracted whether or not the occupants were injured. Where the vehicles were registered in Western Australia, registration details such as make, model, colour, year of manufacture, weight, and number of cylinders were obtained from police vehicle licence records. On the 18,676 casualty records, fields for name, gender, and age as well as injuries (as reported by the informant) graded in severity as fatal, requiring hospitalisation, requiring other medical attention,
or not requiring medical attention were available. Names of passengers were not routinely entered into the ROTARS database, and consequently required additional clerical work to retrieve this information from the original report forms. Ambulance data. Data were obtained from the St. John Ambulance Association which provides comprehensive road ambulance services to both city and country hospitals. Information from the Association included time taken and distances travelled for the various stages of the ambulance cycle from base to accident scene to hospital. As well as name and address identification, priority code of the ambulance trip and the patient’s condition were also recorded. An ambulance trip resulting from a traffic accident was identified by a code on each trip record. There were 5,828 such trips during the study period. Death data. Records of the deaths occurring during the study period, as well as those occurring during the first two months of 1989, were obtained from the office of the Registrar General. Information stored for each registered death included demographic details, date and time of death, whether the death occurred in a hospital, and the cause of death. A total of 11,258 deaths from all causes were recorded during the study period. No attempt was made to select a “traffic accident” subset from these prior to linking, although cause of death for traffic fatalities is usually coded as the external cause of injury (ICD-9-CM). For the study period, the police records had identified 318 road fatalities.
Information not incorporated into the database Property damage only accidents. Noncasualty accidents, i.e. those resulting in “property damage only”, were considered unlikely to link to a hospital inpatient record and were excluded from the Road Injury Database. Restricting linkage to the 23% of police records relating to “casualty” accidents considerably reduced the size of the linkage task. A manual search of police “property damage only” records was carried out for June 1988 to estimate the proportion of links lost through this restriction. Accident and emergency records. Records of casualties who sought or required treatment from an “Accident and Emergency” section of a hospital and who were not admitted to hospital were not incorporated in the database for this study period. At the time of this study, records of these centres were not routinely kept on electronic media. However, since that time, records from the accident and emergency department of one of the three major teaching hospitals in the Perth metropolitan area have been made available for the study period and
Construction
have subsequently of the database.
of a road injury database
been linked to the other sections
Other hospital records. There were 358,528 hospitalizations in Western Australia during 1988, of which 4,480 (I 3%) were specified as resulting from a traffic accident. Restricting the hospital data to records of casualties resulting from a traffic accident minimised the linkage task, but assumed that the coding for external cause of injury was correct. Record
linkage methodology
Special-purpose record-linkage software, obtained from Statistics Canada (Hill 1981; Hill and Pring-Mill 1986), was used to link data from the four sources. The software (GIRLS) uses a probabilistic process to determine whether records from various sources, which do not have unique identifiers, should be matched or “linked.” In the initial phase of the linkage process, “rules” were written to compare similar fields (e.g. gender codes) from the different data sources. Some comparisons were quite sophisticated involving cross-comparisons of parts of some fields with parts of others (e.g. comparing the first four characters of the first given name with the first four characters of the second given name). Surnames were encoded using phonetic systems (Newcombe 1988) that allowed similar sounding surnames and surnames with slight spelling variations to be encoded into the same group. In order to reduce the number of comparisons made in the record linkage process, records were grouped into phonetic “pockets” so that only records that fell into the same pocket were compared. In subsequent phases of the linkage process, records within the same pocket were compared using the rules defined in the earlier phase, creating a database of “potentially” matched pairs of records. Frequency weights were then assigned to these pairs-a positive weight was given when there was agreement, a negative weight when there was disagreement, and no weight when information was missing. A total weight was then aggregated for each “pair” of records based on the outcome from each rule comparison. If the total weight exceeded an upper acceptance threshold as determined by the user, the potentially linked pair was said to be a “definite” link. Those pairs with a total weight falling below a lower rejection threshold or “cutoff” were labelled as “rejected”. The remaining pairs were classed as “possible” links and referred to the next phase. Refinements of rules, weights, and thresholds were made at each phase, resulting in a fine-tuned, data-sensitive record linkage procedure. Finally, the resulting “definite” links were assembled into
661
groups of linked records to form a one-to-one linkage result or, if desired, a one-to-many, many-to-one, or many-to-many result.
The linkage process
Linking of police, hospital, ambulance and death data was undertaken in stages. In the first instance, each of the major data sources was internally linked (i.e. records were matched to other records in the same dataset). This was necessary, as it identified any duplicate records for the same individual as well as instances where an individual had been involved in more than one casualty accident during the study period. When this was completed, each pair of datasets was linked. In all, six major stages were carried out. Hospital to ambulance. Hospital data was linked to St. John’s Ambulance data in a one-to-many type linkage. This linkage strategy was used to capture all ambulance records (trips) for a person involved in an accident(s) during the 15-month period. The hospital-ambulance linkage stage involved three linkage “passes”. In the first pass, records were pocketed by NYSIIS code (Newcombe 1988) and linked using year of birth, gender, first initial, accident date, hospital destination and surname. In the second pass, records that failed to match in the first pass were repocketed by year of birth, and the same link fields (except year of birth) were used to match records. In the third pass, unmatched records were again pocketed by NYSIIS code. The linking criteria remained the same but the acceptance threshold was lowered. Hospital to police. Linking hospital to police records was considered to be the most crucial component of the project. In all, seven “passes” through the data were required to produce the final number of links. As a result of the first two passes, in which records were pocketed by NYSIIS code, 2,83 1 hospital records linked to 2,549 police records. In the third pass, records were pocketed by accident/admission date. In order to incorporate possible time lags between time of accident and time of hospital admission, time frames of two days (48 hours) were formed and records pocketed accordingly. In the fourth pass, records were once again pocketed by 48-hour time slices with a 24-hour offset from the previous pocketing scheme. In subsequent passes, the time slices were extended to 10 days. In the fifth and sixth pass, gender was also used as a pocketing variable (in order to limit pocket size). Link criteria consisted of year of birth, gender, first name (first four characters), second initial (and a cross-comparison of first initial
A. M. FERRANTEet al.
662
with second initial in cases where there was disagreement on first name and second initial), accident date, surname, and road-user type (driver, passenger, etc). In cases where possible links were manually scanned, an additional comparison of hospital location and accident area code was made. Ambulance to police. Ambulance records that had not previously linked to hospital data were pocketed by NYSIIS code and compared to police records using a one-to-many type linkage. More than 90% of the final number of links were achieved in the first pass. In the second pass, unmatched records from the first pass were pocketed by NYSIIS code, threshold levels were lowered, and manual checking of doubtful matches was performed. In the third and fourth passes “accident times slices” were relaxed, first to 10 days and then offset by five days. Year of birth, gender, first and second initials, surname, and accident date were used for comparison. Accident hour was also used because of the proximity of the time of accident to the time of ambulance arrival for many road traffic accidents. In cases where matches were manually checked, a comparison of ambulance destination area code to accident area code was also made. Death to police. Two passes were used to link the 11,258 death records to police records. In the first pass, records were pocketed by NYSIIS code, producing 93% of the links, while in the second pass, the remaining unmatched records were pocketed by date of death and then linked. Death to hospital. Finally, death data were linked to hospital data using a similar strategy to that used for death-police. This was a one-to-many linkage, achieving all but 2% of the links in the first pass. RESULTS Issues addressed in this section include a critical examination of data quality and the “success” of the linkage process, as well as an overview of the resultant database. Data quality Duplicates in each of the source datasets were resolved. Internal linkage of police data identified 438 (2.3%) repeated person identifiers. These were of two types: people who were involved in more than one accident over the 15-month period (147) and duplicate reports for the same accident (81), which had been independently entered into the police computer system. An internal linkage of hospital data detected 463 (8.1%) repeated person identifiers, three of which were duplicates, the remainder being
repeated hospital admissions or transfers between hospitals. Checks for data inconsistencies within the police data were made where possible. Wet or dry roads were tabulated against clear or rainy weather and in only eight accidents were dry roads reported in rainy conditions (wet roads in clear conditions were considered acceptable). Similarly, “time of day” and “light condition” were cross-tabulated and showed 32 accidents where daylight was reported more than one hour after sunset and 241 accidents where dark conditions were reported more than one hour after sunrise. Movement inappropriate to the road user was denoted as “missing” and instances of a pedestrian being the “colliding” rather than the “target” unit were corrected. Some fields in the hospital file measured the same or similar characteristics to those in the Police file. These were examined to determine the level of consistency in these data sources. Where blatant errors were discovered, further examination of the data sometimes allowed corrections to be made. The gender was in dispute for 15 of the 2,947 (0.5%) linked road accident casualties. Where there was such a disagreement, the hospital record of gender was preferred. There was often a discrepancy in the age of a casualty. Where age was recorded in both data sources, the hospital age was also preferred. Exact agreement for age was found in 1,487 males (83.4%) and 849 females (81.7%). A difference of five or more years was found in 121 males and 72 females. Large discrepancies (up to 60 years) were found predominantly in police records where a police officer had attended the accident scene and completed the report form. Most of these appeared to be transcription errors (e.g. 70 years instead of 10 years). A comparison of the level of injury severity coded on the police form with confirmed links to ambulance, hospital, and death records also showed inconsistencies (Rosman and Knuiman 1992). A full analysis of the differences between the (police) linked and unlinked portions of the hospital dataset, showing that the more severe casualties were more likely to link to a police record, is contained in Rosman and Knuiman 1992. Casualties of single vehicle accidents, especially motorcyclists; casualties admitted as a result of “late effects” of a traffic accident; and those in private hospitals were less likely to link to a police record. The linkage process Table 1 highlights the results from the linkage process. The most significant result demonstrated is the achievement of more than 90% of the links for most dataset pairs in the first GIRLS run. The
Construction
of a road injury database
Table 1. Percentage of links achieved in the first pass. (The numbers in parenthesis indicate the number of link passes carried out) Source Hospital Police Ambulance Deaths
Hospital % Police % Ambulance % Deaths % 94 (7) 9377) 78 (3) 95 (2)
91(s) 93 (3)
81 (3) 91 (5) -
98 (2) 93 (3) -
exceptions were for the hospital-ambulance link (the first attempted) and the hospital-police link where the proportion of links achieved in the first pass of the hospital-police link was 81% of police records but 93% of hospital records. Because several hospital admissions could result from one accident or a person could be involved in more than one accident during the study period, the proportion of police records linking at any stage would not necessarily be the same as the proportion of hospital records. Link quality Before embarking on the linkage of the four sources described above, trials were carried out to gauge the level of automatic linkage that could be expected. Using small test datasets, 97% of correct links were achieved automatically using GIRLS procedures. However, for larger datasets (exceeding 60,000 records), about 1% of links had to be checked manually to achieve this result. After the computer linkage process had been completed, data for both hospital and police for June 1988 were extracted in order to manually check a sample of computer links. Searches were made for “false negative” links (i.e. those which should have linked but did not) and “false positive” matches (i.e. those incorrectly linked). Of the 278 hospital records of traffic accident casualties for this month, 175 links from hospital to police had been made by the computer system. A manual check of these identified two extra “possible” links matching on only two of the four match variables. There were no intances of computer generated links that were thought to be false. (Within the full 15month dataset three “false positive” links were uncovered during routine analyses). Database design A database design was required that would store source data together with linkage results in such a way as to enable extraction of linked records while at the same time retaining the integrity of the original data. In addition, since the linkage process had assigned links in a probabilistic manner, links achieved retained a degree of uncertainty. It was therefore
663
decided not to permanently attach records from the various sources in a massive single dataset, but to retain the original data in separate “tables” and use the link numbers that had been assigned as a result of the linking process as a means of joining records together. This database structure conforms to a “relational” design and is readily maintained using a relational database management system (RDMS) such as ORACLE (Date 1986). The database structure developed is displayed in Fig. 1. First, all names were stripped from records to preserve confidentiality. The police dataset was subdivided into accident, vehicle, and casualty sections and the hospital, ambulance, and death data each remained as separate entities. Unique accident identifiers, attached to related records within the police data, enabled “rejoining” of these record sections. The splitting of the police data was necessary to avoid duplication of accident details where several vehicles and casualties belonged to a single accident . The hospital-police, hospital-ambulance, and hospital-death link numbers served to link hospital data for an individual to the information on the accident report form, if one existed, as well as to any ambulance trips and the death record in the case of a fatality. Similarly, link numbers for each other pair of data sources were used to join the separate data items for an individual. Thus each casualty record within the casualty table would contain three link numbers (police-ambulance, police-hospital, and police-death), while each hospital record would also contain three link numbers (hospital-police, hospital-ambulance, and hospital-death). Ambulance (ambulance-police, ambulance-hospital) and death (death-police, death-hospital) records contained two link numbers with linkage to the third source being achieved through hospital links. This relational database design allowed data to be extracted for accidents, vehicles, or casualties, depending on the unit of measurement required. Accidents. Each record in this table represents a single accident. Information such as road environment details, the date and time of the crash, the number of vehicles, and the total number of casualties were included for the 13,781 unique accident records. Units. The details of vehicle make, model, year of manufacture, etc., as well as driver gender and age and the number of casualties in that vehicle were stored in the “unit” table. Where pedestrians were involved, the gender and age as well as movement details were also entered in this table. Bicycle details were also recorded in this unit table. In order to incorporate vehicle details of all
664
A. M. FERRANTE et al.
II
Police data
I-
2911
AMBULANCE 5828 trips “1701
(1636)
1760
ACCIDENTS
UNITS
CASUALTIES -
13781 accidents
26872 vehicles cycles pedestrians
HOSPITAL
2702
18557 casuabies
q@+J, /
/ 2992\
5694 discharges (5182 distinct persons)
(18410 distinct persons)
‘56
(90)
87
DEATHS 325 persons Fig. 1. Structure of the Road Injury Database. (The number of records indicating the number of unique records. The number and direction
parties involved in the casualty accident for later analysis, the information for vehicles of uninjured parties were added from a separate file. Of the total 26,872 records in the unit table, 18,111 records contained useful information about the driver and 17,483 contained vehicle registration details. Casualties. The only casualty variables available from police data were gender, age, road-user type, and injury severity. Year and accident number as well as unit number were also retained as part of each of the 18,557 casualty records available after extraction of duplicates. The three link numberspolice-hospital, police-ambulance, and policedeath-were retained in this table. The remainder of casualty information came from linking to hospital, ambulance, and death records. Of the 18,557 casualty records relating to 18,410 distinct persons, 2,702 could be linked to a hospital record, 2,903 to an ambulance record, and 3 18 to a death record. Because the police-hospital link had been many-tomany, 2,702 police records belonging to 2,664 distinct persons had linked to 2,992 hospital episodes. Hospital. Of the 5,694 hospital discharge records for 5,182 distinct persons, 2,992 were linked to a police casualty record, 1,760 to an ambulance record, and 96 to a death record. For analysis of the complete history of an individual, aggregation over multiple hospital discharges is required. Some of the original hospital-police links had
in each dataset is shown inside each box, with those in parenthesis of links between datasets is shown along the relevant link line.)
to be broken for the cases where casualties were involved in more than one accident during the study period. Although a test set of one month’s data for June 1988 had been checked manually with no evidence of false positive links, three links were found to be false while manually checking age discrepancies and accident dates. Ambulance. The ambulance data remained as one record per ambulance trip. Of the 5,828 trips, 1,701 (1,636 distinct persons) could be linked to a hospital discharge and 2,911 (2,882 distinct persons) to a police casualty record. Since the ambulance dataset had not been internally linked, the number of distinct persons covered by the 5,828 trips could not be determined. Deaths. There were 325 death records of which 302 had police links and 87 hospital links. Only nine of the 23 registered deaths that linked to hospital but not to police showed trauma-related diagnoses, indicating that the death may not have been as a result of the traffic accident but may have occurred coincidentally in the same 15-month period. DISCUSSION The creation of the Road Injury Database has enabled a range of research projects to be undertaken. Studies examining pedestrian crash risk (Arnold et al. 1992), motorcycle conspicuity (Cercarelli et al. 1992), and hospital costs (Hendrie et al. 1992)
Construction
of a road injury database
have utilised the database. Most significantly it has had a major role in the understanding and investigation of casualty accident underreporting in Western Australia (Rosman and Knuiman 1992). In this state an accident that results in injury at any level or causes more than $1,000 damage to property is required to be reported to police. The police dataset has thus been used to produce official road casualty statistics for this state for more than 10 years. This project has not only shown that many accidents resulting in hospital admission are not reported but also that the level of injury reported is often inaccurate. With the benefit of linkage to hospital and death records and later accident and emergency data as well, detailed knowledge of the consequences of different road accident types will be available. Ideally, in order to study the mechanisms that lead to road trauma, there is no substitute for detailed analysis at the scene of the accident with follow-up of casualties over an extended period. However, the cost and logistics associated with this method preclude its use for large numbers of accidents. Personal injury insurance claim data can also be utilised in the surveillance of road injury or death, overcoming the problem of accidents not reported to the police. However, the problem of encoding information not collected for research purposes can be immense and sometimes biased. Databases like the Road Injury Database provide effective alternatives. A database such as the Road Injury Database is not without limitations. There are many problems associated with the use of data derived from routinely collected mass data, whether from police, hospital, ambulance, or insurance sources. Not only are many variables restricted by the coding system imposed by the collecting agency, but data integrity is always a concern. However, the data gathered through the linkage process described here will provide a valuable resource for accident researchers and could indicate areas suitable for more detailed investigation. Expansion of the Road Injury Database to include data from the same sources for further years is planned. This expanded database will be structured so that analyses for each calender year, as well as longitudinal studies over several years, can be carried out. Care with link allocation will be needed since the possibility of multiple accidents for an individual will increase with an expanded time frame. Data quality surveillance and expansion of the police dataset to include blood alcohol level and more details on road-user movements should improve the value and scope of this dataset.
665
Acknowledgements-The Road Accident Prevention Research Unit gratefully acknowledges the cooperation of the Police, Main Roads, and Health Departments of Western Australia, the St. John Ambulance Association and the Registrar General, in providing source data and assisting with numerous enquiries. Discussions with the Accident and Emergency Department of Sir Charles Gairdner Hospital, the State Forensic Laboratory, and the Royal Flying Doctor Service concerning the scope for inclusion of other relevant data are also gratefully acknowledged. Finally, we thank our colleagues in the Road Accident Prevention Research Unit and the Department of Public Health for computing support and numerous discussions concerning various aspects of this project.
REFERENCES Agran, P. F.; Dunkle, D. E. A comparison of reported and unreported noncrash events. Accid. Anal. Prev. 17:7-13; 1985. Arnold, P. K.; Rosman, D. L.; Thornett, M. L. Pedestrian crash risk in Western Australia for both pedestrians and drivers. Road and Transport Research 1:60-75; 1992. Cercarelli, L. R.; Arnold, P. K.; Rosman D. L.; Sleet, D. Travel exposure and choice of comparison crashes for examining motorcycle conspicuity by analysis of crash data. Accid. Anal. Prev. 24:363-368; 1992. Date, C. J. An introduction to database systems. AddisonWesley Systems Programming Series. Reading, MA: Addison-Wesley; 1986. Fife, D. Matching fatal accident reporting system cases with National Center for Health Statistics motor vehicle deaths. Accid. Anal. Prev. 21:79-83; 1989. Gordon, M.; Charlton, G.; Ravazdy, K.; Lam, P.; Langley, I.; Williams, J.; Young, A.; Hardes, G.; Gibberd, R. Hospital based systems for studying road crashes: Hunter region study. Report CR 46. Canberra: Federal Office of Road Safety; 1986. Hendrie, D.; Rosman D. L.; Harris, A. H. Hospital inpatient costs resulting from road crashes in Western Australia. Perth: Road Accident Prevention Research Unit, University of Western Australia; 1992. Hill, T. Generalized Iterative Record Linkage System: GIRLS (Glossary, Concepts, Strategy guide, User guide). Ottawa: Statistics Canada; 1981. Hill, T.; Pring-Mill, F. Generalized Iterative Record Linkage System: GIRLS (revised ed.) Ottawa: Statistics Canada; 1986. Lindeuer, J. E. A trial linkage of the road accident and vehicle registration files. Accid. Anal. Prev. 19:91-104; 1987. National Center for Health Statistics (NCHS). International classification of diseases. 9th revision. Clinical modification. Washington: NCHS; 1988. Newcombe, H. B. Handbook of record linkage. Methods for health and statistical studies, administration, and business. Oxford Medical Publications. New York: Oxford University Press; 1988. Rosman, D. L.; Knuiman, M. W. A comparison of hospital and police data in the Western Australian road injury database. Perth: Road Accident Prevention Research Unit, University of Western Australia; 1992.