An Evaluation of Severity And Outcome Of Injury By Type Of Object Struck (First Object Struck Only) for Motor Vehicle Crashes in Connecticut

Table Of Contents

list of Tables

List of Figures

List Of Abbreviations

Abstract

Introduction

Methods

Results

Discussion

Recommendations

Appendix A

Appendix B

References

List Of Abbreviations
Table of Content

Abstract
Table of Content

A deterministic algorithm was developed which allowed data from Department of Transportation motor vehicle crash records, state mortality registry records, and hospital admission and emergency department records to be linked for analysis of the types of objects struck in motor vehicle crashes in Connecticut in 1995, and their consequences. Only objects listed on the police report as ‘first object struck' were analyzed. Of 132,918 vehicular crash records, 14.4% were thus identified as having struck an object; similarly, fourteen percent of the crashes which resulted in treatment in the ED or hospital admission visits were thus identified as having struck an object. Mean total hospital charges for these visits were $3,021; the mean length of stay for those visits resulting in an admission was slightly less than 5 days. Logistic regression analysis identified the most frequently reported factor correlated with striking an object to be driver illness (identified by the traffic safety officer), followed by dark conditions and speeding. Deer were the first object struck almost 5% of the time, resulting in 1.7% of the hospital visits. The highest mean total charges were for crashes involving a wall as first object struck, $5,986, while the first object struck resulting in the highest frequency of hospital visits were metal beam guide rails, in 13% of the cases (with mean total charges of $3,326). Wire rope guide rails and Jersey barriers resulted in lesser utilization of hospital services, while metal beam guide rails, vehicles off road, and trees resulted in greater medical utilization. Similarly, the largest number of fatalities resulted from crashes involving metal beam guide rails and trees as first objects struck.

This report examines motor vehicle crashes occurring in Connecticut during 1995, using several linked data sets. The findings reported herein illustrate the usefulness of using linked data sets to perform these types of analyses. Alone, each data set could not provide the type and depth of information provided by the group of linked data sets.

Data sets used for the studies include:

• The CHIME^® database, including Inpatient and Emergency Department data
• Ambulatory Surgery data from 31 general acute care facilities
• State of Connecticut, Department of Transportation (DOT) crash file
• State of Connecticut Mortality Data Set (CTMDS).

The CHIMEÒ dataset identifies all people involved in a MVC (motor vehicle crash) who had inpatient, emergency, or ambulatory surgery treatment at a Connecticut facility regardless of the state in which the MVC occurred. The DOT dataset identifies all MVCs and people involved in a crash, regardless of whether or not they had treatment at a hospital. The mortality dataset identifies deaths from MVCs. It includes all deaths from MVCs in Connecticut, whether the fatality was a resident of Connecticut or not, in addition to deaths of Connecticut residents who died in MVCs outside Connecticut which were reported by the state where they died.

Linking these data sets allows in-depth analysis of severity and outcome of injury by type of object struck. For instance, for the evaluation of severity and outcome of injury by type of object struck, the type-of-object-struck variable is contained in the Department of Transportation data set. Linking the CHIME^® and DOT databases contributes diagnoses and procedures resulting from the crash as well as total hospital and Emergency Department charges, while linking the mortality database furnishes information regarding the eventual survival or mortality of the persons involved. Thus the full picture of the effects of crashes with any particular object may be observed.

What follows are a description of the linking, a statistical analysis of the data, and a summary of our findings.

This study was funded in part by the National Highway Traffic Safety Administration as part of the CODES demonstration project, and performed in collaboration by the Connecticut Healthcare Research and Education Foundation (CHREF, a non-profit affiliate of the Connecticut Hospital Association), the State of Connecticut Department of Transportation (DOT), and Hartford Hospital.

The MVC data were obtained from the 1995 Collision Analysis Auxiliary Input (CAAI) Files. This is a database of motor vehicle crash data, owned by the State of Connecticut Department of Transportation.

There are six different record formats in the DOT files, described as follows:

• Record Type 1: Crash Summary Record
• Record Type 2: Traffic Unit Information Record
• Record Type 3: Traffic Unit Pen-Based Only Record
• Record Type 4: Involved Person Record
• Record Type 5: Property Damage Record
• Record Type 6: Crash Narrative Record.

Record Types 1, 2 and 4 were used for this analysis. Record Type 1 contains information pertinent to the crash as a whole, such as date and time, location and other crash-specific information. Record Type 2 identifies each vehicle or pedestrian involved in a crash, defined as a vehicle involved in a crash or a pedestrian who was struck by a vehicle involved in a crash. Record Type 4 contains information about vehicle operators, struck pedestrians, passengers, and witnesses. If more than four persons were involved in a crash, more than one person-record was created. Table 1 summarizes the number of records in these files.

Table 1. Summary Of 1995 Collision Analysis Input Files

The working MVC data file was constructed based on Type 1, 2 and 4 records in the DOT file. Type 1 records were merged with Type 2 records, to produce a file of one record per vehicle or pedestrian involved in a crash. The Type 4 records were converted from one record for each 1 to 4 involved persons into one record per involved person (i.e., if there were 4 people involved in a crash, the original file had one Type 4 record but the converted file has 4 records), then merged with the file of involved vehicles and pedestrians. This process produced one record for each involved person, containing all the data describing that person, as well as the specific crash and the specific vehicle. Table 2 categorizes the records contained in the DOT file.

The CHIME^® database was used for this analysis. Included in the CHIME^® database is demographic, clinical and financial information about each patient visit occurring in Connecticut acute care hospitals.

Data were extracted from this database in a two step process. In the first step, an index file containing information about Connecticut hospital ED visits, ambulatory surgery visits, and inpatient stays during 1995 was created for all patients having an ICD-9-CM code ranging from E810 to E819 (motor vehicle traffic crash E-codes), as detailed in Table 3.

In the second step, a medical history file containing the previous year's hospital visit information for those patients having an MVC in the index year was created. There were 40,883 records in the index CHIME^® database and 12,280 records in the history CHIME^® database.

Mortality data for victims of motor vehicle crashes were derived from the State of Connecticut Mortality Database (CTMDS). This database is offered to individuals and institutions from the State of Connecticut Department of Public Health, Office of Planning & Evaluation, Vital Records Bureau, and offers a comprehensive view of primary causes of mortality in Connecticut.

There were a total of 390 records selected from the state of Connecticut 1995 mortality database as possessing a motor vehicle crash related cause of death. Table 4 details these records by location of residency and location of crash.

A proprietary deterministic matching algorithm was developed in the FOCUS language to merge these databases. Key variables used to link the crash and hospital data were date of crash, date of birth, date of ED visit, date of inpatient admission(s), date of death, gender, and towncode of crash. Because passenger DOT records do not specify a gender, three steps of merging were employed. The first step included only driver and pedestrian records, with gender identified in the DOT database. The second step included passenger records from the DOT database, for which gender cannot be used as a linking variable. The third step included all unmatched records from the first and second steps. This algorithm did not allow for fuzzy or probabilistic linking; however, since crash date and ED or inpatient admission date would not always be expected to match exactly, four levels of date window were allowed within each matching step.

One hundred percent complete linkage is not expected when linking the DOT crash database to the CHIME^® database; for instance, if a motor vehicle crash occurred outside the state of Connecticut and the victim was taken to a Connecticut emergency room, or admitted to a Connecticut hospital, the patient would be included in the CHIME^® database but not the DOT database. Conversely, anyone who had a crash occurring in the state of Connecticut and was admitted to a hospital or ED outside of Connecticut would be included in the DOT database but not in the CHIME^® database. The result of this slight disjunction between the underlying pools of subjects is that the maximum linkage rate attainable will be reduced below 100% by an unknown amount, since we do not have a count of persons involved in either out of state crashes, or out of state hospital visits.

The mortality registry contains some records of Connecticut residents who die in other states, dependent on the other state's reporting them. Therefore, similarly to the above, Connecticut residents who die out of state in a crash might appear in the mortality database, but not in the DOT or CHIME^® databases. Conversely, a person injured in a crash in Connecticut and admitted to a Connecticut hospital, but who eventually dies out of state, might appear in the DOT and CHIME^® databases, but not in the mortality registry. Again, this would reduce the maximum attainable rate of linkage to the mortality registry, by an amount that we are not able to predict.

Table 5 describes the matching steps and levels in the merging algorithm. The output linked-dataset was inspected to verify the quality of the match.

This study was divided into two phases. The first phase analyzed all eligible DOT records to determine the distribution of the variables under examination and identify significant predictors of these variables and their odds ratios. The second phase was restricted to cases that successfully linked or merged, with a primary goal of determining the clinical events after MVCs.

The outcome variable for the first phase was the frequency of object struck, evaluating only the first object struck. The outcome variables for the second phase of the study included mortality, total charges, and length of stay.

Total hospital charge was calculated on an unadjusted basis only, due to lack of cost/charge ratio information. Mortality was categorized as died at the crash site, Emergency Department death (died in hospital with zero length of stay), died as inpatient (died in hospital with length of stay equal to or greater than 1 day), and died after discharge.

Independent variables in this study were drawn from two sources, the DOT data file and the CHIME^® database. Those variables included demographic, geographic, subjective, and objective factors, road and weather/season condition, police judgment/investigation, and clinical variables. Demographic variables included gender (female or male). Geographic variables included location of the crash and location of the object struck. Subjective factors included were speeding, following too closely, violating traffic controls, unsafe use of highway by pedestrian, etc. Objective factors included driver illness, vehicle involved in emergency, etc. Road condition included construction and road surface. Weather/seasonal variables included snow and rain. Police judgment/investigation included whether or not the driver had been drinking, and lighting conditions. Clinical variables included having at least one MVC and a hospital visit and admission diagnosis codes within past one year or 6 months. Other variables included type of motor vehicle, collision type, and injury classification. All categorical variables were converted into binary variables, as required for the analysis.

For the first phase of the study, the frequency for each outcome in the studied cohort was determined. The bivariate associations with outcome of road condition, weather/season condition, police judgment/investigation, demographic, geographic, subjective, objective, and clinical variables were evaluated, then a stepwise logistic regression model with a group of independent variables was developed, to find the significant predictors. Candidate independent variables were selected from the variables identified in the bivariate analysis as having an association with p < 0.10.

All stepwise models were constructed with an entry significance level of 0.01 and an exit significance level of 0.05, chosen to identify a parsimonious set of independent variables in the models. Partial residual plots were used to evaluate potential problematic areas of fit. Goodness-of-fit was evaluated by comparing fitted probabilities with observed value of dependent variables within deciles of probability, and calculating the corresponding observed chi-square statistic. In addition, an area under the receiver operator curve for logistic models was calculated to evaluate the models' predictive power.

An adjusted odds ratio was derived in which each odds ratio was adjusted for all other independent variables listed. An odds ratio less than 1 indicates that a crash event with that characteristic has a lower likelihood of association with the outcome variable than without that characteristic, while an odds ratio higher than 1 indicates that a crash event with that characteristic has a higher likelihood of association with the outcome variable than without that characteristic. For each of the studies, the logistic regression model's odds ratios and 95 percent confidence intervals for predictors were reported. In addition, a chi-square test or non-parametric test was performed for each bivariate analysis.

All calculations were performed using the software systems SASÒ 6.12 (SAS Institute, Cary, NC) and STATAÒ 3.0 (STATA Corporation, College Station, TX).

There were 40,883 records selected from the CHIMEÒ data set as having motor vehicle crash related E-codes, as detailed in Table 3. Of these, 35,832 records (87.6%) were linked and merged. After deleting duplicate records (1,054, 2.9%), 34,778 records remained (85.1%). Of these records, 364 (1%) were excluded from future analysis by reason of unreliable key variables.

Table 6 and Figure 1 show the linkage/merging rate of CHIME^® records for each of the linkage levels described in Table 5, classified by crash severity index in the Type 1 record of the DOT file. Since gender is such a useful linking variable, levels 1 through 4 link drivers and pedestrians only; levels 5 through 12 link passengers (who do not have gender recorded by the DOT) and individuals with gender unrecorded by reason of incomplete or defective records.

One hundred percent complete linkage is not expected when linking DOT files to all Connecticut hospital and emergency department discharges, since, if a motor vehicle crash occurred outside the state of Connecticut and the victim was hospitalized or admitted to a Connecticut hospital, the patient would be included in the CHIME^® database but not the DOT database. Conversely, anyone who had a crash occurring in the state of Connecticut and was admitted to a non-Connecticut hospital or ED would be included in the DOT database but not in the CHIME^® database. If both such cases could be eliminated, the final linked and merged rate would be higher than the current 85.1%.

A total of 329 records (84% of the 390 motor vehicle crash related fatalities) from the Connecticut Mortality dataset were successfully linked and merged with the DOT and CHIME^® files.

Overall, there were a total of 72,639 motor vehicle crashes reported to the DOT in the state of Connecticut during calendar 1995 (38 records of the total 72,677 were excluded due to duplication), involving 136,165 vehicles and pedestrians and 183,358 individual persons (Table 1 and Table 2); of the total persons involved in a crash, 34,778 (19%) were successfully linked to an ED visit or hospitalization (Table 6), and 329 to a mortality entry.

Figure 2. Precentage of Crashes in Connecticut, 1995, by Town or City

Figure 2 shows a geographical view of the percentage of total crashes by town or city, calculated as the number of crashes in the index town or city divided by total crashes in the state. As can be seen, the highest rates occur in towns and cities surrounding Interstate 91 (I-91), Interstate 95 (I-95) between the New York border and New Haven, Route 15, Interstate 84 (I-84), and Interstate 395 (I-395) between I-95 and Route 6.

There are 169 towns or cities recorded in the DOT files, with crash rates ranging from 0.01% to 5.1%. The five lowest towns or cities were Lyme (0.01%), Warren (0.01%), Colebrook (0.02%), Hampton (0.02%), and Hartland (0.02%), while the five highest were New Haven (5.07%), Hartford (5.00%), Bridgeport (4.81%), Stamford (3.20%), and Norwalk (2.93%). Appendix A details the crash rates by town.

Figure 3. Rate of Injury for CT Motor Vehicle Crashes by Town or City

Figure 3 shows the rate of injury by town or city in the state of Connecticut. Presence of injury was determined from the DOT Type 1 record injury severity code, including fatalities or any type of injuries, but excluding property damage only. Rate of injury was determined as number of injuries divided by total crashes in the index town or city.

Overall, the injury rate ranged from 23% to 70%; the five lowest town or cities were Old Lyme (22.89%), Madison (23.71%), Chester (25.00%), Essex (25.25%), and Guilford (27.17%), while the five highest were Sterling (69.57%), Hartford (63.38%), Hampton (62.50%), Windsor Locks (60.81%), and New Haven (59.82%). Appendix A contains detailed data for Figure 3.

Figure 4. Mortality by Position in Motor Vehicle and Place of Death

Figure 4 and Table 7 show mortality by position in vehicle (driver, passenger, or pedestrian) and place of death (at the crash site, emergency department [LOS = 0], inpatient [LOS > 0], or after discharge).

Figure 5 and Table 7 show mean age of fatalities by place of death. Inpatient deaths tended to be older than the other classes of fatalities. There was no significant difference between males and females.

Figure 6. Fatality Rate of Motor Vehicle Crashes by Town or City

Figure 6 shows fatality rate of crashes by town or city, determined as the number of deaths divided by number of crashes in each town or city. The mortality rate ranged from 0 to 10%, the five highest areas being Lyme (10%, 1 killed in 10 crashes), Hampton (6.25%, 1 killed in 16 crashes), Andover (4.76%, 2 killed in 42 crashes), Pomfret (4.23%, 2 killed in 71 crashes), and Canaan (4.12%, 1 killed in 24 crashes).

Figure 7. Percentage of Total Mortality by Town or City

Figure 7 shows mortality by town or city where crash occurred, as a percent of total state mortality. By this measure, Hartford, New Haven, Bridgeport, Waterbury, and Bristol accounted for 29.5% of total state mortality. There were 59 towns or cities where mortality was zero (no one killed by crashes in those areas during 1995).

For the study of factors influencing crashes involving a non-vehicular object, the study sample consisted of 132,918 vehicle records. Of these, 14.4% (19,153) had an object listed as first object struck, as detailed in Table 8.

The mean age of drivers identified as having struck an object was 38 years, with standard deviation of 16.2; 37% were female, 59% male, and 3% with gender unrecorded. There was a large variation between towns in the percentage of crashes identified as having struck an object, ranging from 6.2% to 77.5% of total crashes in each town, as shown in Figure 8. Appendix A enumerates the number of cases with objects struck, by town. As might be expected, the higher frequencies of objects struck appear in towns with lower traffic density.

Figure 8. Rate of Having an Object listed as First Object Struck, by Town/City

Results

The bivariate analysis of characteristics associated with object struck is detailed in Appendix B; the odds ratios are shown in Table 9.

Adjusted odds ratio was derived from a multiple regression analysis in which each odds ratio was adjusted for all other independent variables listed. An odds ratio less than 1 indicates that a crash event with that characteristic has a lower likelihood of being recorded with any first object struck, while an odds ratio higher than 1 indicates that a crash event with that characteristic has a higher likelihood of being recorded with any first object struck.

Driver illness had, by far, the highest correlation with a crash involving an object struck (odds ratio 5.37, 95% confidence interval 4.16 to 6.92), followed by dark conditions (odds ratio 2.38, 95% confidence interval 1.79 to 3.17), and speeding (odds ratio 2.05, 95% confidence interval 1.93 to 2.18).

The study sample was limited to the 34,778 persons in the merged file of DOT records that linked with CHIME^® database records, less the 364 (1%) which had unreliable key variables and were excluded from future analysis. The mean age of the 34,414 persons examined was 32 years, with a standard deviation of 15.8; 44% were female and 56% male. Since this study focused on medical and financial consequences of crashes, records from all involved persons were included in the analysis, rather than just drivers. Of these 34,414, 25,184 (73%) were drivers, 8,446 (25%) were passengers, and 783 (2%) were pedestrians.

A total of 4,885 (14.2%) of the study sample of 34,414 persons were identified as having been in a crash involving striking at least one object. Table 10 and Figure 9 detail the mean length of stay and total hospital charge by type of first object struck. For crashes resulting in hospital charges, the mean total hospital charge was $3,021 with standard deviation of $11,122 (white bars on Figure 9, axis at bottom of chart). Among crashes that had hospital charges, there was a wide variation by first object struck in the percentage resulting in inpatient admissions, possibly reflecting the underlying variation in severity of impact. For crashes resulting in inpatient admissions, the mean length of stay by first object struck varied from 1.3 days (4 cases) to 16 days (1 case), with an overall mean of 4.87 days and a standard deviation of 5.86 (black bars on Figure 9, axis at top of chart; 4 or more admissions only). Since many of the frequencies in this table are low, the means shown should not be regarded as definitive. The most frequent first objects struck which resulted in hospital charges were metal beam guide rails, 626 (13%) cases with mean charge $3,326, while the highest mean charge, $5,986, was for crashes involving a wall as first object struck. Figure 10 shows total mortality by first object struck. The first objects struck associated with the highest number of fatalities were metal beam guide rails and trees, followed by curbing, and banks, ledges, and rocks (off road).

Once again, this analysis only delineates the relationship between mortality and the first object struck, while the actual cause of the injury and mortality could be a subsequent object. This study is somewhat biased by the possibility that subsequent objects struck could be responsible for the majority of the injury, and consequently the majority of the hospital charge, length of stay, and mortality; this is likely to be the case for several of these objects, such as curbing, and possibly guide rails and off road banking.

Figure 10. Total Mortality by Object Struck

Approximately 14% of the motor vehicle crashes and their related hospitalizations studied here involve striking a non-vehicular object. Driver illness, as determined by traffic safety officer at the scene, was strongly associated with crashes involving an object struck, even adjusting for other factors. This finding suggests that drivers should be educated regarding the risks of driving while ill. More detailed study of what kinds of illness and what history of illness are associated with these crashes would be valuable; linkage of the drivers' histories of medication and surgery could identify particular medications which are problematic, or identify general or specific tendencies to release patients for driving too soon after surgery. Other conditions highly correlated with striking a non-vehicular object were darkness, and speeding. Quantifying the costs of speeding, particularly after dark, can justify expending resources on interventions designed to prevent this behavior.

The effects of striking an object varied greatly with the type of object struck. Striking a wire rope guardrail was associated with less severe consequences than striking a metal beam guide rail. Further analysis of the linked data could factor out confounding variables such as differences in average roadway speed, leading to determination of whether there is a general difference in safety between the two types of guard rail, or identifying specific locations which are hazardous. Impact with a tree was also associated with a high degree of severity. Again, using the linked data could determine whether there were any particular locations that were particularly dangerous, or whether this was a general overall risk. Such determinations would allow for informed decisions regarding modifications to existing roadways and design of new ones.

This project demonstrated that the individual data sets (CAAI data, CHIME^® database, ED data, and CTMDS data) can be successfully linked together, permitting sophisticated analyses that would otherwise be impossible. The medical database generates useful information on the type and severity of injury to organ systems that have been damaged, as well as the length of stay in the Emergency Department, the Intensive Care Units, and the hospital. The value and utility of the medical database are greatly enhanced by the ability to identify and correlate specific environmental elements, such as road conditions and time of day or night, physical conditions such as type of car and type of object struck, personal conditions such as the use of seat belts or air bags, and specific injuries to the people involved. It is now possible to examine the impact of environmental and physical variables and determine the differences in cost and outcome.

Linked data can make possible substantial progress in the design and safety improvement of motor vehicles and roadways. The relative risks and consequences of the placement and layout of automobile design features and roadway features can be quantified by coupling the data to crash outcomes in terms of personal injury, loss of independence, and cost. Specific injury prevention and public policy recommendations can now derive from carefully performed studies statistically controlled for extraneous environmental and physical factors, using linked data to compare outcomes of different types and severities of crash with reference to mortality, length of stay, ICU stay, rehabilitation, and cost. For instance, the type of evidence presented here regarding the relative risks of differing types of barriers and guardrails can inform the design and decision making process for highway and roadway planners. Similarly, the type and usage of frontal air bags, side air bags, and rear seat restraints, and how these factors interact with various objects struck to affect type of injury, outcome, cost, and rehabilitation, could be very helpful to both legislative bodies and vehicular design and manufacturing interests.

This data linkage project has demonstrated that large databases from the highway safety domain and the medical domain can be linked successfully. It has shown that mortality, morbidity, cost, and outcome data can be integrated with environmental and physical crash data to yield important information. This information can be helpful in shaping public policy relative to injury prevention.

Crashes with non-vehicular objects make up a considerable fraction of the crash-related morbidity and mortality. Factors influencing the frequency of such crashes, and the types of objects associated with the most severe results, can be identified and action designed to prevent such crashes planned and undertaken. Quantifying the costs associated with these crashes will help allocate the resources required to prevent them.

• Educate people regarding the risk of being overcome by illness while driving.
• Acquire pharmaceutical utilization data and link it with the merged traffic safety and medical/surgical procedure and outcomes data.
• Factor out other confounding variables, in order to identify either any overall high degree of risk represented by certain types of objects (trees, guard beams, etc.), or any specific locations where such objects are especially hazardous.
• Extend the study of type and outcome of injuries by first object struck to include all objects struck.

Appendix A
Table of Content

Crashes, Injuries, And Objects Struck, By Town or City

Table 11. Crashes, Injuries and Objects Struck, By Town or City

(N=132,918, Driver only)