- Email: [email protected]
Accuracy of self-reported data for estimating crash severity
Accident Analysis and Prevention 35 (2003) 833–840
Accuracy of self-reported data for estimating crash severity Michael R. Elliott a,∗ , Kristy B. Arbogast b , Rajiv Menon b , Dennis R. Durbin c , Flaura K. Winston b a
b
Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, 612 Blockley Hall, 423 Guardian Drive, Philadelphia, PA 19104, USA Children’s Hospital of Pennsylvania, 34th Street and Civic Center Boulevard, 3535 Traumalink 10th Floor, Suite 1024, Philadelphia, PA 19104, USA c Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, 711 Blockley Hall, 423 Guardian Drive, Philadelphia, PA 19104, USA Received 24 April 2002; received in revised form 6 July 2002; accepted 11 July 2002
Abstract Estimated traveling speed and speed limit have typically been used in population-based surveillance data to estimate crash severity. The accuracy of these measures in predicting crash severity is unknown. The Partners for Child Passenger Safety (PCPS) surveillance system offers a unique opportunity to compare these measures, as well as a novel measure of crash severity, “self-report” delta-V, to the accepted measure of delta-V estimated during detailed crash-investigations in 118 crashes. This “self-report” delta-V was computed from the estimated traveling speeds and direction of impact obtained from telephone interviews with drivers. These “self-reported” delta-V estimates are modestly associated with crash-investigation delta-V estimates, with the degree of association a function of the direction of impact: when the respondent was struck from the rear, the degree of association is strong; frontal, side, and single-vehicle crashes yield weaker associations. This “self-reported” delta-V measure, however, is a substantial improvement over use of estimated traveling speed or speed limit only. © 2003 Elsevier Science Ltd. All rights reserved. Keywords: Delta-V; Crash-investigations; Momentum; Speed limit
1. Introduction Crash severity is an important variable to account for when measuring the risk of injury in motor vehicle crashes in large population-based studies. In studies using data not collected by crash-investigation, there is no universally accepted or validated way of accounting for crash severity. Measures that have been used include speed at the time of crash estimated either from self-report or speed limit on the roadway (Johnston et al., 1994; Shibata and Fukuda, 1994). In studies where in-depth investigation of the crash is performed, crash severity is estimated by delta-V or the instantaneous change in velocity; however, this method is far more expensive to ascertain. Comparing the relationship between the various crash severity estimates with the crash-investigation, delta-V can evaluate the appropriateness of using either estimated traveling speed or speed limit as a reasonable proxy of crash severity in the absence of crash-investigation measures. In addition to the traditional measures of crash severity, in ∗
Corresponding author. Tel.: +1-215-573-4467; fax: +1-215-573-4865. E-mail address: [email protected] (M.R. Elliott).
this paper we calculate a new measure of crash severity, “self-report” delta-V, by utilizing estimated traveling speed, the speed and type of the other vehicle to strike the case vehicle, and the principle of the conservation of momentum. The objective of this study is to compare the three measures of self-reported crash severity (estimated travel speed, speed limit, and “self-report” delta-V) with delta-V calculated as part of an in-depth crash-investigation for single-vehicle impacts and for front, side, and rear two-vehicle impacts.
2. Methods 2.1. Study population and data collection Data for this study were obtained from the Partners for Child Passenger Safety (PCPS), which consists of a large-scale crash surveillance system created by linking electronic insurance claims data at State Farm Insurance Companies (Bloomington, IL) to telephone survey and crash-investigation data. Crashes qualifying for inclusion in the surveillance system were those involving at least
0001-4575/$ – see front matter © 2003 Elsevier Science Ltd. All rights reserved. PII: S 0 0 0 1 - 4 5 7 5 ( 0 2 ) 0 0 0 8 8 - X
834
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
one child occupant <15 years of age riding in a model year 1990 or newer State Farm-insured vehicle. Qualifying crashes were limited to those that occurred in 15 states and the District of Columbia, representing three large regions of the United States (east: NY, NJ, PA, DE, MD, VA, WV, NC, DC; mid-west: OH, MI, IN, IL; west: CA, NV, AZ). On a daily basis, data from qualifying and consenting claims were transferred electronically from all involved State Farm field offices to researchers at The Children’s Hospital of Philadelphia and University of Pennsylvania (CHOP/Penn). Data in this initial transfer included contact information for the insured, the ages and genders of all child occupants, and a coded variable describing the medical treatment received by all child occupants. The sample used in these analyses consist of reported claims between 1 December 1998 and 30 June 2001. For the telephone survey component of data collection, the surveillance data were subjected to an automated sampling algorithm to select a representative sample of claims to be included in the surveillance system, with those crashes most likely to involve serious injury oversampled to a high degree. Further details of this surveillance sampling scheme are available elsewhere (Durbin et al., 2001; Winston et al., 2000). Once selected for the telephone survey, 30-min interviews were attempted with the driver of the vehicle and parent(s) of the involved children. Only adult drivers and parents were interviewed; 82% of those sampled participated in the survey. The median length of time between State Farm’s notification of a crash and the completion of the interview was approximately 7 days. Respondents were asked to estimate their speed at the time of crash along with the speed of the other vehicle in the crash, if any. (Only single-vehicle and two-vehicle crashes were included due to the complexity of multiple-vehicle crashes.) Direction of first impact from the telephone survey was derived from a series of questions regarding the vehicle parts that were involved in the first collision. Passenger compartment intrusion was also assessed via the telephone interview. For the crash-investigation component of data collection, a series of cases were chosen based on manual review of claims files to identify crashes requiring crash-investigations (e.g. crashes involving at least one child occupant admitted to the hospital, as well as crashes meeting any of several targeted areas of child occupant safety research such as air bag interaction or side impact collisions with intrusion). Cases were screened via telephone with the policyholder to confirm the restraint status and medical details of the case. Contact information from selected cases was then forwarded to a crash-investigation firm (Dynamic Science Inc., Annapolis, MD), and a full-scale on-site crash-investigation was conducted using custom child-specific data collection forms. The data forms were based on those used by the National Automotive Sampling System (NASS) with particular areas enhanced to capture more child-specific information, i.e. more detailed restraint information and child anthropometrics. Crash-investigation teams were dispatched
to the crash scenes within 24 h of notification to measure and document the crash environment, damage to the vehicles involved, and occupant contact points according to a standardized protocol. The on-scene investigations were supplemented by information from witnesses, crash victims, physicians, hospital records, police reports, and emergency medical service personnel. From this information, reports were generated that included estimates of the vehicle dynamics and occupant kinematics during the crash, as well as detailed descriptions of the injuries sustained in the crash by body region, type of injury, and severity of injury. Posted speed limit at the location of the crash is also obtained during the crash-investigations. Delta-V (the instantaneous change in velocity), an accepted measure of crash severity, was calculated using WinSmash (NHTSA, Washington, DC, 1997) and crush measurements of the vehicles. This program assigns a stiffness value to each vehicle make and incorporates this information into the delta-V calculation. A total of 140 crashes in which both crash-investigation and surveillance data were obtained compose the study sample for these analyses. Of these, we excluded seven whose vehicles had multiple points of first contact; in addition, 15 drivers were either unable to supply information about their speed or, in the case of two-vehicle crashes, the other driver’s speed; or the crash-investigations did not provide a delta-V estimate. Thus, 118 crashes were utilized in these analyses. Additional data involving the mass of the subject’s vehicle involved in the crash was obtained from VINDICATOR (Insurance Institute for Highway Safety/Highway Loss Data Institute, Arlington, VA). Because the vehicle identification number was available only for the subject’s vehicle, the mass of the other vehicle in two-vehicle crashes was estimated as the mean mass of the vehicle type (passenger car, minivan, sport utility vehicle, or pickup truck) from the entire telephone survey database, based on 12,659 crashes. The mean weight ± standard deviation by vehicle type were as follows (in kg): passenger cars, 1301 ± 202; minivan, 1689 ± 112; sport utility vehicle, 1849 ± 314; and pickup truck, 1831 ± 313. (Note that we know the other vehicle weight from the crash-investigations, but we wish to construct a self-report measure using only data that could reasonably be obtained through self-report.) 2.2. Data analysis The “self-report” delta-V is estimated from the mass and reported velocities (speed and direction) of the vehicles at impact. For the single-vehicle crashes, it is simply the reported speed at time of impact. For two-vehicle crashes, the delta-V is determined from the reported speed of the other vehicle, the mass of the other vehicle (estimated as the mean of vehicle type), the direction of impact (front, side, and rear), and the law of the conservation of momentum, where it is assumed that, at least for the time necessary for the rapid deceleration to take place, the two vehicles are combined
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
835
Fig. 1. Vector computation of instantaneous change in velocity (delta-V) in multiple-vehicle crashes: m1 = mass of subject’s vehicle; m2 = mass of other vehicle; v1 = reported velocity of subject’s vehicle; v2 = reported velocity of other vehicle. Mass of subject’s vehicle taken from VINDICATOR; mass of other vehicle estimated as mean of vehicle type; velocities based on self-report of driver.
into a single body after impact (see Fig. 1). Angular impacts are assumed to take place at a 45◦ angle. The self-report delta-V is then the length of the vector difference between the initial and final velocity of the target (State Farm) vehicle. For each of these four crash types—single-vehicle and multiple-vehicle: front, side, and rear first impact sites—the “self-report” delta-Vs are then plotted against the delta-Vs calculated based on the crash-investigation, and a linear regression slope is computed using SAS Version 8 (SAS Institute, 2000). That is, we model the relationship between the crash-investigation delta-Vs yi and the self-report delta-Vs xi as yi = α + βxi + εi
(1)
where α and β are unknown constants and εi are unknown, normally-distributed error terms. (The residuals obtained after fitting the least squares regression of the self-report versus crash-investigation delta-Vs are approximately normally distributed.) We use linear regression to estimate the intercept α and slope β: if the self-report and crash-investigation delta-Vs correspond on average, the intercept and slope of this regression line will clearly be 0 and 1, respectively; we test this null hypothesis. We also report the proportion of
variability in the crash-investigation delta-Vs explained by the self-report estimates using (1) (R2 measure). We also consider whether the delta-Vs derived from the self-reports have the same rank–order as the crash-investigation-derived delta-Vs by considering the Spearman’s rho statistics, a measure of rank–order correlation (Sprent, 1990). We then conduct a parallel set of analyses using two simpler measures of crash severity, self-reported speed of the respondent’s vehicle and speed limit of the road on which the crash took place, in place of the self-report delta-V estimates. We compare the results to those obtained using self-report delta-V estimates.
3. Results Table 1 shows the distribution of reported speeds of subject vehicle and other vehicle by point of first contact, together with the posted speed limit and crash-investigation measures of delta-V. The crash-investigation delta-Vs and posted speed limits do not differ significantly by point of first contact (P = 0.79 and 0.92, respectively), whereas the subject vehicle speed is reported to be much lower and the other vehicle speed much higher in side and
836
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
Table 1 Crash-investigation delta-V, subject-reported speed of subject vehicle and other vehicle to strike subject vehicle, and posted speed limit, by point of first contact Point of first contact
Sample size
Crash-investigation delta-V (km/h)
Subject vehicle reported speed (km/h)
Other vehicle reported speed (km/h)
Posted speed limit (km/h)
Mean (S.D.)
Range
Mean (S.D.)
Range
Mean (S.D.)
Range
Mean (S.D.)
Range
Single-vehicle
20
31.3 (13.4)
14–58
59.7 (24.2)
19–104
N/A (N/A)
N/A
69.6 (19.8)
40–89
Two-vehicle Front Side Rear
53 29 16
30.3 (14.1) 27.6 (11.1) 31.1 (16.3)
10–79 12–60 11–69
54.9 (27.1) 33.2 (28.9) 3.8 (6.8)
0–96 5–104 0–24
53.6 (31.6) 76.8 (23.7) 72.5 (17.7)
0–104 8–141 45–104
72.1 (16.4) 71.8 (19.2) 69.1 (17.2)
32–89 40–105 40–89
S.D.: standard deviation.
rear impacts than in front or single-vehicle impacts (both P < 0.001). 3.1. Single-vehicle crashes Comparing the “self-report” delta-Vs with the delta-Vs obtained from the crash-investigations for n = 20 single-vehicle crashes, we find that the least squares regression intercept is 19.6 and the slope is 0.19; 11% of the crash-investigation delta-V variability is explained by the self-report delta-V estimates (see Fig. 2). The crash-investigation delta-Vs tended to be lower than the self-reported speeds. For intrusion crashes (n = 11), the association appeared stronger (intercept = 8.7, slope = 0.27) than for non-intrusion crashes (n = 9; intercept = 44.4, slope = −0.10), although this difference was not
statistically significant (P = 0.14). The rank–order association was relatively weak (Spearman’s rho = 0.20, P = 0.07). The results using reported speed-only are identical since this was the measure used to compute delta-V in single-vehicle crashes. The relationship between crash-investigation delta-V and posted speed limit is actually negative (intercept = 46.9, slope = −0.24; Spearman’s rho = −0.45) (see Table 2). 3.2. Two-vehicle frontal impacts Comparing the “self-report” delta-Vs with the deltaVs obtained from the crash-investigations for two-vehicle crashes where the point of first contact is in the front (n = 53), we find that the least squares regression intercept is 15.9 and the slope is 0.27; 14% of the crash-investigation
Fig. 2. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for single-vehicle crashes. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
837
Table 2 Least squares fit of self-report estimate of delta-V, self-report speed of respondent and posted speed limit vs. crash-investigation estimate of delta-V Point of first contact
Estimated intercept α (S.E.)
Estimated slope β (S.E.)
P-value under Ho (α = 0, β = 1)
R2
Spearman’s rho (P-value)
0.19 (0.13) 0.19 (0.13) −0.24 (0.15)
<0.001 <0.001 <0.001
0.113 0.113 0.148
0.201 (.075) 0.201 (0.075) −0.451 (0.069)
Single-vehicle Delta-V from self-report Speed-only estimate from self-report Posted speed limit
19.6 (8.1) 19.6 (8.1) 46.9 (10.8)
Two-vehicle front Delta-V from self-report Speed-only estimate from self-report Posted speed limit
15.9 (5.4) 27.1 (5.4) 16.0 (8.8)
0.27 (0.09) 0.06 (0.07) 0.20 (0.12)
<0.001 <0.001 <0.001
0.135 0.012 0.053
0.261 (0.059) 0.165 (0.24) 0.180 (0.20)
Two-vehicle side Delta-V from self-report Speed-only estimate from self-report Posted speed limit
14.2 (9.3) 25.8 (3.2) 15.6 (8.2)
0.30 (0.20) 0.06 (0.07) 0.16 (0.11)
<0.001 <0.001 <0.001
0.076 0.022 0.077
0.366 (0.051) 0.082 (0.67) 0.170 (0.40)
Two-vehicle rear Delta-V from self-report Speed-only estimate from self-report Posted speed limit
3.5 (13.6) 36.8 (3.1) −23.0 (9.7)
0.73 (0.35) −1.5 (0.50) 0.80 (0.14)
0.17 <0.001 <0.001
0.241 0.391 0.726
0.624 (0.007) −0.800 (0.002) 0.883 (<0.001)
R2 = proportion of variation in crash-investigation delta-Vs explained by self-report-estimated delta-Vs, self-reported speed and posted speed limit.
delta-V variability is explained by the self-report estimates (see Fig. 3). As with single-vehicle crashes, frontal crashes with intrusion (n = 25) had better self-report and crashinvestigation correspondence (intercept = 18.9, slope = 0.28) than frontal crashes without intrusion (n = 28; intercept = 22.8, slope = 0.06), although this difference was not significant (P = 0.28). The rank–order association was relatively weak (Spearman’s rho = 0.26, P = 0.06).
The association between reported speed-only and crashinvestigation delta-V was far weaker (intercept = 27.1, slope = 0.06; Spearman’s rho = 0.17). The association between posted speed limit and crash-investigation delta-V was intermediate between the self-report delta-V and crashinvestigation delta-V association, and reported speed-only and crash-investigation delta-V association (intercept = 16.0, slope = 0.20; Spearman’s rho = 0.18) (see Table 2).
Fig. 3. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for two-vehicle crashes with frontal point of first contact. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
838
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
Fig. 4. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for two-vehicle crashes with side point of first contact. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
Fig. 5. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for two-vehicle crashes with rear point of first contact. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
3.3. Two-vehicle side impacts Comparing the “self-report” delta-Vs with the deltaVs obtained from the crash-investigations for two-vehicle crashes where the point of first contact is in the side (n = 29), we find that the least squares regression intercept is 14.2 and the slope is 0.30; 8% of the crashinvestigation delta-V variability is explained by the selfreport estimates (see Fig. 4). The rank–order association was moderate (Spearman’s rho = 0.37, P = 0.05). The association between crash-investigation delta-V and reported speed of subject vehicle was again far weaker (intercept = 25.8, slope = 0.06; Spearman’s rho = 0.08). The difference between intrusion and non-intrusion crashes is not considered because of insufficient sample size for non-intrusion crashes (n = 4). As with frontal impacts, the posted speed limit–crash-investigation deltaV association was intermediate between the self-report delta-V–crash-investigation delta-V association and reported speed-only–crash-investigation delta-V association (intercept = 15.6, slope = 0.16; Spearman’s rho = 0.17) (see Table 2). 3.4. Two-vehicle rear impacts Comparing the “self-report” delta-Vs with the delta-Vs obtained from the crash-investigations for n = 16 twovehicle crashes where the point of first contact is in the rear, we find that rear impacts have the greatest correspondence between the self-report-estimated and crash-investigationestimated delta-Vs: the least squares regression intercept is 3.5 and the slope is 0.73; 24% of the crash-investigation delta-V variability is explained by the self-report estimates (see Fig. 5). The null hypothesis of an intercept of 0 and a slope of 1 cannot be rejected (P = 0.17). There does not appear to be a difference in association between the (n = 6) intrusion and (n = 10) non-intrusion crashes, but the sample sizes are small. The rank–order association was moderately strong (Spearman’s rho = 0.63, P = 0.007). The association between crash-investigation delta-V and reported speed of subject vehicle was strong (R2 = 0.39), but in the wrong direction (intercept = 36.8, slope = −1.5; Spearman’s rho = −0.80): many of the worst crashes occurred in vehicles that were not moving. Posted speed limit, however, was highly correlated with crash-investigation delta-V, although it tended to overestimate the crash-investigation delta-V by approximately 20 km/h (intercept = −23.0, slope = 0.80; Spearman’s rho = 0.88) (see Table 2). 4. Discussion This report demonstrates that a novel measure of crash severity (“self-report” delta-V) presents an improved crash severity measure compared to estimated traveling speed and
839
speed limit for use in a large population study of crashes. This “delta-V” measure uses reported speed(s), direction of impact, and mass(es) of the vehicle(s) together with the principles of the conservation of momentum to produce a measure that more closely agrees with delta-V measures determined through crash-investigation. This self-reported measure is far from perfect, however, and errors in the self-reporting measures tend to bias the relationship between the self-report and crash-investigation delta-Vs toward the null (Weisberg, 1985). In the case of single-vehicle crashes, reported speed at the time of crashes was considerably higher than the crash-investigation-estimated delta-V. This is to be expected, as self-reported traveling speed does not take in account any pre-impact braking that would reduce the instantaneous change in velocity at the point of impact. Vehicles also might have struck elastic objects, although in this sample this does not appear to have been the case (of the 18 crashes in which the struck object was recorded, 11 (61%) involved trees or telephone poles and 2 (11%) walls or rocks). Alternatively, crash-investigations may underestimate delta-V in some non-intrusion crashes. Inaccuracy of the delta-V calculation in low severity crashes has been observed by other researchers (Noga and Smith, 1982). Intrusion crashes appeared to be better estimated than non-intrusion crashes, although this difference was not statistically significant. In two-vehicle crashes, rear impacts had the closest correspondence between the delta-Vs estimated from self-report and those estimated by crash-investigation. This is perhaps not surprising since rear impacts to subject vehicles generally occur when the subject is stopped (11 out of 16 crashes occurred with the subject vehicle being motionless), thereby removing the inherent variability associated with reported traveling speed. Frontal impacts also had crash-investigation delta-Vs lower than that estimated from self-reports. Again, this is most likely due to the effect of pre-impact braking and its influence on instantaneous change in velocity. As with single-vehicle crashes, intrusion crashes appeared to be better estimated than non-intrusion crashes, although this difference was not statistically significant. Self-reported delta-V estimates in side impact crashes also tended to be higher than those obtained from crash-investigations; in addition, the percentage of variability explained was the smallest among the crash types. This may reflect the fact that the estimation of crash severity in side impacts by a calculation of delta-V has many limitations, since the WinSmash program was developed using data from frontal crash tests only (Frampton et al., 1998). Perhaps the least one could ask of the self-report data is that it ranks the crash severity in the same order as the crash-investigations. For rear impact two-vehicle crashes, the agreement in ordering using the self-report estimates of the delta-Vs was moderately high (Spearman’s rho = 0.62). However, we found more modest associations for
840
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
single-vehicle, frontal, and side impact crashes (Spearman’s rho = 0.20, 0.26, and 0.37, respectively). Using reported speed alone was generally a far poorer proxy for crash-investigation delta-V than the “self-report” delta-V measure that attempted to include information about the momentum of both vehicles. A strong negative association was found between reported speed and delta-V in rear impacts, where motionless vehicles were often struck from behind at reported high speeds. Finally, posted speed limit had a generally stronger association with crash-investigation delta-V than reported speed, but was not as good a proxy as the “self-report” delta-V; an exception again being rear crashes, where it performed nearly as well as the self-report delta-V.
5. Conclusions We investigated the accuracy of three self-report measures of crash severity, estimated traveling speed, speed limit, and a novel measure of “self-report” delta-V, by comparing these measures to delta-V calculated from crash-investigation. The “self-report” delta-V utilizes all available information about estimated speeds, direction of impact, vehicle mass, and the principle of conservation of momentum. We find that, in general, this self-report delta-V estimate, although not a perfect proxy, provides a better prediction of crash severity than either estimated traveling speed or posted speed limit. This was particularly the case in rear impact collisions, where a reasonable accurate estimate of delta-V may be obtained.
4.1. Study limitations Acknowledgements Self-report data from all drivers in a multi-car impact, and detailed information on the mass of struck objects in single-vehicle accidents, rather than estimated speeds and vehicle types from the driver of the subject vehicle, would undoubtedly improve the correspondence between the self-report delta-V estimates and those obtained through detailed crash-investigations. Although the self-report data are obtained from a known-probability sample, the crash-investigations are not sampled under a known-probability design. In addition, the population from which both the surveillance survey and crash-investigations have been sampled is limited to crashes involving children in State Farm-insured vehicles. Hence, it is somewhat uncertain how generalizable these results are to all drivers. However, there is no reason to believe that the relationship between the “self-report” delta-V and the crash-investigation delta-V would be associated with the probability of selection or the State Farm-insured population itself. Given that complete data to estimate the delta-V from the surveillance survey are required, there is surprisingly little missing data: self-report and crash-investigation delta-Vs are available for 20 of 21 single-vehicle crashes, 53 of 63 front impact two-vehicle crashes, 29 of 31 side impact crashes, and 16 of 17 rear impact crashes. Hence, missing data is unlikely to introduce much bias in these results.
The authors would like to thank State Farm Insurance Companies for their financial support of this work through the Partners for Child Passenger Safety Project. References Durbin, D., Bhatia, E., Holmes, J.H., Shaw, K.N., Werner, J.V., Sorenson, W., Winston, F.K., 2001. Partners for Child Passenger Safety: a unique child-specific crash surveillance system. Accid. Anal. Prev. 33, 407– 412. Frampton, R., Brown, R., Thomas, P., Fay, R., 1998. The importance of non-struck side occupants in side collision. In: Proceedings of the 42nd Annual Conference of the Association for the Advancement of Automotive Medicine, Charlottesville, VA. Johnston, C., Rivara, F.P., Soderberg, R., 1994. Children in car crashes: analysis of data for injury and use of restraints. Pediatrics 93, 960–965. Noga, J.T., Smith, R.A., 1982. Accuracy and sensitivity of CRASH. In: Proceedings of the 26th Stapp Car Crash Conference, Ann Arbor, MI. SAS Institute Inc., 2000. Statistical Analysis System Version 8 for Windows. SAS Institute Inc., Cary, NC. Shibata, A., Fukuda, K., 1994. Risk factors of fatality in motor vehicle traffic accidents. Accid. Anal. Prev. 26, 391–397. Sprent, P., 1990. Applied Non-Parametric Statistical Methods. Chapman & Hall, London, pp. 172–175. Weisberg, S., 1985. Applied Linear Regression, 2nd ed. Wiley, New York, pp. 76–79. Winston, F., Durbin, D.R., Kallan, M.J., Moll, E.K., 2000. The danger of premature graduation to seat belts for young children. Pediatrics 105, 1179–1183.
Accuracy of self-reported data for estimating crash severity Michael R. Elliott a,∗ , Kristy B. Arbogast b , Rajiv Menon b , Dennis R. Durbin c , Flaura K. Winston b a
b
Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, 612 Blockley Hall, 423 Guardian Drive, Philadelphia, PA 19104, USA Children’s Hospital of Pennsylvania, 34th Street and Civic Center Boulevard, 3535 Traumalink 10th Floor, Suite 1024, Philadelphia, PA 19104, USA c Center for Clinical Epidemiology and Biostatistics, University of Pennsylvania School of Medicine, 711 Blockley Hall, 423 Guardian Drive, Philadelphia, PA 19104, USA Received 24 April 2002; received in revised form 6 July 2002; accepted 11 July 2002
Abstract Estimated traveling speed and speed limit have typically been used in population-based surveillance data to estimate crash severity. The accuracy of these measures in predicting crash severity is unknown. The Partners for Child Passenger Safety (PCPS) surveillance system offers a unique opportunity to compare these measures, as well as a novel measure of crash severity, “self-report” delta-V, to the accepted measure of delta-V estimated during detailed crash-investigations in 118 crashes. This “self-report” delta-V was computed from the estimated traveling speeds and direction of impact obtained from telephone interviews with drivers. These “self-reported” delta-V estimates are modestly associated with crash-investigation delta-V estimates, with the degree of association a function of the direction of impact: when the respondent was struck from the rear, the degree of association is strong; frontal, side, and single-vehicle crashes yield weaker associations. This “self-reported” delta-V measure, however, is a substantial improvement over use of estimated traveling speed or speed limit only. © 2003 Elsevier Science Ltd. All rights reserved. Keywords: Delta-V; Crash-investigations; Momentum; Speed limit
1. Introduction Crash severity is an important variable to account for when measuring the risk of injury in motor vehicle crashes in large population-based studies. In studies using data not collected by crash-investigation, there is no universally accepted or validated way of accounting for crash severity. Measures that have been used include speed at the time of crash estimated either from self-report or speed limit on the roadway (Johnston et al., 1994; Shibata and Fukuda, 1994). In studies where in-depth investigation of the crash is performed, crash severity is estimated by delta-V or the instantaneous change in velocity; however, this method is far more expensive to ascertain. Comparing the relationship between the various crash severity estimates with the crash-investigation, delta-V can evaluate the appropriateness of using either estimated traveling speed or speed limit as a reasonable proxy of crash severity in the absence of crash-investigation measures. In addition to the traditional measures of crash severity, in ∗
Corresponding author. Tel.: +1-215-573-4467; fax: +1-215-573-4865. E-mail address: [email protected] (M.R. Elliott).
this paper we calculate a new measure of crash severity, “self-report” delta-V, by utilizing estimated traveling speed, the speed and type of the other vehicle to strike the case vehicle, and the principle of the conservation of momentum. The objective of this study is to compare the three measures of self-reported crash severity (estimated travel speed, speed limit, and “self-report” delta-V) with delta-V calculated as part of an in-depth crash-investigation for single-vehicle impacts and for front, side, and rear two-vehicle impacts.
2. Methods 2.1. Study population and data collection Data for this study were obtained from the Partners for Child Passenger Safety (PCPS), which consists of a large-scale crash surveillance system created by linking electronic insurance claims data at State Farm Insurance Companies (Bloomington, IL) to telephone survey and crash-investigation data. Crashes qualifying for inclusion in the surveillance system were those involving at least
0001-4575/$ – see front matter © 2003 Elsevier Science Ltd. All rights reserved. PII: S 0 0 0 1 - 4 5 7 5 ( 0 2 ) 0 0 0 8 8 - X
834
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
one child occupant <15 years of age riding in a model year 1990 or newer State Farm-insured vehicle. Qualifying crashes were limited to those that occurred in 15 states and the District of Columbia, representing three large regions of the United States (east: NY, NJ, PA, DE, MD, VA, WV, NC, DC; mid-west: OH, MI, IN, IL; west: CA, NV, AZ). On a daily basis, data from qualifying and consenting claims were transferred electronically from all involved State Farm field offices to researchers at The Children’s Hospital of Philadelphia and University of Pennsylvania (CHOP/Penn). Data in this initial transfer included contact information for the insured, the ages and genders of all child occupants, and a coded variable describing the medical treatment received by all child occupants. The sample used in these analyses consist of reported claims between 1 December 1998 and 30 June 2001. For the telephone survey component of data collection, the surveillance data were subjected to an automated sampling algorithm to select a representative sample of claims to be included in the surveillance system, with those crashes most likely to involve serious injury oversampled to a high degree. Further details of this surveillance sampling scheme are available elsewhere (Durbin et al., 2001; Winston et al., 2000). Once selected for the telephone survey, 30-min interviews were attempted with the driver of the vehicle and parent(s) of the involved children. Only adult drivers and parents were interviewed; 82% of those sampled participated in the survey. The median length of time between State Farm’s notification of a crash and the completion of the interview was approximately 7 days. Respondents were asked to estimate their speed at the time of crash along with the speed of the other vehicle in the crash, if any. (Only single-vehicle and two-vehicle crashes were included due to the complexity of multiple-vehicle crashes.) Direction of first impact from the telephone survey was derived from a series of questions regarding the vehicle parts that were involved in the first collision. Passenger compartment intrusion was also assessed via the telephone interview. For the crash-investigation component of data collection, a series of cases were chosen based on manual review of claims files to identify crashes requiring crash-investigations (e.g. crashes involving at least one child occupant admitted to the hospital, as well as crashes meeting any of several targeted areas of child occupant safety research such as air bag interaction or side impact collisions with intrusion). Cases were screened via telephone with the policyholder to confirm the restraint status and medical details of the case. Contact information from selected cases was then forwarded to a crash-investigation firm (Dynamic Science Inc., Annapolis, MD), and a full-scale on-site crash-investigation was conducted using custom child-specific data collection forms. The data forms were based on those used by the National Automotive Sampling System (NASS) with particular areas enhanced to capture more child-specific information, i.e. more detailed restraint information and child anthropometrics. Crash-investigation teams were dispatched
to the crash scenes within 24 h of notification to measure and document the crash environment, damage to the vehicles involved, and occupant contact points according to a standardized protocol. The on-scene investigations were supplemented by information from witnesses, crash victims, physicians, hospital records, police reports, and emergency medical service personnel. From this information, reports were generated that included estimates of the vehicle dynamics and occupant kinematics during the crash, as well as detailed descriptions of the injuries sustained in the crash by body region, type of injury, and severity of injury. Posted speed limit at the location of the crash is also obtained during the crash-investigations. Delta-V (the instantaneous change in velocity), an accepted measure of crash severity, was calculated using WinSmash (NHTSA, Washington, DC, 1997) and crush measurements of the vehicles. This program assigns a stiffness value to each vehicle make and incorporates this information into the delta-V calculation. A total of 140 crashes in which both crash-investigation and surveillance data were obtained compose the study sample for these analyses. Of these, we excluded seven whose vehicles had multiple points of first contact; in addition, 15 drivers were either unable to supply information about their speed or, in the case of two-vehicle crashes, the other driver’s speed; or the crash-investigations did not provide a delta-V estimate. Thus, 118 crashes were utilized in these analyses. Additional data involving the mass of the subject’s vehicle involved in the crash was obtained from VINDICATOR (Insurance Institute for Highway Safety/Highway Loss Data Institute, Arlington, VA). Because the vehicle identification number was available only for the subject’s vehicle, the mass of the other vehicle in two-vehicle crashes was estimated as the mean mass of the vehicle type (passenger car, minivan, sport utility vehicle, or pickup truck) from the entire telephone survey database, based on 12,659 crashes. The mean weight ± standard deviation by vehicle type were as follows (in kg): passenger cars, 1301 ± 202; minivan, 1689 ± 112; sport utility vehicle, 1849 ± 314; and pickup truck, 1831 ± 313. (Note that we know the other vehicle weight from the crash-investigations, but we wish to construct a self-report measure using only data that could reasonably be obtained through self-report.) 2.2. Data analysis The “self-report” delta-V is estimated from the mass and reported velocities (speed and direction) of the vehicles at impact. For the single-vehicle crashes, it is simply the reported speed at time of impact. For two-vehicle crashes, the delta-V is determined from the reported speed of the other vehicle, the mass of the other vehicle (estimated as the mean of vehicle type), the direction of impact (front, side, and rear), and the law of the conservation of momentum, where it is assumed that, at least for the time necessary for the rapid deceleration to take place, the two vehicles are combined
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
835
Fig. 1. Vector computation of instantaneous change in velocity (delta-V) in multiple-vehicle crashes: m1 = mass of subject’s vehicle; m2 = mass of other vehicle; v1 = reported velocity of subject’s vehicle; v2 = reported velocity of other vehicle. Mass of subject’s vehicle taken from VINDICATOR; mass of other vehicle estimated as mean of vehicle type; velocities based on self-report of driver.
into a single body after impact (see Fig. 1). Angular impacts are assumed to take place at a 45◦ angle. The self-report delta-V is then the length of the vector difference between the initial and final velocity of the target (State Farm) vehicle. For each of these four crash types—single-vehicle and multiple-vehicle: front, side, and rear first impact sites—the “self-report” delta-Vs are then plotted against the delta-Vs calculated based on the crash-investigation, and a linear regression slope is computed using SAS Version 8 (SAS Institute, 2000). That is, we model the relationship between the crash-investigation delta-Vs yi and the self-report delta-Vs xi as yi = α + βxi + εi
(1)
where α and β are unknown constants and εi are unknown, normally-distributed error terms. (The residuals obtained after fitting the least squares regression of the self-report versus crash-investigation delta-Vs are approximately normally distributed.) We use linear regression to estimate the intercept α and slope β: if the self-report and crash-investigation delta-Vs correspond on average, the intercept and slope of this regression line will clearly be 0 and 1, respectively; we test this null hypothesis. We also report the proportion of
variability in the crash-investigation delta-Vs explained by the self-report estimates using (1) (R2 measure). We also consider whether the delta-Vs derived from the self-reports have the same rank–order as the crash-investigation-derived delta-Vs by considering the Spearman’s rho statistics, a measure of rank–order correlation (Sprent, 1990). We then conduct a parallel set of analyses using two simpler measures of crash severity, self-reported speed of the respondent’s vehicle and speed limit of the road on which the crash took place, in place of the self-report delta-V estimates. We compare the results to those obtained using self-report delta-V estimates.
3. Results Table 1 shows the distribution of reported speeds of subject vehicle and other vehicle by point of first contact, together with the posted speed limit and crash-investigation measures of delta-V. The crash-investigation delta-Vs and posted speed limits do not differ significantly by point of first contact (P = 0.79 and 0.92, respectively), whereas the subject vehicle speed is reported to be much lower and the other vehicle speed much higher in side and
836
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
Table 1 Crash-investigation delta-V, subject-reported speed of subject vehicle and other vehicle to strike subject vehicle, and posted speed limit, by point of first contact Point of first contact
Sample size
Crash-investigation delta-V (km/h)
Subject vehicle reported speed (km/h)
Other vehicle reported speed (km/h)
Posted speed limit (km/h)
Mean (S.D.)
Range
Mean (S.D.)
Range
Mean (S.D.)
Range
Mean (S.D.)
Range
Single-vehicle
20
31.3 (13.4)
14–58
59.7 (24.2)
19–104
N/A (N/A)
N/A
69.6 (19.8)
40–89
Two-vehicle Front Side Rear
53 29 16
30.3 (14.1) 27.6 (11.1) 31.1 (16.3)
10–79 12–60 11–69
54.9 (27.1) 33.2 (28.9) 3.8 (6.8)
0–96 5–104 0–24
53.6 (31.6) 76.8 (23.7) 72.5 (17.7)
0–104 8–141 45–104
72.1 (16.4) 71.8 (19.2) 69.1 (17.2)
32–89 40–105 40–89
S.D.: standard deviation.
rear impacts than in front or single-vehicle impacts (both P < 0.001). 3.1. Single-vehicle crashes Comparing the “self-report” delta-Vs with the delta-Vs obtained from the crash-investigations for n = 20 single-vehicle crashes, we find that the least squares regression intercept is 19.6 and the slope is 0.19; 11% of the crash-investigation delta-V variability is explained by the self-report delta-V estimates (see Fig. 2). The crash-investigation delta-Vs tended to be lower than the self-reported speeds. For intrusion crashes (n = 11), the association appeared stronger (intercept = 8.7, slope = 0.27) than for non-intrusion crashes (n = 9; intercept = 44.4, slope = −0.10), although this difference was not
statistically significant (P = 0.14). The rank–order association was relatively weak (Spearman’s rho = 0.20, P = 0.07). The results using reported speed-only are identical since this was the measure used to compute delta-V in single-vehicle crashes. The relationship between crash-investigation delta-V and posted speed limit is actually negative (intercept = 46.9, slope = −0.24; Spearman’s rho = −0.45) (see Table 2). 3.2. Two-vehicle frontal impacts Comparing the “self-report” delta-Vs with the deltaVs obtained from the crash-investigations for two-vehicle crashes where the point of first contact is in the front (n = 53), we find that the least squares regression intercept is 15.9 and the slope is 0.27; 14% of the crash-investigation
Fig. 2. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for single-vehicle crashes. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
837
Table 2 Least squares fit of self-report estimate of delta-V, self-report speed of respondent and posted speed limit vs. crash-investigation estimate of delta-V Point of first contact
Estimated intercept α (S.E.)
Estimated slope β (S.E.)
P-value under Ho (α = 0, β = 1)
R2
Spearman’s rho (P-value)
0.19 (0.13) 0.19 (0.13) −0.24 (0.15)
<0.001 <0.001 <0.001
0.113 0.113 0.148
0.201 (.075) 0.201 (0.075) −0.451 (0.069)
Single-vehicle Delta-V from self-report Speed-only estimate from self-report Posted speed limit
19.6 (8.1) 19.6 (8.1) 46.9 (10.8)
Two-vehicle front Delta-V from self-report Speed-only estimate from self-report Posted speed limit
15.9 (5.4) 27.1 (5.4) 16.0 (8.8)
0.27 (0.09) 0.06 (0.07) 0.20 (0.12)
<0.001 <0.001 <0.001
0.135 0.012 0.053
0.261 (0.059) 0.165 (0.24) 0.180 (0.20)
Two-vehicle side Delta-V from self-report Speed-only estimate from self-report Posted speed limit
14.2 (9.3) 25.8 (3.2) 15.6 (8.2)
0.30 (0.20) 0.06 (0.07) 0.16 (0.11)
<0.001 <0.001 <0.001
0.076 0.022 0.077
0.366 (0.051) 0.082 (0.67) 0.170 (0.40)
Two-vehicle rear Delta-V from self-report Speed-only estimate from self-report Posted speed limit
3.5 (13.6) 36.8 (3.1) −23.0 (9.7)
0.73 (0.35) −1.5 (0.50) 0.80 (0.14)
0.17 <0.001 <0.001
0.241 0.391 0.726
0.624 (0.007) −0.800 (0.002) 0.883 (<0.001)
R2 = proportion of variation in crash-investigation delta-Vs explained by self-report-estimated delta-Vs, self-reported speed and posted speed limit.
delta-V variability is explained by the self-report estimates (see Fig. 3). As with single-vehicle crashes, frontal crashes with intrusion (n = 25) had better self-report and crashinvestigation correspondence (intercept = 18.9, slope = 0.28) than frontal crashes without intrusion (n = 28; intercept = 22.8, slope = 0.06), although this difference was not significant (P = 0.28). The rank–order association was relatively weak (Spearman’s rho = 0.26, P = 0.06).
The association between reported speed-only and crashinvestigation delta-V was far weaker (intercept = 27.1, slope = 0.06; Spearman’s rho = 0.17). The association between posted speed limit and crash-investigation delta-V was intermediate between the self-report delta-V and crashinvestigation delta-V association, and reported speed-only and crash-investigation delta-V association (intercept = 16.0, slope = 0.20; Spearman’s rho = 0.18) (see Table 2).
Fig. 3. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for two-vehicle crashes with frontal point of first contact. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
838
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
Fig. 4. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for two-vehicle crashes with side point of first contact. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
Fig. 5. Delta-V estimated based on self-report vs. delta-V estimated from crash-investigation for two-vehicle crashes with rear point of first contact. Solid line shows perfect correspondence between self-report and crash-investigation estimated (intercept = 0, slope = 1); dotted line gives least squares fit.
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
3.3. Two-vehicle side impacts Comparing the “self-report” delta-Vs with the deltaVs obtained from the crash-investigations for two-vehicle crashes where the point of first contact is in the side (n = 29), we find that the least squares regression intercept is 14.2 and the slope is 0.30; 8% of the crashinvestigation delta-V variability is explained by the selfreport estimates (see Fig. 4). The rank–order association was moderate (Spearman’s rho = 0.37, P = 0.05). The association between crash-investigation delta-V and reported speed of subject vehicle was again far weaker (intercept = 25.8, slope = 0.06; Spearman’s rho = 0.08). The difference between intrusion and non-intrusion crashes is not considered because of insufficient sample size for non-intrusion crashes (n = 4). As with frontal impacts, the posted speed limit–crash-investigation deltaV association was intermediate between the self-report delta-V–crash-investigation delta-V association and reported speed-only–crash-investigation delta-V association (intercept = 15.6, slope = 0.16; Spearman’s rho = 0.17) (see Table 2). 3.4. Two-vehicle rear impacts Comparing the “self-report” delta-Vs with the delta-Vs obtained from the crash-investigations for n = 16 twovehicle crashes where the point of first contact is in the rear, we find that rear impacts have the greatest correspondence between the self-report-estimated and crash-investigationestimated delta-Vs: the least squares regression intercept is 3.5 and the slope is 0.73; 24% of the crash-investigation delta-V variability is explained by the self-report estimates (see Fig. 5). The null hypothesis of an intercept of 0 and a slope of 1 cannot be rejected (P = 0.17). There does not appear to be a difference in association between the (n = 6) intrusion and (n = 10) non-intrusion crashes, but the sample sizes are small. The rank–order association was moderately strong (Spearman’s rho = 0.63, P = 0.007). The association between crash-investigation delta-V and reported speed of subject vehicle was strong (R2 = 0.39), but in the wrong direction (intercept = 36.8, slope = −1.5; Spearman’s rho = −0.80): many of the worst crashes occurred in vehicles that were not moving. Posted speed limit, however, was highly correlated with crash-investigation delta-V, although it tended to overestimate the crash-investigation delta-V by approximately 20 km/h (intercept = −23.0, slope = 0.80; Spearman’s rho = 0.88) (see Table 2). 4. Discussion This report demonstrates that a novel measure of crash severity (“self-report” delta-V) presents an improved crash severity measure compared to estimated traveling speed and
839
speed limit for use in a large population study of crashes. This “delta-V” measure uses reported speed(s), direction of impact, and mass(es) of the vehicle(s) together with the principles of the conservation of momentum to produce a measure that more closely agrees with delta-V measures determined through crash-investigation. This self-reported measure is far from perfect, however, and errors in the self-reporting measures tend to bias the relationship between the self-report and crash-investigation delta-Vs toward the null (Weisberg, 1985). In the case of single-vehicle crashes, reported speed at the time of crashes was considerably higher than the crash-investigation-estimated delta-V. This is to be expected, as self-reported traveling speed does not take in account any pre-impact braking that would reduce the instantaneous change in velocity at the point of impact. Vehicles also might have struck elastic objects, although in this sample this does not appear to have been the case (of the 18 crashes in which the struck object was recorded, 11 (61%) involved trees or telephone poles and 2 (11%) walls or rocks). Alternatively, crash-investigations may underestimate delta-V in some non-intrusion crashes. Inaccuracy of the delta-V calculation in low severity crashes has been observed by other researchers (Noga and Smith, 1982). Intrusion crashes appeared to be better estimated than non-intrusion crashes, although this difference was not statistically significant. In two-vehicle crashes, rear impacts had the closest correspondence between the delta-Vs estimated from self-report and those estimated by crash-investigation. This is perhaps not surprising since rear impacts to subject vehicles generally occur when the subject is stopped (11 out of 16 crashes occurred with the subject vehicle being motionless), thereby removing the inherent variability associated with reported traveling speed. Frontal impacts also had crash-investigation delta-Vs lower than that estimated from self-reports. Again, this is most likely due to the effect of pre-impact braking and its influence on instantaneous change in velocity. As with single-vehicle crashes, intrusion crashes appeared to be better estimated than non-intrusion crashes, although this difference was not statistically significant. Self-reported delta-V estimates in side impact crashes also tended to be higher than those obtained from crash-investigations; in addition, the percentage of variability explained was the smallest among the crash types. This may reflect the fact that the estimation of crash severity in side impacts by a calculation of delta-V has many limitations, since the WinSmash program was developed using data from frontal crash tests only (Frampton et al., 1998). Perhaps the least one could ask of the self-report data is that it ranks the crash severity in the same order as the crash-investigations. For rear impact two-vehicle crashes, the agreement in ordering using the self-report estimates of the delta-Vs was moderately high (Spearman’s rho = 0.62). However, we found more modest associations for
840
M.R. Elliott et al. / Accident Analysis and Prevention 35 (2003) 833–840
single-vehicle, frontal, and side impact crashes (Spearman’s rho = 0.20, 0.26, and 0.37, respectively). Using reported speed alone was generally a far poorer proxy for crash-investigation delta-V than the “self-report” delta-V measure that attempted to include information about the momentum of both vehicles. A strong negative association was found between reported speed and delta-V in rear impacts, where motionless vehicles were often struck from behind at reported high speeds. Finally, posted speed limit had a generally stronger association with crash-investigation delta-V than reported speed, but was not as good a proxy as the “self-report” delta-V; an exception again being rear crashes, where it performed nearly as well as the self-report delta-V.
5. Conclusions We investigated the accuracy of three self-report measures of crash severity, estimated traveling speed, speed limit, and a novel measure of “self-report” delta-V, by comparing these measures to delta-V calculated from crash-investigation. The “self-report” delta-V utilizes all available information about estimated speeds, direction of impact, vehicle mass, and the principle of conservation of momentum. We find that, in general, this self-report delta-V estimate, although not a perfect proxy, provides a better prediction of crash severity than either estimated traveling speed or posted speed limit. This was particularly the case in rear impact collisions, where a reasonable accurate estimate of delta-V may be obtained.
4.1. Study limitations Acknowledgements Self-report data from all drivers in a multi-car impact, and detailed information on the mass of struck objects in single-vehicle accidents, rather than estimated speeds and vehicle types from the driver of the subject vehicle, would undoubtedly improve the correspondence between the self-report delta-V estimates and those obtained through detailed crash-investigations. Although the self-report data are obtained from a known-probability sample, the crash-investigations are not sampled under a known-probability design. In addition, the population from which both the surveillance survey and crash-investigations have been sampled is limited to crashes involving children in State Farm-insured vehicles. Hence, it is somewhat uncertain how generalizable these results are to all drivers. However, there is no reason to believe that the relationship between the “self-report” delta-V and the crash-investigation delta-V would be associated with the probability of selection or the State Farm-insured population itself. Given that complete data to estimate the delta-V from the surveillance survey are required, there is surprisingly little missing data: self-report and crash-investigation delta-Vs are available for 20 of 21 single-vehicle crashes, 53 of 63 front impact two-vehicle crashes, 29 of 31 side impact crashes, and 16 of 17 rear impact crashes. Hence, missing data is unlikely to introduce much bias in these results.
The authors would like to thank State Farm Insurance Companies for their financial support of this work through the Partners for Child Passenger Safety Project. References Durbin, D., Bhatia, E., Holmes, J.H., Shaw, K.N., Werner, J.V., Sorenson, W., Winston, F.K., 2001. Partners for Child Passenger Safety: a unique child-specific crash surveillance system. Accid. Anal. Prev. 33, 407– 412. Frampton, R., Brown, R., Thomas, P., Fay, R., 1998. The importance of non-struck side occupants in side collision. In: Proceedings of the 42nd Annual Conference of the Association for the Advancement of Automotive Medicine, Charlottesville, VA. Johnston, C., Rivara, F.P., Soderberg, R., 1994. Children in car crashes: analysis of data for injury and use of restraints. Pediatrics 93, 960–965. Noga, J.T., Smith, R.A., 1982. Accuracy and sensitivity of CRASH. In: Proceedings of the 26th Stapp Car Crash Conference, Ann Arbor, MI. SAS Institute Inc., 2000. Statistical Analysis System Version 8 for Windows. SAS Institute Inc., Cary, NC. Shibata, A., Fukuda, K., 1994. Risk factors of fatality in motor vehicle traffic accidents. Accid. Anal. Prev. 26, 391–397. Sprent, P., 1990. Applied Non-Parametric Statistical Methods. Chapman & Hall, London, pp. 172–175. Weisberg, S., 1985. Applied Linear Regression, 2nd ed. Wiley, New York, pp. 76–79. Winston, F., Durbin, D.R., Kallan, M.J., Moll, E.K., 2000. The danger of premature graduation to seat belts for young children. Pediatrics 105, 1179–1183.