This article was downloaded by: [University of Hong Kong Libraries]On: 12 November 2014, At: 04:06Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,37-41 Mortimer Street, London W1T 3JH, UK

Click for updates

Traffic Injury PreventionPublication details, including instructions for authors and subscription information:http://www.tandfonline.com/loi/gcpi20

Evaluation of NASS-CDS Side Crash Delta-V EstimatesUsing Event Data RecordersNicholas S. Johnsona & Hampton C. Gablera

a Virginia Tech Center for Injury Biomechanics, Blacksburg, VirginiaAccepted author version posted online: 16 Jan 2014.Published online: 15 Jul 2014.

To cite this article: Nicholas S. Johnson & Hampton C. Gabler (2014) Evaluation of NASS-CDS Side Crash Delta-V EstimatesUsing Event Data Recorders, Traffic Injury Prevention, 15:8, 827-834, DOI: 10.1080/15389588.2014.881995

To link to this article: http://dx.doi.org/10.1080/15389588.2014.881995

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) containedin the publications on our platform. However, Taylor & Francis, our agents, and our licensors make norepresentations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of theContent. Any opinions and views expressed in this publication are the opinions and views of the authors, andare not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon andshould be independently verified with primary sources of information. Taylor and Francis shall not be liable forany losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoeveror howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use ofthe Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Traffic Injury Prevention (2014) 15, 827–834Copyright C© Taylor & Francis Group, LLCISSN: 1538-9588 print / 1538-957X onlineDOI: 10.1080/15389588.2014.881995

Evaluation of NASS-CDS Side Crash Delta-V EstimatesUsing Event Data Recorders

NICHOLAS S. JOHNSON and HAMPTON C. GABLER

Virginia Tech Center for Injury Biomechanics, Blacksburg, Virginia

Received 9 October 2013, Accepted 7 January 2014

Objective: Planar crash severity is most commonly defined by delta-V (�V ), which is the change in a vehicle’s velocity vector during acrash. All �V estimates contained in the NASS-CDS are generated using a damage-based program called WinSMASH. WinSMASH�V accuracy in side crashes has not previously been validated against real-world crash data. This study will investigate the accuracyof WinSMASH �V estimates in real-world side crashes.

Methods: This study uses biaxial �V data from event data recorders (EDRs) to assess the accuracy of side crash �V estimatesin NASS-CDS crash cases. Single-event side crashes were identified in the NASS-CDS for which (a) WinSMASH �V had beencoded and (b) biaxial EDR data were available. For these crashes, the WinSMASH-estimated resultant �V was compared with theEDR-recorded resultant �V to assess the accuracy of the former. EDR �V values were adjusted for an assumed average restitutionvalue of 10 percent, based on values reported in the literature. Principal direction of force (PDOF) is the orientation of the net crashimpulse relative to the vehicle and is a key parameter in WinSMASH reconstructions. NASS-CDS PDOF estimates were comparedto the PDOF computed from the EDR data to assess their accuracy as well.

Results: WinSMASH systematically overestimates EDR �V by about 12.9 percent for cars struck by cars and by about 2.4 percentfor cars struck by light trucks and vans (LTVs). �V error was significantly different (at 95% confidence) for crashes to different areasof the vehicle side. The mean discrepancy in NASS-CDS PDOF was −0.9◦ with a standard deviation of 12.6◦.

Discussion: WinSMASH systematically overestimates �V at common velocity by about 13 percent for cars struck by cars andabout 2 percent for cars struck by LTVs. Since WinSMASH’s assumption of zero restitution is accounted for in this analysis, thissuggests that WinSMASH stiffness parameters represent LTV impacts better than car impacts in side crashes. �V error differs byimpacted area, suggesting that side impact stiffness parameters used in damage-based �V reconstruction must closely represent thecrash mode being reconstructed for best results. NASS-CDS estimates of PDOF do not appear to exhibit any systematic discrepancy.The amount of random PDOF discrepancy observed here is consistent with prior studies.

Keywords: event data recorder, NASS-CDS, delta-V , crash reconstruction

Introduction

Planar crash severity is most commonly defined by delta-V(�V ), which is the change in a vehicle’s velocity vector duringa crash. �V is the single best and most widely used correlatefor occupant injury in automobile crashes (Bahouth et al.2004; Gabauer and Gabler 2008). �V is frequently estimatedby correlating postcrash vehicle damage with the amount ofenergy absorbed in the crash and then using conservation ofmomentum and energy to compute velocity change directlywithout any knowledge of the initial or final vehicle speed(Campbell 1974; NHTSA 1981; Prasad 1987).

Managing Editor David Viano oversaw the review of this article.Address correspondence to Hampton C. Gabler, VirginiaTech–Wake Forest University Center for Injury Biomechanics,325 Stanger Street, Kelly Hall 445, Blacksburg, VA 24061. E-mail: gabler@vt.eduColor versions of one or more of the figures in the article can befound online at www.tandfonline.com/gcpi.

All �V estimates contained in the NASS-CDS are gen-erated in this way using a CRASH3-derived program calledWinSMASH (NHTSA 1981; Sharma et al. 2007). The ac-curacy of WinSMASH has significant implications for ve-hicle safety policy and research, both in the United Statesand internationally, because NASS-CDS data is heavily usedin many vehicle safety–related fields. WinSMASH �V es-timate accuracy in real-world crashes has previously beenexamined through comparison with �V from event datarecorders (EDRs) of crashed vehicles (Hampton and Gabler2010; Niehoff and Gabler 2006). Historically, most EDRs inthe U.S. vehicle fleet have recorded only longitudinal �V , be-cause most U.S. vehicles have had only frontal air bags. Con-sequently, WinSMASH has been validated against real-worlddata for frontal crash reconstructions only. Prior validation forside-impact crashes has used only staged crash tests (Johnson2011; Johnson et al. 2009), which are not real-world crashes.However, some EDRs from vehicles equipped with side airbags record biaxial (i.e., both longitudinal and lateral) �V . Asside air bags have become more common in the U.S. fleet, it

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

828 Johnson and Gabler

has become feasible to assess damage-based �V accuracy—inparticular, that of WinSMASH—against real-world datafor side crashes as has been done previously for frontalcrashes.

Objective

This study will use biaxial �V data from newer EDRs to assessWinSMASH �V accuracy for real-world side crashes. WinS-MASH is the source of all �V estimates in the NASS-CDSdatabase, so the accuracy of the program has direct implica-tions for U.S. vehicle safety policy and regulations.

Methods

The NASS-CDS reconstructs crashed vehicle �V using WinS-MASH whenever possible. Conveniently, it also obtains EDRdata for these vehicles—with the consent of the vehicleowner—whenever possible. This analysis examined single-event side-impact crashes from the NASS-CDS for which (a)WinSMASH �V had been coded and (b) biaxial EDR datawere available. For these crashes, the WinSMASH-estimatedresultant �V was compared with the EDR-recorded resul-tant �V to assess the accuracy of WinSMASH. Althoughcrashes can involve multiple events, only single-event crasheswere used because this eliminates any ambiguity as to whichcoded impact the EDR data corresponds to. Additionally,only EDRs from crashes where the air bags actually deployedwere used. Air bag deployment prevents the crash record frombeing overwritten with data from subsequent events that maynot be the event of interest.

This analysis also used biaxial EDR data to assess the ac-curacy of NASS-CDS estimates of principal direction of force.Principal direction of force, or PDOF, is the orientation of thenet crash impulse relative to the vehicle and is a key parameterin WinSMASH reconstructions. In the NASS-CDS, PDOF isvisually estimated by crash investigators using evidence suchas the vehicle deformation pattern and collision trajectories.Having both longitudinal and lateral �V from an EDR allowsfor computation of the actual PDOF in a crash, to which theNASS estimates can be compared.

Extraction of Resultant �V, PDOF from EDRs

Biaxial EDRs sense both longitudinal and lateral accelerationsduring a crash and process them to obtain longitudinal andlateral �V histories. Acceleration from each axis is measuredand processed separately to obtain the �V history for thataxis. In this analysis, the maximum resultant EDR �V wasexamined. To obtain this, resultant �V was first calculatedat each time increment in the record using the longitudinaland lateral �V . The maximum of these resultant �Vs wasused as the resultant EDR �V for that crash. PDOF was thencalculated as the arctangent of the longitudinal and lateralcomponents of this �V .

Many older EDRs record only about 70–80 ms of �V dataafter the recording algorithm is triggered, which is shorterthan most crash pulses. Because of this, older EDRs some-times underestimate actual �V (Niehoff and Gabler 2006).However, newer biaxial EDRs typically record 200–300 ms ofcrash data, with some portion of that being pre-algorithm-trigger. Newer EDRs thus typically record the entire crashpulse. In this analysis, any records with resultant accelera-tion greater than 1 g between the last and second-to-last timesteps were manually inspected for completeness, regardless ofrecord duration. Most EDRs do not record acceleration butonly measure it to compute �V . Consequently, accelerationwas calculated from the change in resultant �V between thefinal and second-to-last data points.

Most newer (i.e., biaxial) EDRs are capable of recordingand storing as many as 2 separate events. Although everyselected crash had only a single event coded in NASS-CDS,this did not preclude the EDR from recording multiple impactpulses as a result of that event. EDRs were therefore checkedto ensure that they contained only a single locked record.In the minority of cases where the EDR contained multiplelocked records, the correct one was identified by examiningeach recorded pulse individually. If the correct record couldnot be determined, the case was discarded.

Comparison of WinSMASH �V and PDOF to ValuesObtained from EDRs

For each case examined, the analysis compared (1) theWinSMASH-estimated resultant �V to the resultant EDR�V , and (2) the investigator’s estimate of PDOF to the EDR-derived PDOF. The analysis also examined WinSMASH �Vand PDOF estimate accuracy with respect to a number ofother parameters, such as vehicle body style and area of thevehicle struck, to determine whether there was any correlation.WinSMASH was never intended to be a forensic reconstruc-tion tool but rather a standard benchmark that is accurateon average. Precision of the �V in individual reconstructionsis therefore of less importance than the accuracy over manycases. For �V , systematic error was computed as the slope ofa linear regression of WinSMASH �V versus actual �V withthe intercept fixed at zero. On a cross-plot of estimated �Vversus actual �V , the amount by which the slope of such a re-gression deviates from unity gives an indication of the amountof systematic error in the estimated values. For example, aslope of 1.050 would indicate a 5 percent overestimation. Pre-cision of the �V estimates, or the random error, was computedas the root mean square (RMS) error about the regression line.

WinSMASH estimates �V assuming that there is no resti-tution in crashes, even though most real crashes do involvesome degree of restitution (Brach and Brach 2005). In crasheswith restitution, the WinSMASH �V estimate correspondsto the �V only up to the point when restitution begins. IfWinSMASH is reconstructing crashes completely accurately,then for any crash with nonzero restitution, the WinSMASH�V estimate should always be less than the EDR-derived �V(which includes restitution); that is, the WinSMASH assump-tion of zero restitution is a known source of systematic �V

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

Side Crash Delta-V Estimates 829

underestimation. This effect could potentially obscure system-atic errors from other sources, so in order to account for it, allEDR �Vs used in the analysis were reduced by a percentagetypical of restitution in side-impact crashes. These reduced�V values, referred to henceforth as “adjusted” EDR �Vs,represent the actual crash �V minus the average contributionfrom restitution.

Ishikawa (1994) computed restitution coefficients for 32staged vehicle-to-vehicle side crash tests of varying configura-tions. In 16 of these tests, the computed restitution coefficientswere significantly negative, indicating that a common velocitywas not reached at the contact point between the 2 vehicles(WinSMASH is not applicable to such crashes). The aver-age restitution coefficient in the remaining 16 crash tests was0.086, or about 9 percent. Johnson and Gabler (2012) foundthat restitution increased �V in NHTSA regulatory side crashtests by about 8 percent on average. The effect of restitutionincreases with decreasing impact severity (Brach and Brach2005) and staged crash tests (particularly NHTSA regulatorytests) are generally more severe than most real-world crashes.Hence, in this analysis, an adjustment of −10 percent was ap-plied to all EDR �Vs to account for the average effect ofrestitution.

Statistical testing was performed using SAS v9.3 (SAS,Cary, NC). The NASS-CDS uses a stratified, clustered, andweighted sample design allowing for nationally representativeestimates to be made from the data. However, no attemptwas made in this study to perform statistical testing using theNASS weights, because the sample size was insufficient. There-fore, this analysis is only representative of the population ofcrashes recorded in the NASS-CDS and is not representativeof all U.S. crashes nationally.

Sample Summary

Table 1 gives a summary of the final data set used in this anal-ysis. The data set consists of 41 single-impact crashes coded inthe NASS-CDS (case years 2008–2011) for which both WinS-MASH �V and biaxial EDR data (locked, air bag deploymentevents) were available. Only crashes where cars were struck inthe side by another vehicle were retained for this analysis.Struck light trucks and vans (LTV) crashes (5 cases: 2010-12-154, 2010-49-65, 2011-09-098, 2011-49-043, and 2011-81-033)and crashes involving fixed objects (1 case: 2006-50-46) werenot available in sufficient numbers for a meaningful analysisof those crash configurations and were excluded. Side impactswere identified by the general area of damage (GAD) coded in

Table 1. Data set composition

Total crashes: 41Struck vehicle make:

Ford 1General Motors 40

Struck vehicle body style:Cars 41

Striking vehicle body style:Cars 19LTVs 22

Fig. 1. WinSMASH �V vs. adjusted EDR �V . Regression equa-tion: y = 1.0722x. RMSerr about regression line; 5.366 km/h.

NASS. GAD is defined as part of the Collision DeformationClassification (CDC) code (SAE 1980); all cases in the samplehave a GAD of “L” (left) or “R” (right).

One case—NASS case 2007-82-045—was excluded fromthe final sample for having an incomplete �V record. TheEDR in this case came from a 2001 Ford Taurus and recordedonly a 78 ms window of �V . Additionally, this EDR wasunusual in that it provided a record of acceleration in additionto a record of �V . The recorded acceleration clearly indicatedthat the crash was not complete by the end of the recording.All other examined cases reported either 220 or 250 ms of �Vdata with some amount of that being pre-trigger �V (exceptin case 2010-75-159), so pulse truncation is unlikely in theretained sample.

Results

WinSMASH �V Accuracy

Figure 1 compares WinSMASH-estimated �V to EDR-measured �V for the data set. WinSMASH appears to sys-tematically overpredict adjusted EDR �V in the examinedside crashes by about 7.2 percent (95% confidence interval[CI], −0.6% to +15.0%). Random error (RMS error about theregression line) is about 5.4 km/h. There was insufficient ev-idence to reject the assumption of normally distributed data.A paired t test indicated that the �V overprediction was sig-nificantly different from zero (P = .0212) in this sample.

Effect of Striking Vehicle Body Type

Figures 2 and 3 compare WinSMASH �V to EDR �V forcars struck by cars and cars struck by LTVs, respectively. Allvehicles in the sample fell into one of these 2 categories. Forcars struck by other cars, WinSMASH �V appears to be

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

830 Johnson and Gabler

Fig. 2. WinSMASH �V vs. adjusted EDR �V for cars struck bycars. Regression equation: y = 1.1291x. RMSerr; 4.260 km/h.

systematically higher than adjusted EDR �V by about 12.9percent on average (95% CI, +3.4% to +22.4%). For carsstruck by LTVs, WinSMASH appears to systematically over-predict �V by about 2.4 percent (95% CI, −10.0% to +14.8%).Observed random error was higher for striking LTVs than forstriking cars (6.06 km/h vs. 4.26 km/h RMS error). The as-sumption of normally distributed data was rejected for strikingLTVs but not for striking cars. A Kruskal-Wallis test did notfind the difference in WinSMASH �V error between strikingbody types to be statistically significant (P = .1824).

Fig. 3. WinSMASH �V vs. adjusted EDR �V for cars struck byLTVs. Regression equation: y = 1.0240x. RMSerr; 6.065 km/h.

Fig. 4. SHL codes defined by SAE standard J224 (SAE 1980) andused in NASS-CDS.

Effect of Impacted Vehicle Region

Figure 4 shows the specific horizontal location (SHL) codesdefined by the Collision Deformation Classification (CDC)standard (Society of Automotive Engineers [SAE] 1980) andused by NASS-CDS to describe the particular horizontal areaof the vehicle which was damaged in an impact. Figure 5 showsWinSMASH �V error broken down by SHL. WinSMASHappears to underestimate �V slightly for crashes involvingdamage to F (front), while D (front, passenger compartmentand back) and possibly P (passenger compartment) appear tobe overestimated. Note that the sample contained no B (back)impacts. One-way analysis of variance provides marginallysignificant evidence (P = .0557) that the mean errors for eachSHL are not universally equal. Tukey’s Honestly SignificantDifference procedure found the difference between D and Fcrashes—and only that difference—to be significant.

Effects of WinSMASH Calculation Type

WinSMASH can perform different types of reconstructioncalculations depending upon what damage information isavailable for a case. The type of reconstruction performedfor a case is coded by the DVBASIS variable in NASS-CDS.“Standard” (n = 31), “Missing Vehicle” (n = 9) and “CDC-Only” (n = 1) reconstructions (see Sharma et al. 2007) arerepresented in the cases studied here. To test for differences in�V error between different calculation types, all 31 Standardreconstructions in the sample were re-reconstructed in WinS-MASH, once treating the EDR-bearing vehicle as a MissingVehicle and again treating it as a CDC-Only vehicle. The 10cases which were originally reconstructed as Missing Vehicleand CDC-Only were not included, because it would not havebeen possible to run them as Standard reconstructions.

For this purpose, the WinSMASH inputs used by the orig-inal NASS case investigators were retrieved from the internal

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

Side Crash Delta-V Estimates 831

Fig. 5. WinSMASH �V error vs. SHL of impact. Black barsindicate the mean �V error for each SHL value.

NASS-CDS Oracle database. Before using these inputs in newreconstructions, their integrity was verified by reconstructingthe original �V values coded in NASS-CDS; this was success-ful in all cases.

Figure 6 compares WinSMASH �V error between Stan-dard, Missing Vehicle, and CDC-Only reconstructions of the31 cases that were originally run as Standard reconstructions.Missing Vehicle reconstructions appear to overestimate �Vmore than Standard and CDC-Only reconstructions, althoughrepeated measures analysis of variance (PROC GLM using theREPEATED option) did not find calculation type to make astatistically significant difference in �V error (P = .2766).On cross-plots analogous to Figure 1, Standard reconstruc-tions would show 11.3 percent overestimation (95% CI, +3.2%to +19.4%) and 5.011 km/h RMS error; CDC-Only wouldshow 11.5 percent overestimation (95% CI, +1.5% to +21.5%)and 6.180 km/h RMS error; Missing Vehicle reconstructionswould show 34.3 percent overestimation (95% CI, +21.8% to+46.9%) and 7.726 km/h RMS error, the largest systematicand random error of the 3.

Investigator PDOF Accuracy

Figure 7 shows discrepancy in NASS investigator estimates ofPDOF with respect to the PDOF computed from EDR data.Mean PDOF discrepancy was −0.9◦; this was not found to besignificantly different from zero (P = .9442, signed rank sumtest, normality rejected for the sample). Note that NASS onlycodes PDOF to the nearest 10◦. Smith and Noga (1982) gave95 percent confidence limits on the accuracy of field-recordedPDOF as ±20◦. Observed standard deviation in PDOF dis-crepancy here was 12.7◦, which equates to 95 percent confi-dence limits of ±24.9◦ assuming normally distributed error(1.96 ∗ standard deviation).

Fig. 6. WinSMASH �V error vs. WinSMASH calculation type.Black bars indicate the mean �V error for each calculation type.

Effect of PDOF Discrepancy on WinSMASH �V Accuracy

Figure 8 shows a cross-plot of WinSMASH �V error mag-nitude against PDOF discrepancy magnitude. There appearsto be no correlation between PDOF discrepancy and WinS-MASH �V error; the extremely small R2 value for the regres-sion (R2 = 0.0001) indicates that it predicts virtually none ofthe observed variance in the data.

Fig. 7. Discrepancy in NASS PDOF vs. EDR-derived PDOF.Mean NASS PDOF discrepancy: −0.907◦. Positive values areclockwise from the front of the vehicle when viewed from over-head; negative values are counterclockwise.

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

832 Johnson and Gabler

Fig. 8. WinSMASH �V error magnitude vs. PDOF discrepancymagnitude. Regression equation: y = −0.00470x + 4.1734, R2 =0.0001.

Discussion

Figures 2 and 3 show that WinSMASH overestimates EDR�V by about 13 percent for cars struck by cars and overes-timates EDR �V by only about 2 percent for cars struck byLTVs. Hence, WinSMASH �V estimates appear to be moreor less accurate on average for LTVs striking cars but appearabout 13 percent high for cars striking cars. There is a greatdeal of variation in accuracy between individual cases (ran-

Fig. 9. WinSMASH �V error vs. SHL of impact, using nonad-justed EDR �V (including restitution). The relative differencesin WinSMASH error between SHL values are qualitatively un-changed.

dom error), but this is to be expected given the constraints andintended purpose of WinSMASH.

Johnson and Gabler (2012) examined WinSMASH �Vaccuracy for NHTSA side crash tests. In that analysis, testinstrumentation made it possible to determine when restitu-tion began in individual crashes and to obtain actual vehicle�V at that point. That study found that for NHTSA sidecrash tests, WinSMASH overestimated prerestitution �V byabout 19 percent for struck cars, which is greater than the 12.9percent overestimation observed here for cars struck by cars.Note that the Johnson and Gabler (2012) study examined onlyone crash configuration (NHTSA side impact crash tests) andthat each crash they examined used an identical impactor (theNHTSA moving deformable barrier). This analysis examinedreal-world crashes with many configurations, so some differ-ences are to be expected. Unfortunately, EDRs do not recordsufficient data to determine prerestitution �V in individualcrashes, so EDR �Vs could not be individually adjusted forrestitution in a way analogous to Johnson and Gabler (2012).

WinSMASH, like all damage-based crash reconstructionprograms, uses vehicle stiffness parameters to estimate theenergy dissipated in crashes from measurements of residualvehicle damage. Different stiffness parameters are used fordamage to the front, side, and rear of individual vehicles inWinSMASH. All side crash stiffnesses used by WinSMASHare derived from damage in NHTSA side crash tests. Thesestiffnesses may not represent crashes with damage to areasdifferent from the tests from which they are derived, becausevehicle side structure is not homogenous. NHTSA side impacttests are Y (SHL) impacts verging on P, so the observed under-estimation of �V for F crashes (Figures 4 and 5) may be dueto the stiffer vehicle structures in this region, absorbing moreenergy for a given amount of crush than do the softer struc-tures of the P region (passenger compartment). The geometryof the striking vehicle itself also influences which areas of theimpacted vehicle are engaged. LTVs tend to be taller and widerthan cars and may or may not have greater ground clearance.Given collisions of otherwise identical configuration, strikingLTVs can engage different or additional structures comparedto striking cars. This could account for the differences betweenFigures 2 and 3 and suggests that WinSMASH stiffness pa-rameters may represent energy absorption due to LTV impactsbetter than energy absorption in car impacts.

WinSMASH calculation type (Figure 6) may make asubstantial difference in the systematic �V error. Standardreconstructions use measured damage for both vehicles to es-timate the absorbed energy. CDC-Only reconstructions esti-mate a crush profile for one vehicle from its CDC in lieu ofactual measurements. Missing Vehicle reconstructions use aforce–balance procedure developed by Prasad (1991) to esti-mate energy absorption by one vehicle with no damage in-formation for that vehicle whatsoever. Missing Vehicle cal-culations are observed to overestimate prerestitution �V byabout 34 percent, while the same cases reconstructed as Stan-dard or CDC-Only show only overestimate by 11–12 percent.Although statistical testing did not find a significant differ-ence in systematic �V error between calculation types, thismay only mean that the effect size was too small to be distin-guished from random variation. WinSMASH reconstruction

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

Side Crash Delta-V Estimates 833

accuracy has a great deal of random variation, and 31 cases (41− 10) is not that many. It is also notable that the random �Verror appears to increase when the calculation type uses lessinformation: Standard reconstructions have 5.0 km/h RMSerror; CDC-Only, which does not use crush, has 6.2 km/h; andMissing Vehicle, which does not use crush, CDC or PDOF,has 7.7 km/h. Standard reconstructions, where both vehicleshave a full damage profile and PDOF estimates, appear toprovide greater accuracy than other calculation types.

PDOF and �V

In our sample, NASS PDOF estimates for side crashes(Figure 7) showed only −0.9◦ of systematic discrepancy, whichis effectively zero. Especially considering that NASS onlycodes PDOF to the nearest 10◦, this seems to indicate thatthere is no systematic PDOF error. RMS (random) PDOF er-ror was observed to be 12.6◦; EDR-derived estimates of PDOFare known to have an RMS error of 4.4◦ and differ from valuesobtained from crash test instrumentation by no more than 10◦(Kusano et al. 2013).

Based on the observed standard deviation in PDOF dis-crepancy, 95 percent confidence limits for side crash NASSPDOF are ±24.9◦. This is similar to the 95 percent confidencelimits of ±20◦ reported by Smith and Noga (1982) for fieldmeasurements of PDOF. The way in which Smith and Nogaarrived at their estimate is somewhat different than the ap-proach used here, and their estimate was derived from investi-gation of vehicle-to-vehicle crash tests, not real-world crashesas here.

Figure 8 shows that there is no correlation between themagnitude of PDOF error and the magnitude of WinSMASH�V error. In the WinSMASH calculations, the relationshipbetween PDOF error and �V error depends on other param-eters describing the crash configuration, so this is perhaps un-surprising. This also indicates that the effects of PDOF errorare being washed out by other sources of error.

Limitations

Ideally, the actual amount of restitution would be determinedfor each case individually, but that was not feasible in thisanalysis. The data available for reconstruction of NASS-CDScrashes are almost never sufficient to determine a coefficientof restitution through planar momentum conservation, andEDR data are not sufficient to determine the point duringan impact at which restitution begins. Thus, this analysis as-sumed that restitution increases side crash �V by 10 percenton average. Because this analysis was primarily concerned withsystematic error, accounting for restitution in this fashion isnot unreasonable. However, though the findings for systematicover-/underestimation of �V should in principal account forWinSMASH’s assumption of zero restitution, assuming thesame value for restitution in all cases could inflate the RMSerror in individual reconstructions because restitution will bedifferent in each individual case. Restitution is affected by anumber of factors, including the precise combination of ve-hicle structures engaged in a crash (both on the striking and

struck vehicles) and the closing speed of the impact. RMS er-ror figures for the nonsystematic component of WinSMASH�V error should therefore be considered upper bounds.

This also has implications for the observed variation in sys-tematic WinSMASH error by SHL. Average restitution likelyvaries by the SHL contacted (Brach and Brach 2005), so −10percent may not be the correct adjustment to apply to EDR�V when disaggregated by SHL. The differences in system-atic WinSMASH �V error in Figure 5 may not be entirelyaccurate. However, the relative differences in error betweenSHLs appears to be a robust finding. Figure 9 shows thatwhen WinSMASH �V is compared to the nonadjusted EDR�V including any restitution, the relative differences in WinS-MASH error between SHL values is essentially unchanged,even though the absolute systematic error values within eachSHL are different.

EDR measurements do not account for the effects of rota-tion. This could potentially skew their �V measurements, butany such skew would not be systematic and it would probablynot be very large in comparison to the measured �Vs. EDRsare generally not mounted far from the vehicle center of grav-ity, which would tend to reduce the effects of rotation on �Vmeasurements. In addition, typical rotation rates in NHTSAside crash tests are a relatively slow 90◦/s (Johnson 2011). Thecrashes in this sample are much less severe than the typicalNHTSA side crash test, and most of them have SHL valuesthat suggest crash impulse moment arms that are not anylarger than those of side crash tests. It therefore seems likelythat rotation rates in these crashes would tend to be lower thanthe already low values observed in NHTSA side crash tests.Additionally, EDR-derived PDOF estimates, which are com-puted from biaxial �V measurements, have been observed tohave a root mean square error of only 4.4◦ in staged crash testsinvolving rotation (Kusano et al. 2013).

The findings of this study do not represent the nationalvehicle fleet but pertain only to the analysis sample, whichis mostly GM vehicles. The limited sample size precludes theuse of statistical techniques necessary to perform significancetests with clustered, stratified, and weighted NASS-CDS data.Small sample size is also a limitation in its own right. Here,it is a result of the limited number of NASS-CDS crashes forwhich EDR data are available, compounded by the additionalrequirements for biaxial EDR data and a single-event crash.The NASS-CDS has only collected EDR data since 2000 and,at first, NASS investigators could only read EDRs from Fordand General Motors. This is reflected in the near-total pro-portion of General Motors vehicles in the data set. Thoughthe lack of makers other than General Motors could be seenas a limitation, it is unlikely to have a significant effect. WinS-MASH vehicle stiffness parameters are derived from tests ofindividual vehicles, so any characteristics particular to a givenmanufacturer are already controlled for.

WinSMASH vehicle stiffness parameters are unique toWinSMASH; other damage-based crash reconstruction pro-grams use different stiffnesses. Thus, the exact �V overesti-mation seen for WinSMASH in this analysis may not gener-alize to other programs that use different stiffness parameters.Though this does limit the scope of the quantitative findingsgiven here, the more important conclusion is that, regardless

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014

834 Johnson and Gabler

of the reconstruction program, stiffness parameters appear tohave a strong effect on �V accuracy in side crashes. Side crashstiffness parameters should therefore be chosen carefully to re-flect the crash mode being reconstructed. In addition, WinS-MASH is perhaps the most influential damage-based recon-struction program currently in use. It is the sole source of �Vfor the NASS-CDS, and therefore WinSMASH accuracy hasdirect implications for road safety policy and vehicle testingprocedures, both in the United States and internationally.

Conclusions

In real-world side crashes, WinSMASH appears to overesti-mate struck vehicle �V for cars struck by cars by about 13percent. For cars struck by LTVs, the observed overestima-tion was only about 2 percent; WinSMASH likely does notactually overestimate these �Vs. However, this analysis didnot examine crashes where the struck vehicle was an LTV, nordid it examine fixed-object crashes of any type. It seems likelythat the side impact stiffness parameters used by WinSMASHrepresent energy absorption in LTV impacts better than incar impacts. Random error in WinSMASH �V appears to beminimized in reconstructions that use the maximum amountof information possible.

NASS field estimates of PDOF do not appear to exhibitany systematic discrepancy. The amount of random PDOFdiscrepancy observed here is consistent with the findings ofprior studies. PDOF discrepancy magnitude showed no corre-lation with WinSMASH �V magnitude, which indicates thatits effect is being washed out by other sources of discrepancy.

References

Bahouth GT, Digges KH, Bedewi NE, et al. Development of UR-GENCY 2.1 for the prediction of crash injury severity. Top EmergMed. 2004;26(2):157–165.

Brach RM, Brach RM. Vehicle Accident Analysis and ReconstructionMethods. Warrendale, PA: Society of Automotive Engineers; 2005.

Campbell K. Energy Basis for Collision Severity. Warrendale, PA: Societyof Automotive Engineers; 1974. SAE 740565.

Gabauer DJ, Gabler HC. Comparison of roadside crash injury metricsusing event data recorders. Accid Anal Prev. 2008;40:548–558.

Hampton CE, Gabler HC. Evaluation of the accuracy of NASS/CDSdelta-V estimates from the enhanced WinSMASH algorithm.In:Proceedings from the 54th Annual Meeting of the Association forthe Advancement of Automotive Medicine in Las Vegas, Nevada, Octo-ber 17–20, 2010. Barrington, IL: Association for the Advancement ofAutomotive Medicine; 2010:241–252.

Ishikawa H. Impact Center and Restitution Coefficients for Accident Re-construction. Warrendale, PA: Society of Automotive Engineers; 1994.SAE 940564.

Johnson N. Assessment of Crash Energy–Based Side Impact Reconstruc-tion Accuracy [thesis]. Blacksburg, VA: Virginia Tech; 2011.

Johnson N, Gabler HC. Accuracy of a damage-based reconstructionmethod in NHTSA side crash tests. Traffic Inj Prev. 2012;13:72–80.

Johnson N, Hampton CE, Gabler HC. Evaluation of the accuracyof side impact crash test reconstructions. Biomed Sci Instrum.2009;45:250–255.

Kusano KD, Kusano SM, Gabler HC. Automated crash notification al-gorithms: evaluation of in-vehicle principal direction of force (PDOF)estimations. Transp Res Part C Emerg Technol. 2013;32:116–128.

National Highway Traffic Safety Administration. CRASH3 (CalspanReconstruction of Accident Speeds on the Highway) User’s Guide andTechnical Manual. Washington, DC: Accident Investigation Division,N.C.S.A., National Highway Traffic Safety Administration; 1981. HS-805 732.

Niehoff P, Gabler HC. The accuracy of WinSMASH delta-V estimates:the influence of vehicle type, stiffness, and impact mode. In: Pro-ceedings from the 50th Annual Meeting of the Association for the Ad-vancement of Automotive Medicine in Chicago, Illinois, October 16–18,2006. Barrington, IL: Association for the Advancement of AutomotiveMedicine; 2006:73–89.

Prasad AK. CRASH3 Damage Model Reformulation. Vols I & II.Washington, DC: US Department of Transportation; 1987. VRTC-87-0053.

Prasad AK. Missing Vehicle Algorithm (OLDMISS) Reformulation.Warrendale, PA: Society of Automotive Engineers; 1991. SAE 910121.

Sharma D, Stern S, Brophy J, Choi E. An overview of NHTSA’s crashreconstruction software WinSMASH. Paper presented at: 20th Inter-national Technical Conference on Enhanced Safety of Vehicles; June18–21, 2007; Lyon, France.

Smith RA, Noga JT. Accuracy and Sensitivity of CRASH. Washington,DC: National Highway Traffic Safety Administration; 1982. DOT-HS-806-152.

Society of Automotive Engineers. Collision Deformation Classification.Warrendale, PA: Society of Automotive Engineers; 1980. SAE SurfaceVehicle Standard J224.

Dow

nloa

ded

by [

Uni

vers

ity o

f H

ong

Kon

g L

ibra

ries

] at

04:

06 1

2 N

ovem

ber

2014