Accident Analysis and Prevention 70 (2014) 267–272
Contents lists available at ScienceDirect
Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap
Short Communication
Crash fatality risk and unibody versus body-on-frame structure in SUVs Eric M. Ossiander a,∗ , Thomas D. Koepsell b , Barbara McKnight c a b c
Washington State Department of Health, PO Box 47812, Olympia, WA 98504-7812, USA Department of Epidemiology, School of Public Health, University of Washington, Seattle, WA 98195, USA Department of Biostatistics, School of Public Health, University of Washington, Seattle, WA 98195, USA
a r t i c l e
i n f o
Article history: Received 9 January 2014 Received in revised form 12 March 2014 Accepted 31 March 2014 Keywords: Compatibility Unibody construction SUVs Body-on-frame construction
a b s t r a c t Background: In crashes between cars and SUVs, car occupants are more likely to be killed than if they crashed with another car. An increasing proportion of SUVs are built with unibody, rather than truck-like body-on-frame construction. Unibody SUVs are generally lighter, less stiff, and less likely to roll over than body-on-frame SUVs, but whether unibody structure affects risk of death in crashes is unknown. Objective: To determine whether unibody SUVs differ from body-on-frame SUVs in the danger they pose to occupants of other vehicles and in the self-protection they offer to their own occupants. Methods: Case–control study of crashes between one compact SUV and one other passenger vehicle in the US during 1995–2008, in which the SUV was model year 1996–2006. Cases were all decedents in fatal crashes, one control was selected from each non-fatal crash. Findings: Occupants of passenger vehicles that crashed with compact unibody SUVs were at 18% lower risk of death compared to those that crashed with compact body-on-frame SUVs (adjusted odds ratio 0.82 (95% confidence interval 0.73–0.94)). Occupants of compact unibody SUVs were also at lower risk of death compared to occupants of body-on-frame SUVs (0.86 (0.72–1.02)). Conclusions: In two-vehicle collisions involving compact SUVs, unibody structure was associated with lower risk of death both in occupants of other vehicles in the crash, and in SUVs’ own occupants. © 2014 Elsevier Ltd. All rights reserved.
1. Introduction The occupants of passenger vehicles that crashed with compact SUVs have been found to have more than twice the risk of dying as the occupants of vehicles that crashed with a car (Ossiander et al., 2014). Although compact SUVs offered better protection to their own occupants than cars did, the excess risk they posed to occupants of other vehicles outweighed the benefit due to better self-protection, and in two-vehicle crashes, SUVs were associated with an excess overall risk of death (Ossiander et al., 2014). Before 1995, nearly all SUV models were built with a body-onframe design, in which the body is bolted onto a strong laddertype frame. Since then an increasing number of SUV models have been built with a unibody design, in which the body and frame are designed and welded together as a single unit (Bradsher, 2002).
∗ Corresponding author. Tel.: +1 360 236 4252; fax: +1 360 236 4245. E-mail addresses: [email protected] (E.M. Ossiander), [email protected] (T.D. Koepsell), [email protected] (B. McKnight). http://dx.doi.org/10.1016/j.aap.2014.03.019 0001-4575/© 2014 Elsevier Ltd. All rights reserved.
Unibody SUVs differ from body-on-frame SUVs in several ways that might affect both the harm they pose to occupants of other vehicles in crashes and the self-protection they offer their own occupants. They are generally lighter and less stiff than body-on-frame SUVs, and may have better-engineered crumple zones (Bradsher, 2002). However, there has been no careful study of the crash fatality risk associated with unibody SUVs compared to body-on-frame SUVs. In a crash between two vehicles, the fatality risk of a vehicle is determined by two components – the crashworthiness of the vehicle (its ability to protect its own occupants), and the crash aggressivity of the vehicle (the hazard it imposes on the other vehicle in the crash) (Huang et al., 2011, 2014). Therefore, we performed two analyses to determine whether crash fatality risk was related to structure type in SUVs. The first analysis assessed the association of SUV structure type with fatality risk of occupants of other passenger vehicles the SUV crashed with (the opposing unibody effect). The second analysis assessed the association of SUV structure type with fatality risk among SUVs’ own occupants (the unibody self-protection effect). There are few full-size unibody SUV models, and only one of them (Honda Pilot) was involved in a substantial number of crashes,
268
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
so we restricted our analysis to compact SUVs. Before the 1996 model year, all unibody SUVs were of a single make (Jeep), so we restricted our analysis to model years 1996 and later.
to minimize confounding and to minimize bias in estimates of effects or their variances that may arise from confounder selection (Harrell, 2001; Greenland, 2008).
2. Materials and methods
2.4. Multiple imputation for missing data
Both analyses used case–control designs. The selected case or control was the index person. The vehicle the index person was riding in was the index vehicle, and the vehicle that they crashed with was the opposing vehicle. For the analysis of the effect of SUV structure type on the occupants of other vehicles (the opposing effect analysis), we used crashes in which the index vehicle was any passenger vehicle, and the opposing vehicle was a compact SUV. For the analysis of SUV structure type on SUV occupants (the self-protection analysis), we used crashes in which the index vehicle was a compact SUV, and the opposing vehicle was any passenger vehicle. Some 1996 model year SUVs crashed in 1995, so we used twovehicle crashes occurring in 1995–2008 in which one vehicle was a model year 1996–2006 compact SUV and one vehicle was a model year 1980 or newer passenger vehicle of any type. We excluded all crashes involving pedestrians.
Information on structure type was missing for 12.5% of the opposing SUVs in the opposing effect analysis and for 13.6% of the index SUVs in the self-protection analysis. Information was also missing for many of the adjustment variables; sometimes for a large percentage of the records (Table 1). We used multiple imputation to create 10 datasets, as described below, each of which included the observed information and imputed values for the missing information (Rubin, 1987; Schafer, 1997). To impute missing values, we used sequential regression multivariate imputation (SRMI) (Raghunathan et al., 2001; Schenker et al., 2006), as implemented in IVEware (IVEware., 2007). In the imputation models, we included variables defining the stratification, clustering, and weighting of the GES sample design, all the covariates we intended to use in the analysis, including the outcome and exposures of interest, and other variables in the data that we thought would improve the prediction of the missing values. In SRMI, for each variable with missing data, a regression model is formed with that variable as the outcome and all other variables as predictors, to produce a predictive distribution for the missing data. Then values are randomly drawn from that predictive distribution. These imputed values are included, along with the observed data, in the predictive regression model for the next variable with missing data. The process cycles through all the variables with missing data, repeating for several iterations. The values imputed in the last iteration are used to form a dataset that includes the observed information and imputed values for the missing information. We repeated this entire process 10 times to create 10 imputed datasets, each containing the same observed data, but with possibly different imputed values across datasets for data that had originally been missing.
2.1. Case and control selection We drew cases from the Fatality Analysis Reporting System (FARS), which is maintained by the National Highway Traffic Safety Administration (NHTSA) (Tessmer, 2006). For the opposing unibody effect analysis, cases were all passenger vehicle occupants who died in a crash in which the opposing vehicle was a compact SUV. For the unibody self-protection analysis, cases were all occupants of a compact SUV who died in a crash with another passenger vehicle. We drew controls from the General Estimates System (GES) database, which is also maintained by NHTSA (NHTSA, 2007; Shelton, 1988). The GES data include a probability sample that is designed to be representative of all police-reported crashes in the U.S. For the opposing unibody effect analysis, we classified the compact SUV as the opposing vehicle, and the other vehicle as the index vehicle. For the self-protection analysis, we classified the compact SUV as the index vehicle. If a collision was between two compact SUVs, we randomly selected one of them to be the index vehicle. In each analysis, we randomly selected one occupant from the index vehicle to be the index person. If there was a fatality in the GES crash, that crash was still eligible, but we selected a surviving occupant as the control. 2.2. Curbweight, wheelbase and track width We obtained vehicle curbweight, wheelbase and track width from the Canadian Vehicle Specifications (CVS) dataset of vehicle measurement data compiled by the Collision Investigation and Research Division of Transport Canada (Transport Canada, 2007). For measurements missing from the CVS dataset, we obtained measurements from an automotive review website (MSN, 2009). 2.3. Adjustment variables We developed a list of adjustment variables by examining the fields in the FARS and GES data and including variables that could be related to both the likelihood of dying in a crash and SUV structure type, based on previous crash studies (Cummings et al., 2002) and biomechanical considerations. We excluded from the list any factors that might mediate the relationship between risk of dying and structure type, such as the amount of damage to the vehicle, and whether it rolled over or not. Nineteen variables were included on the list (Table 1). We adjusted for all of these variables in order
2.5. Statistical analysis We used logistic regression to estimate odds ratios of death according to structure type of the opposing SUV (for the opposing effect analysis) or the index SUV (for the self-protection analysis). For the controls, we computed new survey design weights that accounted for our random selection of the index person, as well as the original GES sampling probabilities. The design weights for cases were all 1, as they were selected with certainty. The design weights for the controls ranged from 1.2 to 3872. An analysis which uses weights that have large variation is often inefficient (Korn and Graubard, 1995). Scott and Wild proposed a method of rescaling the weights that may improve efficiency without producing biased estimates (Scott and Wild, 2001, 2003). Therefore, to obtain better statistical efficiency, we rescaled the weights so that the sum of the weights for the cases and controls equaled the numbers of cases and controls, respectively. We compared the estimates produced with these rescaled weights to design-weighted estimates to verify comparability. For both analyses, we considered whether the odds ratio associated with structure type was modified by each of these factors: vehicle speed, speed limit, crash type, index vehicle type, seating position, and SUV wheelbase. We planned to retain interaction terms if they were statistically significant at the p = 0.05 level. We conducted the analysis in SUDAAN, which automatically averages the coefficient estimates from the 10 imputation sets, and computes variance estimates and confidence intervals in a way that accounts for both the variance within each imputed set and the variance among them (Rubin, 1987; Research Triangle Institute, 2004),
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
269
Table 1 List of adjustment variables, with the percentage of missing data. Potential confounder
Percent missing Opposing unibody analysis
Unibody self-protection analysis
Crash characteristics Type of road Speed limit (linear and quadratic terms) Crash type Number of traffic lanes Crash year Hour of crash (linear and quadratic terms) Day of the week of the crash Crash month
9.6 6.4 0.8 14.0 0.0 0.4 0.0 0.0
10.5 6.9 0.8 15.5 0.0 0.4 0.0 0.0
Vehicle characteristics Size class of index vehicle Size class of opposing vehicle Speed of index vehicle (linear and quadratic terms) Speed of opposing vehicle (linear and quadratic terms) Model year of index vehicle Model year of opposing vehicle Curbweight of the index vehicle Curbweight of the opposing vehicle Wheelbase of the index vehicle Wheelbase of the opposing vehicle (linear and quadratic terms) Track width of the index vehicle Track width of the opposing vehicle (linear and quadratic terms)
0.6 – 56.1 52.5 0.0 – 20.0 – – 11.8 – 11.8
– 0.7 52.3 55.1 – 0.0 – 22.5 13.0 – 13.0 –
3.2 2.1 7.4 0.5
3.4 2.4 6.9 0.5
Index person characteristics Age (linear and quadratic terms) Sex Seat belt use Seating position
so that the resulting confidence intervals reflect both the variability in the observed data, and the uncertainty that arises because some data are missing. The study used only publicly available data, and was classified as exempt from human subjects review. 3. Results The analysis was limited to compact SUVs, but the classification of SUV as compact versus full-size was missing, and therefore imputed, for some vehicles. Therefore the numbers of cases and controls varied across the ten imputation sets. For the analysis of the opposing unibody effect, there were an average of 8725 cases and 33,884 controls. The cases were drawn from an average of 7787 fatal crashes. For the unibody self-protection analysis, the number of cases was 4265 in each of the imputation sets, and there were an average of 34,008 controls. The cases were drawn from 3805 fatal crashes. The controls came from the same set of crashes in both analyses, but since the index vehicle in the analysis of the opposing unibody effect was the opposing vehicle in the unibody self-protection analysis, the control persons were different in the two analyses. Even though the crashes from which the controls were selected were the same, variables common to the crash, such as speed limit or type of road, may show slight differences between the two analyses because of the imputation of missing data. In the analysis of the opposing unibody effect, there were 5757 deaths among occupants of passenger vehicles that crashed with compact body-on-frame SUVs. In the unibody self-protection analysis, there were 2825 deaths among occupants of compact body-on-frame SUVs. In the analysis of the opposing unibody effect, the opposing SUV in crashes involving fatal cases was more likely to be body-onframe than in crashes involving controls, and the index vehicles in which cases were riding were generally lighter, shorter, and more narrow than those carrying controls (see Table 2). In the self-protection analysis, the index SUV for cases was more likely
to be body-on-frame than the index SUV among controls, and the opposing vehicles in case crashes were generally heavier, longer, and wider than among controls. In both analyses, both the index and the opposing vehicles were more likely to be traveling at high speed in crashes involving cases than in crashes involving controls, and case crashes were more likely to occur on roads with high speed limits than control crashes. Cases were less likely to be wearing a seatbelt than controls. In each analysis we tested for effect modification as described in Section 2; none of them was significant at the p = 0.05 level. In the opposing unibody analysis, unibody structure was associated with a reduced risk of death among occupants of index vehicles (adjusted odds ratio (OR) 0.82, 95% confidence interval (CI) 0.73–0.94) (Table 3). When we also adjusted for curbweight of the opposing SUVs, the OR for unibody structure was attenuated slightly to 0.86, suggesting that some, but not all, of the reduced fatality risk associated with unibody SUVs was associated with their lower curbweight, compared to body-on-frame SUVs. In the unibody self-protection analysis, unibody structure was associated with a reduced risk of death among SUV occupants (OR 0.86, 95% CI 0.72–1.02) (Table 3). When we also adjusted for curbweight of the index SUVs, the OR was further reduced to 0.80, suggesting that if unibody SUVs were as heavy as body-on-frame SUVs, unibody structure might be associated with a greater reduction in risk. An analysis that used the original survey design weights instead of the rescaled weights provided similar estimated odds ratios, but wider confidence intervals, suggesting that using the rescaled weights did not produce biased estimates, but did improve the statistical efficiency of the estimates. 4. Discussion Our results suggest that among compact SUVs, unibody structure was associated with lower fatality risk both among occupants of vehicles that SUVs crashed with, and among SUVs’ own
270
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
Table 2 Characteristics of cases and controls. For the controls, the reported percentages are weighted using the survey design weights. Characteristic
Number of crashes
Opposing unibody study
Unibody self-protection study
Cases a (N = 8725) %
Controlsa (N = 33,884) %
Cases a (N = 4265) %
Controlsa (N = 34,008) %
7787
33,884
3805
34,008
66 34
58 42 66 34
59 41
Opposing vehicle structure type Body-on-frame Unibody Index vehicle structure type Body-on-frame Unibody Index vehicle type Car Compact SUV Full-size SUV Minivan Full-size van Compact pickup Full-size pickup Opposing vehicle type Car Compact SUV Full-size SUV Minivan Full-size van Compact pickup Full-size pickup
75 7 1 5 1 6 5
59 15 4 7 2 5 8
100
100
100
41 14 5 5 4 7 24
100
60 15 4 6 1 5 9
Curbweight of index vehicle 1299 kg or less 1300–1499 kg 1500–1799 kg 1800–2199 kg 2200 kg or more
41 23 25 9 2
25 24 30 16 6
5 18 47 30 1
3 15 47 34 2
Curbweight of opposing vehicle 1299 kg or less 1300–1499 kg 1500–1799 kg 1800–2199 kg 2200 kg or more
1 12 46 39 2
3 15 47 34 1
16 17 26 25 16
25 24 29 16 6
Wheelbase of index vehicle 219–259 cm 260–269 cm 270–279 cm 280–299 cm 300 cm or more Wheelbase of opposing vehicle 219–259 cm 260–269 cm 270–279 cm 280–299 cm 300 cm or more
19 25 22 30 4
14 37 30 17 2
Estimated speed of index vehicle (miles per hour) 0 10 13 1–19 21 20–39 33 40–59 23 60+
22 28 25 22 3
21 37 29 11 1
15 33 35 15 2
23 29 25 21 3
Front track width of index vehicle 94–144 cm 145–149 cm 150–154 cm 150–159 cm 160 cm or more Front track width of opposing vehicle 94–144 cm 145–149 cm 150–154 cm 150–159 cm 160 cm or more
25 25 24 23 2
14 34 35 14 2 32 27 27 11 2
9 6 16 37 31
39 27 23 9 2
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
271
Table 2 (Continued) Characteristic
Opposing unibody study
Unibody self-protection study
a
a
Cases (N = 8725) %
Controls (N = 33,884) %
Cases a (N = 4265) %
Controlsa (N = 34,008) %
Estimated speed of opposing vehicle (miles per hour) 6 0 1–19 4 20–39 18 40–59 47 60+ 25
37 27 25 9 2
10 5 14 36 35
31 27 29 11 2
Speed limit (miles per hour) 0–30 35–40 45–50 55+
7 18 24 51
22 37 26 15
6 14 18 63
22 37 26 15
Seat belt use Seatbelt used Not used
60 40
97 3
51 49
98 2
a
Percents may not sum to 100 because of rounding.
Table 3 Crude and adjusted odds ratios for the effect of unibody structure compared to body-on-frame structure among compact SUVs. The “opposing unibody effect” is the effect of unibody structure on fatality risk among the occupants of other passenger vehicles that SUVs crash with. The “unibody self-protection effect” is the effect of unibody structure on fatality risk among the SUV’s own occupants. Opposing unibody effect
Crude odds ratio Adjusted odds ratioa Also controlling for curbweightb
Unibody self-protection effect
OR
95% CI
OR
95% CI
0.71 0.82 0.86
0.65–0.77 0.73–0.94 0.76–0.99
0.72 0.86 0.80
0.65–0.80 0.72–1.02 0.66–0.97
a The odds ratio for the opposing unibody effect was adjusted for the adjustment variables listed in Table 1 for the opposing unibody analysis. The odds ratio for the unibody self-protection effect was adjusted for the adjustment variables listed in Table 1 for the unibody self-protection analysis. b The odds ratio for the opposing unibody effect was adjusted for the items listed above plus the curbweight of the opposing SUV. The odds ratio for the unibody self-protection effect was adjusted for the items listed above plus the curbweight of the index SUV.
occupants, compared to body-on-frame structure. After adjusting for the smaller average size of unibody SUVs and other possible confounders, the risk of death among passenger vehicle occupants was 18% (95% CI 6–27%) lower if the other vehicle in the crash was a compact unibody SUV, compared to crashes in which the other vehicle was a compact body-on-frame SUV. The risk of death among occupants of compact SUVs was 14% (95% CI 2–28%) lower when they were riding in unibody SUVs, compared to body-on-frame SUVs. We found that unibody SUVs were, on average, smaller (as measured by wheelbase) than body-on-frame SUVs, and for a given size, were also lighter than body-on-frame SUVs. In our analysis, we controlled for size as a potential confounder, assuming that it is a marketing feature. For example, if buyers are choosing between a Ford Escape and a Ford Explorer (the Escape is a unibody SUV which is smaller than the body-on-frame Explorer), we assume that size would be one of the features they would use in deciding which SUV to buy, and that Ford produces both models so it can appeal to a greater variety of buyers. The average size difference between unibody and body-on-frame SUVs may change if buyer preferences change, or as manufacturers design new models. We assume that after considering size and other features, curbweight will have little influence on a buyer’s choice of SUV model, and that manufacturers do not manipulate curbweight to appeal to buyers. Structure type is, in part, a determinant of curbweight, so we could not control for curbweight without also controlling for the effect of structure type, which is the effect we wanted to estimate. Therefore, we did not treat curbweight as a confounder. We did, however, control for size of the vehicle. However, in addition to our main estimates of the unibody effect, we also produced estimates in which we controlled for curbweight, in order to help describe the influence of mass on unibody SUVs’ effect on fatality risk. We estimated that if unibody SUVs
were as heavy, for a given size, as body-on-frame SUVs, they would be still be associated with a 14% (95% CI 1–24%) lower fatality risk among occupants of vehicles they crashed with. This suggests that unibody SUVs’ lower weight may contribute to the lower risk they pose to occupants of other vehicles, but that other features also contribute. We also estimated that if unibody SUVs were as heavy as bodyon-frame SUVs, they would be associated with a 20% (95% CI 3–34%) lower risk of death among their own occupants. Our results suggest that unibody structure can enhance SUV occupant protection without increasing curbweight, and perhaps even while lowering curbweight. In a study involving model year 1980–2009 compact SUVs of both structure types, we found that occupants of vehicles that crashed with compact SUVs were 2.3 times as likely to be killed as occupants of vehicles that crashed with cars (Ossiander et al., 2014). Other studies of crash compatibility between cars and light trucks have provided roughly similar results (Huang et al., 2011; Kahane, 1995; Fredette et al., 2008). Unibody structure is associated with only a small reduction in this excess risk, suggesting that other features of compact SUVs are more important than structure type in determining risk to occupants of other vehicles. 4.1. Limitations Information on SUV structure type was missing for 12–14% of the SUVs in the analysis, and there was a substantial amount of missing data on other covariates, most notably vehicle speed. An analysis of the subset of records with complete data can produce biased estimates unless the records with missing data are like a random subset of the data – the missing values are not correlated with either the outcome or the predictors in the model. We used
272
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
multiple imputation, which allowed us to produce valid estimates under the more plausible assumption that the missing values may be correlated with either the outcome or the predictors, but are not correlated with unmeasured data after conditioning on the observed data (Schafer, 1997). Estimates of vehicle speed before the crash were missing for nearly 60% of the vehicles. Since speed is strongly related to the risk of death in a crash, and may be related to structure type, inadequate control for the confounding effects of speed may bias our estimates. However, we found that when we removed the terms for vehicle speed from the model (but retained the speed limit terms), our estimates changed only very slightly. 4.2. Conclusions In this large population-based study of crashes between compact SUVs and other passenger vehicles, we found that unibody SUVs posed a lower threat to occupants of other vehicles than bodyon-frame SUVs did. We also found that unibody SUVs may protect their own occupants better than body-on-frame SUVs do. Acknowledgements We thank Drs Frederick Rivara and Peter Cummings for advice on study design and comments on an earlier version of the manuscript. References Bradsher, K., 2002. High and Mighty: SUVs – The World’s Most Dangerous Vehicles and How They Got That Way. Public Affairs, New York. Cummings, P., Koepsell, T.D., Rivara, F.P., McKnight, B., Mack, C., 2002. Air bags and passenger fatality according to passenger age and restraint use. Epidemiology 13 (5), 525–532. Fredette, M., Mambu, L.S., Chouinard, A., Bellavance, F., 2008. Safety impacts due to the incompatibility of SUVs: minivans, and pickup trucks in two-vehicle collisions. Accid. Anal. Prev. 40 (6), 1987–1995.
Greenland, S., 2008. Invited commentary: variable selection versus shrinkage in the control of multiple confounders. Am. J. Epidemiol. 167 (5), 523–529. Harrell Jr., F.E., 2001. Regression Modeling Strategies. Springer-Verlag, New York. Huang, H., Hu, S., Abdel-Aty, M., 2014. Indexing crash worthiness and crash aggressivity by major car brands. Saf. Sci. 62, 339–347. Huang, H., Siddiqui, C., Abdel-Aty, M., 2011. Indexing crash worthiness and crash aggressivity by vehicle type. Accid. Anal. Prev. 43 (4), 1364–1370. IVEware, 2007. Imputation and variance estimation software, Available at: http://www.isr.umich.edu/src/smp/ive/ (accessed 28.05.12). Kahane, C.J., 2003. Vehicle weight, fatality risk and crash compatibility of model year 1991–99 passenger cars and light trucks. Tech. Rep. DOT HS 809 662, NHTSA. Korn, E.L., Graubard, B.I., 1995. Analysis of large health surveys: accounting for the sampling design. J. R. Stat. Soc. A 158 (2), 263–295. MSN, 2009. msn.autos, Available at: autos.msn.com (accessed 28.05.12). NHTSA, 2007. National Automotive Sampling System (NASS); General Estimates System (GES) Analytical User’s Manual 1988–2006. National Highway Traffic Safety Administration, Washington, DC. Ossiander, E.M., Koepsell, T.D., McKnight, B., 2014. Crash fatality and vehicle incompatibility in collisions between cars and light trucks or vans. Injury Prevention (in press). Raghunathan, T.E., Lepkowski, J.M., Van Hoewyk, J., Solenberger, P., 2001. A multivariate technique for multiply imputing missing values using a sequence of regression models. Surv. Methodol. 27 (1), 85–95. 2004. Research Triangle Institute SUDAAN Language Manual, Release 9.0. Research Triangle Institute, Research Triangle Park, NC. Rubin, D.B., 1987. Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons, New York. Schafer, J.L., 1997. Analysis of Incomplete Multivariate Data. Chapman & Hall, Boca Raton. Schenker, N., Raghunathan, T.E., Chiu, P., Makuc, D.M., Zhang, G., Cohen, A.J., 2006. Multiple imputation of missing income data in the National Health Interview Survey. J. Am. Stat. Assoc. 101 (475), 924–933. Scott, A., Wild, C., 2001. Case–control studies with complex sampling. Appl. Statist. 50 (3), 389–401. Scott, A., Wild, C., 2003. Fitting logistic regression models in case–control studies with complex sampling. In: Chambers, R.L., Skinner, C.J. (Eds.), Analysis of Survey Data. John Wiley & Sons, Chichester, pp. 109–122. Shelton, T.S.T., 1991. National Accident Sampling System; General Estimates System Technical Note, 1988 to 1990. National Highway Traffic Safety Administration, Washington, DC. Tessmer, J.M., 2007. FARS Analytic Reference Guide 1975 to 2006. National Highway Traffic Safety Administration, Washington, DC. Transport Canada, 2007. Canadian vehicle specifications, Available at: http://www.carsp.ca/documents/47#cvs, (version 2007.1, accessed 28.05.12).
Contents lists available at ScienceDirect
Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap
Short Communication
Crash fatality risk and unibody versus body-on-frame structure in SUVs Eric M. Ossiander a,∗ , Thomas D. Koepsell b , Barbara McKnight c a b c
Washington State Department of Health, PO Box 47812, Olympia, WA 98504-7812, USA Department of Epidemiology, School of Public Health, University of Washington, Seattle, WA 98195, USA Department of Biostatistics, School of Public Health, University of Washington, Seattle, WA 98195, USA
a r t i c l e
i n f o
Article history: Received 9 January 2014 Received in revised form 12 March 2014 Accepted 31 March 2014 Keywords: Compatibility Unibody construction SUVs Body-on-frame construction
a b s t r a c t Background: In crashes between cars and SUVs, car occupants are more likely to be killed than if they crashed with another car. An increasing proportion of SUVs are built with unibody, rather than truck-like body-on-frame construction. Unibody SUVs are generally lighter, less stiff, and less likely to roll over than body-on-frame SUVs, but whether unibody structure affects risk of death in crashes is unknown. Objective: To determine whether unibody SUVs differ from body-on-frame SUVs in the danger they pose to occupants of other vehicles and in the self-protection they offer to their own occupants. Methods: Case–control study of crashes between one compact SUV and one other passenger vehicle in the US during 1995–2008, in which the SUV was model year 1996–2006. Cases were all decedents in fatal crashes, one control was selected from each non-fatal crash. Findings: Occupants of passenger vehicles that crashed with compact unibody SUVs were at 18% lower risk of death compared to those that crashed with compact body-on-frame SUVs (adjusted odds ratio 0.82 (95% confidence interval 0.73–0.94)). Occupants of compact unibody SUVs were also at lower risk of death compared to occupants of body-on-frame SUVs (0.86 (0.72–1.02)). Conclusions: In two-vehicle collisions involving compact SUVs, unibody structure was associated with lower risk of death both in occupants of other vehicles in the crash, and in SUVs’ own occupants. © 2014 Elsevier Ltd. All rights reserved.
1. Introduction The occupants of passenger vehicles that crashed with compact SUVs have been found to have more than twice the risk of dying as the occupants of vehicles that crashed with a car (Ossiander et al., 2014). Although compact SUVs offered better protection to their own occupants than cars did, the excess risk they posed to occupants of other vehicles outweighed the benefit due to better self-protection, and in two-vehicle crashes, SUVs were associated with an excess overall risk of death (Ossiander et al., 2014). Before 1995, nearly all SUV models were built with a body-onframe design, in which the body is bolted onto a strong laddertype frame. Since then an increasing number of SUV models have been built with a unibody design, in which the body and frame are designed and welded together as a single unit (Bradsher, 2002).
∗ Corresponding author. Tel.: +1 360 236 4252; fax: +1 360 236 4245. E-mail addresses: [email protected] (E.M. Ossiander), [email protected] (T.D. Koepsell), [email protected] (B. McKnight). http://dx.doi.org/10.1016/j.aap.2014.03.019 0001-4575/© 2014 Elsevier Ltd. All rights reserved.
Unibody SUVs differ from body-on-frame SUVs in several ways that might affect both the harm they pose to occupants of other vehicles in crashes and the self-protection they offer their own occupants. They are generally lighter and less stiff than body-on-frame SUVs, and may have better-engineered crumple zones (Bradsher, 2002). However, there has been no careful study of the crash fatality risk associated with unibody SUVs compared to body-on-frame SUVs. In a crash between two vehicles, the fatality risk of a vehicle is determined by two components – the crashworthiness of the vehicle (its ability to protect its own occupants), and the crash aggressivity of the vehicle (the hazard it imposes on the other vehicle in the crash) (Huang et al., 2011, 2014). Therefore, we performed two analyses to determine whether crash fatality risk was related to structure type in SUVs. The first analysis assessed the association of SUV structure type with fatality risk of occupants of other passenger vehicles the SUV crashed with (the opposing unibody effect). The second analysis assessed the association of SUV structure type with fatality risk among SUVs’ own occupants (the unibody self-protection effect). There are few full-size unibody SUV models, and only one of them (Honda Pilot) was involved in a substantial number of crashes,
268
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
so we restricted our analysis to compact SUVs. Before the 1996 model year, all unibody SUVs were of a single make (Jeep), so we restricted our analysis to model years 1996 and later.
to minimize confounding and to minimize bias in estimates of effects or their variances that may arise from confounder selection (Harrell, 2001; Greenland, 2008).
2. Materials and methods
2.4. Multiple imputation for missing data
Both analyses used case–control designs. The selected case or control was the index person. The vehicle the index person was riding in was the index vehicle, and the vehicle that they crashed with was the opposing vehicle. For the analysis of the effect of SUV structure type on the occupants of other vehicles (the opposing effect analysis), we used crashes in which the index vehicle was any passenger vehicle, and the opposing vehicle was a compact SUV. For the analysis of SUV structure type on SUV occupants (the self-protection analysis), we used crashes in which the index vehicle was a compact SUV, and the opposing vehicle was any passenger vehicle. Some 1996 model year SUVs crashed in 1995, so we used twovehicle crashes occurring in 1995–2008 in which one vehicle was a model year 1996–2006 compact SUV and one vehicle was a model year 1980 or newer passenger vehicle of any type. We excluded all crashes involving pedestrians.
Information on structure type was missing for 12.5% of the opposing SUVs in the opposing effect analysis and for 13.6% of the index SUVs in the self-protection analysis. Information was also missing for many of the adjustment variables; sometimes for a large percentage of the records (Table 1). We used multiple imputation to create 10 datasets, as described below, each of which included the observed information and imputed values for the missing information (Rubin, 1987; Schafer, 1997). To impute missing values, we used sequential regression multivariate imputation (SRMI) (Raghunathan et al., 2001; Schenker et al., 2006), as implemented in IVEware (IVEware., 2007). In the imputation models, we included variables defining the stratification, clustering, and weighting of the GES sample design, all the covariates we intended to use in the analysis, including the outcome and exposures of interest, and other variables in the data that we thought would improve the prediction of the missing values. In SRMI, for each variable with missing data, a regression model is formed with that variable as the outcome and all other variables as predictors, to produce a predictive distribution for the missing data. Then values are randomly drawn from that predictive distribution. These imputed values are included, along with the observed data, in the predictive regression model for the next variable with missing data. The process cycles through all the variables with missing data, repeating for several iterations. The values imputed in the last iteration are used to form a dataset that includes the observed information and imputed values for the missing information. We repeated this entire process 10 times to create 10 imputed datasets, each containing the same observed data, but with possibly different imputed values across datasets for data that had originally been missing.
2.1. Case and control selection We drew cases from the Fatality Analysis Reporting System (FARS), which is maintained by the National Highway Traffic Safety Administration (NHTSA) (Tessmer, 2006). For the opposing unibody effect analysis, cases were all passenger vehicle occupants who died in a crash in which the opposing vehicle was a compact SUV. For the unibody self-protection analysis, cases were all occupants of a compact SUV who died in a crash with another passenger vehicle. We drew controls from the General Estimates System (GES) database, which is also maintained by NHTSA (NHTSA, 2007; Shelton, 1988). The GES data include a probability sample that is designed to be representative of all police-reported crashes in the U.S. For the opposing unibody effect analysis, we classified the compact SUV as the opposing vehicle, and the other vehicle as the index vehicle. For the self-protection analysis, we classified the compact SUV as the index vehicle. If a collision was between two compact SUVs, we randomly selected one of them to be the index vehicle. In each analysis, we randomly selected one occupant from the index vehicle to be the index person. If there was a fatality in the GES crash, that crash was still eligible, but we selected a surviving occupant as the control. 2.2. Curbweight, wheelbase and track width We obtained vehicle curbweight, wheelbase and track width from the Canadian Vehicle Specifications (CVS) dataset of vehicle measurement data compiled by the Collision Investigation and Research Division of Transport Canada (Transport Canada, 2007). For measurements missing from the CVS dataset, we obtained measurements from an automotive review website (MSN, 2009). 2.3. Adjustment variables We developed a list of adjustment variables by examining the fields in the FARS and GES data and including variables that could be related to both the likelihood of dying in a crash and SUV structure type, based on previous crash studies (Cummings et al., 2002) and biomechanical considerations. We excluded from the list any factors that might mediate the relationship between risk of dying and structure type, such as the amount of damage to the vehicle, and whether it rolled over or not. Nineteen variables were included on the list (Table 1). We adjusted for all of these variables in order
2.5. Statistical analysis We used logistic regression to estimate odds ratios of death according to structure type of the opposing SUV (for the opposing effect analysis) or the index SUV (for the self-protection analysis). For the controls, we computed new survey design weights that accounted for our random selection of the index person, as well as the original GES sampling probabilities. The design weights for cases were all 1, as they were selected with certainty. The design weights for the controls ranged from 1.2 to 3872. An analysis which uses weights that have large variation is often inefficient (Korn and Graubard, 1995). Scott and Wild proposed a method of rescaling the weights that may improve efficiency without producing biased estimates (Scott and Wild, 2001, 2003). Therefore, to obtain better statistical efficiency, we rescaled the weights so that the sum of the weights for the cases and controls equaled the numbers of cases and controls, respectively. We compared the estimates produced with these rescaled weights to design-weighted estimates to verify comparability. For both analyses, we considered whether the odds ratio associated with structure type was modified by each of these factors: vehicle speed, speed limit, crash type, index vehicle type, seating position, and SUV wheelbase. We planned to retain interaction terms if they were statistically significant at the p = 0.05 level. We conducted the analysis in SUDAAN, which automatically averages the coefficient estimates from the 10 imputation sets, and computes variance estimates and confidence intervals in a way that accounts for both the variance within each imputed set and the variance among them (Rubin, 1987; Research Triangle Institute, 2004),
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
269
Table 1 List of adjustment variables, with the percentage of missing data. Potential confounder
Percent missing Opposing unibody analysis
Unibody self-protection analysis
Crash characteristics Type of road Speed limit (linear and quadratic terms) Crash type Number of traffic lanes Crash year Hour of crash (linear and quadratic terms) Day of the week of the crash Crash month
9.6 6.4 0.8 14.0 0.0 0.4 0.0 0.0
10.5 6.9 0.8 15.5 0.0 0.4 0.0 0.0
Vehicle characteristics Size class of index vehicle Size class of opposing vehicle Speed of index vehicle (linear and quadratic terms) Speed of opposing vehicle (linear and quadratic terms) Model year of index vehicle Model year of opposing vehicle Curbweight of the index vehicle Curbweight of the opposing vehicle Wheelbase of the index vehicle Wheelbase of the opposing vehicle (linear and quadratic terms) Track width of the index vehicle Track width of the opposing vehicle (linear and quadratic terms)
0.6 – 56.1 52.5 0.0 – 20.0 – – 11.8 – 11.8
– 0.7 52.3 55.1 – 0.0 – 22.5 13.0 – 13.0 –
3.2 2.1 7.4 0.5
3.4 2.4 6.9 0.5
Index person characteristics Age (linear and quadratic terms) Sex Seat belt use Seating position
so that the resulting confidence intervals reflect both the variability in the observed data, and the uncertainty that arises because some data are missing. The study used only publicly available data, and was classified as exempt from human subjects review. 3. Results The analysis was limited to compact SUVs, but the classification of SUV as compact versus full-size was missing, and therefore imputed, for some vehicles. Therefore the numbers of cases and controls varied across the ten imputation sets. For the analysis of the opposing unibody effect, there were an average of 8725 cases and 33,884 controls. The cases were drawn from an average of 7787 fatal crashes. For the unibody self-protection analysis, the number of cases was 4265 in each of the imputation sets, and there were an average of 34,008 controls. The cases were drawn from 3805 fatal crashes. The controls came from the same set of crashes in both analyses, but since the index vehicle in the analysis of the opposing unibody effect was the opposing vehicle in the unibody self-protection analysis, the control persons were different in the two analyses. Even though the crashes from which the controls were selected were the same, variables common to the crash, such as speed limit or type of road, may show slight differences between the two analyses because of the imputation of missing data. In the analysis of the opposing unibody effect, there were 5757 deaths among occupants of passenger vehicles that crashed with compact body-on-frame SUVs. In the unibody self-protection analysis, there were 2825 deaths among occupants of compact body-on-frame SUVs. In the analysis of the opposing unibody effect, the opposing SUV in crashes involving fatal cases was more likely to be body-onframe than in crashes involving controls, and the index vehicles in which cases were riding were generally lighter, shorter, and more narrow than those carrying controls (see Table 2). In the self-protection analysis, the index SUV for cases was more likely
to be body-on-frame than the index SUV among controls, and the opposing vehicles in case crashes were generally heavier, longer, and wider than among controls. In both analyses, both the index and the opposing vehicles were more likely to be traveling at high speed in crashes involving cases than in crashes involving controls, and case crashes were more likely to occur on roads with high speed limits than control crashes. Cases were less likely to be wearing a seatbelt than controls. In each analysis we tested for effect modification as described in Section 2; none of them was significant at the p = 0.05 level. In the opposing unibody analysis, unibody structure was associated with a reduced risk of death among occupants of index vehicles (adjusted odds ratio (OR) 0.82, 95% confidence interval (CI) 0.73–0.94) (Table 3). When we also adjusted for curbweight of the opposing SUVs, the OR for unibody structure was attenuated slightly to 0.86, suggesting that some, but not all, of the reduced fatality risk associated with unibody SUVs was associated with their lower curbweight, compared to body-on-frame SUVs. In the unibody self-protection analysis, unibody structure was associated with a reduced risk of death among SUV occupants (OR 0.86, 95% CI 0.72–1.02) (Table 3). When we also adjusted for curbweight of the index SUVs, the OR was further reduced to 0.80, suggesting that if unibody SUVs were as heavy as body-on-frame SUVs, unibody structure might be associated with a greater reduction in risk. An analysis that used the original survey design weights instead of the rescaled weights provided similar estimated odds ratios, but wider confidence intervals, suggesting that using the rescaled weights did not produce biased estimates, but did improve the statistical efficiency of the estimates. 4. Discussion Our results suggest that among compact SUVs, unibody structure was associated with lower fatality risk both among occupants of vehicles that SUVs crashed with, and among SUVs’ own
270
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
Table 2 Characteristics of cases and controls. For the controls, the reported percentages are weighted using the survey design weights. Characteristic
Number of crashes
Opposing unibody study
Unibody self-protection study
Cases a (N = 8725) %
Controlsa (N = 33,884) %
Cases a (N = 4265) %
Controlsa (N = 34,008) %
7787
33,884
3805
34,008
66 34
58 42 66 34
59 41
Opposing vehicle structure type Body-on-frame Unibody Index vehicle structure type Body-on-frame Unibody Index vehicle type Car Compact SUV Full-size SUV Minivan Full-size van Compact pickup Full-size pickup Opposing vehicle type Car Compact SUV Full-size SUV Minivan Full-size van Compact pickup Full-size pickup
75 7 1 5 1 6 5
59 15 4 7 2 5 8
100
100
100
41 14 5 5 4 7 24
100
60 15 4 6 1 5 9
Curbweight of index vehicle 1299 kg or less 1300–1499 kg 1500–1799 kg 1800–2199 kg 2200 kg or more
41 23 25 9 2
25 24 30 16 6
5 18 47 30 1
3 15 47 34 2
Curbweight of opposing vehicle 1299 kg or less 1300–1499 kg 1500–1799 kg 1800–2199 kg 2200 kg or more
1 12 46 39 2
3 15 47 34 1
16 17 26 25 16
25 24 29 16 6
Wheelbase of index vehicle 219–259 cm 260–269 cm 270–279 cm 280–299 cm 300 cm or more Wheelbase of opposing vehicle 219–259 cm 260–269 cm 270–279 cm 280–299 cm 300 cm or more
19 25 22 30 4
14 37 30 17 2
Estimated speed of index vehicle (miles per hour) 0 10 13 1–19 21 20–39 33 40–59 23 60+
22 28 25 22 3
21 37 29 11 1
15 33 35 15 2
23 29 25 21 3
Front track width of index vehicle 94–144 cm 145–149 cm 150–154 cm 150–159 cm 160 cm or more Front track width of opposing vehicle 94–144 cm 145–149 cm 150–154 cm 150–159 cm 160 cm or more
25 25 24 23 2
14 34 35 14 2 32 27 27 11 2
9 6 16 37 31
39 27 23 9 2
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
271
Table 2 (Continued) Characteristic
Opposing unibody study
Unibody self-protection study
a
a
Cases (N = 8725) %
Controls (N = 33,884) %
Cases a (N = 4265) %
Controlsa (N = 34,008) %
Estimated speed of opposing vehicle (miles per hour) 6 0 1–19 4 20–39 18 40–59 47 60+ 25
37 27 25 9 2
10 5 14 36 35
31 27 29 11 2
Speed limit (miles per hour) 0–30 35–40 45–50 55+
7 18 24 51
22 37 26 15
6 14 18 63
22 37 26 15
Seat belt use Seatbelt used Not used
60 40
97 3
51 49
98 2
a
Percents may not sum to 100 because of rounding.
Table 3 Crude and adjusted odds ratios for the effect of unibody structure compared to body-on-frame structure among compact SUVs. The “opposing unibody effect” is the effect of unibody structure on fatality risk among the occupants of other passenger vehicles that SUVs crash with. The “unibody self-protection effect” is the effect of unibody structure on fatality risk among the SUV’s own occupants. Opposing unibody effect
Crude odds ratio Adjusted odds ratioa Also controlling for curbweightb
Unibody self-protection effect
OR
95% CI
OR
95% CI
0.71 0.82 0.86
0.65–0.77 0.73–0.94 0.76–0.99
0.72 0.86 0.80
0.65–0.80 0.72–1.02 0.66–0.97
a The odds ratio for the opposing unibody effect was adjusted for the adjustment variables listed in Table 1 for the opposing unibody analysis. The odds ratio for the unibody self-protection effect was adjusted for the adjustment variables listed in Table 1 for the unibody self-protection analysis. b The odds ratio for the opposing unibody effect was adjusted for the items listed above plus the curbweight of the opposing SUV. The odds ratio for the unibody self-protection effect was adjusted for the items listed above plus the curbweight of the index SUV.
occupants, compared to body-on-frame structure. After adjusting for the smaller average size of unibody SUVs and other possible confounders, the risk of death among passenger vehicle occupants was 18% (95% CI 6–27%) lower if the other vehicle in the crash was a compact unibody SUV, compared to crashes in which the other vehicle was a compact body-on-frame SUV. The risk of death among occupants of compact SUVs was 14% (95% CI 2–28%) lower when they were riding in unibody SUVs, compared to body-on-frame SUVs. We found that unibody SUVs were, on average, smaller (as measured by wheelbase) than body-on-frame SUVs, and for a given size, were also lighter than body-on-frame SUVs. In our analysis, we controlled for size as a potential confounder, assuming that it is a marketing feature. For example, if buyers are choosing between a Ford Escape and a Ford Explorer (the Escape is a unibody SUV which is smaller than the body-on-frame Explorer), we assume that size would be one of the features they would use in deciding which SUV to buy, and that Ford produces both models so it can appeal to a greater variety of buyers. The average size difference between unibody and body-on-frame SUVs may change if buyer preferences change, or as manufacturers design new models. We assume that after considering size and other features, curbweight will have little influence on a buyer’s choice of SUV model, and that manufacturers do not manipulate curbweight to appeal to buyers. Structure type is, in part, a determinant of curbweight, so we could not control for curbweight without also controlling for the effect of structure type, which is the effect we wanted to estimate. Therefore, we did not treat curbweight as a confounder. We did, however, control for size of the vehicle. However, in addition to our main estimates of the unibody effect, we also produced estimates in which we controlled for curbweight, in order to help describe the influence of mass on unibody SUVs’ effect on fatality risk. We estimated that if unibody SUVs
were as heavy, for a given size, as body-on-frame SUVs, they would be still be associated with a 14% (95% CI 1–24%) lower fatality risk among occupants of vehicles they crashed with. This suggests that unibody SUVs’ lower weight may contribute to the lower risk they pose to occupants of other vehicles, but that other features also contribute. We also estimated that if unibody SUVs were as heavy as bodyon-frame SUVs, they would be associated with a 20% (95% CI 3–34%) lower risk of death among their own occupants. Our results suggest that unibody structure can enhance SUV occupant protection without increasing curbweight, and perhaps even while lowering curbweight. In a study involving model year 1980–2009 compact SUVs of both structure types, we found that occupants of vehicles that crashed with compact SUVs were 2.3 times as likely to be killed as occupants of vehicles that crashed with cars (Ossiander et al., 2014). Other studies of crash compatibility between cars and light trucks have provided roughly similar results (Huang et al., 2011; Kahane, 1995; Fredette et al., 2008). Unibody structure is associated with only a small reduction in this excess risk, suggesting that other features of compact SUVs are more important than structure type in determining risk to occupants of other vehicles. 4.1. Limitations Information on SUV structure type was missing for 12–14% of the SUVs in the analysis, and there was a substantial amount of missing data on other covariates, most notably vehicle speed. An analysis of the subset of records with complete data can produce biased estimates unless the records with missing data are like a random subset of the data – the missing values are not correlated with either the outcome or the predictors in the model. We used
272
E.M. Ossiander et al. / Accident Analysis and Prevention 70 (2014) 267–272
multiple imputation, which allowed us to produce valid estimates under the more plausible assumption that the missing values may be correlated with either the outcome or the predictors, but are not correlated with unmeasured data after conditioning on the observed data (Schafer, 1997). Estimates of vehicle speed before the crash were missing for nearly 60% of the vehicles. Since speed is strongly related to the risk of death in a crash, and may be related to structure type, inadequate control for the confounding effects of speed may bias our estimates. However, we found that when we removed the terms for vehicle speed from the model (but retained the speed limit terms), our estimates changed only very slightly. 4.2. Conclusions In this large population-based study of crashes between compact SUVs and other passenger vehicles, we found that unibody SUVs posed a lower threat to occupants of other vehicles than bodyon-frame SUVs did. We also found that unibody SUVs may protect their own occupants better than body-on-frame SUVs do. Acknowledgements We thank Drs Frederick Rivara and Peter Cummings for advice on study design and comments on an earlier version of the manuscript. References Bradsher, K., 2002. High and Mighty: SUVs – The World’s Most Dangerous Vehicles and How They Got That Way. Public Affairs, New York. Cummings, P., Koepsell, T.D., Rivara, F.P., McKnight, B., Mack, C., 2002. Air bags and passenger fatality according to passenger age and restraint use. Epidemiology 13 (5), 525–532. Fredette, M., Mambu, L.S., Chouinard, A., Bellavance, F., 2008. Safety impacts due to the incompatibility of SUVs: minivans, and pickup trucks in two-vehicle collisions. Accid. Anal. Prev. 40 (6), 1987–1995.
Greenland, S., 2008. Invited commentary: variable selection versus shrinkage in the control of multiple confounders. Am. J. Epidemiol. 167 (5), 523–529. Harrell Jr., F.E., 2001. Regression Modeling Strategies. Springer-Verlag, New York. Huang, H., Hu, S., Abdel-Aty, M., 2014. Indexing crash worthiness and crash aggressivity by major car brands. Saf. Sci. 62, 339–347. Huang, H., Siddiqui, C., Abdel-Aty, M., 2011. Indexing crash worthiness and crash aggressivity by vehicle type. Accid. Anal. Prev. 43 (4), 1364–1370. IVEware, 2007. Imputation and variance estimation software, Available at: http://www.isr.umich.edu/src/smp/ive/ (accessed 28.05.12). Kahane, C.J., 2003. Vehicle weight, fatality risk and crash compatibility of model year 1991–99 passenger cars and light trucks. Tech. Rep. DOT HS 809 662, NHTSA. Korn, E.L., Graubard, B.I., 1995. Analysis of large health surveys: accounting for the sampling design. J. R. Stat. Soc. A 158 (2), 263–295. MSN, 2009. msn.autos, Available at: autos.msn.com (accessed 28.05.12). NHTSA, 2007. National Automotive Sampling System (NASS); General Estimates System (GES) Analytical User’s Manual 1988–2006. National Highway Traffic Safety Administration, Washington, DC. Ossiander, E.M., Koepsell, T.D., McKnight, B., 2014. Crash fatality and vehicle incompatibility in collisions between cars and light trucks or vans. Injury Prevention (in press). Raghunathan, T.E., Lepkowski, J.M., Van Hoewyk, J., Solenberger, P., 2001. A multivariate technique for multiply imputing missing values using a sequence of regression models. Surv. Methodol. 27 (1), 85–95. 2004. Research Triangle Institute SUDAAN Language Manual, Release 9.0. Research Triangle Institute, Research Triangle Park, NC. Rubin, D.B., 1987. Multiple Imputation for Nonresponse in Surveys. John Wiley & Sons, New York. Schafer, J.L., 1997. Analysis of Incomplete Multivariate Data. Chapman & Hall, Boca Raton. Schenker, N., Raghunathan, T.E., Chiu, P., Makuc, D.M., Zhang, G., Cohen, A.J., 2006. Multiple imputation of missing income data in the National Health Interview Survey. J. Am. Stat. Assoc. 101 (475), 924–933. Scott, A., Wild, C., 2001. Case–control studies with complex sampling. Appl. Statist. 50 (3), 389–401. Scott, A., Wild, C., 2003. Fitting logistic regression models in case–control studies with complex sampling. In: Chambers, R.L., Skinner, C.J. (Eds.), Analysis of Survey Data. John Wiley & Sons, Chichester, pp. 109–122. Shelton, T.S.T., 1991. National Accident Sampling System; General Estimates System Technical Note, 1988 to 1990. National Highway Traffic Safety Administration, Washington, DC. Tessmer, J.M., 2007. FARS Analytic Reference Guide 1975 to 2006. National Highway Traffic Safety Administration, Washington, DC. Transport Canada, 2007. Canadian vehicle specifications, Available at: http://www.carsp.ca/documents/47#cvs, (version 2007.1, accessed 28.05.12).
