Evaluation and Comparison of Roadside Crash Injury Metrics (2023)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

124 8 Build and Analyze Harm Analysis Dataset for Frontal, Oblique, and Side Crashes Introduction The purpose of this chapter is to present the Harm cost prediction models associated with each of the crash severity metrics in frontal, side, and oblique crashes. The IAD was composed of NASS/CDS crashes with available EDR data. The frontal, side, and oblique IADs were used to train linear regression models based on five candidate metrics: MDV, OIV, OLC, ASI, and VPI. The linear models were evaluated with equivalent test datasets composed of frontal, side, and oblique crashes extracted from CISS. Each candidate injury metric model was then ranked according to its RMSE. Methods Occupants in real-world frontal, side, and oblique impacts from the IAD were used to build and test the Harm models. The NASS/CDS cases were used to construct the models, and the CISS cases were used to test the models. 8.2.1 Case Selection Criteria Specific body region data were necessary to calculate the Harm cost for each occupant, so the datasets analyzed in Chapter 5 and Chapter 6 were used to form the analysis datasets for the frontal and side Harm models, respectively. For the oblique models, the same process used to obtain the injury data in Chapter 5 and Chapter 6 was used to obtain injury data for the oblique dataset. Occupants who suffered only injuries that had an unknown AIS value or an unknown body region were removed from the dataset. This is because they pose an additional amount of unknown variability to the dataset, as all unknown injuries are treated as a fixed dollar amount. If an occupant had several injuries and only a subset of them had an unknown AIS value or body location, the occupant was included in the analysis and the unknown injuries were assigned the cost associated with an unknown injury. 8.2.2 Consideration of Alternative Injury Measurement Methods In addition to the Harm metric, the FCI and ISS scores were identified as injury scales that could potentially be used to build injury prediction models using the IAD. FCI ranges from 0 to 100 and quantifies a person’s level of function over the course of 12 months following the crash. ISS ranges from 1 to 75 and is the sum of the squares of the highest AIS code in a person’s three most severely injured body regions. The current study implements AIS 1998, which does not provide a direct translation from AIS values to FCI values. Additionally, ISS is a transformation of the AIS. As a result, if thresholds equivalent to an MAIS2+F injury were chosen to differentiate the negative versus positive cases, the models yielded were identical to the models that implemented AIS. If a higher threshold was chosen on the ISS scale, there were not enough injury cases to produce reliable models. This is also why we were unable to construct MAIS3+F models when implementing the AIS.

125 8.2.3 Calculating Harm Cost The Harm metric is a means of measuring the societal cost of traffic crashes and is frequently used in the evaluation of impact injury countermeasures. This societal cost includes both medical costs and indirect costs, such as loss of wages. The original Harm metric was first developed by Malliaris et al. (1985) as a means of balancing number of injuries with the severity or cost of an injury, where severity is determined using the AIS. The improved Harm metric (multi-Harm) developed by Fildes et al. (1992) assigns a societal cost to each injury to estimate a total societal cost of injury. Equation 13 is the equation used to calculate the multi-Harm cost for an occupant. Table 8-1 tabulates the cost in thousands of U.S. dollar for each injury based on severity and body region. 𝐻𝐻𝑅𝑅𝑃𝑃𝑚𝑚 = � 𝐶𝐶𝑃𝑃𝑃𝑃𝑙𝑙𝑖𝑖(𝑏𝑏𝑃𝑃𝑏𝑏𝐴𝐴𝑃𝑃𝑃𝑃𝑙𝑙𝑃𝑃𝑃𝑃𝑃𝑃,𝐴𝐴𝑀𝑀𝑀𝑀) 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑖𝑖𝑁𝑁𝑁𝑁 𝑖𝑖=1 (12) Table 8-1. Average cost per injury in thousands of U.S. dollars. Injury Severity Body Region Minor Moderate Serious Severe Critical Maximum Unknown (AIS = 1) (AIS = 2) (AIS = 3) (AIS = 4) (AIS = 5) (AIS = 6) -- External 3 16 45 73 106 644 3 Head 4 19 78 180 636 644 3 Face 4 19 78 103 211 644 3 Neck 4 19 78 103 211 644 3 Chest 3 16 45 73 106 644 3 Abdomen-Pelvis 3 16 45 73 106 644 3 Spine 3 16 105 905 1,082 644 3 Upper Extremity 4 28 66 -- -- -- 3 Lower Extremity 3 28 84 124 211 -- 3 8.2.4 Injury Risk Modeling The models were developed using linear regression. Since Harm costs can span several orders of magnitude, the models were constructed to predict the square root of Harm. Several predictor variables were used within the models to predict the square root of Harm: • Injury Metrics. The injury metric values for each case were the same values computed for use in the injury risk models. Since the Harm metric varies over a range of order of magnitudes, the square root of Harm was taken to transform the data and fit it to a linear model. • Belt Status. Belt status was modeled as a categorical variable, where a value of 1 indicates an occupant using a three-point belt restraint, and 0 indicates the occupant was unbelted. Belt status was determined using the EDR belt status variable. • Age. Age was modeled as a categorical variable, where 1 indicates ≥ 65 years old and 0 indicates ≥ 13 years old and < 65 years old. • Sex. Sex was modeled as a categorical variable, where 1 indicates male and 0 indicates female.

126 • BMI. Body mass index (units kg/m2) was modeled as a categorical variable, where 1 indicates obese and 0 indicates not obese. • Occupant Seating Location. The occupant’s seating location was modeled as a categorical variable, where a value of 1 indicates the occupant was the driver, and 0 indicates the occupant was in the right front passenger seat. • General Area of Damage (GAD). GAD describes where the greatest area of deformation occurred. It was modeled as a categorical variable, where 0 indicates side damage and 1 indicates frontal damage. • Vehicle Type. This variable was defined using NHTSA’s Vehicle Body Type Classification. All the vehicles in this dataset fall into the category of either Passenger Car (PC; body types: 1-11, 17) or Light Trucks and Vans (LTV; body types: 14-16, 19-22, 24, 25, 28-41, 45-49). All of the vehicles in our dataset fell into one of these categories. Vehicle type was modeled as a categorical variable, where a value of 1 indicates a PC and 0 indicates an LTV. Our approach was to first develop linear regression models using the full set of covariates. These initial regression models were then examined to determine which of the covariates were statistically significant. Based on those results, a second set of regression models was developed in which only the statistically significant covariates were included. If a covariate was significant in the initial model and insignificant in the final model, it was removed. The linear regression models for each set were developed using the survey package and SVYDESIGN function in R (Lumley 2020). The R survey package was designed based on SAS and is capable of computing confidence limits that account for the complex NASS-CDS stratified sampling scheme. 8.2.5 Predictive Capability The RMSE and coefficient of determination values were used to quantify the predictive capability of the models. The RMSE measures the differences between the observed values and the values predicted by the model. The coefficient of determination, often referred to as the R2 value, ranges from 0 to 1 and is a measure of what percentage of variation in the independent variable is accounted for by the model covariates. Final Linear Predictive Models Based on the initial regression models, a second set of injury models was developed. These models use only the covariates that were statistically significant in the initial model. Table 8-2 through Table 8-16 show the regression coefficients for each of the final models. Each set of initial models used a separate set of covariates. The logit expanded containing each of the initial covariates is provided in each sub-section. 8.3.1 Final Frontal Crash Harm Models Equation 14 is the logit expanded for the frontal crash Harm models. 𝑌𝑌𝑓𝑓𝑁𝑁𝑓𝑓𝑁𝑁𝑓𝑓 = 𝛽𝛽0 + 𝛽𝛽1 ⋅ (𝑃𝑃𝑃𝑃𝑖𝑖𝐴𝐴𝑃𝑃𝐴𝐴 𝑚𝑚𝑃𝑃𝑙𝑙𝑃𝑃𝑃𝑃𝑃𝑃) + 𝛽𝛽2 ⋅ 𝑏𝑏𝑃𝑃𝑅𝑅𝑙𝑙_𝑃𝑃𝑙𝑙𝑅𝑅𝑙𝑙𝐴𝐴𝑃𝑃 + 𝛽𝛽3 ⋅ 𝑃𝑃𝑃𝑃𝑠𝑠 + 𝛽𝛽4 ⋅ 𝑅𝑅𝑙𝑙𝑃𝑃 + 𝛽𝛽5 ⋅ 𝑃𝑃𝑏𝑏𝑃𝑃𝑃𝑃𝑃𝑃 + 𝛽𝛽6 ⋅ 𝑃𝑃𝑃𝑃𝑅𝑅𝑙𝑙𝑃𝑃𝑃𝑃𝑙𝑙𝑙𝑙𝑓𝑓𝑙𝑙𝑙𝑙𝑓𝑓𝑖𝑖𝑓𝑓𝑁𝑁 + 𝛽𝛽7 ⋅ 𝑉𝑉𝑃𝑃ℎ𝑃𝑃𝑃𝑃𝑅𝑅𝑃𝑃 𝑇𝑇𝐴𝐴𝑡𝑡𝑃𝑃 + 𝛽𝛽8 ⋅ 𝑃𝑃𝑏𝑏𝑖𝑖𝑃𝑃𝑃𝑃𝑙𝑙 𝐶𝐶𝑃𝑃𝑃𝑃𝑙𝑙𝑅𝑅𝑃𝑃𝑙𝑙𝑃𝑃𝑏𝑏 (13)

127 Table 8-2. Parameters for the MDV final frontal linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept -32.093 36.438 0.380 MDV β1, MDV (m/s) 19.102 3.238 < 0.001** Belt Status β2, Belted -35.901 14.012 0.012** Sex β3, Male -29.488 11.948 0.015** Age β4, ≥ 65 years 57.169 18.633 0.003** BMI β5, < 25 kg/m2 25.175 9.823 0.012** Seating Location β6, Driver 28.075 10.648 0.009** Object Contacted β8, Not a tree or pole -38.536 13.634 0.005** Table 8-3. Parameters for the OIV final frontal linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept -32.939 36.604 0.370 OIV β1, OIV (m/s) 20.065 3.463 < 0.001** Belt Status β2, Belted -37.577 14.189 0.009** Sex β3, Male -28.804 11.719 0.015** Age β4, ≥ 65 years 56.808 17.765 0.002** BMI β5, < 25 kg/m2 26.117 9.869 0.009** Seating Location β6, Driver 29.144 11.041 0.009** Object Contacted β8, Not a tree or pole -41.664 13.252 0.002** Table 8-4. Parameters for the OLC final frontal linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept -61.973 25.934 0.018** OLC β1, OLC (g) 9.054 1.692 < 0.001** Age β4, ≥ 65 years 42.291 17.253 0.016** BMI β5, < 25 kg/m2 28.174 11.760 0.018** Seating Location β6, Driver 26.754 10.853 0.015** Object Contacted β8, Not a tree or pole 32.673 11.769 0.006** Table 8-5. Parameters for the ASI final frontal linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept -51.586 33.807 0.130** ASI β1, ASI 150.104 26.415 < 0.001** Belt Status β2, Belted -33.324 16.037 0.040** Sex β3, Male -29.074 12.565 0.022** Age β4, ≥ 65 years 54.275 16.355 0.001** BMI β5, < 25 kg/m2 23.322 9.693 0.018** Seating Location β6, Driver 28.808 11.199 0.011** Table 8-6. Parameters for the VPI final frontal linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05).

128 Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept -51.586 33.807 0.078** VPI β1, VPI (m/s2) 150.104 26.415 < 0.001** Belt Status β2, Belted -33.324 16.037 0.040** Sex β3, Male -29.074 12.565 0.022** Age β4, ≥ 65 years 54.275 16.355 0.001** BMI β5, < 25 kg/m2 23.322 9.693 0.018** Seating Location β6, Driver 28.808 11.199 0.011** 8.3.2 Final Oblique Crash Harm Models 𝑌𝑌𝑓𝑓𝑜𝑜𝑙𝑙𝑖𝑖𝑜𝑜𝑁𝑁𝑁𝑁 = 𝛽𝛽0 + 𝛽𝛽1 ⋅ (𝑃𝑃𝑃𝑃𝑖𝑖𝐴𝐴𝑃𝑃𝐴𝐴 𝑚𝑚𝑃𝑃𝑙𝑙𝑃𝑃𝑃𝑃𝑃𝑃) + 𝛽𝛽2 ⋅ 𝑏𝑏𝑃𝑃𝑅𝑅𝑙𝑙_𝑃𝑃𝑙𝑙𝑅𝑅𝑙𝑙𝐴𝐴𝑃𝑃 + 𝛽𝛽3 ⋅ 𝑃𝑃𝑃𝑃𝑠𝑠 + 𝛽𝛽4 ⋅ 𝑅𝑅𝑙𝑙𝑃𝑃 + 𝛽𝛽5 ⋅ 𝑃𝑃𝑏𝑏𝑃𝑃𝑃𝑃𝑃𝑃 + 𝛽𝛽6 ⋅ 𝑃𝑃𝑃𝑃𝑅𝑅𝑙𝑙𝑃𝑃𝑃𝑃𝑙𝑙𝑙𝑙𝑓𝑓𝑙𝑙𝑙𝑙𝑓𝑓𝑖𝑖𝑓𝑓𝑁𝑁 + 𝛽𝛽7 ⋅ 𝑉𝑉𝑃𝑃ℎ𝑃𝑃𝑃𝑃𝑅𝑅𝑃𝑃 𝑇𝑇𝐴𝐴𝑡𝑡𝑃𝑃 + 𝛽𝛽8 ⋅ 𝑃𝑃𝑏𝑏𝑖𝑖𝑃𝑃𝑃𝑃𝑙𝑙 𝐶𝐶𝑃𝑃𝑃𝑃𝑙𝑙𝑅𝑅𝑃𝑃𝑙𝑙𝑃𝑃𝑏𝑏 + 𝛽𝛽9 ⋅ 𝐺𝐺𝐴𝐴𝑃𝑃 (14)

129 Table 8-7. Parameters for the MDV final oblique linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 127.196 36.210 0.001** MDV β1, MDV (m/s) 11.663 4.106 0.007** Belt Status β2, Belted -75.894 19.136 < 0.001** Object Contacted β8, Not a tree or pole -101.248 12.157 < 0.001** GAD β9, Frontal damage 35.300 9.956 0.001** Table 8-8. Parameters for the OIV final oblique linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 146.715 33.166 < 0.001** OIV β1, OIV (m/s) 7.359 2.650 0.008** Belt Status β2, Belted -74.025 22.059 0.002** Object Contacted β8, Not a tree or pole -83.220 8.873 < 0.001** Table 8-9. Parameters for the OLC final oblique linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 173.479 24.097 < 0.001** OLC β1, OLC (g) 2.507 0.934 0.010** Belt Status β2, Belted -73.006 22.453 0.002** Object Contacted β8, Not a tree or pole -74.466 8.039 < 0.001** Table 8-10. Parameters for the ASI final oblique linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 159.865 25.849 < 0.001** ASI β1, ASI 39.728 13.642 0.005** Belt Status β2, Belted -71.274 21.631 0.002** Object Contacted β8, Not a tree or pole -78.401 8.788 < 0.001** Table 8-11. Parameters for the VPI final oblique linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 161.960 26.992 < 0.001** VPI β1, VPI (m/s2) 0.132 0.050 0.011** Belt Status β2, Belted -74.409 21.640 0.001** Object Contacted β8, Not a tree or pole -76.365 8.771 < 0.001** 8.3.3 Final Side Crash Harm Models 𝑌𝑌𝑁𝑁𝑖𝑖𝑠𝑠𝑁𝑁 = 𝛽𝛽0 + 𝛽𝛽1 ⋅ (𝑃𝑃𝑃𝑃𝑖𝑖𝐴𝐴𝑃𝑃𝐴𝐴 𝑚𝑚𝑃𝑃𝑙𝑙𝑃𝑃𝑃𝑃𝑃𝑃) + 𝛽𝛽2 ⋅ 𝑏𝑏𝑃𝑃𝑅𝑅𝑙𝑙_𝑃𝑃𝑙𝑙𝑅𝑅𝑙𝑙𝐴𝐴𝑃𝑃 + 𝛽𝛽3 ⋅ 𝑃𝑃𝑃𝑃𝑠𝑠 + 𝛽𝛽4 ⋅ 𝑅𝑅𝑙𝑙𝑃𝑃 + 𝛽𝛽5 ⋅ 𝑃𝑃𝑏𝑏𝑃𝑃𝑃𝑃𝑃𝑃 + 𝛽𝛽6 ⋅ 𝑃𝑃𝑃𝑃𝑅𝑅𝑙𝑙𝑃𝑃𝑃𝑃𝑙𝑙𝑙𝑙𝑓𝑓𝑙𝑙𝑙𝑙𝑓𝑓𝑖𝑖𝑓𝑓𝑁𝑁 + 𝛽𝛽7 ⋅ 𝑉𝑉𝑃𝑃ℎ𝑃𝑃𝑃𝑃𝑅𝑅𝑃𝑃 𝑇𝑇𝐴𝐴𝑡𝑡𝑃𝑃 + 𝛽𝛽8 ⋅ 𝑃𝑃𝑏𝑏𝑖𝑖𝑃𝑃𝑃𝑃𝑙𝑙 𝐶𝐶𝑃𝑃𝑃𝑃𝑙𝑙𝑅𝑅𝑃𝑃𝑙𝑙𝑃𝑃𝑏𝑏 + 𝛽𝛽9 ⋅ 𝑀𝑀𝑚𝑚𝑡𝑡𝑅𝑅𝑃𝑃𝑙𝑙 𝑇𝑇𝐴𝐴𝑡𝑡𝑃𝑃 (15) Table 8-12. Parameters for the MDV final side linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 59.429 32.473 0.072 MDV β1, MDV (m/s) 8.645 2.599 0.001** Belt Status β2, Belted -84.582 29.720 0.006** Age β4, < 65 years old 58.839 23.942 0.017**

130 Side Impact Type β9, Nearside impact 45.543 16.598 0.008** Table 8-13. Parameters for the OIV final side linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 61.325 32.242 0.062 OIV β1, OIV (m/s) 8.605 2.624 0.002** Belt Status β2, Belted -84.465 29.761 0.006** Age β4, < 65 years old 58.215 23.961 0.018** Side Impact Type β9, Nearside impact 84.465 16.729 0.010** Table 8-14. Parameters for the OLC final side linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 88.003 30.647 0.006** OLC β1, OLC (g) 2.728 1.122 0.018** Belt Status β2, Belted -86.556 33.270 0.012** Age β4, < 65 years old 55.400 24.122 0.025** Side Impact Type β9, Nearside impact 47.329 16.072 0.005** Table 8-15. Parameters for the ASI final side linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 58.375 32.310 0.076 ASI β1, ASI 41.816 12.156 0.001** Belt Status β2, Belted -80.792 31.076 0.012** Age β4, < 65 years old 60.395 23.932 0.014** Side Impact Type β9, Nearside impact 48.063 16.502 0.005** Table 8-16. Parameters for the VPI final side linear regression model used to predict √Harm. ** indicates statistical significance (p-value < 0.05). Predictor Variable Parameter Coefficient Std. Error p-Value --- β0, Intercept 56.509 32.977 0.092 VPI β1, VPI (m/s2) 0.202 0.059 0.001** Belt Status β2, Belted -81.255 30.749 0.010** Age β4, < 65 years old 58.638 24.097 0.018** Side Impact Type β9, Nearside impact 47.944 16.330 0.005** 8.3.4 Linear Harm Prediction Curves The predict function in R was used to obtain √Harm values for each of the five final models. To generate the curves, the cost associated with injury was evaluated for belted and unbelted occupants as the crash severity metric increased by 0.01 from the minimum value of the metric in the dataset to the maximum value of the metric in the dataset. The horizontal black line on each plot represents the cost of a fatal injury. Figure 8-1 through Figure 8-15 plot the risk curves for all five metrics for all three models. Because the original values for these metrics, besides VPI, were either zero or negative, they were multiplied by -1 prior to plotting to maintain a positive x-axis. The shaded regions represent 95% confidence intervals.

131 8.3.4.1 Frontal Crash Harm Prediction Curves Figure 8-1. Frontal crash MDV Harm prediction curves for male drivers, under the age of 65 with a BMI less than 25 kg/m2, who did not strike a tree or pole. MDV, belt status, sex, age, BMI, seating location, and object contacted explain 52% of the variance associated with the model. Figure 8-2. Frontal crash OIV Harm prediction curves for passenger car drivers under the age of 65 with a BMI less than 25 kg/m2. OIV, belt status, sex, age, BMI, seating location, and object contacted explain 52% of the variance associated with the model.

132 Figure 8-3. Frontal crash OLC Harm prediction curves for passenger car drivers under the age of 65 with a BMI less than 25 kg/m2. OLC, age, BMI, seating location, and vehicle body type explain 46% of the variance associated with the model. Figure 8-4. Frontal crash ASI Harm prediction curves for male drivers under the age of 65 with a BMI less than 25 kg/m2. ASI, belt status, sex, age, BMI, and seating location type explain 51% of the variance associated with the model. Figure 8-5. Frontal crash VPI Harm prediction curves for male drivers under the age of 65 with a BMI less than 25 kg/m2. VPI, belt status, sex, age, BMI, and seating location type explain 49% of the variance associated with the model.

133 8.3.4.2 Oblique Crash Harm Prediction Curves Figure 8-6. Oblique crash MDV Harm prediction curves for front row occupants in vehicles that did not strike a tree or pole and have their GAD to the front of the vehicle. MDV, belt status, GAD, and object contacted explain 42% of the variance associated with the model. Figure 8-7. Oblique crash OIV Harm prediction curves for front row occupants in vehicles that did not strike a tree or pole. OIV, belt status, and object contacted explain 36% of the variance associated with the model. Figure 8-8. Oblique crash OLC Harm prediction curves for front row occupants in vehicles that did not strike a tree or pole. OLC, belt status, and object contacted explain 35% of the variance associated with the model.

134 Figure 8-9. Oblique crash ASI Harm prediction curves for front row occupants in vehicles that did not strike a tree or pole. ASI, belt status, and object contacted explain 36% of the variance associated with the model. Figure 8-10. Oblique crash VPI Harm prediction curves for front row occupants in vehicles that did not strike a tree or pole. VPI, belt status, and object contacted explain 35% of the variance associated with the model. 8.3.4.3 Side Crash Harm Prediction Curves Figure 8-11. Side crash MDV Harm prediction curves for front row occupants younger than 65 years old. MDV, belt status, side impact type, and age explain 34% of the variance associated with the model.

135 Figure 8-12. Side crash OIV Harm prediction curves for front row occupants younger than 65 years old. OIV, belt status, side impact type, and age explain 34% of the variance associated with the model. Figure 8-13. Side crash OLC Harm prediction curves for front row occupants younger than 65 years old. OLC, belt status, side impact type, and age explain 31% of the variance associated with the model. Figure 8-14. Side crash ASI Harm prediction curves for front row occupants younger than 65 years old. ASI, belt status, side impact type, and age explain 34% of the variance associated with the model.

136 Figure 8-15. Side crash VPI Harm prediction curves for front row occupants younger than 65 years old. VPI, belt status, side impact type, and age explain 34% of the variance associated with the model. Predictive Capability Evaluation and Validation The data were validated using a subset of real-world crashes from the CISS database; 339, 179, and 101 cases were available for validation for the frontal, oblique, and side models, respectively. The RMSE was used to compare the models (Table 8-17). For each crash type, all the final models yield similar RSME values. Overall, the model for side crashes yields the least amount of error. Table 8-17. RMSE for the frontal, oblique, and side Harm models for each metric. Crash Type Metric Model RMSE R2 Frontal MDV 65.91 52% OIV 67.49 52% OLC 64.40 46% ASI 65.71 51% VPI 66.40 49% Oblique MDV 75.75 42% OIV 70.83 36% OLC 71.41 35% ASI 69.30 36% VPI 71.48 35% Side MDV 62.89 34% OIV 62.40 34% OLC 62.86 31% ASI 62.90 34% VPI 63.26 34% Discussion Belt status, side impact type, age, and the metric were significant in each of the final side crash models. The oblique models each had belt status, object contacted, and the metric as significant. Additionally, GAD was significant in the MDV model. The frontal models showed more variation in their final set of covariates. The metric, occupant age, BMI, and seating location were significant in every model. Sex and belt status were significant in all but OLC, while object contacted was significant in MDV, OIV, and OLC. For each severity metric holding all covariates constant, the Harm value was lower for a belted occupant than for an unbelted occupant. Additionally, Harm

137 was greater for occupants in nearside crashes than for those in far-side crashes. Using RSME as a metric of model accuracy, the side crash models perform better than the front and oblique crash models. OLC, ASI, and OIV are the best metrics for the frontal, oblique, and side crash models, respectively. Currently, the U.S. and other roadside safety hardware crash test procedures prescribe preferred and maximum thresholds for the OIV and ASI metrics (Table 8-18). The models presented in this chapter can provide Harm costs associated with these thresholds for the best- and worst-case scenario occupant (Table 8-18). For example, for the frontal OIV model, a best-case scenario occupant would be a belted, male occupant in the right front passenger seat, less than 65 years old and with a BMI less than 30, not in a collision with a narrow object. The best case-scenario for the frontal ASI model would be the same type of occupant; however, the object contacted would not be a factor. This is because the object contacted covariate was not significant in the final ASI model. While other populations could be explored as well, this chapter examines only the two extremes. Calculating the Harm costs for the OIV and ASI thresholds makes it possible to compare the level of predicted cost across the two metrics. For example, for the best-case scenario frontal crash occupant for OIV, the Harm cost is just under US$11,000 for the maximum threshold but over US$29,000 for the ASI maximum threshold.

138 Table 8-18. Summary of real-world occupant Harm cost associated with current roadside hardware injury metric thresholds: Overall MAIS2+F injury in frontal crashes. Crash and Occupant Conditions Harm (US$100) Scenario Belt Status Age (years) Sex BMI (kg/m2) Seating Location Object Contacted OIV Threshold (m/s) ASI Threshold (--) 9.1 12.2 1.0 1.4 1.9 Best Case Belted < 65 Male < 30 RF Not a narrow object 17.3 107.8 13.0 92.5 293.1 Worst Case Unbelted ≥ 65 Female ≥ 30 Driver Narrow object 685.0 1049.3 419.9 702.1 1156.1 Table 8-19. Summary of real-world occupant Harm cost associated with current roadside hardware injury metric thresholds: Overall MAIS2+F injury in side crashes. Crash and Occupant Conditions Harm (US$100) Scenario Belt Status Age (years) Impact Type OIV Threshold (m/s) ASI Threshold (--) 9.1 12.2 1.0 1.4 1.9 Best Case Belted < 65 years Far-Side 30.4 66.9 3.8 13.0 32.5 Worst Case Unbelted ≥ 65 Nearside 587.6 723.9 435.3 507.9 606.5 Table 8-20. Summary of real-world occupant Harm cost associated with current roadside hardware injury metric thresholds: Overall MAIS2+F injury in oblique crashes. Crash and Occupant Conditions Harm (US$100) Scenario Belt Status Object Contacted OIV Threshold (m/s) ASI Threshold (--) 9.1 12.2 1.0 1.4 1.9 Best Case Belted Not a narrow object 43.9 86.0 23.2 39.7 66.7 Worst Case Unbelted Narrow object 342.0 446.9 264.9 315.6 385.2 Limitations One of the limitations of this work is how few occupants suffered MAIS2+F injuries in both the training and test datasets. The value of the models lies in their ability to be able to predict costs associated with occupants who are injured. The less injured occupants they are exposed to in the training data, the weaker they will be when it comes to accurately predicting the associated costs in the test data. Additionally, there were too few MAIS3+F injuries in the dataset to build MAIS3+F injury risk curves. This prevents extensive comparison between this work and previous MAIS3+F injury work. Another limitation to this study is that the inclusion criteria do not consider vehicle incompatibility. The size and weight differences between two colliding vehicles were not restricted. Finally, some cases had to be removed from the dataset due to unreasonably high case weights. Case weights are assigned to each case in these databases to help estimate the national incidence of various crash types. Despite this, most subsets of these data frames contain a few cases that have very large weights (Brumbelow 2019). Often, the cases assigned these large weights share similar outcomes to other cases with weights that are much smaller. To avoid largely weighted cases dominating and skewing potentially meaningful results, this analysis excludes any cases that have a case weight greater than 5,000 (Kononen 2011). The third limitation to this study is that the Harm values tabulated in Table 8-1 are the average societal costs per injury in the early 1990s. Since then, these values have increased by some

139 percentage due to inflation over time. In a study by Gabler et al. (2005), the average cost per injury was normalized to the cost of a fatal injury. Normalizing the injury costs this way means they can be easily adapted over time as the cost of a fatal injury changes. Conclusions For each of the crash severity metrics, Harm linear regression curves were constructed for frontal, side, and oblique crashes. These models were constructed using the NASS/CDS cases in the IAD and were tested with the CISS cases in the IAD. In general, the higher the crash severity metric, the higher the Harm cost for the occupant. Based on the RMSE, the OLC metric was the model that produced the least amount of error for predicting occupant Harm costs in frontal crashes. For oblique and side crashes, ASI and OIV were the models with the lowest RMSE. However, there was very little variation in the RMSE values for the metrics in side crashes. Factors other than the crash severity metric are important predictors of Harm costs. For every crash type, belted occupants were predicted to have lower Harm costs than unbelted occupants. In side crashes, younger drivers and occupants in far-side impacts were also likely to yield lower Harm costs. In oblique crashes, occupants in a vehicle that did not contact a narrow object (such as poles and trees) were more likely to yield lower Harm costs. More factors yielded significance in the frontal crash models than in the oblique and side crash models. For frontal crashes, occupants were more likely to yield lower Harm costs if they were less than 65 years old, male, had a lower BMI, were the right-front passenger, and/or did not collide with a narrow object. All of these covariates can drastically affect the Harm outcome even when the crash severity metric is held constant. At the maximum OIV, the best-case scenario was a Harm cost ranging from $6,700 to $10,800, depending on the crash type. However, at the worst-case scenario, Harm ranged from $44,700 to $104,900.