Provisional THOR Chest Injury Risk Function with Frontal and

Oblique Sled Tests

Timothy L. McMurry∗ Gerald S. Poplin† Jason Forman Joseph Ash
Gregory Shaw Jeff Crandall

January 8, 2016

This document explains the functional form of the chest injury risk function relative to frontal
crash test conditions described in Appendix A. Results are preliminary and presented only to
provide a general impression for the metrics under current consideration as well as the likely risk
function.

1 Data and Model Description

The data consists of 44 sled test experiments on 40 post-mortem human subjects (PMHS). The
PMHS were then evaluated for chest injuries, specifically ribcage fractures, and classified by the
severity of injury using the Abbreviated Injury Scale (AIS), 2005 edition (’08 update).
Outcomes of the PMHS tests were matched to identical sled test conditions using the THOR-
metric dummy. THOR records x, y, z, and d (cord) deflections at 4 chest locations: upper left
(UL), upper right (UR), lower left (LL), and lower right (LR). These deflections, in addition to the
derived resultant deflection, were recorded in time over the course of the test.
The goal of this analysis is to associate the 4 THOR deflection measurements with the proba-
bility of AIS 3+ injury. In the development of this injury risk function, three modeling constraints
were considered.

1. There is an approximate rule of thumb that each considered predictor requires 10 events (AIS
3+ injuries) and 10 non-events for a reliable model fit (Vittinghoff et al., 2012, p.422). These
data contain 25 subject observations with AIS 3+ injuries resulting and 19 observations not
sustaining AIS 3+ injuries, thus limiting the model to two independent predictors.

2. Age is considered an important predictor of PMHS injury and should be included in the
model.

3. The remaining predictor should be left/right symmetric. That is if the deflection patterns were
reflected across the dummy’s left/right aspect, the estimated injury risk should be unchanged.
∗
[email protected]
†
[email protected]

1
 3.0
2.5
Variances

2.0
1.5
1.0
0.5

87.1% 6.4% 5.4% 1.1%

0.0

First Second Third Fourth

Component

Figure 1: Variance (and percentage of variance) explained by the 4 principal components.

1.1 Model Details

The left/right symmetry requirement led us to consider 4 measures derived from the resultant
deflections measured at the 4 dummy sensors.
uptot ≡ (|U L|max + |U R|max )
lowtot ≡ (|LL|max + |LR|max )
(1)
updif ≡ |U L − U R|max
lowdif ≡ |LL − LR|max
where |U L|max indicates the maximum resultant displacement observed at the upper left sensor over
the course of the sled test, and |U L − U R|max indicates the maximum observed difference between
the upper left and upper right resultant displacements; the other terms are defined similarly.
In order to reduce these 4 measures down to a single composite measure, we used principal
components analysis (PCA), a standard technique to reduce the dimensionality of a dataset while
retaining as much information as possible. PCA works by an eigenvector decomposition of the
correlation matrix. The eigenvectors give the principal component directions, and the eigenvalues
describe the variance explained by each component.
The loadings (or weighting factors) for the individual components in the resulting PCA are
shown in Table 1.
Comp.1 Comp.2 Comp.3 Comp.4
up tot 0.486 0.718 0.454 -0.206
low tot 0.492 0.132 -0.833 -0.216
up dif 0.496 -0.664 0.309 -0.466
low dif 0.526 -0.160 0.069 0.833

Table 1: Loadings for each variable in each of the 4 principal components.

Figure 1 shows that almost all the variability in the test data can be explained by a single
principal component. Moreover, that principal component is an approximately equally weighted
sum of the 4 inputs scaled to have mean 0 and variance 1.

2
1.2 Calculating the Principal Components Score
Since one principal component explains almost all of the deflection patterns observed in the data
set (87.1%, Fig. 1), loadings from the first principal component are used as the basis for the primary
predictor of injury. The first principal component can be calculated by:
updif − m3
       
uptot − m1 lowtot − m2 lowdif − m4
P C1 = l1 + l2 + l3 + l4 . (2)
s1 s2 s3 s4

The variables in equation (2) are defined as follows. uptot , lowtot , updif , and lowdif are the variables
described in (1).
l1 , . . . , l4 are the principal components loadings (first column in Table 1). Their values are

l1 l2 l3 l4
0.486 0.492 0.496 0.526

m1 , . . . , m4 are the means of the 4 input deflection metrics (described in (1)). Their values are

m1 m2 m3 m4
59.856 47.417 27.389 26.345

s1 , . . . , s4 are the standard deviations of the 4 input deflection metrics (described in (1)). Their
values are

s1 s2 s3 s4
17.439 14.735 9.672 12.384

1.3 Calculating the Translated Principal Components Score

Unfortunately, the principal components scores calculated by equation (2) are not guaranteed to be
non-negative, which makes them inappropriate for use with a survival model. However, since the
principal component loadings are all positive, there is a natural translation to guarantee every score
is non-negative and to ensure that no deflection corresponds to zero injury risk. The translated PC
Score is given by:
updif − m3
       
trans uptot − m1 lowtot − m2 lowdif − m4
P C1 = l1 + l2 + l3 + l4
s1 s2 s3 s4
m1 m2 m3 m4
+ l1 + l2 + l3 + l4
s1 s2 s3 s4
uptot lowtot updif lowdif
= l1 + l2 + l3 + l4 . (3)
s1 s2 s3 s4
Equation (3) is used as the primary independent predictor for the logistic regression and survival
models.
Equation (3) also describes how injury risk predictor for a new THOR-metric dummy test series
should be calculated. The quantities described in (1) would be calculated, and then substituted
into (3) in place of uptot , lowtot , updif , and lowdif . Then the loading values l1 , . . . , l4 and standard
deviations s1 , . . . , s4 (given above) are substituted, and the final score calculated.

3
2 Preliminary Fitted Models
Statistical modeling was conducted in two stages. The first stage, presented in this section, is
to assess the fitted model for goodness of fit, etc. There is a failed assumption with the models
presented in this section with regard to the independence of observations, as 4 of the cadavers were
tested twice, the first time at a slow, non-injurious impact speed, and the second at a higher speed.
These repeated observations on the same subject cannot be treated as independent. Remedying
this failed assumption is addressed in Section 3.1.
Two types of models were fit using the predictor described by (3) and age of the PMHS as
a potential confounding factor. The first is a logistic regression including age and the translated
principal components score as predictors. The second is a parametric survival model, specifically,
an accelerated failure time (AFT) model with Weibull distribution, treating the deflection metric
P C1trans as the failure “time”, and age as a covariate. In the present setting the two models are
expected to produce results of similar quality due to the double-censored nature of the data. The
first approach (i.e., logistic modeling) has the advantage of being more commonly used and having
well-established diagnostic procedures. The logistic regression model output is:

Call:
glm(formula = ais3 ~ respcaTrans + age, family = binomial, data = thor)

Deviance Residuals:
Min 1Q Median 3Q Max
-2.1762 -0.7262 0.2952 0.7209 1.5378

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -12.76301 4.19781 -3.040 0.00236 **
respcaTrans 1.04992 0.33120 3.170 0.00152 **
age 0.11757 0.04638 2.535 0.01125 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 60.176 on 43 degrees of freedom

Residual deviance: 38.407 on 41 degrees of freedom
AIC: 44.407

Number of Fisher Scoring iterations: 5

The second modeling approach (AFT) has the advantage of producing an estimated 0 risk
corresponding to 0 deflection, and it makes it easier to account for repeated measures. The fitted
accelerated failure time model is:

Call:
survreg(formula = Surv(respcaTrans, time2 = NULL, event = 2 *
ais3, type = "interval") ~ age, data = thor, dist = "weibull")
Value Std. Error z p
(Intercept) 2.931 0.47419 6.18 6.37e-10

4
age -0.019 0.00786 -2.41 1.58e-02
Log(scale) -1.216 0.32432 -3.75 1.76e-04

Scale= 0.296

Weibull distribution
Loglik(model)= -19.6 Loglik(intercept only)= -23.5
Chisq= 7.88 on 1 degrees of freedom, p= 0.005
Number of Newton-Raphson Iterations: 5
n= 44

The diagnostics for the logistic regression showed one potential outlier, a 40 year old subject
whose cause of death was deemed respiratory failure as a result of scleroderma. Upon review, the
severity of the scleroderma was deemed to increase the risk of fracture, so the subject was removed
for all analysis.

Residuals vs Fitted Normal Q−Q

2

2
●● ● ● ●
● ●
●● ●
Std. deviance resid.
1

●● ●●
1
●●
●
● ●●●●
●
●● ●
●● ●●●
●●● ●●
Residuals

● ● ●●●●●
●●●
0

●●●●
● ●●●
● ●●
●
●●
●● ●●
−1

●
● ●
●●●
−1

●
● ●●
● ●●
●1 ●
4
−2

4 ●1 ●
−2

26 ●
● 26

−4 −2 0 2 −2 −1 0 1 2

Predicted values Theoretical Quantiles

Scale−Location Residuals vs Leverage

1.5

26 ●
●4
1 ● 39 22 ●
●● ●
1
Std. deviance resid.

● ●
● ● ●
Std. Pearson resid.

●● ● ●
● ●
● ●● ●
●●
1.0

● ●
●●● ●
●
●
● ●
0

●● ●●
● ●●●
● ● ●
● ● ● ●
● ● ●
●
● ●
●● ● ● ●
−1

●● ● ●
●●
● ●
●
0.5

●
●
−2

●●
●● ●

0.5
−3

● 26
Cook's distance
0.0

−4 −2 0 2 0.00 0.05 0.10 0.15

Predicted values Leverage

Figure 2: Diagnostics for the logistic model.

Since the fitting processes for both models are similar, the diagnostics for the AFT model will
be similar, and are therefore omitted.
Both logistic and AFT models produce non-significant Hosmer-Lemeschow goodness of fit
(GOF) statistics, which is an indicator of good model fit. Presented below are the results for
logistic regression.

Hosmer and Lemeshow goodness of fit (GOF) test

5
 1.0
●
●
● ●●●
●●

Logistic Model Probability of Injury

●
●● ●

0.8
●●
●
●
●
●●
●

0.6
●
●●
●
0.4 ●

● ●
●
●
●
● ●
●
0.2

● ●

●
0.0

● ●
●

0.0 0.2 0.4 0.6 0.8 1.0

Survival Model Probability of Injury

Figure 3: The logistic and survival models produce very similar estimates of the probability of
injury.

data: thor$ais3, predict(fla, type = "response")

X-squared = 6.9977, df = 6, p-value = 0.3211

Here are the results for AFT.

Hosmer and Lemeshow goodness of fit (GOF) test

data: thor$ais3, predictResponse(fsa, thor)

X-squared = 2.6651, df = 6, p-value = 0.8495

The Hosmer-Lemeschow test requires grouping the data into g groups by quantile. In both cases,
the test results hold for all reasonable choices of g. Test results for g = 8 are shown above.

2.1 Comparison of Logistic and AFT Models

The logistic and accelerated failure time models produce qualitatively similar results, as shown in
Figure 3. Figure 3 displays the predicted injury generated by the two models for each of our 44
tests; the results are seen to be almost the same.
The two models also produce similar areas under receiver operating characteristic curve (AUC).
For the logistic regression the AUC is 0.872, while for the AFT model the AUC is 0.857.

6
 ROC Curves
1.0 ● ● ● ● ● ● ● ● ● ● ● ●

● ● ● ●
● ● ●
● ● ● ●
0.8

●
●
probability of detection

● ●
●
●
0.6
(Sensitivity)

●
●
● ●
●
●
0.4

●
● ●
● ●

● ●
0.2

●
●
●
●
AFT
●
●
●
Logistic
0.0

0.0 0.2 0.4 0.6 0.8 1.0

probability of false alarm

(1−Specificity)

3 Model Fit Accounting For Repeated Measures

3.1 Repeated Measures Fit
The original data set had 4 subjects tested twice. This repeated measure can be accounted for by
treating the outcomes from these 4 subjects as interval censored. If a subject endured an impact
with a translated principal components score of 2.3 uninjured and was injured by an impact with
score 4.1, we treat the impact required for injury to be interval censored between 2.3 and 4.1. If
the subject had also survived the impact of 4.1, then we would consider that subject right censored
at 4.1.
With this adjustment, the final fitted model is:

Call:
survreg(formula = ss ~ age, data = thorRM, dist = "weibull")
Value Std. Error z p
(Intercept) 2.8677 0.47407 6.05 1.46e-09
age -0.0181 0.00785 -2.30 2.15e-02
Log(scale) -1.1975 0.32246 -3.71 2.04e-04

7
Scale= 0.302

Weibull distribution
Loglik(model)= -19.2 Loglik(intercept only)= -22.6
Chisq= 6.74 on 1 degrees of freedom, p= 0.0095
Number of Newton-Raphson Iterations: 5
n= 40

It is seen from Figure 4 that the two models produce fitted values that are quite similar.

1.0

●
●●
●
●
●
●●
●
●
●
0.8

●
●●
●
●
●
0.6

●
RM Model

●
●

●
0.4

●●
●
●
●●
●●
●
●
0.2

●●
0.0

●
●

0.0 0.2 0.4 0.6 0.8 1.0

Original Model

Figure 4: Comparison between the repeated measures survival model and the original independent
data survival model. The two models perform similarly.

3.2 Repeated Measures Model Diagnostics

In addition to the diagnostics performed in Section 2, we examined the dfbeta statistics for the
fitted model. Observations on 4 subjects were identified as potentially influential. No basis for

id Max dfbeta
323 0.31
433 0.36
444 0.33
481 0.31
674 0.37
736 0.37

removing these observations was found, so they were retained.

8
 1.0
● ● ●● ●
● ● ●

0.8
P(AIS 3+Injury)
0.6
0.4
0.2
0.0

● ● ● ● ● ● ●
●●

0 5 10 15
PC Score

Figure 5: Injury risk function for a 55 year old (solid) with 95% pointwise horizontal confidence
limits (dashed). The confidence interval for a 50% chance of injury is illustrated by the arrow.

3.3 Functional Form of the Provisional Model

The current provisional model given in Section 3.1 is an accelerated failure time model with a
Weibull distribution. It estimates the probability a subject is injured in an event with deflection
x, where x is given by equation (3), to be

P [AIS 3+ Injury|age, x] = 1 − exp[− {x exp[−(2.8677 − 0.0181 × age)]}1/0.3019 ]. (4)

3.4 Confidence Intervals

Pointwise 95% confidence intervals were calculated for the translated PC Score to produce a speci-
fied probability of injury. For example the impact required to produce a 50% probability of injury
in a 55 year old passenger has a translated PC Score of 5.84, and the 95% confidence interval is
(4.89, 6.96). The risk function and confidence intervals are shown in Figure 5.

4 Example
Suppose a series of frontal crash tests were run using the THOR-metric dummy. The associated
AIS 3+ injury risk can be estimated through the following stepwise process:

1. Calculate uptot , lowtot , updif , and lowdif as described in (1). For the remainder of the
example, we use the values
uptot = 75.206
lowtot = 51.751
updif = 33.461
lowdif = 36.753

2. Calculate the translated PC Score as given by equation (3). Use the values for uptot , lowtot ,

9
 updif , and lowdif calculated in Step 1, and the the values for li and si given under equation (2).

75.206 51.751 33.461 36.753

P C1trans = 0.486 × + 0.492 × + 0.496 × + 0.526 × = 7.101.
17.439 14.735 9.672 12.384

3. Calculate the probability of injury using equation (4). Substitute P C1trans calculated in Step
2 for x, and input the occupant’s age. For this example, we assume the occupant is 60 years
old.

P [AIS 3+ Injury|age, x = 7.101] = 1 − exp[− {7.101 × exp[−(2.868 − 0.018 × 60)]}1/0.302 ]

= 0.833.

We estimate that a 60 year old subject has an 83.3% chance of injury given the crash test
conditions.

5 Other Potential THOR Injury Risk Functions

We compared the PCA model with the equivalent accelerated failure time models based instead
on the overall maximum resultant deflection (the maximum deflection model below), and the sum
of the 4 resultant deflections (the total deflection model below). The maximum deflection model
produces an identical log likelihood to the PCA model and a slightly higher AUC (although still
less than Hybrid III). The total deflection model appears to be slightly worse than the PCA model
in terms of both log likelihood and AUC.

5.1 Maximum Deflection Model

Call:
survreg(formula = ssOA ~ age, data = thorRM, dist = "weibull")
Value Std. Error z p
(Intercept) 4.7775 0.46819 10.20 1.90e-24
age -0.0171 0.00773 -2.21 2.71e-02
Log(scale) -1.2101 0.31508 -3.84 1.23e-04

Scale= 0.298

Weibull distribution
Loglik(model)= -19.8 Loglik(intercept only)= -22.9
Chisq= 6.3 on 1 degrees of freedom, p= 0.012
Number of Newton-Raphson Iterations: 5
n= 40

The resulting AUC is 0.861.

10
 1.0
●
●
●●
● ●●
●
●

Cmax estimated P(AIS 3+ Injury)

●
●
●●

0.8
●● ●

●
●
●

0.6
●
●●
●
●
●
●

0.4
●
●
●●
●●
● ●●

0.2
●
●
●
●
0.0
●

0.0 0.2 0.4 0.6 0.8 1.0

PC Score estimated P(AIS 3+ Injury)

5.1.1 Maximum Deflection Model – Logistic Regression

(This model is presented only for comparison, and is not used elsewhere.)
Call:
glm(formula = ais3 ~ overall_max + age, family = binomial, data = thor)

Deviance Residuals:
Min 1Q Median 3Q Max
-2.0873 -0.7540 0.3397 0.6459 1.6057

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -11.72791 3.78122 -3.102 0.00192 **
overall_max 0.14170 0.04424 3.203 0.00136 **
age 0.10398 0.04211 2.469 0.01354 *
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

(Dispersion parameter for binomial family taken to be 1)

Null deviance: 60.176 on 43 degrees of freedom

Residual deviance: 39.908 on 41 degrees of freedom
AIC: 45.908

Number of Fisher Scoring iterations: 5

The resulting AUC is 0.872.

5.2 Total Deflection Model

Call:
survreg(formula = ssTot ~ age, data = thorRM, dist = "weibull")

11
 Value Std. Error z p
(Intercept) 5.6211 0.41735 13.47 2.40e-41
age -0.0151 0.00688 -2.20 2.81e-02
Log(scale) -1.3027 0.35838 -3.64 2.78e-04

Scale= 0.272

Weibull distribution
Loglik(model)= -20.8 Loglik(intercept only)= -23.9
Chisq= 6.34 on 1 degrees of freedom, p= 0.012
Number of Newton-Raphson Iterations: 5
n= 40

The resulting AUC is 0.825.

1.0

●
●
●●
●
●
●●
●
Total Deflection AIS 3+ Injury

0.8

●
●
●● ●
●
●
0.6

●
● ● ● ●●
●
● ● ●
0.4

●
●●
●●
● ●●
0.2

●
●
●
0.0

●
●

0.0 0.2 0.4 0.6 0.8 1.0

THOR PC Score AIS 3+ Injury

6 Comparison of Test Condition Ranking Between THOR PC

Score and THOR Cmax
There is a strong correlation between the way THOR PC Score and THOR Cmax rank the severity
of the test conditions, but the measured severities are not identical. The plots below show this
relationship. Points in the plot below are labeled with testing condition numbers; testing conditions
are described in Table 2.

12
 14
●

8
7 ●
●
7

12
● 4 ●
● 6 ●
●
4

THOR PC Score Rank

●8 2 6

10
9 ● ●
THOR PC Score

● 2
6

5 11 ●
●
8

8
●
14 9 ●
●
11

6
4

● ●
●
5
1312 14
●
●

4
3 13 ●
12 ●

2
2

●
3
●
10 10

0
20 30 40 50 0 2 4 6 8 10 12 14

THOR Cmax Deflection THOR Cmax Deflection Rank

Condition Position Buck Restraint Delta V

2 Front Passenger 97 Ford Taurus 3 point belt with FL + AB 48
3 Front Passenger 97 Ford Taurus Lap belt with AB 48
4 Front Passenger 97 Ford Taurus 3 point standard belt with AB 48
5 Front Passenger 97 Ford Taurus 3 point standard belt 29
6 Front Passenger 97 Ford Taurus 3 point standard belt 38
7 Rear Passenger 04 Ford Taurus 3 point standard belt 48
8 Front Passenger Gold Standard 1 3 point standard belt 40
9 Rear Passenger 04 Ford Taurus 3 point belt with FL + PT 48
10 Front Driver Gold Standard Driver 3 point standard belt 10
11 Front Driver Gold Standard Driver 3 point standard belt 40
12 Rear Passenger 04 Ford Taurus 3 point belt with FL + PT + belt bag 48
13 Front Passenger Gold Standard 2 3 point belt with custom FL 30
14 Front Passenger Gold Standard 3 3 point belt with custom FL 30

Table 2: Summary of Testing Conditions

References
Eric Vittinghoff, David V Glidden, Stephen C Shiboski, and Charles E McCulloch. Regression
methods in biostatistics: linear, logistic, survival, and repeated measures models. Springer, New
York, 2nd edition, 2012.

13